JP6918735B2 - Monitoring device, monitoring method and monitoring program - Google Patents

Description

The present invention relates to a monitoring device, a monitoring method and a monitoring program.

A self-encoder, which is one of neural networks, is a model that learns to reproduce itself by using input data as teacher data. This self-encoder is characterized in that the number of nodes in the input layer and the output layer is the same, and the number of nodes in the intermediate layer is smaller than that in the input / output layer. There are several ways to apply the self-encoder. For example, data features can be extracted from the output of the intermediate layer of the model after training, as in principal component analysis. Further, as an abnormality detection in a factory, it is used to detect an abnormality by comparing the difference between input data and output data with respect to a model in which normal sensor data or images are trained.

As an example of using anomaly detection, first, a model is created in which only normal data such as plant sensors are trained. Since this model can reproduce the learned data, that is, the normal data, the reproduction error is reduced. On the other hand, since it is not possible to reproduce abnormal data that has not been learned, the reproduction error becomes large. In this way, abnormality detection can be performed by observing the reproduction error.

In addition, there is clustering and the like as a method of performing anomaly detection, state estimation, and the like without using a neural network. In these methods, anomalies are extracted from the data distance between normal and anomalous.

JP-A-2017-142654

The conventional method as described above has a problem that the importance of each input data to the output data cannot be confirmed dynamically and easily. For example, in the above-mentioned method of extracting anomalies from the data distance between normal and abnormal, important features can be indirectly obtained from the data distance between normal and abnormal, but the relationship between features, etc. It is difficult to obtain detailed features that also take into consideration. Also, as deep learning technology is called a black box, it was difficult to extract how the model made predictions and judgments. The self-encoder is no exception, and it is difficult to see which input affected the output.

For example, in the case of anomaly detection, it is difficult to identify which input affects the abnormal data and causes a reproduction error, so it is sufficient to control the items related to which sensor to prevent the abnormality. I couldn't connect to actions such as Ka.

In order to solve the above-mentioned problems and achieve the object, the monitoring device of the present invention uses a collecting means for collecting a plurality of data related to the monitored object and a plurality of data collected by the collecting means as input data. The detection means for detecting the state of the monitored object using the encoder, each input data input to the self-encoder, and output data output from the self-encoder are used to describe the above. It is characterized by having a calculation means for calculating the importance of each input data input to the self-encoder with respect to the output data.

Further, the monitoring method of the present invention is a monitoring method executed by a monitoring device, in which a collection step of collecting a plurality of data related to a monitoring target and a plurality of data collected by the collection step are used as input data, and the self-monitoring method is performed. The detection step of detecting the state of the monitored object using the encoder, each input data input to the self-encoder, and the output data output from the self-encoder are used to describe the above. It is characterized by including a calculation step of calculating the importance of each input data input to the self-encoder with respect to the output data.

In addition, the monitoring program of the present invention collects a plurality of data acquired by the monitored equipment, a collection step of collecting a plurality of data related to the monitored target, and a plurality of data collected by the collection step. As input data, a detection step for detecting the state of the monitored object using a self-encoder, each input data input to the self-encoder, and output data output from the self-encoder. Is used to cause a computer to perform a calculation step of calculating the importance of each input data input to the self-encoder with respect to the output data.

According to the present invention, it is possible to dynamically and easily confirm the importance of each input data with respect to the output data.

FIG. 1 is a block diagram showing a configuration example of the monitoring device according to the first embodiment. FIG. 2 is a diagram illustrating learning of a self-encoder. FIG. 3 is a diagram illustrating a process of determining an abnormality from a reproduction error of input data and output data. FIG. 4 is a diagram illustrating a process of determining an abnormality from a reproduction error of input data and output data. FIG. 5 is a diagram illustrating an example of the importance calculation process for the self-encoder. FIG. 6 is a diagram illustrating an example in the case of calculating the importance of N × N. FIG. 7 is a diagram illustrating an example of a process of narrowing down the importance calculation target according to the reproduction error. FIG. 8 is a diagram illustrating an example of a process of narrowing down the nodes in the middle layer as the importance calculation target in the neural network. FIG. 9 is a diagram illustrating an outline of the entire process executed by the monitoring device. FIG. 10 is a flowchart showing an example of the flow of abnormality prediction processing in the monitoring device according to the first embodiment. FIG. 11 is a flowchart showing an example of the flow of the importance calculation process in the monitoring device according to the first embodiment. FIG. 12 is a flowchart showing an example of a flow of graph display processing in the monitoring device according to the first embodiment. FIG. 13 is a diagram showing a computer that executes a monitoring program.

Hereinafter, embodiments of a monitoring device, a monitoring method, and a monitoring program according to the present application will be described in detail with reference to the drawings. The monitoring device, monitoring method, and monitoring program according to the present application are not limited by this embodiment.

[First Embodiment]
In the following embodiments, the configuration of the monitoring device 10 and the processing flow of the monitoring device 10 according to the first embodiment will be described in order, and finally, the effects of the first embodiment will be described. In the following, an example will be described in which the monitoring device 10 collects data from the monitored equipment such as a factory or a plant and uses a self-encoder to calculate the degree of abnormality of the monitored equipment.

[Monitoring device configuration]
First, the configuration of the monitoring device 10 will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration example of the monitoring device according to the first embodiment. The monitoring device 10 collects a plurality of data acquired by sensors installed in monitored equipment such as a factory or a plant, and uses the collected data as input data by using a self-encoder. Output data that reproduces the input data is output, the input data is compared with the output data, the reproduction error of the output data with respect to the input data is detected, and the degree of abnormality of the monitored equipment is determined according to the reproduction error. Is calculated.

Further, the monitoring device 10 uses the input data input to the self-encoder and the output data output from the self-encoder to the output data of each input data input to the self-encoder. Calculate the importance. Here, the importance indicates how much each input contributed to the output, in other words, how important each input was to reproduce the output, and is an absolute value of importance. The larger the value, the greater the influence of the input on the output.

As shown in FIG. 1, the monitoring device 10 includes a collection unit 11, a detection unit 12, an output visualization unit 13, a calculation unit 14, an importance visualization unit 15, a process data buffer 16, an analysis frame buffer 17, and a frame abnormality evaluation value. It has a buffer 18 and a model buffer 19. The processing of each part of the monitoring device 10 will be described below.

The collection unit 11, the detection unit 12, the output visualization unit 13, the calculation unit 14, and the importance visualization unit 15 are, for example, electronic circuits such as a CPU (Central Processing Unit), an MPU (Micro Processing Unit), and a GPU (Graphical Processing Unit). It is an integrated circuit such as ASIC (Application Specific Integrated Circuit) and FPGA (Field Programmable Gate Array). Further, the process data buffer 16, the analysis frame buffer 17, the frame abnormality evaluation value buffer 18, and the model buffer 19 are storage devices such as semiconductor memory elements such as RAM (Random Access Memory) and flash memory (Flash Memory), for example. ..

The collection unit 11 collects a plurality of data regarding the monitoring target acquired by the monitoring target equipment. For example, the collecting unit 11 receives data periodically (for example, every minute) from a sensor installed in a monitored facility such as a factory or a plant, and stores the data in the process data buffer 16. Here, the data acquired by the sensor is, for example, various data such as temperature, pressure, sound, and vibration of the equipment to be monitored, the equipment in the plant, and the reactor. The time series data acquired by the sensor will be referred to as process data as appropriate below. Further, the data collected by the collecting unit 11 is not limited to the data acquired by the sensor, and may be, for example, numerical data manually input, label data such as a material type or brand, or the like.

The detection unit 12 detects the state of the equipment to be monitored by using a self-encoder using a plurality of data collected by the collection unit 11 as input data. Specifically, the detection unit 12 outputs output data that reproduces each input data, compares the input data with the output data, and detects a reproduction error of the output data with respect to the input data. For example, the detection unit 12 detects a reproduction error of the output data with respect to the input data input to the neural network of the self-encoder by using the process data and the self-encoder, and responds to the reproduction error. Calculate the degree of abnormality of the monitored equipment. In the following, a case where the reproduction error is detected by using a self-encoder and the state of the monitored equipment is detected by calculating the degree of abnormality of the monitored equipment according to the reproduction error will be described as an example. The condition of the monitored equipment is not limited to this, and the state of the monitored equipment may be detected by using a self-encoder. For example, the state of the monitored equipment may be detected by visualizing the intermediate layer after learning with a self-encoder. The detection unit 12 includes an analysis frame extraction unit 12a and a frame abnormality evaluation value calculation unit 12b. The abnormality degree calculation process described below is an example, and is not limited to this.

The analysis frame extraction unit 12a extracts the process data included in the analysis frame having a predetermined time point or a predetermined time width from the process data collected by the collection unit 11. Specifically, the analysis frame extraction unit 12a extracts and reads the process data for the frame from the process data buffer 16, and stores the read process data in the analysis frame buffer 17. The predetermined time point may be the current time point or a preset time point such as 10 seconds before the present time point.

For example, the analysis frame extraction unit 12a collects the process data every minute by the collection unit 11, extracts the process data acquired at the current time t from the process data buffer 16, and reads the process data. Further, for example, when the time width is "5", the analysis frame extraction unit 12a has the process data acquired at the current time t, the process data acquired one minute before the current time, and 2 of the current time. The process data acquired 3 minutes before the current time, the process data acquired 3 minutes before the current time, and the process data acquired 4 minutes before the current time are extracted from the process data buffer 16 and read out.

The process data buffer 16 stores the process data collected by the collection unit 11. The process data buffer 16 stores at least the latest process data for a predetermined time point or a predetermined time width frame as process data. The analysis frame buffer 17 stores the process data extracted by the analysis frame extraction unit 12a.

The frame abnormality evaluation value calculation unit 12b uses the process data extracted by the analysis frame extraction unit 12a as an input, and uses a self-encoder to output data for the input data input to the neural network of the self-encoder. The reproduction error of the above is detected, and the degree of abnormality of the monitored equipment is calculated according to the reproduction error.

For example, the frame abnormality evaluation value calculation unit 12b receives the data of the sensor A, the sensor B, the sensor C, the sensor D, and the sensor E as the process data of the time t extracted from the analysis frame extraction unit 12a. Then, the frame abnormality evaluation value calculation unit 12b inputs the data of the sensor A, the sensor B, the sensor C, the sensor D, and the sensor E at the time t into the self-encoder, and calculates the reproduction error between the input data and the output data. It is detected, and the frame abnormality evaluation value, which is the degree of abnormality of the monitored equipment at time t, is calculated according to the reproduction error.

Further, the frame abnormality evaluation value is, for example, a probability value at which an abnormality has occurred in the monitored equipment, and may be a numerical value represented by “0” to “1”. In this case, for example, if the probability that an abnormality of the monitored equipment has occurred at a certain point in time is predicted to be "40%", the frame abnormality evaluation value will be "0.4". Further, the frame abnormality evaluation value is not limited to this, and for example, "0" or "1" is used as a value indicating whether or not there is a possibility that an abnormality has occurred in the monitored equipment. It may be a numerical value expressed by any of.

The frame abnormality evaluation value buffer 18 stores the output data output from the self-encoder, the reproduction error of the output data with respect to the input data detected by the frame abnormality evaluation value calculation unit 12b, and the calculated frame abnormality evaluation value. do. The frame abnormality evaluation value buffer 18 stores at least the latest frame abnormality evaluation value for a frame at a predetermined time point or a predetermined time width as a frame abnormality evaluation value. The model buffer 19 stores a self-encoder used by the frame abnormality evaluation value calculation unit 12b described above, that is, a trained model for detecting an abnormality in the monitored equipment.

Here, it is a figure explaining the learning of a self-encoder using FIG. As illustrated in FIG. 2, in the self-encoder neural network, the number of nodes in the input layer and the output layer is the same, and the number of intermediate nodes is smaller than that in the input / output layer. Then, when the input data is input to the self-encoder, the input data is used as teacher data to learn to reproduce itself. Here, the data to be learned by the self-encoder is only the data acquired by the sensor when the monitored equipment is normal (hereinafter, appropriately referred to as "normal data"). That is, by training only the normal data as input data, it is trained so that it can be reproduced well only when the input data is normal data, and it cannot be reproduced well when the input data is abnormal data.

In addition, by learning only with normal data in this way, abnormal data is less likely to occur during normal times, such as equipment in factories and plants, so normal data is overwhelmingly compared to abnormal data. Even when there are many and few abnormal data, it is possible to train appropriately. In the monitoring device 10, the case where the learned self-encoder is stored in the model buffer 19 in advance will be described as an example, but the monitoring device 10 may perform the learning process.

Next, the process of determining an abnormality from the reproduction error of the input data and the output data will be described with reference to FIG. 3 and 4 are diagrams for explaining a process of determining an abnormality from a reproduction error of input data and output data. In the example of FIG. 3, normal data is input as input data, and in the example of FIG. 4, the case where abnormal data is input as input data is illustrated. As illustrated in FIG. 3, for example, the detection unit 12 inputs normal data as input data to the self-encoder and outputs output data. Since the self-encoder learns so that only normal data can be reproduced, the reproduction error between the input data and the output data is small.

On the other hand, as illustrated in FIG. 4, for example, the detection unit 12 inputs the abnormal data as input data to the self-encoder and outputs the output data. Since the self-encoder learns so that only normal data can be reproduced, the reproduction error between the input data and the output data is large.

That is, the detection unit 12 detects the reproduction error of the output data with respect to the input data, calculates so that the larger the reproduction error is, the larger the abnormality degree of the equipment to be monitored is, and the smaller the reproduction error is, the more the monitoring target is. Calculate so that the degree of abnormality of the equipment is small.

Returning to the description of FIG. 1, the output visualization unit 13 outputs a warning sign based on the abnormality prediction evaluation value, and outputs the time series data of the abnormality prediction evaluation value as a chart screen. The output visualization unit 13 includes an abnormality determination unit 13a and a chart display unit 13b.

The abnormality determination unit 13a determines whether or not the frame abnormality evaluation value calculated by the frame abnormality evaluation value calculation unit 12b is equal to or more than a predetermined threshold value, and if the abnormality prediction evaluation value is equal to or more than a predetermined threshold value, Outputs a warning about the occurrence of an error. For example, when the abnormality prediction evaluation value is equal to or higher than a predetermined threshold value, the abnormality determination unit 13a outputs a warning message as a warning sign indicating that an abnormality may occur in the monitored equipment after a certain period of time. Alternatively, a sound for notifying a warning may be output.

The chart display unit 13b displays the time-series data of the frame abnormality evaluation value calculated by the frame abnormality evaluation value calculation unit 12b as a chart screen. For example, when the chart display unit 13b receives a display request from the user, the chart display unit 13b displays the time-series data of the abnormality prediction evaluation value as a chart screen.

The calculation unit 14 uses the input data input to the self-encoder and the output data output from the self-encoder, and the importance of each input data input to the self-encoder to the output data. Is calculated. For example, the calculation unit 14 calculates the importance of each input data with respect to the output data by using an importance feature method such as Salience Map. The calculation unit 14 includes an importance calculation unit 14a and an importance noise removal unit 14b.

The importance calculation unit 14a reads the input data (process data) input to the self-encoder from the analysis frame buffer 17, and at the same time, the output data output from the self-encoder and the reproduction error of the output data with respect to the input data. And the frame abnormality evaluation value are read from the frame abnormality evaluation value buffer 18, and each input data input to the self-encoder and the output data output from the self-encoder are used in the self-encoder. Calculate the importance of each input data to the output data.

Here, a specific example for calculating the importance will be described. For example, in the trained model that calculates the output value from the input value, the importance calculation unit 14a uses the partial differential value or its approximate value for each input value of the output value to determine the importance for each sensor at each time. calculate. As an example, the importance calculation unit 14a calculates the importance for each sensor at each time using the Saliency Map. Saliency Map is a technique used in image classification of a neural network, and is a technique for extracting a partial differential value for each input of an output of a neural network as an importance contributing to the output.

The importance may be calculated by a method other than Saliency Map. For example, a method of obtaining the importance of the input by propagating the output value to each layer of the neural network (see Reference 1) or a simple neural network. It may be a method of finding the importance of input to data by approximating a model (decision tree, etc.) (see Reference 2).
Reference 1: Bach, Sebastian, et al. "On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation." PloS one 10.7 (2015): e0130140.
Reference 2: Shrikumar, Avanti, Peyton Greenside, and Anshul Kundaje. "Learning important features through propagating activation differences." ArXiv preprint arXiv: 1704.02685 (2017).

Here, an example of the importance calculation process for the self-encoder will be described with reference to the example of FIG. FIG. 5 is a diagram illustrating an example of the importance calculation process for the self-encoder. As illustrated in FIG. 5, when the importance calculation unit 14a inputs the input data of x1 to xn into the self-encoder, the importance calculation unit 14a outputs the output data of y1 to yn. Here, focusing on y2 in the output data, the importance calculation unit 14a calculates the importance of each input data x1 to xn that is important for reproducing y2.

Here, the importance calculation unit 14a may calculate the importance of each input data x1 to xn for all the output data. In this case, as illustrated in FIG. 6, when N input data are input, the output data is also N, so the importance of N × N is calculated. FIG. 6 is a diagram illustrating an example in the case of calculating the importance of N × N. Therefore, under a situation where the input dimension is large, the processing load of the monitoring device 10 is heavy, and the number may be too large for humans to visually perform analysis and control judgment.

Therefore, the importance calculation unit 14a extracts the output data whose reproduction error detected by the detection unit 12 is equal to or greater than a predetermined threshold value, and uses each input data input to the self-encoder and the extracted output data. Therefore, the importance of each input data with respect to the extracted output data may be calculated. For example, to explain using the example of FIG. 5 above, the importance calculation unit 14a has a reproduction error between the input data x1 and the output data y1, a reproduction error between the input data x2 and the output data y2, ... It is determined whether or not the reproduction error between xn and the output data yn is equal to or greater than a predetermined threshold value. As a result, the importance calculation unit 14a extracts only the output data for which the reproduction error is determined to be equal to or higher than a predetermined threshold value, and calculates the importance of each input data with respect to the extracted output data.

The process of narrowing down the importance calculation target according to the reproduction error will be described with reference to the example of FIG. 7. FIG. 7 is a diagram illustrating an example of a process of narrowing down the importance calculation target according to the reproduction error. As illustrated in FIG. 7, when N input data are input, the output data is also N, but the result of calculating the importance of only the output data having a large reproduction error among the output data. When the three output data are the targets of the importance calculation, the importance of 3 × N is calculated, and the processing of the monitoring device 10 is compared with the calculation of the importance of N × N. It reduces the load and makes it easier for humans to make analysis and control decisions.

Further, the importance calculation unit 14a may calculate the importance of each input data for each intermediate layer in the neural network. For example, the importance calculation unit 14a uses each input data input to the self-encoder and the output data of the node of the intermediate layer in the neural network of the self-encoder, and each input to the output data of the intermediate layer. Calculate the importance of the data.

An example of the process of narrowing down the nodes in the middle layer as the importance calculation target in the neural network will be described with reference to the example of FIG. FIG. 8 is a diagram illustrating an example of a process of narrowing down the nodes in the middle layer as the importance calculation target in the neural network. As illustrated in FIG. 8, when N input data are input, the output data is also N, but the nodes in the intermediate layer, which has fewer nodes than the input / output layer, are the targets of the importance calculation. As a result, when the output data of the nodes of the three intermediate layers is the target of the importance calculation, the importance of 3 × N is calculated, which is compared with the calculation of the importance of N × N. The processing load of the monitoring device 10 is reduced, and it is easy for humans to perform analysis and control judgment.

The importance noise removing unit 14b removes noise by smoothing the importance calculated by the importance calculating unit 14a within a predetermined time width. For example, the importance noise removing unit 14b may smooth the importance by using the technique of Smooth Grad as a technique for removing noise. Smooth Grad is a method of removing noise by creating dozens of duplicates in which random Gaussian noise is applied to an input image and averaging the importance extracted from each. The above-mentioned noise removing process may be omitted. In that case, the monitoring device 10 may not have the functional unit of the importance noise removing unit 14b.

The importance visualization unit 15 displays a graph showing the transition of the importance of each sensor data, and notifies important sensors in synchronization with the determination result of the abnormality determination unit 13a. The importance visualization unit 15 includes an acquisition unit 15a, a creation unit 15b, an importance display unit 15c, and a notification unit 15d. In addition, each function of the importance visualization unit 15 may be provided in a display device different from the monitoring device 10.

The acquisition unit 15a acquires the importance of each input data input to the self-encoder with respect to the output data. For example, the acquisition unit 15a acquires the importance of each sensor whose noise has been removed by the importance noise removal unit 14b.

The creation unit 15b creates a graph showing the transition of the importance of each input item by using the importance acquired by the acquisition unit 15a. The importance display unit 15c displays the graph created by the creation unit 15b.

When an abnormality in the monitored equipment is detected, the notification unit 15d notifies an input item whose importance is equal to or higher than a predetermined threshold value among the importance levels acquired by the acquisition unit 15a. For example, the notification unit 15d acquires the determination result by the abnormality determination unit 13a, and detects the abnormality when an abnormality of the monitored equipment is detected, that is, when the abnormality prediction evaluation value is equal to or more than a predetermined threshold value. Among the importance of each sensor corresponding to the set time, the sensor whose importance is equal to or higher than a predetermined threshold value is notified. As a result, when an abnormality occurs, it is possible to notify the relationship between the sensors having a particularly high importance.

Here, FIG. 9 is a diagram illustrating an outline of an abnormality prediction process executed by the monitoring device 10. FIG. 9 is a diagram illustrating an outline of an abnormality prediction process executed by the monitoring device.

FIG. 9 shows that a sensor, a device for collecting operation signals, and the like are attached to a reactor, a device, and the like in the plant, and data is collected at regular intervals. Then, FIG. 9 shows a transition of the process data collected from each sensor A to E by the collecting unit 11 (see (A) of FIG. 9), and the detecting unit 12 shows the process data and the process data. Using the self-encoder, the reproduction error of the output data with respect to the input data input to the neural network of the self-encoder is detected, and the abnormality prediction evaluation value of the monitored equipment is calculated according to the reproduction error. (See (B) in FIG. 9). Then, the output visualization unit 13 outputs the time-series data of the calculated abnormality prediction evaluation value as a chart screen (see (C) of FIG. 9).

Further, the calculation unit 14 uses the input data input to the self-encoder and the output data output from the self-encoder to the output data of each input data input to the self-encoder. Calculate the importance (see (D) in FIG. 9). Then, the importance visualization unit 15 displays a graph showing the transition of the importance of each sensor data (see (E) of FIG. 9). In the example of FIG. 9, a graph showing the transition of the importance of the input data for each of the sensors A to E with respect to the output data for the sensor A is displayed.

[Processing procedure of monitoring device]
Next, an example of the processing procedure by the monitoring device 10 according to the first embodiment will be described with reference to FIGS. 10 to 12. FIG. 10 is a flowchart showing an example of the flow of abnormality prediction processing in the monitoring device according to the first embodiment. FIG. 11 is a flowchart showing an example of the flow of the importance calculation process in the monitoring device according to the first embodiment. FIG. 12 is a flowchart showing an example of a flow of graph display processing in the monitoring device according to the first embodiment. The process illustrated in FIG. 11 is a process in which only the input item (sensor) having a large reproduction error is narrowed down as the target of the importance calculation.

First, the flow of the abnormality prediction process by the monitoring device 10 will be described with reference to FIG. As illustrated in FIG. 10, when the collecting unit 11 collects the process data (affirmation in step S101), the process data in the analysis frame having a predetermined time point or a predetermined time width is obtained from the process data collected by the collecting unit 11. Extract (step S102).

Then, the frame abnormality evaluation value calculation unit 12b reproduces the input data by using the autoencoder (self-encoder) (step S103), and outputs the output data for the input data input to the neural network of the autoencoder. The reproduction error is detected, and the frame abnormality evaluation value, which is the degree of abnormality of the monitored equipment, is calculated according to the reproduction error (step S104).

After that, the abnormality determination unit 13a of the output visualization unit 13 determines whether or not the frame abnormality evaluation value calculated by the frame abnormality evaluation value calculation unit 12b is equal to or greater than a predetermined threshold value (step S105). As a result, if the abnormality prediction evaluation value is less than a predetermined threshold value (step S105 negative), the abnormality determination unit 13a ends the process as it is. Further, when the abnormality prediction evaluation value is equal to or higher than a predetermined threshold value (step S105 affirmative), the abnormality determination unit 13a outputs a warning sign (step S106).

Next, the flow of the importance calculation process by the monitoring device 10 will be described with reference to FIG. As illustrated in FIG. 11, the importance calculation unit 14a acquires the process data input to the trained model from the analysis frame buffer 17, and obtains the frame abnormality evaluation value output from the trained model as the frame abnormality evaluation value. When it is acquired from the buffer 18 (affirmation in step S201), the input item (sensor) having a large reproduction error is extracted as the target of the importance calculation (step S202).

Then, the importance calculation unit 14a calculates the importance of each input data with respect to the output data of the items extracted as the target of the importance calculation (step S203). Then, the importance noise removing unit 14b smoothes the importance calculated by the importance calculating unit 14a within a predetermined time width to remove noise (step S204). For example, the importance noise removing unit 14b smoothes the importance by using the technique of Smooth Grad as a technique for removing noise.

Next, the flow of the graph display process by the monitoring device 10 will be described with reference to FIG. As illustrated in FIG. 12, when the acquisition unit 15a acquires the importance of each input data input to the self-encoder with respect to the output data (step S301 affirmative), the creation unit 15b is acquired by the acquisition unit 15a. A graph showing the transition of the importance of each sensor is created by using the importance (step S302). Then, the importance display unit 15c displays the graph created by the creation unit 15b (step S303).

(Effect of the first embodiment)
The monitoring device 10 according to the first embodiment collects a plurality of data related to the equipment to be monitored, uses the collected data as input data, and uses a self-encoder to reproduce each input data. Is output, the input data is compared with the output data, and a reproduction error of the output data with respect to the input data is detected. Then, the monitoring device 10 uses the input data input to the self-encoder and the output data output from the self-encoder to the output data of each input data input to the self-encoder. Calculate the importance. Therefore, the monitoring device 10 can dynamically and easily confirm the importance of each input data with respect to the output data.

That is, in the monitoring device 10 according to the first embodiment, normal abnormality detection, state determination, etc. are performed by passing the trained model of the self-encoder from the input value of the sensor or the like, and the output is in that state. You can dynamically check which input value is important. As a result, it becomes easy to determine which input value should be operated according to the output value.

For example, in the monitoring device 10 according to the first embodiment, for each time and each sample input, inputs important for the output of the self-encoder can be extracted and presented one by one together with their importance. As a result, it has become possible to present an important sensor for the state at that time, and it has become possible to respond to a system in which the state changes. In addition, even if you are not an expert who is familiar with the state, it is easier to specify which item to operate for that state.

Further, in the monitoring device 10 according to the first embodiment, for example, when only the item of the output data having a large reproduction error is targeted for the importance calculation, the processing load is reduced and the processing is speeded up, which is important. The input can be specified efficiently.

(System configuration, etc.)
Further, each component of each of the illustrated devices is a functional concept, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of the device is functionally or physically distributed / physically in arbitrary units according to various loads and usage conditions. Can be integrated and configured. Further, each processing function performed by each device is realized by a CPU or GPU and a program that is analyzed and executed by the CPU or GPU, or as hardware by wired logic. It can be realized.

Further, among the processes described in the present embodiment, all or part of the processes described as being automatically performed can be manually performed, or the processes described as being manually performed. It is also possible to automatically perform all or part of the above by a known method. In addition, the processing procedure, control procedure, specific name, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

(program)
It is also possible to create a program in which the processing executed by the monitoring device described in the above embodiment is described in a language that can be executed by a computer. For example, it is possible to create a monitoring program in which the processing executed by the monitoring device 10 according to the embodiment is described in a language that can be executed by a computer. In this case, the same effect as that of the above embodiment can be obtained by executing the monitoring program by the computer. Further, the same processing as that of the above-described embodiment may be realized by recording the monitoring program on a computer-readable recording medium, reading the monitoring program recorded on the recording medium into the computer, and executing the monitoring program.

FIG. 13 is a diagram showing a computer that executes a monitoring program. As illustrated in FIG. 13, the computer 1000 has, for example, a memory 1010, a CPU 1020, a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. However, each of these parts is connected by a bus 1080.

The memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012, as illustrated in FIG. The ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System). The hard disk drive interface 1030 is connected to the hard disk drive 1090, as illustrated in FIG. The disk drive interface 1040 is connected to the disk drive 1100 as illustrated in FIG. For example, a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100. The serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120, as illustrated in FIG. The video adapter 1060 is connected, for example, to the display 1130, as illustrated in FIG.

Here, as illustrated in FIG. 13, the hard disk drive 1090 stores, for example, the OS 1091, the application program 1092, the program module 1093, and the program data 1094. That is, the above-mentioned monitoring program is stored in, for example, the hard disk drive 1090 as a program module in which instructions executed by the computer 1000 are described.

Further, the various data described in the above embodiment are stored as program data in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 as needed, and executes various processing procedures.

The program module 1093 and program data 1094 related to the monitoring program are not limited to the case where they are stored in the hard disk drive 1090, and may be stored in, for example, a removable storage medium and read by the CPU 1020 via a disk drive or the like. good. Alternatively, the program module 1093 and the program data 1094 related to the monitoring program are stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.), and are stored via the network interface 1070. It may be read by the CPU 1020.

The above-described embodiments and modifications thereof are included in the inventions described in the claims and the equivalent scope thereof, as are included in the technology disclosed in the present application.

10 Monitoring device 11 Collection unit 12 Detection unit 12a Analysis frame extraction unit 12b Frame abnormality evaluation value calculation unit 13 Output visualization unit 13a Abnormality judgment unit 13b Chart display unit 14 Calculation unit 14a Importance calculation unit 14b Importance Noise removal unit 15 Importance Visualization unit 15a Acquisition unit 15b Creation unit 15c Importance display unit 15d Notification unit 16 Process data buffer 17 Analysis frame buffer 18 Frame abnormality evaluation value buffer

Claims (8)

A collection method that collects multiple data about the monitoring target,
A detection means for detecting the state of the monitoring target by using a self-encoder using a plurality of data collected by the collection means as input data, and a detection means.
Using each input data input to the self-encoder and output data output from the self-encoder, the importance of each input data input to the self-encoder to the output data is calculated. and a calculation means to possess,
The detection means uses a plurality of data collected by the collection means as input data, and uses the self-encoder to output output data that reproduces each input data, and outputs the input data and the output data. By comparison, the reproduction error of the output data with respect to the input data is detected.
The calculation means extracts output data in which the reproduction error detected by the detection means is equal to or greater than a predetermined threshold value, and extracts using each input data input to the self-encoder and the extracted output data. A monitoring device characterized by calculating the importance of each of the input data with respect to the output data. A collection method that collects multiple data about the monitoring target,
A detection means for detecting the state of the monitoring target using a self-encoder using a plurality of data collected by the collection means as input data, and a detection means.
Using each input data input to the self-encoder and output data output from the self-encoder, the importance of each input data input to the self-encoder to the output data is calculated. With the calculation method
Have,
The calculation means uses the input data input to the self-encoder and the output data of the nodes of the intermediate layer in the neural network of the self-encoder, and the input data with respect to the output data of the intermediate layer. monitoring apparatus you and calculates the importance of. The monitoring device according to claim 1 , wherein the detection means detects a reproduction error of the output data with respect to the input data, and calculates an abnormality degree of the monitoring target according to the reproduction error. Using the importance calculated by the calculation means, a creating means for creating a graph showing the transition of the importance of each input data with respect to the output data, and
The monitoring device according to claim 1 or 2 , further comprising a display means for displaying a graph created by the creation means. A monitoring method performed by a monitoring device,
A collection process that collects multiple data about what is being monitored,
A detection step of detecting the state of the monitored object using a self-encoder using a plurality of data collected by the collection step as input data, and a detection step.
Using each input data input to the self-encoder and output data output from the self-encoder, the importance of each input data input to the self-encoder to the output data is calculated. and a calculation step of viewing including,
In the detection step, a plurality of data collected by the collection step are used as input data, and the self-encoder is used to output output data that reproduces each input data, and the input data and the output data are combined. By comparison, the reproduction error of the output data with respect to the input data is detected.
In the calculation step, output data having a reproduction error detected by the detection step of a predetermined threshold value or more is extracted, and each input data input to the self-encoder and the extracted output data are used for extraction. A monitoring method characterized in that the importance of each of the input data with respect to the output data is calculated. A monitoring method performed by a monitoring device,
A collection process that collects multiple data about what is being monitored,
A detection step of detecting the state of the monitored object using a self-encoder using a plurality of data collected by the collection step as input data, and a detection step.
Using each input data input to the self-encoder and output data output from the self-encoder, the importance of each input data input to the self-encoder to the output data is calculated. With the calculation process
Including
In the calculation step, the input data input to the self-encoder and the output data of the nodes of the intermediate layer in the neural network of the self-encoder are used, and the input data for the output data of the intermediate layer is used. A monitoring method characterized by calculating the importance of. A collection step that collects multiple data acquired by the monitored equipment,
A collection step that collects multiple data about what is being monitored,
A detection step of detecting the state of the monitoring target using a self-encoder using a plurality of data collected by the collection step as input data, and a detection step.
Using each input data input to the self-encoder and output data output from the self-encoder, the importance of each input data input to the self-encoder to the output data is calculated. Let the computer perform the calculation steps and
In the detection step, a plurality of data collected by the collection step are used as input data, and the self-encoder is used to output output data that reproduces each input data, and the input data and the output data are combined. By comparison, the reproduction error of the output data with respect to the input data is detected.
In the calculation step, output data whose reproduction error detected by the detection step is equal to or greater than a predetermined threshold is extracted, and each input data input to the self-encoder and the extracted output data are used for extraction. A monitoring program characterized by calculating the importance of each of the input data with respect to the output data. A collection step that collects multiple data acquired by the monitored equipment,
A collection step that collects multiple data about what is being monitored,
A detection step of detecting the state of the monitoring target using a self-encoder using a plurality of data collected by the collection step as input data, and a detection step.
Using each input data input to the self-encoder and output data output from the self-encoder, the importance of each input data input to the self-encoder to the output data is calculated. With the calculation steps to be done
Let the computer run
In the calculation step, each input data input to the self-encoder and the output data of the node of the intermediate layer in the neural network of the self-encoder are used, and each input data with respect to the output data of the intermediate layer is used. A monitoring program characterized by calculating the importance of.

Images