JP7767181B2 - Data processing system and data processing method - Google Patents

Description

The present invention relates to data processing technology for detecting signs of abnormalities occurring in an object, and in particular to an invention for data processing for detecting signs of failure in industrial machinery.

An evaluation model is constructed based on time-series data measured from the object of analysis, and the evaluation model is then used to evaluate the condition of the object of analysis. For example, Condition Based Maintenance (CBM) is known as a maintenance method for maintenance targets such as facilities, machinery, equipment, and devices.

This involves monitoring the condition of equipment, even if it is currently operating stably, and maintaining it before it fails depending on the level of deterioration. For example, the operating status of equipment, such as vibration and heat generation in wind power generation facilities and gas turbines, is monitored, and failure predictions are made based on the results.

Predictive maintenance systems use machine learning to create a learning model based on time-series data obtained from devices and equipment, and then apply time-series data from sensors to this model to predict equipment failures. The system incorporates normal data from the sensor data as learning data, builds a detection model based on the learning data, and verifies the model's detection performance by applying the device's past failure history as validation data.

In this type of system, improving the suitability of the training data is desirable to improve the performance of the evaluation model. For example, JP 2020-102001 A discloses a training data verification support device that uses machine learning to detect anomalies in industrial machinery, with the aim of making it easier to confirm whether the data was measured using the same operation when acquiring measurement data from industrial machinery. The device facilitates checking for inappropriate data when acquiring training data consisting only of normal data in advance, and determines whether inappropriate data is mixed in by matching it with past data when generating a training model.

Furthermore, Japanese Patent Application Laid-Open No. 2021-33554 discloses a method for refining training data that improves the predictive accuracy of a model by identifying harmful training data for each of multiple training data sets based on a score that indicates the strength of the impact that the harmful training data has on the predictive accuracy of the model for one sample data set, and removing this harmful training data from the training data.

Japanese Patent Application Laid-Open No. 2020-102001 Japanese Patent Application Laid-Open No. 2021-33554

Conventional systems select normal data from sensor data from periods when the equipment is not malfunctioning, and use the selected data as training data to generate a predictive maintenance model for the equipment. However, determining whether data is suitable for use as training data is not easy in the first place. For example, even data from periods when the equipment is not malfunctioning may contain unrecorded failures, and there is also the possibility that the equipment is operating abnormally, even if it is not an obvious failure. Traditionally, this type of data has been removed from the training data based on predetermined rules, but past rules are often no longer valid for new equipment or new operating environments for the equipment.

On the other hand, if the selection of training data is left to humans, when the time series data spans a long period, not only will the number of data selection patterns be enormous, imposing a heavy burden, but the selection criteria will be arbitrary and lack versatility. Therefore, the present invention aims to improve the detection accuracy of models by assisting in the selection of training data when performing machine learning based on data measured from a target to create a detection model for signs of abnormalities in the target.

To achieve the above objective, the present invention provides a data processing system including a controller that receives time-series data from a measurement sensor of an object and executes a program to perform data processing based on the time-series data to detect signs of abnormality in the object. The controller collects learning data for machine learning from the received time-series data, and the learning data includes multiple pieces of measurement data. The controller performs machine learning based on the learning data to construct a model for detecting signs of abnormality in the object. The model is applied to validation data to calculate the degree of abnormality of the validation data. Measurement data included in the learning data that affects the degree of abnormality is identified as data that should be noted, and the controller presents the data that should be noted. The present invention also provides a data processing method relating to this feature.

According to the present invention, when machine learning is performed based on data measured from a target to create a detection model for predicting abnormalities in the target, the detection accuracy of the model can be improved by assisting in the selection of training data.

1 is a block diagram of an overall system that includes an embodiment of a data processing system and detects a sign of an abnormality in a target. 10 is an example of a recording table of sensor data. 10 is an example of a failure history table. FIG. 4 is a waveform diagram showing an example of time-series data output from a measurement sensor of the device. 10 is a flowchart illustrating an example of a data processing program. 10 is a waveform diagram for explaining division of time-series data output from a measurement sensor of the device at predetermined time intervals. FIG. FIG. 10 is a diagram for explaining how time-series data output from a measurement sensor of the device is divided by clustering. FIG. 10 is a diagram showing the relationship between division of time series data by clustering and a waveform diagram of the time series data. This is a block that explains the process of identifying impact data (data that should be noted) from data for machine learning. 10 is a graph showing the change over time of the target abnormality degree and the reference abnormality degree. 10 is another graph showing the change over time of the target abnormality degree and the reference abnormality degree. 10 is yet another graph showing the change over time of the target abnormality degree and the reference abnormality degree. 10 is yet another graph showing the change over time of the target abnormality degree and the reference abnormality degree. FIG. 10 is a block diagram illustrating how the data refinement module determines the priority of a data block. 10 is an example of a display list as one of the information display modes of impact data. 10 is an example of a screen showing impact data related information output to a user device.

Next, we will explain an embodiment of a data processing system according to the present invention. Figure 1 is a hardware block diagram of an overall system 100 that includes this data processing system 101 and detects signs of abnormalities in an object. The object may be industrial machinery that requires detection of signs of failure, such as wind power generation equipment or gas turbine engines.

The overall system 100 comprises a device group 110, each of which comprises a plurality of target devices 111; a sensor group 120, each of which comprises sensors 121 (sensor 1 to sensor N) for each of the devices; storage 130, which stores sensor data 131, which is measurement data from the sensors, and device failure history data 132; and a data processing system 101.

The data processing system 101 includes a data refining server (data refining device) 150 and a failure prediction server (failure sign detection device) 160 that detects signs of failure. The device group 110, sensor group 120, storage 130, data refining server 150, and failure prediction server 160 are connected to each other via a network 165.

When the storage 130 receives sensor data from the sensor group, it records the measurement data 203, 204 for each device ID 201 of the device group and for each sensor of the sensor group in the sensor data (storage area) 131, as shown in Figure 2 (sensor data recording table). 202 is the measurement time. The sensors output time series data measured from the devices, such as time series data of control variables related to device control, time series data of state variables related to device control, and time series data of the environment in which the devices are operating.

When the storage 130 receives data related to the failure from the device, as shown in FIG. 3 (failure history table), it records, for each failure ID 301, the device ID 302 on which the failure occurred, the failure date 303 on which the failure occurred, the symptoms of the failure 304, and the countermeasures taken against the failure 305 in a table in the failure history (recording area) 132. As shown in FIG. 4, the data processing system 101 sets normal data that does not contain abnormal data such as output data at the time of device failure 406 from the output data 400 of the device 111 as training data 402, in other words, data that is only normal data or mainly normal data, and performs machine learning based on the training data 402 to construct a detection model for signs of failure in the device.

Because the distinction between normal and abnormal data is relative and depends on the state of the device and the operating environment of the device, the data processing system 101 preferably performs unsupervised learning based on the sensor data to create a learning model. The data processing system 101 applies the failure sign detection model to validation data 404, which includes sensor data from before the device failure 406, calculates the degree of anomaly of the validation data 404, and obtains an evaluation index for the detection performance of the failure sign detection model.

The data refinement server 150 comprises a data refinement module 151 for optimizing learning data collected from sensor data, and a model construction module 152 for performing machine learning based on the optimized learning data to construct a failure sign detection model for the device. The data refinement server 150 realizes the data refinement module 151 and the model construction module 152 by having a controller execute a program in memory. Note that the word "module" may also be referred to as a means, part, or unit. The failure prediction server 160 also realizes a module 161 for applying a failure sign detection model to sensor data by having a controller execute a program in memory.

The data refinement module 151 provides support processing to administrators, users, etc. of the data processing system 101 for optimizing the training data 402. Within the system of sensor data collected as training data 402, the data refinement module 151 relatively identifies data that has a predetermined or greater impact on the accuracy of failure sign detection, and presents this to the user as impact data, i.e., data that should be brought to the user's attention, along with its degree of impact.

Upon receiving this presentation, the user can decide whether to perform machine learning while keeping the impact data included in the learning data, or to perform machine learning after excluding the impact data from the learning data. The model construction module 152 receives instructions from the user and performs machine learning using either method.

The operation of the data processing system 101 will be explained based on the flowchart in Figure 5. The flowchart is executed by a processor or controller that implements the data refinement module 151, etc., by executing a program. To clarify the relationship between the module and the flowchart, the flowchart will continue to be explained with the module as the subject.

The data refinement module 151 reads sensor data for model learning from the storage area table (Figure 2) for sensor data 131 in storage 130. The type of sensor data for model learning and the range of data to be read may be based on settings made by the user, etc.

The data refinement module 151 divides the time-series data loaded as training data into multiple data blocks as preprocessing (Figure 5: S500). Methods for dividing the training data include temporal division and classification by clustering. Figure 6A is a waveform diagram of the sensor time-series data, which is the training data 402. Reference numeral 600 denotes a sliding window, which the data refinement module 151 moves regularly along the time axis, allowing one or more pieces of sensor data within the range enclosed by the window to be extracted as a single data block. This is an example of a method for dividing training data over time. The feature quantity of a single data block may be, for example, a statistical quantity such as the average value or standard deviation of the multiple pieces of data included in the dataset.

In contrast, clustering involves classifying training data into multiple clusters according to its characteristics. Figure 6B shows that the training data has been classified into four clusters. Of these, the cluster designated by the reference numeral 610 contains three data blocks. As shown in Figure 6C, the three data 610A, 610B, and 610C each exist discretely in the training data 402, which is time-series data.

Figure 7 is a block diagram illustrating the data refinement process performed by the data refinement module 151, and will be explained in conjunction with the flowchart in Figure 5. As previously described, the data refinement module 151 divides the training data 402 into data blocks 7000 of equal size over time (Figure 5: S500).

Next, the data refinement module 151 narrows down all data blocks 402BL of the training data to the data blocks to be searched for to determine whether they are the impact data described above (S502, Figures 5 and 7). In Figure 7, the data blocks indicated by 7001 to 7004 are the search targets. Since checking whether all data blocks of the training data are impact data would result in a large computational load, only a portion of the data blocks are selected as search targets. The method for determining the data blocks to be searched for will be described later. Note that this does not prevent all data blocks from being search targets.

Next, the data refinement module 151 selects one dataset 7000 from the datasets to be searched (Figs. 5 and 7: S504). The model construction module 152 removes the selected data block from the training data (Fig. 7: S5041), and performs unsupervised machine learning based on the remaining data block (404BL) (Figs. 5 and 7: S506), creating a failure sign detection model (target model) (Fig. 7: 7010).

Furthermore, the model construction module 152 performs unsupervised machine learning based on all data blocks (402BL) of the training data (Figure 7: S507) to create a failure sign detection model (reference model: Figure 5: S508, Figure 7, 7012).

Next, the controller of the failure prediction server starts the model application module 161, which reads the target model 7010 from the model configuration module 152 and further reads the aforementioned verification data 404 from the storage 130, i.e., time-series data that is sensor data from a certain period before the device failure, and applies the target model 7010 to this to calculate the anomaly level (target anomaly level) 7014 of the verification data (Figures 5 and 7: S510), which is then stored in memory. In the same manner, the model application module 161 applies the reference model 7012 to the verification data 404 to calculate the anomaly level (reference anomaly level) 7016 of the verification data (Figures 5 and 7: S512).

Next, the model application module 161 compares the target anomaly degree 7014 with the reference anomaly degree 7016 and evaluates the difference between them (Figures 5 and 7: S514). This evaluation can be performed as follows. 800 is the period (increase period) during which the former increases more than the latter, and 802 is the period (decrease period) during which the former decreases less than the latter. The model application module 161 sets the evaluation index for the increase period to the integral of the increase period width 800 of the anomaly score increase amount 800A, and sets the evaluation index for the decrease period to the integral of the decrease period width 802 of the anomaly score decrease amount 802A, and sets the difference between the two to the evaluation result. Note that the evaluation index may also be the number of data items with an increase in abnormal values and the number of data items with a decrease in abnormal values, and the difference between the two to be the evaluation result.

In step S514, the model application module 161 finishes evaluating the impact of the selected data block on the degree of anomaly, and then repeats this process until evaluation of all data blocks is complete (S516). It then determines the data block with the greatest impact of data blocks 7001 to 7004 as the impact data, and determines the impact of the impact data on the detection accuracy of the failure sign detection model (S518).

Figure 9A is a graph comparing the time changes in the anomaly score of the target anomaly degree 7016 and the anomaly score of the reference anomaly degree 7014. The anomaly score of the target anomaly degree 7016 increases earlier before the device failure date than the reference anomaly degree 7014, and the anomaly score on the failure date is also larger. Therefore, the model application module 161 determines that, in order to improve the accuracy of the failure sign detection model, it is more appropriate to construct the failure sign detection model by excluding impact data from the learning data.

On the other hand, in Figure 9B, the target anomaly level 7016 increases in anomaly score later than the reference anomaly level 7014 before the equipment failure date, and the anomaly score on the failure date is also smaller. Therefore, the model application module 161 determines that not excluding impact data from the learning data is more suitable for maintaining the accuracy of the failure sign detection model.

In Figure 9C, there is almost no difference in the degree of change in the anomaly score between the target anomaly level 7016 and the reference anomaly level 7014, including on the day of the device failure, so the model application module 161 determines that the impact data does not affect the accuracy of the failure sign detection model.

The data processing system 101 completes the flowchart in Figure 5 after step S518. Note that multiple data blocks may be determined as impact data in descending order of impact. The data processing system may start the flowchart in Figure 5 when building a failure sign detection model or when performing maintenance on it.

In the explanation of Figure 7, it was explained that the data refinement module 151 narrows down the search targets for impact data. Therefore, the data refinement module 151 uses the priority of the data block 7000 that belongs to the training data 402. The data refinement module 151 prioritizes data blocks with a priority level or higher (Figure 7: 7001 to 7004) when searching for impact data.

The data refinement module 151 calculates the priority of a data block based on the degree of anomaly (self-anomaly degree) that the data block occupies in the learning data. The higher the self-anomaly degree of a data block, the more likely it is to be impact data.

The data refinement module 151 calculates the degree of self-anomaly of a data block according to the method described in FIG. 10, as an example. The data refinement module 151 divides all data blocks 402BL included in the learning data into two groups BL800 and BL802, performs machine learning 810 based on the data blocks BL800 of one group to create a learning model, and calculates 812 the degree of anomaly of the data blocks BL802 of the other group.

The data refinement module 151 calculates the degree of self-anomalousness of each data block for all patterns, even though there are multiple ways to divide all data blocks contained in the training data into two groups. As a result, the degree of self-anomalousness is calculated for all data blocks.

The data refinement module 151 determines the priority of a data block by adding up multiple anomaly levels for the same data block for all patterns, or by averaging multiple anomaly levels for the same data block. The model application module 161 searches for impact data on a predetermined number of data blocks in descending order of priority. This reduces the load of determining impact data compared to when all data blocks are searched for impact data.

Once the model application module 161 has determined the impact data, it displays it to the user of the data processing system 101. Figure 11 is an example of a display list as one form of displaying impact data information. This display example indicates that there are four data blocks of impact data, distinguished by identifiers 1 to 4; the accuracy of each piece of impact data, i.e., an indicator of whether deleting the impact data from the training data will improve or worsen the accuracy of failure sign detection; and the occurrence rate of the impact data relative to the training data, i.e., the ratio of the size of the impact data to the size of the training data.

Figure 12 is another example of a screen 1000 of impact data-related information output to a user device, and includes the distribution of impact data for each of sensors 1 and 2. The horizontal axis is the scale of the sensor output value, and the vertical axis is the count number of the output value. 1202 indicates the sensor output, and 1200 indicates the impact data within the sensor output. It can be seen that impact data appears in areas where the sensor output value is large for both sensor 1 and sensor 2. Furthermore, screen 1000 includes the temporal distribution of the appearance of impact data. The user can learn more about the characteristics and attributes of the impact data based on the display modes of Figures 11 and 12.

By knowing the impact data and the various analytical information related to the impact data described above, the administrative user can easily determine whether it is better to remove the impact data from the training data and build a failure sign detection model in order to improve the accuracy of the machine learning model's failure sign detection, since the impact data is based on, for example, equipment malfunctions or changes in the equipment's operating environment, or whether it is better to build a failure sign detection model without removing the impact data from the training data, since the impact data accurately represents the equipment's condition, based on the timing of the impact data's appearance, etc.

The present invention is not limited to the above-described embodiments and includes various modifications. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and are not necessarily limited to those including all of the described configurations. Furthermore, it is possible to replace part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. Furthermore, it is possible to add, delete, or replace part of the configuration of each embodiment with other configurations.

In the above-described embodiments, the detection of signs of failure in industrial machinery was described, but the present invention can also be applied to detecting signs of abnormalities in the vital signs of the human body, not just industrial machinery, as long as signs of some kind of abnormality can be detected. Also, in the above-described embodiments, impact data was described as consisting of one data block, but impact data may also be a combination of multiple data blocks.

101: Data processing system, 130: Storage, 150: Data refinement server, 151: Data refinement module, 152: Model construction module, 160: Failure prediction server

Claims (6)

receiving time series data from a target measurement sensor;
a controller that executes a program to perform data processing for detecting a sign of abnormality in the target based on the time-series data;
1. A data processing system comprising:
The controller
collecting learning data for machine learning from the received time series data, the learning data including a plurality of measurement data;
Performing machine learning based on the learning data to construct a model for detecting signs of abnormality in the target;
Applying the model to validation data to calculate the degree of anomaly of the validation data;
Identifying measurement data that affects the degree of anomaly from among the measurement data included in the learning data as data that should be noted;
Present the relevant noteworthy data,
identifying the measurement data as data of interest;
selecting at least one measurement data item from the training data;
the degree of anomaly resulting from performing the machine learning after excluding the selected measurement data from the learning data is set as a first degree of anomaly, the degree of anomaly resulting from performing the machine learning without excluding the selected measurement data is set as a second degree of anomaly, the first degree of anomaly is compared with the second degree of anomaly, and the data to be noted is determined based on the comparison result;
This is done by
Furthermore, the controller
The verification data is time-series data at the time of a failure as an abnormality of the target.
The learning data is collected from time series data of a period that does not include a failure of the target.
Data processing system. The controller
determining, as the data to be noted, the measurement data that maximizes the difference between the first abnormality degree and the second abnormality degree;
10. The data processing system of claim 1. The controller
selecting a plurality of measurement data items from the learning data in a preferential manner;
determining the data to be noted from the selected plurality of measurement data;
3. The data processing system of claim 2. The controller
The machine learning is performed by unsupervised learning.
10. The data processing system of claim 1. The controller
Detecting signs of failure of the target industrial machinery based on time-series data from the industrial machinery;
10. The data processing system of claim 1. The controller
Dividing the training data into a plurality of blocks;
Identifying the data to be noted from the plurality of blocks;
10. The data processing system of claim 1.