Data Processing System

Abstract

To provide a data processing system that can improve the accuracy of machine learning models. [Solution] The data processing system according to the present disclosure is a data processing system that generates training data for a machine learning model to learn the cause of anomalies in time-series data about the operation of equipment, and comprises a training data generation unit that generates training data by suppressing the influence of the pre-processing training data on the operation interruption time, which is the interruption time of the equipment's operation due to external factors of the equipment, and master data that does not include the operation interruption time.

Description

Technical Field

[0001] The present disclosure relates to a data processing system.

Background Art

[0002] In recent years, the development of technologies for detecting anomalies in equipment and diagnosing phenomena occurring in the equipment has been underway. For example, Patent Document 1 discloses a technique for referring to maintenance history information for anomalies detected from the output signals of multi-dimensional sensors attached to equipment and notifying necessary diagnoses and measures. The anomaly detection / diagnosis system according to Patent Document 1 cuts out sections of sensor data during learning, extracts and classifies feature quantities according to the phenomena and countermeasure information taught during learning. In the actual operation stage, the anomaly detection / diagnosis system extracts the features of the time-series signals acquired by the multi-dimensional sensors, discriminates them by a plurality of discriminators together with learning data mainly consisting of normal cases, and determines an anomaly measure.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] The operation of equipment equipped with multidimensional sensors may vary depending on the situation, even if there is no abnormality in the operation. For example, this could include a temporary interruption of the equipment's operation within a normal range. In this case, the accuracy of a machine learning model trained on training data exhibiting such operation may be insufficient. That is, in such a machine learning model, even two time series data that should be considered to be the same category may not be considered the same category because the equipment's operation in each time series data differs within a normal range. In Patent Document 1, sensor data is used directly as training data during training, which may result in insufficient accuracy for the machine learning model.

[0005] This disclosure is made to solve these problems and aims to provide a data processing system that can improve the accuracy of machine learning models. [Means for solving the problem]

[0006] The data processing system relating to this disclosure is a data processing system that generates training data for a machine learning model to learn the causes of anomalies in time-series data about the operation of equipment, and comprises a training data generation unit that generates the training data by suppressing the influence of the pre-processing training data on the operation interruption time, which is the interruption time of the operation of the equipment due to external factors of the equipment, and master data that does not include the operation interruption time. As a result, the data processing system can mitigate variations in the time axis of the pre-processing training data. By using the training data generated by the data processing system, the machine learning model can more easily classify data with similar magnitudes at the same time when the equipment performed the same operation into the same group. As a result, the data processing system can improve the accuracy of the machine learning model. [Effects of the Invention]

[0007] According to this disclosure, it is possible to provide a data processing system that can improve the accuracy of machine learning models. [Brief explanation of the drawing]

[0008] [Figure 1] Figure 1 is a block diagram showing the configuration of the data processing system 1 related to this disclosure. [Figure 2] Figure 2 is a schematic diagram showing an example of time-series data. [Figure 3] Figure 3 is a schematic diagram illustrating an example of the flow of how a machine learning model determines the cause of an anomaly. [Figure 4] Figure 4 shows an example of the hardware configuration of the data processing system 2 related to this disclosure. [Modes for carrying out the invention]

[0009] (Embodiment 1) Embodiments of the present disclosure will be described below with reference to the drawings. Figure 1 is a block diagram showing the configuration of the data processing system 1 according to the present disclosure. The data processing system 1 is a system that generates training data for a machine learning model to learn about the causes of anomalies in time-series data about the operation of equipment. Equipment refers to, for example, production machinery. Production machinery refers to actuators in a production line or presses in press dies. Equipment may be rotating machinery or power generation equipment, as long as it can collect time-series data about its operation. The equipment in question is equipped with sensors, and time-series data about the operation of the equipment can be acquired via these sensors. For example, the sensors can acquire the torque of a servo motor in the equipment.

[0010] The cause of an anomaly determined by machine learning is the direct or indirect cause of the malfunction in the equipment in question. A direct cause of an anomaly is one that can be resolved by repairing or replacing the equipment itself. An indirect cause of an anomaly is one that is not due to a malfunction in the equipment itself, but is due to a malfunction in other equipment or piping to which the equipment is electrically or mechanically connected.

[0011] The training data used by a machine learning model includes time-series data of the equipment under consideration, but may also include the causes of anomalies corresponding to that time-series data. If the equipment under consideration is a production machine on a production line, the training data is typically time-series data for one production cycle, but it may also include time-series data spanning multiple production cycles or time-series data for one day.

[0012] Machine learning models include deep learning models. Through training, machine learning models can determine the causes of anomalies in time-series data, which are the target data for equipment. Machine learning models may also estimate the causes of anomalies. Typically, machine learning models determine the causes of anomalies by classifying or clustering the target data, but they may also predict the causes of anomalies through regression. In other words, machine learning models may use supervised learning or unsupervised learning.

[0013] In supervised learning, the machine learning model classifies the target data and determines the cause of anomalies according to the classified group. The determination of the cause of anomalies by the machine learning model may be achieved by outputting the group classified by the machine learning model, or by outputting the cause of anomalies themselves. When the machine learning model outputs the classified group, a separately constructed system determines past cases and specific causes of anomalies belonging to that group by referring to accumulated maintenance history information. When the machine learning model outputs the cause of anomalies themselves, the machine learning model outputs the cause of anomalies corresponding to the classified group. In unsupervised learning, the machine learning model determines the cause of anomalies by grouping time-series data using clustering, for example.

[0014] The data processing system 1 comprises a learning data generation unit 11 and a target data generation unit 12. The learning data generation unit 11 generates learning data by suppressing the influence of the pre-processing learning data during the operation interruption time, based on pre-processing learning data including the operation interruption time and master data excluding the operation interruption time. Pre-processing learning data refers to time-series data input to the data processing system 1, which is learning data in the stage before processing by the learning data generation unit 11. Pre-processing learning data is typically time-series data relating to the past operation of the target equipment, but it may also be time-series data relating to the past operation of other equipment that is not the target equipment. For example, if the number of past time-series data acquired for the target equipment is insufficient, the learning data generation unit 11 may use time-series data relating to the past operation of other equipment as pre-processing learning data. Also, while there are typically multiple pre-processing learning data sets, there may be only one.

[0015] Operational downtime refers to the downtime of the equipment due to external factors affecting the equipment in question. In other words, operational downtime refers to downtime that is not caused by a direct or indirect malfunction of the equipment in question. Operational downtime may also be referred to as temporary downtime of the equipment within the normal range. Here, equipment downtime caused by a direct or indirect malfunction refers to equipment downtime caused by a malfunction related to the cause of the malfunction determined by a machine learning model. In other words, operational downtime is downtime that cannot be resolved by repairing or replacing the equipment in question or other equipment or piping electrically or mechanically connected to the equipment in question.

[0016] Operational interruption time is, for example, the interruption time caused by an interlock in the equipment in question. That is, if an interlock is activated due to the interruption of other equipment or the activation of a safety mechanism due to a person entering the vicinity of the equipment in question, the equipment interruption time caused by such activation is part of the operational interruption time. This is because such interruption time is due to external factors, not a malfunction of the equipment in question.

[0017] Operation interruptions may occur in the pre-processing training data, for example, between an arbitrary operation by the equipment and the next operation, or in the middle of an arbitrary operation. Multiple operation interruptions may occur in a single pre-processing training data. Furthermore, operation interruptions may occur before or after an operation by the equipment in the pre-processing training data. Since operation interruptions can occur at any point in the pre-processing training data, the timing of predetermined operations by the equipment may differ between pre-processing training data. In other words, the occurrence of operation interruptions may result in differences in the time axis of the pre-processing training data.

[0018] Operational interruptions typically occur at different time points in the pre-processing training data. This is because operational interruptions such as interlocks typically do not have periodicity. On the other hand, some operational interruptions may occur at the same time point in the pre-processing training data.

[0019] Part or all of the pre - learning data input to the data processing system 1 is time - series data including operation interruption time. Part of the pre - learning data may be time - series data that does not include operation interruption time.

[0020] The master data is time - series data that does not include operation interruption time in the operation of the target equipment. The master data may also be referred to as the master waveform. The master data is, for example, time - series data among the pre - learning data that does not include operation interruption time. That is, the master data is, for example, time - series data related to the actual operation of the target equipment. The master data may be time - series data composed of theoretical values of the operation of the equipment.

[0021] The learning data generation unit 11 generates learning data for the machine learning model to learn by suppressing the influence of the pre - learning data during the operation interruption time. Suppressing the influence of the pre - learning data during the operation interruption time means removing part or all of that influence.

[0022] The learning data generation unit 11 generates learning data by comparing each pre - learning data with the master data. Specifically, the learning data generation unit 11 cuts the master data into a plurality of time windows, and calculates the difference between the master data in the time window and the data in the section similar to the master data in the time window among the pre - learning data in the time window to generate learning data.

[0023] The method for generating training data by the training data generation unit 11 will now be described. First, the training data generation unit 11 cuts the master data into multiple time windows. A time window refers to a single interval in time series data. The length of a time window is typically a length with multiple units of time, but it may also be a length with only one unit of time. The training data generation unit 11 can cut out overlapping intervals in the master data as time windows. For example, the training data generation unit 11 cuts out a portion of the master data as a time window, and then cuts out another time window with the same time window length but advanced by a unit of time. Here, the time windows cut out by the training data generation unit 11 do not necessarily have to contain overlapping intervals.

[0024] Next, the training data generation unit 11 searches for time series data similar to the master data in each extracted time window from the pre-processing training data. Here, the training data generation unit 11 considers, for example, the time series data in a predetermined time window and the data of an arbitrary interval in the pre-processing training data as vectors, and searches for intervals similar to the master data by calculating the difference between these vectors. Typically, the training data generation unit 11 searches for similar intervals by calculating the difference between the master data and all intervals in the pre-processing training data, but it is not necessary to calculate the difference for all intervals. The training data generation unit 11 defines the intervals with small calculated differences as similar intervals to the intervals in the master data. Here, calculating the difference between vectors is also referred to as calculating the similarity between the time series data, or calculating the distance between the data.

[0025] Typically, the training data generation unit 11 defines the data in the pre-processing training data that is most similar to the master data in the time window as the similar interval. However, the training data generation unit 11 does not have to define the data in the most similar interval as the similar interval. That is, the training data generation unit 11 may define the second or third most similar interval in the pre-processing training data as the similar interval.

[0026] The training data generation unit 11 generates training data by calculating the difference between the master data and the similarity interval of the pre-processing training data in each time window. The training data may also be a collection of distances between vectors, where the master data and the similarity interval of the pre-processing training data in each time window are considered as vectors. Alternatively, the training data may be a collection of differences per unit time between the master data and the similarity interval of the pre-processing training data in each time window. When the distance between vectors is used as training data, the generated training data will consist of data equal to the number of time windows.

[0027] Here, the process by which the learning data generation unit 11 generates learning data is shown in the diagram. Figure 2 is a schematic diagram showing an example of time-series data. Figure 2(A) shows master data, (B) shows pre-processing learning data, and (C) shows learning data. As shown in Figure 2, the pre-processing learning data (B) has a time interruption period, so the time axis is different from that of the master data (A). Therefore, the timing at which the equipment operates in the pre-processing learning data (B) is different from the timing at which the equipment operates in the master data (A).

[0028] The training data generation unit 11 (A) cuts the master data into multiple time windows and searches for similar intervals in the pre-processing training data (B) for the master data in each time window. Here, the training data generation unit 11 searches for the most similar interval. The training data generation unit 11 considers the intervals in the master data (A) and the similar intervals in the pre-processing training data (B) as vectors, calculates their distance as the difference between the data in that time window, and uses this difference as training data. That is, the training data (C) consists of one data for each time window. In Figure 2, since the training data generation unit 11 cuts out intervals advanced by each unit of time as time windows, the number of data in the master data (A) and the number of data in the training data (C) are the same. As a result, the training data generation unit 11 can generate training data in which the influence of the pre-processing training data during the operation interruption time is suppressed.

[0029] Let's continue the explanation of the training data generation unit 11 shown in Figure 1. The training data generation unit 11 can generate training data using the k-nearest neighbors method. For example, by setting k=1 in the k-nearest neighbors method, the training data generation unit 11 can search for the interval most similar to the master data in a given time window from the pre-processed training data. Also, by setting k=2, the training data generation unit 11 can search for the second most similar interval from the pre-processed training data. By setting an arbitrary k and calculating the difference between the master data and the similar interval in a time window, the training data generation unit 11 can generate training data. In this case, the k-nearest neighbors method is specifically called the time window k-nearest neighbors method.

[0030] The training data generation unit 11 may generate training data in which the influence of pre-processing training data during operation interruption is suppressed by other methods. For example, the training data generation unit 11 cuts the master data into multiple time windows and searches for intervals similar to the master data in each time window from the pre-processing training data. The training data generation unit 11 may use the pre-processing training data of the similar intervals as the training data for that time window, and generate training data by concatenating similar pre-processing training data for each time window.

[0031] Alternatively, the learning data generation unit 11 may compare the master data with the pre-processing learning data to determine the operation interruption time in the pre-processing learning data, delete the time-series data during the operation interruption time, and generate the learning data by concatenating the data before and after the interruption time.

[0032] Furthermore, the training data generation unit 11 may use other trained models to search for pre-processing training data for intervals similar to the master data in the time window. That is, the training data generation unit 11 may use a trained model that uses predetermined time series data as training data to search for similar intervals in the pre-processing training data, and generate training data based on the data in those similar intervals.

[0033] Typically, the training data generation unit 11 generates one training data for one pre-processing training data, but is not limited to this; it may also generate one training data for multiple pre-processing training data. For example, the training data generation unit 11 may generate average data by calculating the average value of multiple pre-processing training data at each time point, and then generate training data by comparing this average data with the master data.

[0034] The target data generation unit 12 generates target data that allows a machine learning model trained using the learning model generated by the learning data generation unit 11 to determine the cause of an anomaly. Specifically, the target data generation unit 12 generates target data by suppressing the influence of the pre-processing target data during the operation interruption time, based on the pre-processing target data including the operation interruption time and the master data. Here, pre-processing target data refers to time-series data input to the data processing system 1, and is the target data in the stage before processing is performed by the target data generation unit 12. Typically, the pre-processing target data is time-series data of the target equipment acquired at the present time, but it may also be past time-series data of the target equipment. Also, while the pre-processing target data is typically one time-series data, there may be multiple pre-processing target data.

[0035] The master data used by the target data generation unit 12 is typically the same time-series data as the master data used by the training data generation unit 11, but it may also be different time-series data from the master data of the training data generation unit 11. By using the same time-series data as the master data used by the training data generation unit 11, the accuracy of the machine learning model that determines the cause of anomalies is improved.

[0036] The target data generation unit 12 generates target data by comparing the pre-processing target data with the master data, similar to the processing performed by the learning data generation unit 11. For example, the target data generation unit 12 extracts the master data into multiple time windows and generates target data by calculating the difference between the master data in each time window and the data in the pre-processing target data that is similar to the master data. Specifically, the target data generation unit 12 uses the difference with the similar interval as the target data.

[0037] The target data generation unit 12 may search the pre-processing target data for the interval most similar to the master data in the time window, define that interval as the similar interval, and calculate the difference with the master data. Alternatively, the target data generation unit 12 may define the second and third most similar intervals to the master data in the time window from the pre-processing training data as similar intervals and calculate the difference with the master data.

[0038] The target data generation unit 12 may generate target data by other methods, similar to the training data generation unit 11. However, the generation of training data by the training data generation unit 11 and the generation of target data by the target data generation unit 12 are performed using the same generation method. This allows the machine learning model to appropriately classify the target data.

[0039] Next, the process by which the machine learning model determines the cause of an anomaly using the data processing system 1 will be explained with diagrams. Figure 3 is a schematic diagram showing an example of the flow of how the machine learning model determines the cause of an anomaly. The training data generation unit 11 generates training data based on the pre-processing training data and the master data. In Figure 3, the training data generation unit 11 cuts the master data into multiple time windows and uses the difference between the master data in each time window and the data from the pre-processing training data that is most similar to the master data as the training data. The training data generation unit 11 also generates one training data for each pre-processing training data. The machine learning model learns using the training model generated by the training data generation unit 11 as training data.

[0040] In the actual operation phase, the target data generation unit 12 generates target data based on the pre-processing target data and the master data. Similar to the learning data generation unit 11, the target data generation unit 12 cuts the master data into multiple time windows and uses the difference between the master data in each time window and the data from the pre-processing target data that is most similar to the master data as the target data.

[0041] The machine learning model receives the target data generated by the target data generation unit 12 as input. The machine learning model extracts waveform groups similar to the target data from the training data. That is, the machine learning model in Figure 3 outputs training data groups that have waveforms similar to the target data. Subsequently, a separately constructed system can determine the cause of the abnormality of the target equipment by extracting similar cases corresponding to the similar waveform groups output by the machine learning model from the maintenance history information stored regarding the cause of the abnormality of the target equipment. Specifically, the system can narrow down the maintenance history based on the date information of the waveforms belonging to the similar waveform group and output a proposed maintenance response.

[0042] Thus, according to the data processing system 1, training data can be generated that suppresses the influence of the pre-processing training data on the operational interruption time, based on the pre-processing training data and the master data. When a machine learning model attempts to determine the cause of an anomaly from the time-series data of the target equipment, it is mainly necessary to train using the time-series data of that equipment. However, the measured time-series data may include operational interruption times caused by equipment interlocks, etc. When such operational interruption times are included, the time axes of the time-series data may differ. That is, when comparing time-series data, the timing of a predetermined operation by the equipment may differ between the time-series data. This can also occur between time-series data that include operational interruption times and time-series data that do not include operational interruption times.

[0043] When time-series data like this is used as training data for a machine learning model, it may be difficult to create a highly accurate machine learning model. In other words, the accuracy of the machine learning model may be insufficient. The main reasons for this are as follows: When classifying time-series data, the machine learning model should classify data where the magnitude of the values ​​at the time the equipment performed the same operation is similar into the same group. However, if the interruption time occurs at different times in the time-series data, simply comparing the data from the beginning of the time-series data will cause the machine learning model to compare values ​​at different times when the operation occurred. That is, a machine learning model trained in this way may group time-series data where the interruption time occurred at the same time into the same group. Therefore, data where the magnitude of the values ​​at the time the equipment performed the same operation is similar may not be classified into the same group in the time-series data.

[0044] The data processing system 1 can improve the accuracy of machine learning models. This is because the data processing system 1 can generate training data that suppresses the influence of pre-processing training data during operational interruption times. In other words, the data processing system 1 can mitigate the variation in the time axis of pre-processing training data by comparing pre-processing training data with different time axes with master data that does not include operational interruption times. By using the training data generated by the data processing system 1, machine learning models can more easily classify data with similar magnitudes at the same time when the equipment performed the same operation into the same group.

[0045] Furthermore, the data processing system 1 may include a target data generation unit 12. In the operational phase, the target data generation unit 12 generates target data by processing the pre-processing target data in the same way as the training data generation method used by the training data generation unit 11. By having the data processing system 1 generate both training data and target data, the accuracy of the machine learning model can be further improved.

[0046] Furthermore, the data processing system 1 can generate training data by cutting the master data into multiple time windows and calculating the difference between the master data in each time window and the data in the pre-processing training data that is similar to the master data in that window. The operation interruption time may differ for each piece of pre-processing training data. Even in such cases, training data that avoids the operation interruption time in the pre-processing training data can be generated by using the master data that does not include the operation interruption time as a reference and searching for data in the same time window as the master data.

[0047] Furthermore, the data processing system 1 can use the difference between the master data in each time window and the pre-processing training data for similar intervals as training data. By using the difference between data itself as training data, it is possible to easily generate training data that avoids operational interruptions.

[0048] Furthermore, the data processing system 1 can generate training data by calculating the difference between the master data in each time window and the pre-processing training data for the interval most similar to the master data. This makes it easier for the data processing system 1 to compare pre-processing training data related to the same operation as the equipment operation related to the master data in that time window with the master data, thereby enabling the generation of accurate training data while avoiding operation interruptions.

[0049] Furthermore, the data processing system 1 can be used when the target equipment is actuators in a production line. Since actuators in a production line perform repetitive operations for each production cycle, the time window for comparison with master data can be divided for each repetitive operation. Therefore, the time window length can be shortened, and the computational cost for comparing data can be reduced.

[0050] (Example hardware configuration) Figure 4 shows an example of the hardware configuration of a data processing system 2 according to this disclosure. In Figure 4, the data processing system 2 includes a processor 21 and a memory 22. The processor 21 may be, for example, a microprocessor, an MPU (Micro Processing Unit), or a CPU (Central Processing Unit). The processor 21 may include multiple processors. The memory 22 is composed of a combination of volatile memory and non-volatile memory. The memory 22 may include storage located away from the processor 21. In this case, the processor 21 may access the memory 22 via an I / O (Input / Output) interface, which is not shown.

[0051] In the above example, the program can be stored and provided to the computer using various types of non-transitory computer-readable medium. Non-transitory computer-readable medium includes various types of tangible storage medium. Examples of non-transitory computer-readable medium include magnetic storage media (e.g., magneto-optical disks), CD-ROMs, CD-Rs, CD-R / Ws, and semiconductor memory (e.g., mask ROMs, PROMs (Programmable ROMs), EPROMs (Erasable PROMs), flash ROMs, RAMs). Alternatively, the program may be provided to the computer using various types of transient computer-readable medium. Examples of transient computer-readable medium include electrical signals, optical signals, and electromagnetic waves. Transitory computer-readable medium can supply the program to the computer via wired communication channels such as electric wires and optical fibers, or via wireless communication channels. Computers include various information processing devices such as PCs, servers, CPUs, MPUs, FPGAs (Field Programmable Gate Arrays), and ASICs (Application Specific Integrated Circuits).

[0052] This disclosure is not limited to the embodiments described above, and can be modified as appropriate without departing from the spirit of the invention. For example, the data processing system 1 according to Embodiment 1 may include only the learning data generation unit 11 and not the target data generation unit 12. Alternatively, the data processing system 1 may include only the target data generation unit 12 and not the learning data generation unit 11. [Explanation of Symbols]

[0053] 1 Data processing system, 2 Data processing system, 11 Training data generation unit, 12 Target data generation unit, 21 Processor, 22 Memory

Claims

1. A data processing system that generates training data for a machine learning model to train on determining the cause of anomalies in time-series data regarding the operation of equipment, The system includes a learning data generation unit that generates learning data by suppressing the influence of the pre-processing learning data on the operation interruption time, based on pre-processing learning data including operation interruption time, which is the interruption time of the operation of the equipment due to external factors, and master data that does not include the operation interruption time. Data processing system.

2. The machine learning model further comprises a target data generation unit that generates target data for determining the cause of the anomaly, The target data generation unit generates the target data by suppressing the influence of the pre-processing target data during the operation interruption time, based on the pre-processing target data including the operation interruption time and the master data. The data processing system according to claim 1.

3. The master data is cut into multiple time windows, and the learning data is generated by calculating the difference between the master data in each time window and the data from the pre-processing learning data that is similar to the master data. The data processing system according to claim 1.

4. The multiple differences across the multiple time windows are used as the training data. The data processing system according to claim 3.

5. The training data is generated by calculating the difference between the master data in the aforementioned time window and the data from the pre-processing training data that is most similar to the master data. The data processing system according to claim 3 or 4.

Images