JP7656863B2 - MODEL GENERATION DEVICE, MODEL GENERATION METHOD, ABNORMALITY SIGN DETECTION DEVICE, ABNORMALITY SIGN DETECTION METHOD, AND ABNORMALITY SIGN DETECTION SYSTEM - Google Patents

Description

The present disclosure relates to a model generation device, a model generation method, an abnormality sign detection device, an abnormality sign detection method, and an abnormality sign detection system.

In order to prevent deterioration or failure of a certain piece of equipment and maintain normal operation, it is common to provide an anomaly sign detection device that uses data acquired from the equipment. Patent Document 1 discloses a condition monitoring device that performs machine learning using current values continuously acquired from a temperature control device to detect anomalies or signs of anomalies.

JP 2018-204940 A

In the technology disclosed in Patent Document 1, machine learning is performed to determine anomalies or signs of anomalies. However, since there are generally few cases in which anomalies occur in equipment, there is also little abnormal data that can be used for machine learning to generate a model. Even if abnormal data is obtained, the data may be mixed with data in which features related to anomalies or signs of anomalies in the equipment are not present. There is no problem if learning is performed after accumulating a large amount of data, as is done with general machine learning. However, data accumulation takes a long time. In addition, production may have ended by the time sufficient data has been accumulated. Therefore, learning after accumulating a large amount of data is not practical. Therefore, it is desirable to learn data that appropriately shows features related to anomalies or signs of anomalies when there is little abnormal data, and to generate a highly accurate trained anomaly sign detection model.

The purpose of this disclosure is to provide technology that can detect abnormalities and signs of abnormalities in production equipment.

A model generation device according to one aspect of the present disclosure includes a preprocessing unit that acquires feature data of production performance data for training, an anomaly degree calculation unit that calculates an anomaly degree of the acquired feature data by utilizing an anomaly degree calculation model that represents a group of features in a normal state, and a training unit that determines a state of the feature data based on the calculated anomaly degree, and trains an anomaly sign detection model that detects the state of the feature data from the feature data using training data composed of the feature data and the state of the feature data, wherein each feature data of the training data is composed of a group of features that overlap on a time axis .

A model generation method according to one aspect of the present disclosure includes the steps of: acquiring feature data of production performance data for training; calculating the degree of anomaly of the acquired feature data using an anomaly calculation model that represents a group of features in a normal state; determining a state of the feature data based on the calculated degree of anomaly; and training an anomaly sign detection model that detects the state of the feature data from the feature data using training data composed of the feature data and the state of the feature data, wherein each feature data of the training data is composed of a group of features that overlap on a time axis .

An anomaly sign detection device according to one aspect of the present disclosure includes a pre-processing unit that acquires feature data of production performance data to be detected, and an anomaly sign detection unit that detects a state of the feature data from the feature data by using a trained anomaly sign detection model, wherein the anomaly sign detection model is trained with training data composed of feature data of training production performance data and a state of the feature data, the state of the feature data is decided based on a degree of anomaly of the feature data calculated by using an anomaly degree calculation model that represents a group of features in a normal state , and each piece of feature data of the training data is composed of a group of features that overlap on a time axis .

An anomaly sign detection method according to one aspect of the present disclosure includes a step of acquiring feature data of production performance data to be detected, and a step of detecting a state of the feature data from the feature data by using a trained anomaly sign detection model, the anomaly sign detection model is trained by training data composed of feature data of training production performance data and a state of the feature data, the state of the feature data is decided based on a degree of anomaly of the feature data calculated by using an anomaly degree calculation model that represents a group of features in a normal state , and each piece of feature data of the training data is composed of a group of features that overlap on a time axis .

This disclosure provides technology that can detect abnormalities and signs of abnormalities in production equipment.

Schematic diagram showing an anomaly detection system FIG. 2 is a block diagram showing the hardware configuration of a model generating device and an anomaly sign detecting device. Schematic diagram showing a configuration example of a winding section Schematic diagram of a wound body FIG. 13 is a diagram showing an example of correspondence between measurement results and determination results of a first sheet material; FIG. 1 is a block diagram illustrating a functional configuration of a model generating device according to a first embodiment; An example of a production performance database An example of a feature database An example of an anomaly database A diagram showing an example of a state label database. FIG. 1 is a block diagram illustrating a functional configuration of an abnormality sign detection device according to a first embodiment; Flowchart illustrating an example of pre-processing 1 is a flowchart illustrating an example of a process for generating an anomaly degree calculation model. 1 is a flowchart illustrating an example of an abnormality degree calculation process. A flowchart for explaining an example of an adoption determination process for an anomaly degree calculation model. Flowchart illustrating an example of a state label setting process 1 is a flowchart illustrating an example of a training process for an anomaly sign detection model. A diagram to explain the learning data list A flowchart for explaining an example of an adoption determination process for an anomaly sign detection model. 1 is a flowchart for explaining an abnormality sign detection process according to a first embodiment. FIG. 11 is a block diagram illustrating a functional configuration of a model generating device according to a second embodiment; 11 is a flowchart for explaining a state label setting process according to a second embodiment. FIG. 13 is a block diagram illustrating a functional configuration of a model generating device according to a third embodiment. 13 is a flowchart for explaining a state label setting process according to the third embodiment.

Each embodiment of the present disclosure will be described in detail below with reference to the drawings. However, more detailed explanations than necessary, such as detailed explanations of matters that are already well known or duplicate explanations of substantially identical configurations, may be omitted.

Note that the following description and the referenced drawings are provided to enable those skilled in the art to understand the present disclosure, and are not intended to limit the scope of the claims of the present disclosure.

(First embodiment)
FIG. 1 is a schematic diagram showing an abnormality sign detection system 10 according to a first embodiment.

As shown in FIG. 1, the anomaly sign detection system 10 includes a model generation device 60 that generates an anomaly sign detection model, and an anomaly sign detection device 100 that performs anomaly sign detection using the generated anomaly sign detection model.

The abnormality sign detection system 10 uses a model generation device 60 to generate an abnormality sign detection model in advance to detect whether there is an abnormality or a sign of an abnormality in the production equipment 200. After generating the abnormality sign detection model, the abnormality sign detection system 10 uses an abnormality sign detection device 100 that can access the trained abnormality sign detection model to detect abnormalities and signs of an abnormality in the production equipment 200 based on the production performance data of the detection target collected while the production equipment 200 is in operation.

The anomaly sign detection model according to the following embodiment distinguishes the state of the production equipment 200 as a normal period, a warning period, or an abnormal period based on production performance data. However, the anomaly sign detection system 10 according to the present disclosure is not limited to this, and may distinguish any other appropriate state of the production equipment 200.

The model generation device 60 extracts training production performance data from the training database 20, and uses the extracted training production performance data to train an anomaly sign detection model through supervised machine learning.

For example, the training production performance data may be time-series production performance data on products, etc., collected while the production equipment 200 is in operation. In reality, it is desirable to be able to generate an anomaly sign detection model before the production equipment 200 starts operating, but it is generally difficult to obtain production performance data that indicates anomalies or signs of anomalies due to aging or the like before the production equipment 200 starts operating.

Therefore, immediately after the operation of the production equipment 200 (for example, one week after the start of operation), it is assumed that the production equipment 200 is in a normal state, and an anomaly calculation model for calculating the degree of anomaly as the deviation from a normal sample is derived based on the production performance data for that period, as described in detail below. Such an anomaly calculation model can be derived by applying an anomaly prediction detection technique that detects anomalies and signs of anomaly based on normal samples, such as the well-known MT method (Mahalanobis Taguchi Method). The model generation device 60 uses this anomaly calculation model to label the production performance data with "normal period," "premonitory period," and "abnormal period," even during a period when the number of collected abnormal samples and sign samples is considered to be relatively small, such as immediately after the operation of the production equipment, and can efficiently generate an anomaly prediction detection model using the labeled production performance data.

In one embodiment, the training database 20 stores unlabeled production performance data, and as described below, the model generation device 60 uses an anomaly calculation model to calculate the degree of anomaly of the unlabeled production performance data, and labels the state of the production performance data as training data for the anomaly sign detection model. Typically, according to a known anomaly sign detection technique such as the MT method, although the training data for training the anomaly sign calculation model needs to be state-labeled, the anomaly sign calculation model can be generated with significantly less training data than the anomaly sign detection model. Therefore, even if only a relatively small amount of state-labeled production performance data can be prepared, it is possible to generate an anomaly sign detection model.

Such an anomaly calculation model may be generated by a method described in detail below and may be provided in advance in the model generation device 60 or may be accessible via a network or the like. However, the present disclosure is not limited to this, and the model generation device 60 may generate not only an anomaly sign detection model but also an anomaly calculation model.

Once training of the anomaly sign detection model is complete, the model generation device 60 provides the trained anomaly sign detection model to the anomaly sign detection device 100.

The abnormality sign detection device 100 detects an abnormality or an abnormality sign of the production facility 200 based on the production performance data of the production facility 200, using the trained abnormality sign detection model generated by the model generation device 60. Specifically, the abnormality sign detection device 100 inputs the production performance data of the product acquired during the operation of the production facility 200 into the trained abnormality sign detection model, acquires the detection result (e.g., "normal period", "sign period", or "abnormal period") from the abnormality sign detection model, and notifies a user such as the manager of the production facility 200 of various information including the detection result for the maintenance of the production facility 200. In this embodiment, one abnormality sign detection device 100 is applied to one production facility 200, but one abnormality sign detection device 100 may be applied to multiple production facilities 200.

In addition, in this embodiment, the model generating device 60 and the anomaly sign detection device 100 are described as computing devices such as a server, but each component may be realized in the form of a server system connected via a network. Through the description of the configuration of the model generating device 60 and the anomaly sign detection device 100, the method of generating an anomaly sign detection model and the method of generating training data according to the present disclosure are explained and understood.

<Hardware configuration of the model generating device 60 and the abnormality sign detection device 100>
FIG. 2 is a block diagram showing the hardware configuration of the model generation device 60 and the anomaly sign detection device 100 according to an embodiment of the present disclosure.

As shown in FIG. 2, the model generating device 60 and the abnormality sign detection device 100 are each realized by any computing device or information processing device such as a server, and have a memory unit 11, a processing unit 12, a communication unit 13, and an input/output unit 14 interconnected via a bus B. Note that the model generating device 60 and the abnormality sign detection device 100 may each be composed of multiple computers that perform distributed processing, or may be realized by a virtual computer virtually constructed by software.

The storage unit 11 includes a main storage device such as a ROM (Read Only Memory) or a RAM (Random Access Memory), and an auxiliary storage device such as a HDD (Hard Disk Drive), an SSD (Solid State Drive), or a flash memory. The storage unit 11 stores programs, instructions, data, and the like for executing the functions and processing of the model generation device 60 and the anomaly sign detection device 100 described below.

The processing unit 12 includes a hardware processor, such as a CPU (Central Processing Unit), and executes the programs, instructions, etc. stored in the memory unit 11 to perform the functions and processing of the model generation device 60 and the anomaly sign detection device 100 described below.

The communication unit 13 executes various communication processes for wired or wireless communication with external devices (e.g., the training database 20, the production equipment 200, etc.) via a network such as the Internet, a LAN (Local Area Network), or a WLAN (Wireless LAN).

The input/output unit 14 has various interfaces such as a Universal Serial Bus (USB) and a High-Definition Multimedia Interface (HDMI (registered trademark)) for connecting to various peripheral devices such as a keyboard, mouse, microphone, display, printer, speaker, sensor, and alarm (e.g., buzzer, lamp, etc.). For example, the input/output unit 14 may notify a user of the abnormality sign detection device 100 of the detection of an abnormality and an abnormality sign. In this embodiment, the user of the abnormality sign detection device 100 may include an administrator of the abnormality sign detection device 100, or an operator who produces the wound body 204 (see FIG. 3 described later) using the production facility 200, or an operator who maintains the production facility 200.

The memory unit 11 and the processing unit 12 do not have to be configured as an integrated server or computer. In other words, the memory unit 11 and the processing unit 12 may be configured separately from each other and located at separate locations as long as they are configured to be able to communicate with each other. In addition, the model generating device 60 and the anomaly sign detection device 100 may have an operation unit not shown in FIG. 2 and may accept operation input from outside. Details of the memory unit 11 and the processing unit 12 will be described later.

<Production facility 200>
In the present embodiment, the production facility 200 is a production facility for lithium ion secondary batteries. As shown in Fig. 3, the production facility 200 has a winding unit 201 and an inspection machine 207. However, the production facility 200 according to the present disclosure is not limited to this and may be a production facility for any other product, etc.

Figure 3 is a schematic diagram showing an example of the configuration of the winding unit 201 and the inspection machine 207. The winding unit 201 has a first supply reel 50, a second supply reel 51, a cutting unit 208, a first laminating roll 205A, a second laminating roll 205B, a winding core 206, and a winding core rotation drive unit 206M. The inspection machine 207 has a first sheet material post-cut length measurement sensor 40 and a second sheet material post-cut length measurement sensor 41.

The winding section 201 laminates a first sheet material 202 supplied from a first supply reel 50 and a second sheet material 203 supplied from a second supply reel 51 using a first laminating roll 205A and a second laminating roll 205B, and winds them up around a winding core 206 to produce a wound body 204.

Figure 4 is a schematic diagram of the wound body 204 produced in the winding section 201. For convenience, FIG. 4 shows the state in which the ends of the first sheet material 202 and the second sheet material 203 constituting the wound body 204 are not wound. As shown in FIG. 4, the second sheet material 203 is formed to have a larger width (length along the axial direction of the wound body 204) than the first sheet material 202. The first sheet material 202 is, for example, a sheet-like member (positive electrode sheet) coated with a positive electrode material, and the second sheet material 203 is, for example, a sheet-like member (negative electrode sheet) coated with a negative electrode material.

Returning to FIG. 3, the explanation will be continued. The core rotation drive unit 206M drives the core 206 at a desired rotation speed. The cutting unit 208 cuts the first sheet material 202 and the second sheet material 203 when winding on the core 206 is completed. Note that while FIG. 3 shows the cutting unit 208 being positioned upstream of the first laminating roll 205A and the second laminating roll 205B, the cutting unit 208 may also be positioned downstream of these laminating rolls.

The inspection machine 207 is a device that inspects the wound body 204 produced by the winding section 201.

When the wound body 204 is produced in the winding section 201, defective products may occur. Defective products may occur, for example, when the first sheet material 202 or the second sheet material 203 wound around the winding core 206 is too long or too short in length after cutting. If the first sheet material 202 or the second sheet material 203 is too long or too short in length, the quality of the lithium ion secondary battery deteriorates. Therefore, the lengths of the first sheet material 202 and the second sheet material 203 are measured using the first sheet material post-cut length measurement sensor 40 and the second sheet material post-cut length measurement sensor 41.

Figure 3 shows that the first sheet material post-cut length measuring sensor 40 and the second sheet material post-cut length measuring sensor 41 are positioned downstream of the first laminating roll 205A and the second laminating roll 205B, but each sensor may be positioned upstream of these laminating rolls.

The measurement results of the first sheet material post-cut length measuring sensor 40 and the second sheet material post-cut length measuring sensor 41 are judged as either "good" or "bad". If the measurement result is within a predetermined target range, it is judged as "good", and if it is not within the target range, it is judged as "bad". Figure 5 shows an example of the correspondence between the measurement result of the first sheet material 202 and the judgment result. That is, the measurement data I1 to I3 in which the length after cutting of the first sheet material 202 is equal to or greater than the target length is judged as "good", and the measurement data I4 and I5 in which the length after cutting of the first sheet material 202 is less than the target length are judged as "bad". In this embodiment, the result is judged as either "good" or "bad", but it may be judged as any of three or more types.

In FIG. 5, the length measurement data I1 to I5 are shown in a time series. That is, FIG. 5 shows the state in which the length of the first sheet material 202 after cutting gradually shortens with the deterioration of the production equipment 200 over time. However, in reality, due to various factors other than the deterioration of the production equipment 200 over time, the way in which the length of the first sheet material 202 changes becomes complicated, and it is not desirable to judge the abnormality of the production equipment 200 simply by the length. For this reason, it is desirable for the model generation device 60 to train a machine learning model, specifically, to read time series data in which the change in the length of the first sheet material 202 is characterized using a neural network model or the like, thereby training the machine learning model in advance as to which of the time series data corresponds to the "normal period", "premonitory period", and "abnormal period". In this case, the abnormality sign detection device 100 uses a machine learning model trained by the model generation device 60 to detect abnormalities and signs of abnormalities in the production equipment 200 based on time-series data in which changes in the length of the first sheet material 202 and the second sheet material 203 are characterized. For the sake of convenience, the following description will be given of a case in which changes in the length of the first sheet material 202 are used, but it goes without saying that changes in the length of the second sheet material 203 or changes in the lengths of both sheet materials may also be used.

<Details of the model generating device 60>
Next, the functional configuration and operation of the model generating device 60 according to an embodiment of the present disclosure will be described in detail.

The model generation device 60 generates an anomaly sign detection model that detects signs of anomaly and anomalies in the production facility 200 based on actual production data related to the production of products by the production facility 200. When the anomaly sign detection model is generated, the model generation device 60 provides the generated anomaly sign detection model to the anomaly sign detection device 100 that detects signs of anomaly and anomalies in the production facility 200 during operation. Note that in this embodiment, anomaly sign detection means detecting signs of anomaly before an anomaly occurs in the production facility 200 in order to prevent the occurrence of anomalies such as deterioration and failure of the production facility 200 and to maintain normal operation.

Figure 6 is a block diagram illustrating the functional configuration of a model generation device 60 according to the first embodiment of the present disclosure.

As shown in the figure, the model generating device 60 has a preprocessing unit 61, an anomaly calculation unit 62, and a training unit 63. For example, one or more of the functional units of the preprocessing unit 61, the anomaly calculation unit 62, and the training unit 63 may be realized by a processing unit 12 having one or more processors executing a program or instructions stored in a storage unit 11 having one or more memories.

The pre-processing unit 61 acquires feature data of the training production performance data. Specifically, the pre-processing unit 61 may receive the training production performance data from the training database 20, and perform a predetermined pre-processing on the received production performance data to extract feature data. As an example, the pre-processing unit 61 may extract production performance unit data of a predetermined number of units or unit time from the time-series production performance data, perform statistical processing on each of the extracted production performance unit data, and acquire one or more acquired statistics as feature data.

The preprocessing unit 61 can perform similar preprocessing on both types of production performance data, labeled and unlabeled production performance data, to derive feature data.

[Production performance database 71]
The pre-processing unit 61 may store, for example, the production performance data acquired from the production facility 200 in a production performance database 71 of the storage unit 11. That is, the production performance database 71 may be a database in which the production performance data PD relating to the production performance of the production facility 200 is cumulatively registered. The production performance data PD of a plurality of production facilities 200 may be registered in the production performance database 71.

Figure 7 is a diagram showing an example of a production history database 71. In the example shown in Figure 7, the production history database 71 registers "production date and time", "equipment", "first sheet material measurement value", "second sheet material measurement value", "first sheet material measurement value judgment result", and "second sheet material measurement value judgment result" in association with each other. These data are associated one by one to form one piece of production history data PD.

"Production date and time" is data regarding the production date and time when the wound body 204 was produced.

"Facility" is data for identifying the facility that produced the wound body 204 when there are multiple production facilities 200. As an example, Figure 7 shows "A", "B", and "C", which are identifiers of different production facilities 200.

The "first sheet material length measurement value" and the "second sheet material length measurement value" are the length measurement value data of the first sheet material 202 and the second sheet material 203 used in the production of the roll 204. These length measurement value data are acquired by the first sheet material post-cut length measurement sensor 40 and the second sheet material post-cut length measurement sensor 41.

The "first sheet material measurement value judgment result" and the "second sheet material measurement value judgment result" are data showing the judgment results of the "first sheet material measurement value" and the "second sheet material measurement value" (see FIG. 5). For example, if the target length of the first sheet material 202 is 761 mm, the first sheet material measurement values of 762.2 mm and 761.4 mm are judged as "good", and the first sheet material measurement value of 760.1 mm is judged as "bad". Also, for example, if the target length of the second sheet material 203 is 692 mm, the second sheet material measurement values of 693.4 mm and 693.5 mm are judged as "good", and the first sheet material measurement value of 691.1 mm is judged as "bad". These judgments are made automatically, for example, by the processing unit 120 or the inspection machine 207.

Such production performance data PD is entered automatically or manually by an operator, for example, each time a wound body 204 is produced in the production facility 200, and is registered in the training database 20. Typically, since it is desired to generate an abnormality sign detection model before the production facility 200 starts full-scale operation, the production performance data for training is collected during an experimental stage, a test operation stage, etc., before the production facility 200 starts full-scale operation, and is stored in the training database 20.

In the above-described embodiment, the training production performance data is stored in the training database 20, and the pre-processing unit 61 extracts the production performance data from the training database 20 and extracts feature data from the extracted production performance data, but the present disclosure is not limited to this. For example, the training database 20 may store training feature data extracted from the training production performance data. In this case, the pre-processing unit 51 does not need to hold the production performance database 61.

[Feature database 72]
The pre-processing unit 61 may store, for example, feature amounts extracted from the production performance data in a feature amount database 72 of the storage unit 11. That is, the feature amount database 72 is a database in which the feature amount data FD is registered. The feature amount data FD may be data obtained by collecting a plurality of consecutive production performance data PD by a predetermined number of counting units W and performing data processing such as statistical processing on the collected production performance data PD to convert the collected production performance data PD into feature amounts. The feature amount database 72 may register feature amount data FD obtained from the production performance data PD of a plurality of production facilities 200.

Figure 8 is a diagram showing an example of the feature database 72. In the example shown in Figure 8, the feature database 72 registers "production date and time", "equipment", "counting unit", "maximum value of the first sheet material measurement value", "minimum value of the first sheet material measurement value", "average value of the first sheet material measurement value", "standard deviation of the first sheet material measurement value", and "number of defective first sheet material measurement values" in association with each other. These data are associated one by one to form one piece of feature data FD.

The "production date and time" is data relating to the production date and time of the production history data PD with the latest production date and time among the collected W pieces of production history data PD, and is the production date and time that represents the production date and time of the W pieces of production history data PD. In other words, the "production date and time" is the production date and time of the production history data PD corresponding to the last produced wound body 204 among the W wound bodies 204 to which the data-processed W pieces of production history data PD correspond.

"Equipment" is data for identifying the production equipment 200 that produced the wound body 204 corresponding to the collected W pieces of production performance data PD.

The "aggregation unit" is data related to the number of pieces of production performance data PD collected. In Figure 8, "100" is shown as an example.

The "maximum value of the first sheet material measurement value", "minimum value of the first sheet material measurement value", "average value of the first sheet material measurement value" and "standard deviation of the first sheet material measurement value" are data indicating each statistical quantity of the first sheet material measurement value contained in the collected W pieces of production performance data PD. In addition to the statistical quantities shown in FIG. 8, the feature database 112 may also register statistical quantities such as the median, first quartile, and third quartile. These statistical quantities are feature quantities.

The "number of defective first sheet material measurement values" is data indicating the number of wound bodies 204 for which the "first sheet material measurement value judgment result" was "defective" and which are included in the W wound bodies 204 corresponding to the collected W pieces of production performance data PD.

Each time production performance data PD of the number of aggregated units is registered in the production performance database 71, the pre-processing unit 61 may execute a predetermined pre-processing to extract such feature data FD, and the extracted feature data may be registered in the feature database 72.

The anomaly calculation unit 62 uses the anomaly calculation model to calculate the anomaly of the acquired feature data.

Here, the anomaly calculation model may be generated from a group of features in a normal state based on any anomaly sign detection technology. When feature data is received as input, the anomaly calculation model outputs the degree of anomaly of the feature data. The anomaly calculation model according to the present disclosure may be generated according to any known anomaly sign detection technology (e.g., MT method, etc.) using normal state feature data derived from production performance data collected during the period immediately after operation, under the assumption that the production equipment 200 is in a normal state immediately after operation. For example, the degree of anomaly of the feature calculated by the anomaly calculation model may be the degree of deviation from the group of features in a normal state (e.g., Mahalanobis distance, etc.).

As an example, the anomaly calculation model M1 may be a model that expresses the deviation of the production equipment 200 from the normal state as a numerical value called anomaly degree A. The anomaly degree A indicates the degree of anomaly of the production equipment 200. In other words, the anomaly degree A is an index indicating how much the feature amount data FD of the anomaly calculation target deviates from the feature amount data FD in the normal period. If the anomaly calculation model M1 is appropriately generated, the greater the anomaly degree A, the greater the degree of anomaly such as deterioration and failure in the production equipment 200, and the higher the probability that a defective product of the winding body 204 will be newly produced. The anomaly calculation model M1 calculates the deviation from the normal state as the anomaly degree A by defining the normal state using the feature amount data FD in the normal period in which no defects occur. As a result, even if the production performance data for training the anomaly calculation model M1 is collected immediately after the operation of the production equipment 200, the frequency of occurrence of anomalies due to aging deterioration, etc. is low, and a sufficient number of sample data of the abnormal state cannot be obtained, it is possible to calculate the anomaly degree. A normal period is a period during which no deterioration or failure occurs in the production equipment 200 and no defects occur in the wound body 204 produced by the production equipment 200.

The anomaly calculation model M1 may be generated using a combination of multiple types of features included in the feature data FD. The selection of the features to be combined may be performed by a user of the anomaly sign detection device 100, or may be performed automatically by the anomaly calculation unit 62 based on predetermined criteria.

Furthermore, by changing the combination of features, multiple different anomaly calculation models M1 can be generated. Note that only one anomaly calculation model M1 may be generated.

The abnormality calculation model M1 is generated based on the production performance data PD produced during a period when the production equipment 200 is assumed to be in a normal state, such as immediately after the production equipment 200 starts operating.

Here, the anomaly calculation model M1 may be appropriately selected from a number of different formats, such as a statistical model or a machine learning model generated by machine learning (e.g., a neural network model).

As an example, a known MT method may be used when generating the anomaly calculation model M1. This makes it possible to generate an anomaly calculation model M1 that can calculate an anomaly A that more accurately indicates the degree of deviation from the normal state. When the degree of deviation from the normal state is used as the degree of anomaly, the degree of anomaly calculated for the feature data is compared with an anomaly threshold that indicates an abnormal state, and the state of the feature data is determined according to the comparison result.

As another example, when generating the anomaly calculation model M1, a known distance-based anomaly sign detection method such as OC-SVM (One Class Support Vector Machine), an autoencoder using a neural network, or the K-nearest neighbor method may be used.

The anomaly calculation model M1 thus generated may be verified as to whether the calculated anomaly is appropriate. For example, the generated anomaly calculation model M1 may be verified based on the anomaly calculated for the feature data and the defect rate in the production performance data corresponding to the feature data. In other words, if the anomaly derived from the anomaly calculation model M1 for the input feature data exceeds the anomaly threshold, the feature data is assumed to correspond to the production performance data of the production equipment 200 during the abnormal period, and should have a positive correlation with the defect rate for the production performance data. Here, the anomaly threshold may be set based on the anomaly at which the production equipment 200 is determined to be in an abnormal state.

Therefore, if the correlation coefficient between the degree of abnormality equal to or greater than the anomaly threshold derived from the anomaly calculation model M1 and the corresponding defect rate is equal to or greater than a predetermined threshold, it is determined that the anomaly calculation model M1 is capable of appropriately distinguishing between a normal state and an abnormal state of the feature data. In this case, the anomaly calculation model M1 is determined to have been appropriately trained, and may be used in the subsequent labeling process of the feature data.

On the other hand, if the correlation coefficient between the degree of abnormality equal to or greater than the abnormality threshold derived from the anomaly calculation model M1 and the corresponding defect rate is less than a predetermined threshold, it is determined that the anomaly calculation model M1 is not capable of properly distinguishing between a normal state and an abnormal state of the feature amount data. In this case, the anomaly calculation model M1 is determined to have not been properly trained, and may be regenerated, for example, using other training data.

The verification of the above-mentioned anomaly calculation model M1 may be performed in advance (manually, for example) before use by the anomaly calculation model 62, or may be performed by the anomaly calculation unit 62.

In the above-described embodiment, the anomaly calculation model M1 generated in advance is provided in the anomaly calculation unit 62, but the present disclosure is not limited to this. For example, the anomaly calculation unit 62 itself may generate an anomaly calculation model using training data. Specifically, the anomaly calculation unit 62 may similarly generate an anomaly calculation model M1 according to the MT method described above, using labeled production performance data or feature amount data stored in the training database 20.

The format of the anomaly calculation model M1 may be selected by a user of the anomaly sign detection device 100, or may be selected automatically by the anomaly calculation unit 62.

[Anomaly Database 73]
The anomaly degree calculation unit 62 may store the calculated degree of anomaly in an anomaly degree database 73 in the storage unit 11. That is, the anomaly degree database 73 is a database in which anomaly degree data AD related to the degree of anomaly A calculated by the anomaly degree calculation model M1 is registered. The anomaly degree database 73 may register anomaly degree data AD obtained based on production performance data PD of a plurality of production facilities 200.

Figure 9 is a diagram showing an example of the abnormality database 73. In the example shown in Figure 9, the abnormality database 73 registers "production date and time," "equipment," "abnormality calculation model ID," "abnormality," "counting unit," and "number of defective first sheet material measurement values" in association with each other. These data are associated one by one to form one piece of abnormality data AD.

"Production date and time" is the production date and time included in the feature data FD for which the degree of anomaly is to be calculated.

The "anomaly calculation model ID" is data for identifying the anomaly calculation model M1 used to calculate the anomaly. As an example, FIG. 9 shows "001", "002", and "003", which are IDs (identification numbers) of different anomaly calculation models M1.

"Abnormality level" is data indicating the abnormality level A calculated using the anomaly level calculation model M1 from the feature amount data FD of the target for anomaly level calculation.

The "aggregation unit" is the aggregation unit included in the feature data FD for which the degree of anomaly is to be calculated.

The "number of defective first sheet material measurement values" is data indicating the number of defective first sheet material measurement values contained in the feature data FD for which the degree of anomaly is to be calculated. This number of defectives is used to evaluate the reliability of the degree of anomaly calculation model, as described in detail below.

Such anomaly data AD may be generated by the anomaly calculation unit 62 and registered in the anomaly database 73 each time feature data FD is registered in the feature database 72.

The training unit 63 determines the state of the feature data based on the calculated degree of anomaly, and trains an anomaly sign detection model that detects the state of the feature data from the feature data using training data composed of the feature data and the state of the feature data.

Here, the anomaly sign detection model accepts feature data as input and outputs the state of the input feature data. The training data used to generate the anomaly sign detection model may be composed of feature data derived from production performance data and the state of the feature data determined using the anomaly degree calculation model. In other words, the training data used to generate the anomaly sign detection model may be production performance data whose state is not labeled in the training database 20, and the training unit 63 may determine the state of the feature data based on the degree of anomaly calculated by the anomaly degree calculation unit 62, and label the production performance data or the feature data according to the determined state.

In one embodiment, the training unit 63 may set the warning period based on the average value of the abnormality level in the normal period and the average value of the abnormality level in the abnormal period. Specifically, the training unit 63 first identifies the period in which the defect rate is equal to or greater than the abnormality threshold as the abnormal period, and calculates the average value A of the abnormality level in the normal period and the average value B of the abnormality level in the abnormal period used to train the abnormality calculation model M1. The training unit 63 may then calculate the average value C of the two calculated average values A and B, and set all periods in which the abnormality level exceeds the average value C between the normal period and the abnormal period as warning periods. By setting the normal period, warning period, and abnormal period in this manner, the training unit 63 can label each piece of feature data as a normal period, warning period, or abnormal period.

In one embodiment, the training unit 63 may set the warning period based on the start point and length of the abnormal period. Specifically, the training unit 63 first identifies a period in which the defect rate is equal to or greater than the abnormal threshold as the abnormal period, and identifies a period of the same length as the abnormal period by going back from the start point of the abnormal period. The training unit 63 then calculates an average value A of the degree of abnormality in the normal period and an average value B of the degree of abnormality in the abnormal period, and may set the identified period as the warning period if the average value D of the degree of abnormality in the identified period is within the range between the average value A and the average value B. By setting the normal period, warning period, and abnormal period in this manner, the training unit 63 can label each piece of feature data as a normal period, warning period, or abnormal period.

In one embodiment, the training unit 63 may set the warning period based on the start point of the abnormal period, the average value of the abnormality in the normal period, and the average value of the abnormality in the abnormal period. Specifically, the training unit 63 first identifies the period in which the defect rate is equal to or greater than the abnormality threshold as the abnormal period, and calculates the average value A of the abnormality in the normal period and the average value B of the abnormality in the abnormal period. The training unit 63 may then set the warning period to a period going back from the start point of the abnormal period and in which the average value E of the abnormality in that period is equal to or greater than the average value C of the two calculated average values A and B. By setting the normal period, warning period, and abnormal period in this manner, the training unit 63 can label each piece of feature data as a normal period, warning period, or abnormal period.

The above-mentioned labeling of normal periods, warning periods, and abnormal periods is merely an example, and labeling of training data for an anomaly prediction detection model according to the present disclosure is not limited thereto. For example, the training unit 63 may perform labeling based on a comparison of the degree of abnormality with any appropriately set warning threshold and abnormality threshold. Specifically, when the degree of abnormality calculated for the feature data to be discriminated is less than the warning threshold, the training unit 63 may discriminate the state of the feature data as a normal period and label the feature data as a "normal period."

In addition, if the degree of abnormality calculated for the feature data to be discriminated is equal to or greater than the prognostic threshold and less than the abnormality threshold, the training unit 63 may discriminate the state of the feature data as being in a prognostic period and label the feature data as a "prognostic period."

In addition, if the degree of abnormality calculated for the feature data to be discriminated is equal to or greater than the abnormality threshold, the training unit 63 may discriminate the state of the feature data as being in an abnormal period and label the feature data as an "abnormal period."

The training unit 63 stores each feature data labeled in this manner as state label data, and uses the state label data as training data for the anomaly sign detection model M2.

[Status Label Database 74]
The training unit 63 may store the generated state label data in the state label database 74 of the storage unit 11. That is, the state label database 74 is a database in which state label data LD is registered to time-series data in which the abnormality degree data AD registered in the abnormality degree database 73 is arranged in order of production date and time. The time-series data is divided for each production facility 200.

Figure 10 is a diagram showing an example of the status label database 74. In the example shown in Figure 10, the status label database 74 registers "status label," "equipment," "start date and time," "end date and time," and "anomaly degree calculation model ID" in association with each other. These data are associated one by one to form one status label data.

The "status label" is data that indicates one of three status stages: "normal period," "premonitory period," and "abnormal period." Note that the premonitory period is a period during which a premonitory abnormality appears in the production equipment 200, and the abnormal period is a period during which an abnormality occurs in the production equipment 200.

"Equipment" is the identifier of the production equipment 200 corresponding to the time series data to which the state label L is assigned.

The "start date and time" and "end date and time" are the date and time of the first (oldest) data and the date and time of the last (newest) data of the anomaly data AD that constitutes the time-series data to which the state label L is assigned. In other words, they are the date and time when the period of the state indicated by the state label L starts and ends.

The state label data LD is generated by the training unit 63 and registered in the state label database 74.

The abnormality sign detection model M2 according to an embodiment of the present disclosure is a machine learning model that predicts whether the state of the production equipment 200 is a "normal period", a "premonition period", or an "abnormal period", and receives feature data as input and outputs the state of the feature data. For example, the abnormality sign detection model M2 is a machine learning model for predicting what kind of time-series changes the production performance data PD has undergone before the abnormality occurred in the production equipment 200 in which many defective products have occurred due to abnormalities such as deterioration and failure. In other words, the abnormality sign detection model M2 can be used to detect whether an abnormality such as deterioration and failure has occurred in the production equipment 200, or whether a premonition of an abnormality has appeared. When the production performance data PD of the production equipment 200 is input to the production performance database 71, the abnormality sign detection model M2 outputs a result of determining whether the state of the production equipment 200 at the time when the production performance data PD was obtained is a "normal period", a "premonition period", or an "abnormal period" based on the feature data derived from the input production performance data.

For example, the training unit 63 uses training data consisting of pairs of feature data and the state of the feature data obtained from the production performance data extracted from the training database 20 to train the abnormality sign detection model M2 to be trained by supervised learning. For example, when the abnormality sign detection model M2 is realized as a neural network, the training unit 63 inputs the feature data of each training data to the neural network and compares the output result from the neural network with the state of the feature data of the training data. Then, the training unit 63 adjusts the parameters of the neural network according to backpropagation or the like depending on the error between the output result and the state of the training feature data. The training unit 63 repeats the parameter adjustment process until a predetermined termination condition is satisfied. The predetermined termination condition may be, for example, that the parameter adjustment process has been performed on all prepared training data, that the error has become less than a predetermined threshold, that the error has converged within a predetermined threshold range, or the like.

When the training of the anomaly sign detection model M2 is completed, the training unit 63 provides the trained anomaly sign detection model M2 to the anomaly sign detection device 100.

In addition, one abnormality sign detection model M2 may be generated using the production performance data PD of multiple production facilities 200, or one abnormality sign detection model M2 may be generated using the production performance data PD of one production facility 200.

The format of the anomaly sign detection model M2 is not particularly limited, but in order to further improve the accuracy of the determination, it is desirable to adopt a machine learning model such as a neural network model. The model to be adopted for the anomaly sign detection model M2 may be selected by the user of the anomaly sign detection device 100, or may be selected automatically by the training unit 63.

<Details of the abnormality sign detection device 100>
Next, an abnormality sign detection device 100 according to an embodiment of the present disclosure will be described with reference to Fig. 11. The abnormality sign detection device 100 uses the abnormality sign detection model generated by the model generation device 60 to detect an abnormality or a sign of an abnormality that has occurred in the production facility 200 based on production performance data of the production facility 200 to be detected.

Figure 11 is a block diagram showing the functional configuration of an anomaly sign detection device 100 according to one embodiment of the present disclosure.

As shown in FIG. 11, the abnormality sign detection device 100 has a preprocessing unit 101 and an abnormality sign detection unit 102.

The pre-processing unit 101 acquires the feature quantities of the production performance data of the detection target. Specifically, when the pre-processing unit 101 acquires the production performance data related to the products, etc., produced by the production equipment 200, the pre-processing unit 101 executes a predetermined pre-processing on the acquired production performance data to acquire the feature quantity data of the production performance data.

Here, the production history data acquired by the pre-processing unit 101 is typically of the same type as the data used by the model generating device 60 to generate the anomaly sign detection model. For example, if the model generating device 60 uses the production history data of six data types, "production date and time," "equipment," "first sheet material measurement value," "second sheet material measurement value," "first sheet material measurement value judgment result," and "second sheet material measurement value judgment result," as training data, the pre-processing unit 101 also acquires the production history data of the detection targets of these six data types from the production equipment 200.

Then, the preprocessing unit 101 performs a predetermined preprocessing on the acquired production performance data to acquire feature data of the production performance data. Specifically, the preprocessing unit 101 performs a process similar to the preprocessing in the model generating device 60 on the production performance data of the detection target, and derives feature data of the production performance data. For example, if the preprocessing in the model generating device 60 is a statistical process and the feature data is composed of eight data types, namely, "production date and time", "equipment", "aggregation unit", "maximum value of the first sheet material measurement value", "minimum value of the first sheet material measurement value", "average value of the first sheet material measurement value", "standard deviation of the first sheet material measurement value", and "number of defects of the first sheet material measurement value", the preprocessing unit 101 also performs a similar statistical process on the production performance data of the detection target, and derives feature data of the same type.

When the feature data is acquired in this manner, the preprocessing unit 101 passes the acquired feature data to the anomaly sign detection unit 102. Note that when the preprocessing unit 101 acquires the feature data of the production performance data instead of the production performance data to be detected, the preprocessing unit 101 does not need to execute the process for deriving the feature data described above, and it is sufficient to simply pass the acquired feature data to the anomaly sign detection unit 102.

The anomaly sign detection unit 102 detects the state of the feature data from the feature data using a trained anomaly sign detection model. Here, the anomaly sign detection model is trained with training data consisting of feature data of training production performance data and the state of the feature data. In addition, the state of the feature data is determined based on the degree of anomaly of the feature data calculated using an anomaly degree calculation model that represents a group of features in a normal period.

Specifically, the anomaly sign detection unit 102 inputs the feature data received from the pre-processing unit 101 into a trained anomaly sign detection model, and acquires the state of the input feature data as an output from the trained anomaly sign detection model. As described above in relation to the model generation device 60, the anomaly sign detection model may be any type of machine learning model, such as a neural network, trained by supervised learning using feature data labeled using a trained anomaly degree calculation model as training data.

In one embodiment, the anomaly calculation unit 62 may evaluate the reliability of the anomaly calculation model by using the anomaly data AD registered in the anomaly database 63. For example, the anomaly calculation unit 62 calculates a correlation coefficient C between the defect rate DR, which is obtained by dividing the number of defects in the anomaly data AD by the number of aggregated units W, and the anomaly A. When the correlation coefficient C is greater than a correlation threshold Th _c , which is a predetermined threshold, the anomaly calculation unit 62 can determine that the anomaly A calculated by the anomaly calculation model M1 appropriately represents the degree of anomaly of the production facility 200. Therefore, the anomaly calculation unit 62 may adopt the anomaly calculation model M1 for which the correlation coefficient C is greater than the correlation threshold Th _c as the anomaly calculation model M1 for the production facility 200.

The correlation threshold value Th _c may be the minimum value of the correlation coefficient C at which a correlation is recognized between the degree of abnormality A calculated by the abnormality calculation model M1 and the number of defects of the wound bodies 204 produced by the production facility 200. In other words, it is the minimum value of the correlation coefficient C assumed when a combination of the feature quantities of the feature quantity data FD used to generate the abnormality calculation model M1 represents the characteristics of an abnormality in the production facility 200. In other words, it is the minimum value of the correlation coefficient C at which it is possible to regard the abnormality calculation model M1 as appropriately calculating the degree of abnormality in the production facility 200. The predetermined correlation threshold value Th _c may be empirically determined based on, for example, past production performance data PD or the like.

Then, the training unit 63 performs the following process every time feature data FD is registered in the feature database 72 (i.e., every time abnormality degree data AD is registered in the abnormality degree database 73). That is, the training unit 63 obtains a defect rate DR by dividing the number of defects in the registered feature data FD (or abnormality degree data AD) by the number of aggregated units W. The training unit 63 compares the defect rate DR with a defect rate threshold Th _dr , which is a predetermined threshold. If the defect rate DR is greater than the defect rate threshold Th _dr , the training unit 63 generates condition label data LD by assigning a condition label L of "normal period", "premonitory period", or "abnormal period" to time-series data in which the abnormality degree data AD registered in the abnormality degree database 73 are arranged in order of production date and time. Then, the training unit 63 registers the condition label data LD in the condition label database 74.

Then, the training unit 63 learns the time series data in which the feature data FD is arranged in chronological order of production date and time, and generates an abnormality sign detection model M2. The time series data used for learning starts from feature data FD whose production date and time is included between the start date and time and the end date and time of the state label data LD of each of the "normal period", "premonition period", and "abnormal period", and arranges a predetermined number of past feature data FD from the starting point. The abnormality sign detection model M2 generated by the training unit 63 is provided to the abnormality sign detection device 100, as described above.

In one embodiment, the training unit 63 may compare the determination accuracy of the anomaly degree calculation model M1 adopted by the anomaly degree calculation unit 62 with the anomaly sign detection model M2 generated by the training unit 63, and adopt the anomaly sign detection model M2 if the determination accuracy of the anomaly sign detection model M2 is higher. In other words, the training unit 63 continues to generate the anomaly sign detection model (learn the learning data list) until the determination accuracy of the anomaly sign detection model M2 exceeds the determination accuracy of the anomaly degree calculation model M1. The adopted anomaly sign detection model M2 is used in the anomaly sign detection device 100 to detect an anomaly or a sign of an anomaly in the production equipment 200.

Then, the abnormality sign detection unit 102 judges the state of the production facility 200 using the feature amount data FD generated from the production performance data PD of the wound body 204 newly produced in the production facility 200 and the abnormality sign detection model M2 adopted by the model generation device 60. Specifically, the abnormality sign detection unit 102 judges the production performance data of the detection target into three stages: "normal period", "sign period", and "abnormal period". If the judgment result is "abnormal period" or "sign period", the abnormality sign detection unit 102 may notify the user who manages the production facility 200 of the detection of an abnormality or a sign. This notification may include an alarm to attract the user's attention, a display informing which production facility 200 an abnormality or a sign of an abnormality occurred and when, etc.

<Details of each process for model generation>
Below, a detailed description will be given of each process for generating an anomaly sign detection model by one implementation of the model generation device 60 shown in Fig. 6. Note that the process described below is merely an example, and the generation of an anomaly sign detection model according to the present disclosure may be realized by a generation process according to the production facility 200, the product, etc.

(Pretreatment)
The following describes the pre-processing executed by the pre-processing unit 61. Fig. 12 is a flowchart for explaining the pre-processing according to an embodiment.

In step S11, the pre-processing unit 61 determines whether or not a predetermined number, S, of new production performance data PD have been registered in the production performance database 71. If S pieces of new production performance data PD have not been registered (step S11: NO), the pre-processing unit 61 repeats step S11. If S pieces of new production performance data PD have been registered (step S11: YES), the pre-processing unit 61 advances the process to step S12.

In step S12, the pre-processing unit 61 extracts production performance data PD of the number W of counting units from the newly registered S pieces of production performance data PD, and generates a production performance data list PL in which the extracted production performance data PD is listed. That is, the pre-processing unit 61 extracts the production performance data list PL. The production performance data list PL is a list of production performance data PD extracted from the latest production performance data PD _{last last} registered in the production performance database 71, going back in time from the production date and time of the production performance data PD last to the production performance data PD of the number W of counting units. That is, the production performance data list PL includes at least the latest production performance data PD _last last registered. The number W of counting units is a number equal to or less than a predetermined number S.

In step S13, the preprocessing unit 61 calculates statistics such as the maximum, minimum, average, and standard deviation of the first sheet material measurement value from the extracted production performance data list PL, and sets these statistics as feature quantities. In addition to the statistics shown here, statistics such as the median, first quartile, and third quartile may be calculated and set as feature quantities. Furthermore, when generating feature quantities, preprocessing such as normalization, binning, logarithmic transformation, and difference transformation may be performed in advance, or lag features may be generated by shifting the data a predetermined number of times in the past.

In step S14, the pre-processing unit 61 counts the number of production performance data PD in which the first sheet material measurement value judgment result is "defective" and included in the extracted production performance data list PL, and sets this as the number of defective first sheet material measurement values.

In step S15, the preprocessing unit 61 registers the feature amount data FD including the calculated feature amount and the number of defects in the feature amount database 72. The production date and time of the feature amount data FD is the production date and time of the latest production result data PD _last .

(Generation process of abnormality calculation model)
Next, a description will be given of the generation process of an anomaly degree calculation model executed by the anomaly degree calculation unit 62. Note that, when the anomaly degree calculation unit 62 uses a trained anomaly degree calculation model, the generation process of the anomaly degree calculation model does not need to be executed by the anomaly degree calculation unit 62. Fig. 13 is a flowchart for describing the generation process of an anomaly degree calculation model according to one embodiment.

In step S21, the anomaly calculation unit 62 determines whether or not a predetermined number or more of feature data FD containing a production date and time within a normal period (e.g., one week or a period longer than one week) are registered in the feature database 72. The normal period is a period during which no deterioration or failure occurs in the production equipment 200 and no defects occur in the wound body 204 produced by the production equipment 200.

If the number of feature data FD containing the production date and time within the normal period is less than the predetermined number (step S21: NO), the abnormality calculation unit 62 repeats step S21. If the number of feature data FD containing the production date and time within the normal period is greater than the predetermined number (step S21: YES), the abnormality calculation unit 62 proceeds to step S22. In other words, the abnormality calculation unit 62 waits for the predetermined number of feature data FDs with the "number of defective measurement values of the first sheet material" of 0 to be registered in the feature database 72, and then proceeds to step S22.

In step S22, the abnormality degree calculation unit 62 extracts a predetermined number of feature data FD from the feature database 72, and generates a feature data list FL that lists the extracted feature data FD. In other words, the abnormality degree calculation unit 62 extracts the feature data list FL. The feature data list FL is a list of a predetermined number of feature data FD that are registered in the feature database 72 and whose production date and time fall within the normal period described above.

In step S23, the anomaly calculation unit 62 generates an anomaly calculation model M1 by using (e.g., learning) a combination of multiple features possessed by the feature data FD included in the feature data list FL. For example, the anomaly calculation unit 62 may generate the anomaly calculation model M1 according to any known anomaly sign detection technology (e.g., MT method, etc.). For example, the anomaly calculation model M1 may calculate the degree of deviation (e.g., Mahalanobis distance, etc.) from a group of features in a normal state as the degree of anomaly of the feature data FD. In this way, the anomaly calculation unit 62 may generate multiple anomaly calculation models M1 by changing the way in which multiple features are combined.

In step S24, the anomaly calculation unit 62 registers the multiple anomaly calculation models M1 that have been generated.

(Abnormality Degree Calculation Process)
Next, a description will be given of the abnormality degree calculation process executed by the abnormality degree calculation unit 62. Fig. 14 is a flowchart for explaining the abnormality degree calculation process according to one embodiment.

In step S31, the abnormality degree calculation unit 62 determines whether or not new feature amount data FD _new has been registered in the feature amount database 72. If new feature amount data FD _new has not been registered (step S31: NO), the abnormality degree calculation unit 62 repeats step S31. If new feature amount data FD _new has been registered (step S31: YES), the abnormality degree calculation unit 62 advances the process to step S32.

In step S32, the abnormality degree calculation unit 62 extracts the newly registered feature amount data FD _new from the feature amount database 72.

In step S33, the abnormality degree calculation unit 62 reads out a plurality of registered abnormality degree calculation models M1, and calculates a plurality of abnormality degrees A using each read out abnormality degree calculation model M1 and the feature contained in the extracted feature amount data FD _new . That is, a plurality of abnormality degrees A are calculated from one feature amount data FD _new . The type of feature amount used to calculate the abnormality degree A is the same as the type of feature amount used (e.g., learned) when generating the abnormality degree calculation model M1 used to calculate the abnormality degree A.

In step S34, the abnormality degree calculation unit 62 generates abnormality degree data AD new including the calculated abnormality degree A and the number of defects included in the feature amount data FD _new , and registers the generated abnormality degree data AD _new in the abnormality degree database 73. The production date and time of the abnormality degree data AD _new is the production date and time of the feature amount data FD _new used in the calculation of the abnormality degree A.

The abnormality degree data AD _new includes the ID of the abnormality degree calculation model M1 used to calculate the abnormality degree A. Therefore, it is possible to identify which abnormality degree calculation model M1 was used to calculate the abnormality degree A.

(Processing for determining whether anomaly calculation model is adopted)
Next, a description will be given of the adoption determination process of an anomaly degree calculation model executed by the anomaly degree calculation unit 62. Note that, if the anomaly degree calculation unit 62 is equipped with only one anomaly degree calculation model, the adoption determination process does not need to be executed by the anomaly degree calculation unit 62. Fig. 15 is a flowchart for describing the adoption determination process of an anomaly degree calculation model according to one embodiment.

In step S41, the abnormality degree calculation unit 62 judges whether or not new abnormality degree data AD _new in which the number of defects in the first sheet material measurement value is 1 or more has been registered in the abnormality degree database 73. If new abnormality degree data AD _new in which the number of defects in the first sheet material measurement value is 1 or more has not been registered (step S41: NO), the abnormality degree calculation unit 62 repeats step S41. If new abnormality degree data AD _new in which the number of defects in the first sheet material measurement value is 1 or more has been registered (step S41: YES), the abnormality degree calculation unit 62 advances the process to step S42.

In step S42, the abnormality degree calculation unit 62 extracts a predetermined number of abnormality degree data AD from the abnormality degree database 73, and generates an abnormality degree data list AL in which the extracted abnormality degree data AD is listed. That is, the abnormality degree calculation unit 62 extracts the abnormality degree data list AL. The abnormality degree data list AL is a list of the abnormality degree data AD that is extracted from the production date and time of the abnormality degree data AD _new up to a predetermined number of abnormality degree data AD, with the abnormality degree data AD _new newly registered in the abnormality degree database 73 being the first data. The abnormality degree data AD extracted at this time is the abnormality degree data AD that has the same abnormality degree calculation model ID as the abnormality degree calculation model ID of the abnormality degree data AD _new . In other words, the abnormality degree calculation unit 62 extracts the abnormality degree data list AL for each abnormality degree calculation model ID that has calculated the abnormality degree A. The abnormality level data list AL includes at least newly registered abnormality level data AD _new in which the number of defects in the first sheet material measurement value is one or more.

In step S43, the anomaly calculation unit 62 calculates a defect rate DR for each anomaly data AD included in the anomaly data list AL. The defect rate DR is a value obtained by dividing the number of defects in the first sheet material measurement value included in the anomaly data AD by the number of counting units W. Next, the anomaly calculation unit 62 calculates a correlation coefficient C between the defect rate DR and the anomaly A. The calculation of the correlation coefficient C is performed for each anomaly calculation model M1, that is, for each anomaly calculation model ID. The anomaly calculation unit 62 selects the maximum correlation coefficient C from the multiple correlation coefficients C calculated for each anomaly calculation model ID.

In step S44, the abnormality calculation unit 62 determines whether the selected maximum correlation coefficient C is greater than the correlation threshold Th _c . If the maximum correlation coefficient C is equal to or less than the correlation threshold Th _c (step S44: NO), the abnormality calculation unit 62 returns the process to step S41. If the maximum correlation coefficient C is greater than the correlation threshold Th _c (step S44: YES), the abnormality calculation unit 62 advances the process to step S45. As described above, the correlation threshold Th _c is the minimum value of the correlation coefficient C at which a correlation is recognized between the abnormality A calculated by the abnormality calculation model M1 and the number of defects of the wound body 204 produced by the production facility 200.

In step S45, the anomaly calculation unit 62 adopts the anomaly calculation model M1 that calculated the maximum correlation coefficient C.

Note that even after adopting the anomaly calculation model M1 in step S45, the anomaly calculation unit 62 may repeat the processes from step S41 onwards. As a result, there is a possibility that an anomaly calculation model M1' different from the adopted anomaly calculation model M1 will calculate the anomaly A used to calculate the maximum correlation coefficient C. In this case, the anomaly calculation unit 62 may adopt the anomaly calculation model M1' instead of the anomaly calculation model M1. In this case, each process executed by the training unit 63 after the process executed by the anomaly calculation unit 62 is executed based on the replaced anomaly calculation model M1'.

(State label setting process)
Next, a description will be given of the state label setting process executed by the training unit 63. Fig. 16 is a flowchart for explaining the state label setting process.

In step S51, the training unit 63 judges whether or not feature data FD with a defect rate DR greater than the defect rate threshold value Th _dr has been registered in the feature database 72. If feature data FD with a defect rate DR greater than the defect rate threshold value Th _dr has not been registered (step S51: NO), the training unit 63 repeats step S51. If feature data FD with a defect rate DR greater than the defect rate threshold value Th _dr has been registered (step S51: YES), the training unit 63 advances the process to step S52.

In step S52, the training unit 63 sets the period of production dates and times of the feature amount data FD used when the abnormality calculation model M1 adopted in the abnormality calculation model adoption determination process was generated by the abnormality calculation unit 62 as the normal period. The identifier of the target production equipment 200, the start date and time and the end date and time of the set normal period, and the abnormality calculation model ID of the adopted abnormality calculation model M1 are associated with each other to generate state label data LD _normal for the normal period. In addition, a period other than the abnormal period or the predictive period described below can be set as the normal period. For example, a period that was not used in generating the abnormality calculation model M1 but that precedes the period set as the predictive period may be set as the normal period.

In step S53, the training unit 63 sets, as the abnormal period, the production date and time period of the feature amount data FD in which the defect rate DR is greater than the defect rate threshold value Th _dr , which is registered in the feature amount database 72. The training unit 63 generates state label data LD anomaly for the abnormal period by associating the identifier of the target production equipment 200, the start date and time and the end date and time of the set abnormal period, and the abnormality degree calculation model ID of the adopted _abnormality degree calculation model M1.

In step S54, the training unit 63 sets the start date and time of the abnormal period condition label data LD _anomaly as the end date and time, and sets the warning period to a period whose start date and time is a date and time going back a predetermined period from the end date and time. The training unit 63 generates the warning period condition label data LD sign by associating the _identifier of the target production equipment 200, the start date and time and the end date and time of the set warning period, and the abnormality degree calculation model ID of the adopted abnormality degree calculation model M1.

In step S55, the training unit 63 performs the following process based on the abnormality degree data AD that matches the identifier of the target equipment and the abnormality degree calculation model ID. That is, the training unit 63 extracts from the abnormality degree database 73 all the abnormality degree data AD whose production date and time are included between the start date and time and the end date and time of _{the state} label data LD normal in the normal period, and calculates the average value of the abnormality degree A (i.e., the average value of the abnormality degree A in the normal period) MA _normal . The training unit 63 also extracts from the abnormality degree database 73 all the abnormality degree data AD whose production date and time are included between the start date and time and the end date and time of the state label data LD _sign in the predictive period, and calculates the average value of the abnormality degree A (i.e., the average value of the abnormality degree A in the predictive period) MA _sign . Furthermore, the training unit 63 extracts from the abnormality database 73 all the abnormality data AD whose production date and time is included between the start date and time and the end date and time of the condition label data LD _anomaly of the abnormal period, and calculates the average value of the abnormality degrees A (i.e., the average value of the abnormality degrees A during the abnormal period) MA _anomaly .

In step S56, the training unit 63 judges whether the magnitude relationship between the average value MA _normal of the abnormality degree A in the normal period, the average value MA _sign of the abnormality degree A in the predictive period, and the average value MA _abnormality of the _abnormality degree A in the abnormal period is MA _normal < MA _sign < MA abnormal. If MA _normal < MA _sign < MA _abnormal is not satisfied (step S56: NO), the training unit 63 ends the process and discards the generated state label data LD _normal , LD _sign , and LD _abnormal .

On the other hand, if MA _normal < MA _sign < MA _anomaly (step S56: YES), the training unit 63 proceeds to step S57. In step S57, the training unit 63 registers the generated state label data LD _normal , LD _sign , and LD _anomaly in the state label database 74.

(Generation process of anomaly sign detection model)
Next, a description will be given of a generation process of an abnormality sign detection model executed by the training unit 63. Fig. 17 is a flowchart for describing a generation process of an abnormality sign detection model according to an embodiment.

In step S61, the training unit 63 extracts all state label data LD from the state label database 74.

In step S62, the training unit 63 extracts a predetermined number of feature data FD including a production date and time between the start date and time and the end date and time of the extracted normal period condition label data LD _normal from the feature database 72. The training unit 63 generates a normal feature data list FL _{normal using the extracted predetermined number of feature data FD. If the feature data FD does not reach the predetermined number, the training unit 63 generates a normal feature data list FL normal} using all feature data FD including a production date and time between the start date and time and the end date and time of the extracted _normal period condition label data LD _normal .

In step S63, the training unit 63 extracts a predetermined number of feature data FD including a production date and time between the start date and time and the end date and time of the extracted state label data LD _sign of the predictive period from the feature database 62. The training unit 63 generates a predictive feature data list FL _{sign using the extracted predetermined number of feature data FD. If the feature data FD does not reach the predetermined number, the training unit 63 generates a predictive feature data list FL sign using} all feature data FD including a production date and time between the start date and time and the end date and time of the extracted state label data LD _sign of the predictive period.

In step S64, the training unit 63 extracts a predetermined number of feature data FD including a production date and time between the start date and time and the end date and time of the extracted condition label data LD _anomaly from the feature database 72. The training unit 63 generates an abnormal feature data list FL _anomaly using the extracted predetermined number of feature data FD. If the feature data FD does not reach the predetermined number, the training unit 63 extracts all feature data FD including a production date and time between the start date and time and the end date and time of the extracted condition label data LD _anomaly , and generates an abnormal feature data list FL _anomaly .

In step S65, the training unit 63 generates feature data FD _labeled by assigning three-level state _labels L, namely, "normal period", "premonitory period" and "abnormal period", to all feature data FD included in the normal feature data list FL _normal , the predictive feature data list FL sign and the abnormal feature data _list FL abnormal . The training unit 63 extracts a learning data list TL from the feature database 62 based on each feature data FD _labeled . The learning data list TL is a list obtained by arranging the feature data FD in chronological order and extracting and listing feature data FD from a predetermined number of feature data FD before the feature data FD _labeled to which the state label L has been assigned, to the feature data FD _labeled to which the state label L has been assigned. That is, the feature amount data FD _labeled to which the condition label L is assigned is the latest feature amount data FD among the feature amount data FD included in the learning data list TL. Note that the learning data list TL may include feature amount data FD with a production date and time newer than that of the feature amount data FD _labeled . In any case, the learning data list TL includes at least the feature amount data FD _labeled . Furthermore, the number of learning data lists TL to be generated may be equal to the number of feature amount data FD ( _i.e. , feature amount data FD _labeled ) included in the normal feature amount data list FL _normal , the predictive feature amount data list FL _sign , and the abnormal feature amount data list FL anomaly.

The learning data list TL will be described with reference to Fig. 18, taking as an example a case where feature amount data FD _labeled to which a state label L of "premonition period" is assigned is generated. Fig. 18 is a diagram for explaining the learning data list.

In the illustrated example, the predictive feature data list FL _sign includes a plurality of feature data FD. A predictive period state label L is assigned to all of the plurality of feature data FD, and all of these feature data FD become feature data FD _labeled . That is, each feature data labeled for training data may be composed of a group of feature amounts that overlap with respect to the time axis. In this way, feature data labeled with a plurality of "predictive periods" and "abnormal periods" can be generated from one predictive period and abnormal period, and the number of training data corresponding to a state that occurs less frequently can be increased. A predetermined number of feature data FD arranged in time series are selected from the feature data FD arranged in time series, regardless of which feature data list FL they are included in. Among the feature data FD selected at this time, the feature data FD with the latest production date and time is the feature data FD _labeled . The learning data list TL is composed of the predetermined number of feature data FD selected in this way.

The feature data FD included in the normal feature data list FL _normal and the abnormal feature data list FL _anomaly , like the feature data FD included in the predictive feature data list FL _sign , are all set as feature data FD _labeled . Then, a learning data list TL including these feature data FD _labeled is generated and extracted.

That is, the learning data list TL includes multiple pieces of feature amount data FD arranged in a time series, which are generated from the multiple pieces of production result data PD arranged in a time series. The learning data list TL also includes feature amount data FD _labeled . The feature amount data FD _labeled is representative feature amount data representing the multiple pieces of feature amount data FD included in the learning data list TL. The representative feature amount data FD _labeled is assigned a status label L of "normal period", "premonitory period", or "abnormal period".

17 again. In step S66, the training unit 63 performs machine learning using all of the learning data list TL to generate an anomaly sign detection model M2. The features and their combinations used at this time are the features and their combinations used at the time of generating the anomaly degree calculation model M1 adopted by the anomaly degree calculation unit 62. In learning, supervised learning is performed in which the state label L assigned to the representative feature data FD _labeled is taken as the correct answer and all of the feature data FD included in the learning data list TL is accepted as input.

As described above, the abnormality sign detection model M2 is a model that learns the characteristics of time-series changes in the production equipment 200 leading up to an abnormality or a sign of an abnormality. The abnormality sign detection model M2 is used to detect abnormalities and signs of an abnormality in the production equipment 200.

In addition, a single abnormality sign detection model M2 may be generated by learning the learning data list TL generated based on the production performance data PD of multiple production facilities 200, or a different abnormality sign detection model M2 may be generated for each production facility 200.

In step S67, the training unit 63 registers one or more generated anomaly sign detection models M2.

(Processing for determining whether anomaly prediction detection model is adopted)
Next, a description will be given of the adoption determination process for an anomaly sign detection model executed by the training unit 63. Note that if the training unit 63 generates only one anomaly sign detection model, this adoption determination process is not necessary. FIG. 19 is a flowchart for describing the process executed by the anomaly sign detection model adoption determination unit 127.

In step S71, the training unit 63 extracts all state label data LD from the state label database 74.

In step S72, the training unit 63 extracts a predetermined number of feature data FD including a production date and time between the start date and time and the end date and time of the extracted "normal period" state label data LD _normal from the feature database 72. The training unit 63 generates a normal feature data list EFL _normal for evaluation using the extracted predetermined number of feature data FD. If the feature data FD does not reach the predetermined number, the training unit 63 generates a normal feature data list EFL normal for evaluation using all feature data FD including a production date and time between the start date and time and the end date and time of the extracted normal period state label data LD _{normal .}

It is preferable that the feature data FD included in the normal feature data list EFL _normal for evaluation generated in step S72 be different from the feature data FD included in the normal feature data list FL _normal generated in step S62 of FIG. 17.

In step S73, the training unit 63 extracts a predetermined number of feature data FD including a production date and time between the start date and time and the end date and time of the extracted state label data LD _sign of the predictive period from the feature database 72. The training unit 63 generates a predictive feature data list EFL _sign for evaluation using the extracted predetermined number of feature data FD. If the feature data FD does not reach the predetermined number, the training unit 63 generates a predictive feature data list EFL sign for evaluation using all feature data FD including a production date and time between the start date and time and the end date and time of the extracted state label _data LD _sign of the predictive period.

It is preferable that the feature data FD included in the predictive feature data list EFL _sign for evaluation generated in step S73 be different from the feature data FD included in the predictive feature data list FL _sign generated in step S63 of FIG. 17.

In step S74, the training unit 63 extracts a predetermined number of feature data FD including a production date and time between the start date and time and the end date and time of the extracted "abnormal period" condition label data LD _anomaly from the feature database 72. The training unit 63 generates an abnormal feature data list EFL _{anomaly for evaluation using the extracted predetermined number of feature data FD. If the feature data FD does not reach the predetermined number, the training unit 63 generates an abnormal feature data list EFL anomaly} for evaluation using all feature data FD including a production date and time between the start date and time and the end date and time of the extracted condition label data LD _anomaly for the _abnormal period.

It is preferable that the feature data FD included in the abnormal feature data list EFL _anomaly for evaluation generated in step S74 be different from the feature data FD included in the abnormal feature data list FL _anomaly generated in step S64 in FIG. 17.

In step S75, the training unit 63 generates feature data FD labeled by assigning three-level state labels L, namely, " _normal period", "premonitory period" and "abnormal period", to all feature data FD included in the normal feature data _list for evaluation EFL _normal , the predictive feature data list for evaluation EFL _sign and the abnormal feature data list for evaluation EFL abnormality. The training unit 63 extracts an evaluation data list EL from the feature database 72 based on each feature data FD _labeled . The evaluation data list EL is a list obtained by arranging the feature data FD in chronological order and extracting and listing feature data FD from a predetermined number of feature data FDs before the feature data FD _labeled to which the state label L has been assigned, to the feature data FD _labeled to which the state label L has been assigned. That is, the evaluation data list EL includes at least the feature amount data FD _labeled . The number of evaluation data lists EL to be generated may be equal to the number of feature amount data FD _(i.e. , feature amount data FD labeled) included in the evaluation normal feature amount data list EFL _normal , the evaluation predictive feature amount data list EFL _sign , and the evaluation abnormal feature amount data _list EFL anomaly.

That is, the evaluation data list EL is a list of the same format as the learning data list TL described with reference to Fig. 18. The evaluation data list EL is also common to the learning data list TL in that it includes feature amount data FD _labeled to which a state label L is assigned.

In step S76, the training unit 63 inputs all of the extracted evaluation data lists EL to the newly registered abnormality sign detection model M2 _new . The abnormality sign detection model M2 _new then outputs state labels L equal to the number of evaluation data lists EL input. The training unit 63 counts the number of matches (i.e., correct answers) C2 between the state labels L assigned to the feature data FD _labeled included in the evaluation data list EL and the state labels L output from the abnormality sign detection model M2 _new .

In step S77, the training unit 63 extracts the anomaly degree calculation model M1 that calculated the anomaly degree A used by the anomaly sign detection model M2 _new for learning. The anomaly degree A used by the anomaly sign detection model M2 _new for learning is the anomaly degree A included in the anomaly degree data AD extracted by the training unit 63 in step S55. The training unit 63 inputs the feature amount data FD _labeled included in all the evaluation data lists EL to the extracted anomaly degree calculation model M1. Then, the anomaly degree calculation model M1 calculates the anomaly degree A for the number of the input feature amount data FD _labeled . The anomaly degree A calculated by the anomaly degree calculation model M1 is classified as follows. If the anomaly degree A is greater than the anomaly degree threshold ThA _anomaly , which is a predetermined threshold, the anomaly degree A is classified as an "abnormal period". When the degree of abnormality A is greater than a predictive threshold _ThAsign , which is a predetermined threshold, and is equal to or less than the abnormality threshold _ThAanomaly , the degree of abnormality A is classified as a "predictive period." When the degree of abnormality A is equal to or less than the predictive threshold _TAsign , the degree of abnormality A is classified as a "normal period." The training unit 63 counts the number of cases C1 in which the state label L assigned to the feature data _FDlabeled matches the classification result of the degree of abnormality A calculated by the abnormality calculation model M1 (i.e., the number of correct answers).

For the abnormality threshold ThA _anomaly , for example, MA _anomaly , which is the average value of the abnormality A in the abnormal period calculated in step S55, may be used. For the predictor threshold ThA _sign , for example, MA _sign , which is the average value of the abnormality A in the predictor period calculated in step S55, may be used. This allows the training unit 63 to compare the number of matches C1 and the number of matches C2 using the same criteria in step S78 described later.

In step S78, the training unit 63 compares whether the number of matches C2 is greater than the number of matches C1. If the number of matches C2 is greater than the number of matches C1 (step S78: YES), the training unit 63 proceeds to step S79. If the number of matches C2 is equal to or less than the number of matches C1 (step S78: NO), the anomaly sign detection model M2 _new is not adopted, and the training unit 63 ends the process.

In step S79, the training unit 63 adopts the abnormality sign detection model M2 _new .

The adopted abnormality sign detection model M2 _new is provided to the abnormality sign detection device 100 and used to detect abnormalities and signs of abnormalities in the production equipment 200.

If the anomaly sign detection model M2 _new is not adopted (step S78: NO), a new anomaly sign detection model M2 _new is generated again by the training unit 63 as the feature amount data FD accumulates, and the training unit 63 determines whether or not to adopt this new anomaly sign detection model M2 _new . That is, machine learning for generating the anomaly sign detection model M2 is continued until the number of matches C2 becomes larger than the number of matches C1. In this way, the model generating device 60 can collect production performance data even after the production equipment 200 is in operation, and retrain the anomaly sign detection model M2 to improve detection accuracy by using the production performance data collected during operation.

<Details of each process for abnormality sign detection process>
The following describes in detail the abnormality sign detection process according to one embodiment of the abnormality sign detection device 100 shown in Fig. 11. Here, the pre-processing of the abnormality sign detection device 100 is similar to that of the model generation device 60, and therefore will not be described to avoid duplication. Note that the process described below is merely an example, and the abnormality sign detection process according to the present disclosure may be realized by different processes depending on the abnormality sign detection model generated by the model generation device 60, the production facility 200, the product, etc.

(Abnormality sign detection process)
Next, a description will be given of the abnormality sign detection process executed by the abnormality sign detection unit 102. Fig. 20 is a flowchart for explaining the abnormality sign detection process according to one embodiment.

In step S81, the abnormality sign detection unit 102 determines whether or not new feature data FD _new has been registered in the feature database 72. If new feature data FD _new has not been registered (step S81: NO), the abnormality sign detection unit 102 repeats step S81. If new feature data FD _new has been registered (step S81: YES), the abnormality sign detection unit 102 proceeds to step S82.

In step S82, the abnormality sign detection unit 102 extracts a feature data list FL from the feature database 72 based on the newly registered feature data FD _new . The feature data list FL is a list obtained by arranging the feature data FD in chronological order and extracting the feature data FD from a predetermined number of feature data _FD before the newly registered feature data FD new up to the newly registered feature data FD _new . In other words, the feature data list FL includes at least the newly registered feature data FD _new .

In other words, the feature data list FL is a data list in the same format as the learning data list TL described with reference to FIG. 18.

In step S83, the abnormality sign detection unit 102 inputs the feature amount data list FL to the abnormality sign detection model M2, which then outputs a plurality of state labels L _new .

In step S84, the abnormality sign detection unit 102 determines whether the status label L _new is a "premonition period". If the status label L _new is a "premonition period" (step S84: YES), the abnormality sign detection unit 102 proceeds to step S85. Note that the premonition period is a period during which no abnormality has occurred in the production equipment 200, but a premonition of an abnormality is expected to appear.

In step S85, the abnormality sign detection unit 102 notifies the manager or the like of the production equipment 200 that a sign of an abnormality has appeared. In step S85, the abnormality sign detection unit 102 may display the production equipment 200 that has been determined to be in the "sign period" on the terminal of the manager or the like. For example, the abnormality sign detection unit 102 may display the production equipment 200 that is determined to have a sign of an abnormality along with a message such as "The following production equipment has a sign of an abnormality!".

On the other hand, if the status label L _new is not "premonition period" (step S84: NO), the abnormality sign detection unit 102 proceeds to step S86.

In step S86, the abnormality sign detection unit 102 determines whether the state label L _new is an “abnormal period.” Note that the abnormal period is a period in which an abnormality is assumed to occur in the production equipment 200 beyond the stage of an abnormality sign.

If the state label L _new is "abnormal period" (step S86: YES), the abnormality sign detection unit 102 advances the process to step S87. In step S87, the abnormality sign detection unit 102 may display the production equipment 200 determined to be in the "abnormal period" on the terminal of the manager or the like. For example, the abnormality sign detection unit 102 may display the production equipment 200 determined to have an abnormality together with a message such as "An abnormality has occurred in the following production equipment!". In addition, the abnormality sign detection unit 102 may issue an alarm to notify the manager or the like that an abnormality has occurred in the target production equipment 200. A situation in which an abnormality has occurred in the production equipment 200 is a situation requiring emergency maintenance work. For this reason, the abnormality sign detection unit 102 may promptly notify the user of the abnormality sign detection device 100 of the occurrence of an abnormality not only by displaying a message but also by outputting a sound or light by an alarm.

If the status label L _new is not an "abnormal period" (step S86: NO), the abnormality sign detection unit 102 ends the process.

<Functions and Effects of the Model Generation Device 60 and the Anomaly Sign Detection Device 100 of the First Embodiment>
As described above, the model generation device 60 and the anomaly sign detection device 100 according to the first embodiment perform the following steps: generating feature amount data FD from the production performance data PD of the production facility 200; generating an anomaly degree calculation model M1 that calculates the anomaly degree A of the production facility 200 from the feature amount data FD by learning or the like; determining whether the generated anomaly degree calculation model M1 appropriately calculates the degree of anomaly of the production facility 200; and generating a state label that sets the state of the production facility 200 into three stages, namely, a “normal period”, a “premonition period”, and an “abnormal period”, based on the anomaly degree A calculated by the anomaly degree calculation model M1. a process of generating, by learning from the feature data FD, an abnormality sign detection model M2 that detects an abnormality or a sign of an abnormality in the production equipment 200 based on the set state label L; a process of comparing the determination accuracy of the anomaly degree calculation model M1 and the abnormality sign detection model M2, and adopting the abnormality sign detection model M2 when the determination accuracy of the abnormality sign detection model M2 is higher; and a process of detecting whether or not an abnormality or a sign of an abnormality has occurred in the production equipment 200 using the adopted abnormality sign detection model M2, and notifying the user when an abnormality or a sign of an abnormality has occurred.

Therefore, even when there are few abnormalities such as deterioration and failure of the production equipment 200, feature amount data that appropriately represents features related to abnormalities such as deterioration and failure of the production equipment 200 or signs of abnormalities can be extracted, and a highly accurate trained model (anomaly sign detection model M2) can be generated using the extracted feature amount data. Furthermore, when an abnormality of the production equipment 200 is detected using the trained model, an alarm can be issued to allow the user to take emergency action. Furthermore, when a sign of an abnormality is detected using the trained model, the user can be notified of the production equipment 200 that has detected a sign of an abnormality, and maintenance work can be performed to prevent the deterioration and failure of the production equipment 200 from progressing while the incidence rate of defective products is low.

The abnormality prediction detection model M2 is a trained model generated by machine learning using the learning data list TL. When a feature data list FL, which is a data list in the same format as the learning data list TL, is input, the abnormality prediction detection model M2 outputs the result of determining whether the production equipment 200 is in a "normal period," a "premonition period," or an "abnormal period."

According to the present disclosure, by labeling unlabeled training data using an anomaly degree calculation model that can be generated from a relatively small amount of labeled training data, an anomaly sign detection model M2 can be generated before a large amount of production performance data PD for the warning period and the abnormal period is accumulated. Therefore, even if a warning sign of an anomaly appears or an anomaly occurs relatively soon after the production equipment 200 starts producing products, a trained anomaly sign detection model that can reliably detect such an anomaly can be used. Therefore, it is possible to detect the appearance of a warning sign of an anomaly and the occurrence of an anomaly relatively soon after the production equipment 200 starts producing products.

Second Embodiment
A second embodiment of the present disclosure will be described below. Fig. 21 is a block diagram illustrating a functional configuration of a model generation device 60A according to the second embodiment. The model generation device 60A according to the second embodiment differs from the model generation device 60 according to the first embodiment in the following respects. That is, the process performed by the training unit 63A is different from the process performed by the training unit 63.

The following describes the differences from the first embodiment. Configurations similar to those in the first embodiment are described with the same reference numerals as in the first embodiment, and configurations different from those in the first embodiment are described with the reference numeral followed by the letter "A."

In the first embodiment, the training unit 63 sets a certain warning period immediately before the abnormality period, on the assumption that warning signs of an abnormality appear during a certain period immediately before the occurrence of an abnormality such as deterioration or a breakdown in the production equipment 200. This method is effective when the deterioration pattern of the production equipment 200 is always the same. However, in reality, due to various factors, deterioration may progress more rapidly than expected, and the period during which warning signs of an abnormality appear may become shorter. In this case, the feature amount data FD in which no warning signs of an abnormality appear will be included in the feature amount data FD in the warning period more than expected, and the generated abnormality warning detection model M2 will lack the performance to detect warning signs of an abnormality. The second embodiment generates an abnormality warning detection model M2 that can handle such cases.

Figure 22 is a flowchart explaining the state label setting process according to the second embodiment.

In step S91, the training unit 63A judges whether or not feature data FD with a defect rate DR greater than the defect rate threshold value Th _dr has been registered in the feature database 72. If feature data FD with a defect rate _DR greater than the defect rate threshold value Th dr has not been registered (step S91: NO), the training unit 63A repeats step S91. If feature data FD with a defect rate DR greater than a predetermined defect rate threshold value Th _dr has been registered (step S91: YES), the training unit 63A advances the process to step S92.

In step S92, the training unit 63A sets the period of production dates and times of the feature amount data FD used when the abnormality degree calculation model M1 adopted by the abnormality degree calculation unit 62 was generated by the abnormality degree calculation unit 62 as the normal period. The identifier of the target production equipment 200, the start date and time and the end date and time of the set normal period, and the abnormality degree calculation model ID of the adopted abnormality degree calculation model M1 are associated with each other to generate state label data LD _normal for the normal period. In addition, a period other than the abnormal period or the predictive period described below can be set as the normal period. For example, a period that was not used in generating the abnormality degree calculation model M1 but that precedes the period set as the predictive period can be set as the normal period.

In step S93, the training unit 63A sets, as the abnormal period, the production date and time period of the feature amount data FD in which the defect rate DR is greater than the defect rate threshold value Th _dr , which is registered in the feature amount database 72. The training unit 63A generates state label data LD anomaly for the abnormal period by associating the identifier of the target production equipment 200, the start date and time and the end date and time of the set abnormal period, and the abnormality degree calculation model ID of the adopted _abnormality degree calculation model M1.

In step S94, the training unit 63A performs the following process based on the abnormality degree data AD that matches the identifier of the target equipment and the abnormality degree calculation model ID. That is, the training unit 63A extracts from the abnormality degree database 73 all the abnormality degree data AD whose production date and time are included between the start date and time and the end date and time of _{the state label data LD normal} in the normal period, and calculates the average value of the abnormality degree A (average value of the abnormality degree A in the normal period) MA _normal . The training unit 63A also extracts from the abnormality degree database 73 all the abnormality degree data AD whose production date and time are included between the start date and time and the end date and time of the state label data LD _abnormality in the abnormal period, and calculates the average value of the abnormality degree A (average value of the abnormality degree A in the abnormal period) MA _anomaly .

In step S95, the training unit 63A calculates the average value of the average value MA _normal of the abnormality degree A in the normal period and the average value MA _anomaly of the abnormality degree A in the abnormal period. The training unit 63A sets the calculated average value as the precursor threshold value ThA _sign for the precursor period.

In step S96, the training unit 63A sets the start date and time of the abnormal period as the end date and time, sets a period whose start date and time is a date going back from the end date and time as the warning period, and generates state label data LD _sign for the warning period. The training unit 63A sets the warning period to the longest period during which the average value MA _sign of the abnormality levels A of all the abnormality level data AD whose production date and time is included between the start date and time and the end date and time of the warning period is greater than the warning threshold value ThA _sign for the warning period. In other words, the training unit 63A sets the start date and time of the warning period so as to go back as far as possible from the start date and time of the abnormal period as a starting point.

In step S97, the training unit 63A registers the generated condition label data LD _normal , LD _sign , and LD _abnormality in the condition label database 74.

As described above, according to the model generation device 60A of the second embodiment, even if the period during which signs of abnormality appear is not constant and can be short, it is possible to appropriately set the period during which signs of abnormality appear and generate a highly accurate trained model (anomaly sign detection model M2).

Third Embodiment
A third embodiment of the present disclosure will be described below. Fig. 23 is a block diagram illustrating a functional configuration of a model generation device 60B according to the third embodiment. The model generation device 60B according to the third embodiment differs from the model generation device 60 according to the first embodiment in the following respects. That is, the process performed by the training unit 63B is different from the process performed by the training unit 63.

The following describes the differences from the first embodiment. Configurations similar to those in the first embodiment are described with the same reference numerals as in the first embodiment, and configurations different from those in the first embodiment are described with the reference numeral followed by the letter "B."

In the first embodiment, the training unit 63 sets a certain warning period immediately before the abnormality period, on the assumption that a warning sign of an abnormality appears during a certain period immediately before the occurrence of an abnormality such as deterioration or a breakdown in the production equipment 200. This method is effective when the warning sign of an abnormality becomes clear as the deterioration of the production equipment 200 progresses, and eventually leads to the occurrence of an abnormality. However, in reality, due to various factors, the period immediately before the abnormality period may include a period in which no warning signs of an abnormality appear. In this case, the feature amount data FD in which no warning signs of an abnormality appear will be included in the feature amount data FD of the warning period, and the generated abnormality warning detection model M2 will lack the performance to detect warning signs of an abnormality. The third embodiment generates an abnormality warning detection model M2 that can handle such a case.

Figure 24 is a flowchart explaining the state label setting process according to the third embodiment.

In step S101, the training unit 63B judges whether or not feature data FD with a defect rate DR greater than a defect rate threshold Th _dr has been registered in the feature database 62. If feature data with a defect rate DR greater than a predetermined defect rate threshold Th _dr has not been registered (step S101: NO), the training unit 63B repeats step S101. If feature data FD with a defect rate DR greater than a predetermined defect rate threshold Th _dr has been registered (step S101: YES), the training unit 63B advances the process to step S102.

In step S102, the training unit 63B sets the period of production dates and times of the feature amount data FD used when the abnormality degree calculation model M1 adopted by the abnormality degree calculation unit 62 was generated by the abnormality degree calculation unit 62 as the normal period. The identifier of the target production equipment 200, the start date and time and the end date and time of the set normal period, and the abnormality degree calculation model ID of the adopted abnormality degree calculation model M1 are associated with each other to generate state label data LD _normal for the normal period. In addition, a period other than the abnormal period or the predictive period described below can be set as the normal period. For example, a period that was not used in generating the abnormality degree calculation model M1 but that precedes the period set as the predictive period can be set as the normal period.

In step S103, the training unit 63B sets, as the abnormal period, the production date and time period of the feature amount data FD in which the defect rate DR is greater than the defect rate threshold value Th _dr , which is registered in the feature amount database 72. The training unit 63B generates state label data LD anomaly for the abnormal period by associating the identifier of the target production equipment 200, the start date and time and the end date and time of the set abnormal period, and the abnormality degree calculation model ID of the adopted _abnormality degree calculation model M1.

In step S104, the training unit 63B performs the following process based on the abnormality degree data AD that matches the identifier of the target equipment and the abnormality degree calculation model ID. That is, the training unit 63B extracts from the abnormality degree database 73 all the abnormality degree data AD whose production date and time are included between the start date and time and the end date and time of _{the state label data LD normal} in the normal period, and calculates the average value of the abnormality degree A (average value of the abnormality degree A in the normal period) MA _normal . The training unit 63B also extracts from the abnormality degree database 73 all the abnormality degree data AD whose production date and time are included between the start date and time and the end date and time of the state label data LD _abnormality in the abnormal period, and calculates the average value of the abnormality degree A (average value of the abnormality degree A in the abnormal period) MA _anomaly .

In step S105, the training unit 63B calculates the average value of the average value MA _normal of the abnormality degree A in the normal period and the average value MA _anomaly of the abnormality degree A in the abnormal period. The training unit 63B sets the calculated average value as the precursor threshold value ThA _sign for the precursor period.

In step S106, the training unit 63B extracts abnormality degree data AD whose abnormality degree A is greater than the predictive threshold value ThA _sign and whose production date and time is not included in the normal period or the abnormal period. The training unit 63B groups the extracted abnormality degree data AD in units of a predetermined period, for example, one day. The training unit 63B generates state label data LD _sign for the predictive period by assuming that a period of a group in which a predetermined number or more of abnormality degree data AD exists is a predictive period.

In step S107, the training unit 63B registers the generated state label data LD _normal , LD _sign , and LD _abnormality in the state label database 74.

As described above, according to the model generation device 60B of the third embodiment, even if the period immediately preceding the abnormal period includes a period in which no signs of abnormality appear, the abnormality warning period can be appropriately set, and a highly accurate trained model (anomaly warning detection model M2) can be generated.

The detection performance of each of the anomaly sign detection models M2 generated in the first, second, and third embodiments may be compared in the training units 63, 63A, and 63B, and the anomaly sign detection model M2 with the highest detection performance may be adopted.

Furthermore, it goes without saying that the production facility 200 to which the model generation device 60, 60A, 60B and the abnormality sign detection device 100 according to the present disclosure are applied may be a production facility that produces products other than lithium ion secondary batteries.

The application of the abnormality sign detection model M2 is not limited to the production equipment 200 that generated the production performance data PD used in generating the abnormality sign detection model M2, but may be other equipment that produces the same product as the production equipment 200. Furthermore, if there is a common tendency for abnormalities to occur, the application of the abnormality sign detection model M2 may be equipment that produces a product different from the production equipment 200.

This disclosure is useful for an anomaly detection device that detects anomalies or signs of anomalies in production equipment.

10 Abnormality sign detection system 20 Training database 60, 60A, 60B Model generation device 61 Pre-processing unit 62 Abnormality degree calculation unit 63, 63A, 63B Training unit 71 Production record database 72 Feature amount database 73 Abnormality degree database 74 Condition label database 100 Abnormality sign detection device 101 Pre-processing unit 102 Abnormality sign detection unit 200 Production equipment 201 Winding unit 202 First sheet material 203 Second sheet material 204 Winding body 205A First laminating roll 205B Second laminating roll 206 Winding core 206M Winding core rotation drive unit 207 Inspection machine 208 Cutting unit 40 First sheet material post-cut length measurement sensor 41 Second sheet material post-cut length measurement sensor 50 First supply reel 51 Second supply reel

Claims (10)

A pre-processing unit that acquires feature quantity data of the production performance data for training;
an anomaly degree calculation unit that calculates an anomaly degree of the acquired feature amount data by using an anomaly degree calculation model that represents a group of feature amounts in a normal state;
a training unit that determines a state of the feature data based on the calculated degree of anomaly, and trains an anomaly sign detection model that detects the state of the feature data from the feature data using training data composed of the feature data and the state of the feature data;
having
A model generating device in which each feature data of the training data is composed of a group of feature values that overlap with each other on a time axis . The model generating device according to claim 1, wherein the anomaly prediction detection model judges the state of the feature data as a normal period, a prediction period, or an abnormal period. The model generating device according to claim 2, wherein the training unit determines the state of the feature data based on a comparison of the calculated degree of anomaly with a predictive threshold and an anomaly threshold. The model generating device according to any one of claims 1 to 3, wherein the anomaly calculation model calculates the degree of deviation from the group of features in the normal state as the degree of anomaly of the feature data. The model generating device according to any one of claims 1 to 4, wherein the preprocessing unit determines the feature data based on statistics of the training production performance data. The model generating device according to claim 1 , wherein the anomaly degree calculation unit generates the anomaly degree calculation model using training data configured of the feature amount data and a state of the feature amount data acquired from a training database. A step of acquiring feature quantity data of production performance data for training;
calculating an anomaly degree of the acquired feature amount data by using an anomaly degree calculation model that represents a group of feature amounts in a normal state;
determining a state of the feature data based on the calculated degree of anomaly, and training an anomaly sign detection model that detects the state of the feature data from the feature data using training data composed of the feature data and the state of the feature data;
The computer executes
A model generation method in which each feature data of the training data is composed of a group of features that overlap with each other on a time axis . a pre-processing unit for acquiring feature quantity data of the production performance data to be detected;
an anomaly sign detection unit that detects a state of the feature data from the feature data by using a trained anomaly sign detection model;
having
The anomaly sign detection model is trained using training data that is configured from feature amount data of training production performance data and a state of the feature amount data,
the state of the feature amount data is determined based on an anomaly degree of the feature amount data calculated using an anomaly degree calculation model that represents a group of feature amounts in a normal state ;
An anomaly sign detection device , wherein each feature data of the training data is composed of a group of feature amounts that overlap on a time axis . A step of acquiring feature amount data of production performance data of a detection target;
Detecting a state of the feature data from the feature data by using a trained anomaly sign detection model;
The computer executes
The anomaly sign detection model is trained using training data that is configured from feature amount data of training production performance data and a state of the feature amount data,
the state of the feature amount data is determined based on an anomaly degree of the feature amount data calculated using an anomaly degree calculation model that represents a group of feature amounts in a normal state ;
The method for detecting an anomaly , wherein each feature data of the training data is composed of a group of feature amounts that overlap with each other on a time axis . A model generating device according to any one of claims 1 to 6 ;
The abnormality sign detection device according to claim 8 ;
An abnormality detection system having the above-mentioned function.

Images