Anomaly prediction detection system and anomaly prediction detection model generation method

Abstract

To provide an abnormality sign detection technique capable of suppressing false detection caused by dummy correlations.SOLUTION: An abnormality sign detection system 1 includes one or more computers 5 configured to perform machine learning of an abnormality-sign detection-model M that detects at least one of an abnormality in an object facility to be monitored and a sign of the abnormality. The computer 5 is configured to: acquire a plurality of process amounts P generated at the object facility 2; classify each of the process amounts P into either correlation data for which correlation between the plurality of process amounts is learned or decorrelation data for which correlation between the plurality of process amounts is not learned; generate, depending on this classification, learning input data in which each of the process amounts P is associated as the correlation data or the decorrelation data; and perform the machine learning by inputting the learning input data to the abnormality-sign detection-model M.SELECTED DRAWING: Figure 2

Description

Technical Field

[0001] Embodiments of the present invention relate to abnormal sign detection technology.

Background Art

[0002] In large-scale plants such as nuclear power plants and thermal power plants, a large number of process signals are monitored for the purpose of monitoring the soundness of various systems and devices that make up the plant. In recent years, as a technology for supporting plant monitoring, the practical application of abnormal sign detection using artificial intelligence (AI: Artificial Intelligence) realized by machine learning has been progressing. In this detection of abnormal changes, attempts have been made to detect signs before abnormalities become apparent by using machine learning techniques that have rapidly developed in recent years. For example, there is a known technique for predicting failures by predicting the correlation of a process quantity group consisting of several hundred to several thousand and the movement of the process quantity during normal operation using a machine learning model learned with the movement of the process quantity during normal operation as teacher data.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] During training with training data, it can occur that the trends in change over time are coincidentally similar even though there is no physical correlation between multiple process quantities. This is called spurious correlation. When AI learns this spurious correlation, a phenomenon occurs where, when one process quantity changes, the predicted values ​​of other process quantities that are not physically affected also change simultaneously. This is the cause of false detection. To counter this, multiple process quantities that are not physically correlated are classified into different groups. Specifically, when the probability of a change in one process quantity having a physical impact on other process quantities is small, that process quantity is separated in advance into a separate model, for example, a model that monitors process quantities that do not learn correlations between process quantities and monitors for anomaly predictions. By performing training in this separated state, it is necessary to avoid learning spurious correlations and to perform plant anomaly predictions with high accuracy.

[0005] Furthermore, there are no clear criteria for classifying models that do not learn correlations between process quantities, leaving it to the user's discretion. This lack of clear classification criteria results in a tremendous amount of work being done in classification decisions. Moreover, because it depends on the skill of the user performing the classification, there is variability in the accuracy of the classification, making it impossible to sort appropriately. As a result, the performance of anomaly prediction detection devices equipped with machine learning models may not improve as intended. It should be noted that this is due to spurious correlations between process quantities, and therefore, this problem exists not only in the aforementioned technology but also in all algorithms that learn correlations between process quantities.

[0006] The embodiments of the present invention have been made in consideration of these circumstances, and aim to provide an anomaly prediction detection technology that can suppress false detections caused by spurious correlations. [Means for solving the problem]

[0007] An embodiment of the present invention provides an anomaly prediction detection system comprising at least one computer that detects anomalies or anomaly predictions of a target facility to be monitored using an anomaly prediction detection model, wherein the at least one computer acquires a plurality of process quantities occurring at the target facility, and the plurality of process quantities predetermined Classify into multiple groups, and according to predetermined classification conditions related to switching the operating conditions of the target facilities, select the same group from among the multiple groups In one process, the amount of processing is different from other Process amount and correlation When Correlation data and Classify them as follows: The same group The amount of the first process is the amount of the other Process amount and correlation When not Uncorrelated data and do Classify the above multiple process quantities, For each of the aforementioned groups, The system is configured to generate training input data linked as either correlated or uncorrelated data, input the training input data into the anomaly prediction detection model to perform machine learning, and determine the anomaly or anomaly prediction of the target facility based on the difference between the input data and output data of the anomaly prediction detection model after machine learning has been performed. [Effects of the Invention]

[0008] Embodiments of the present invention provide an anomaly prediction detection technology that can suppress false detections caused by spurious correlations. [Brief explanation of the drawing]

[0009] [Figure 1] A block diagram showing the hardware configuration of the anomaly prediction detection system. [Figure 2] A functional block diagram showing the processing flow of the anomaly prediction detection system. [Figure 3] A functional block diagram showing the processing flow of the data classification unit. [Figure 4] A diagram showing the configuration of equipment that operates using fixed values. [Figure 5](A) is a graph showing the opening signal of the flow rate regulating valve of the first system piping, and (B) is a graph showing the flow rate of the second system piping. [Figure 6] Configuration diagram showing a system having redundant devices. [Figure 7] (A) is a graph showing the discharge pressure of the first pump, (B) is a graph showing the discharge pressure of the second pump, and (C) is a graph showing the flow rate of the piping. [Figure 8] Configuration diagram showing a system having devices that perform intermittent operation. [Figure 9] (A) is a graph showing the discharge pressure of the first pump, (B) is a graph showing the discharge pressure of the second pump, and (C) is a graph showing the water level in the tank. [Figure 10] Configuration diagram showing devices that vary under the influence of the external environment. [Figure 11] (A) is a graph showing the pressure difference between inside and outside the building, and (B) is a graph showing the indoor pressure difference. [Figure 12] Configuration diagram showing devices that cause sudden fluctuations. [Figure 13] (A) is a graph showing the flow rate of the first system, and (B) is a graph showing the flow rate of the second system. [Figure 14] Functional block diagram showing the processing flow of the abnormal sign detection system of the modification example.

Embodiments for Carrying Out the Invention

[0010] Hereinafter, embodiments of the abnormal sign detection system and the abnormal sign detection model generation method will be described in detail with reference to the drawings.

[0011] Reference numeral 1 in FIG. 1 is the abnormal sign detection system of the present embodiment. This abnormal sign detection system 1 uses the data obtained from the target facility as input data to detect an abnormality or a sign of an abnormality in the target facility. The target facility to be monitored is, for example, a nuclear power plant, a thermal power plant, factory equipment, or production equipment. As the target facility of the present embodiment, such a plant 2 is exemplified.

[0012] In addition, a large number of sensors 3 are provided in the plant 2. These sensors 3 are, for example, predetermined measuring instruments attached to predetermined devices such as pipes, pumps, and valves. Further, the sensors 3 acquire measurement values (actual measured values) including information indicating the states of these devices. A large number of measurement values obtained from these sensors 3 are referred to as process values V (Figure 2). Note that control signals output from a control device that controls a device are also included in the process value V. Also, the opening degree of a valve, etc. is included in the process value V.

[0013] In the abnormal sign detection technique, an abnormality or its sign is detected by detecting a slight change in the process value V. For this purpose, it is necessary to accurately determine the normal state of the plant 2. An incorrect determination causes false detection and generates unnecessary work for the operator. Also, in order to detect a slight change in the process value V, it is necessary to perform accurate determination including minute electrical noise signals that are difficult to remove from the process value V.

[0014] Also, the amount of process values V acquired in the plant 2 becomes enormous. Therefore, in the present embodiment, in order to determine an abnormality or its sign from this enormous amount of process values V, artificial intelligence (AI) realized by machine learning is used.

[0015] For example, a learning model generated by machine learning using a neural network, a learning model generated by other machine learning, a deep learning algorithm, a mathematical algorithm such as regression analysis, etc. can be used. Also, forms of machine learning include forms such as clustering and deep learning.

[0016] For example, this abnormal sign detection system 1 may be composed of one computer equipped with a neural network, or may be composed of a plurality of computers equipped with a neural network.

[0017] Here, a neural network is a mathematical model that represents the characteristics of brain function through computer simulation. For example, it shows a model in which artificial neurons (nodes) that form a network through synaptic connections change the strength of their synaptic connections through learning and acquire problem-solving abilities. Furthermore, neural networks acquire problem-solving abilities through deep learning.

[0018] For example, a neural network may have multiple layers, each consisting of several units. By pre-training a multi-layer neural network with training data (supervised data), it is possible to automatically extract certain features from patterns of changes in the state of a circuit or system. Furthermore, the number of hidden layers, units, learning rate, number of training iterations, and activation function of a multi-layer neural network can be set arbitrarily via the user interface.

[0019] This embodiment describes an anomaly prediction detection technology using an autoencoder (encoder-decoder network). The learning model in this embodiment is implemented using this autoencoder. Note that other algorithms besides the autoencoder may also be applied to the machine learning in this embodiment.

[0020] Furthermore, during machine learning, spurious correlations occur where the trends of change over time are coincidentally similar even though there is no physical correlation between multiple process quantities. If the AI ​​learns these spurious correlations, it can cause false detections. However, in this embodiment, the learning of spurious correlations is avoided, and abnormality predictions for Plant 2 are performed with high accuracy.

[0021] As shown in Figure 1, the anomaly prediction detection system 1 comprises a data input computer 4, a learning computer 5, and a detection computer 6. These computers have hardware resources such as a CPU, ROM, RAM, and HDD, and the CPU executes various programs, thereby enabling software-based information processing using hardware resources. Furthermore, the anomaly prediction detection model generation method of this embodiment is realized by having the computers execute various programs.

[0022] The data input computer 4 collects process values ​​V acquired by sensors 3 installed in plant 2. This data input computer 4 is, for example, a server for storing plant information. The collected process values ​​V are then sent to the learning computer 5 or the detection computer 6.

[0023] The learning computer 5 generates an anomaly prediction detection model M (Figure 2) that detects at least one of an anomaly or its precursor in plant 2. The generated anomaly prediction detection model M is sent to the detection computer 6.

[0024] The detection computer 6 uses an anomaly prediction detection model M to detect at least one of an anomaly or a precursor to an anomaly in plant 2.

[0025] Although each component of the Anomaly Prediction Detection System 1 is implemented on a separate computer, these components do not necessarily need to be implemented on multiple computers. For example, each component of the Anomaly Prediction Detection System 1 may be implemented on a single computer.

[0026] The learning computer 5 comprises an input unit 7, an output unit 8, a communication unit 9, a storage unit 10, and a processing circuit 11.

[0027] The input unit 7 receives predetermined information in response to the user's operations on the learning computer 5. This input unit 7 includes input devices such as a mouse or keyboard. In other words, predetermined information is input to the input unit 7 in response to the operation of these input devices.

[0028] The output unit 8 outputs predetermined information. For example, the learning computer 5 includes a device for displaying images, such as a display that outputs the analysis results. In other words, the output unit 8 controls the images displayed on the display. The display may be separate from the computer body or integrated into it.

[0029] The communication unit 9 communicates with the data input computer 4 or the detection computer 6 via a predetermined communication line. In this embodiment, the data input computer 4, the learning computer 5, and the detection computer 6 are connected to each other via a LAN (Local Area Network).

[0030] The memory unit 10 stores various information necessary for generating the anomaly prediction detection model M. For example, the memory unit 10 stores the process value V (Figure 2) sent from the data input computer 4.

[0031] The processing circuit 11 includes a data classification unit 12, a classification result processing unit 13, and a learning model generation unit 14. These are implemented by the CPU executing a program stored in memory or on the HDD.

[0032] Next, with reference to Figure 2, the processing flow of the anomaly prediction detection system 1 will be explained.

[0033] Note that the arrows in Figure 2 are just one example of a processing flow, and there may be other processing flows besides those indicated by the arrows. Also, the order of each process is not necessarily fixed, and the order of some processes may be reversed. Furthermore, some processes may be executed in parallel with other processes. In addition, the anomaly prediction detection system 1 may include components other than those shown in Figure 2, and some of the components shown in Figure 2 may be omitted.

[0034] First, each sensor 3 installed in plant 2 acquires the respective process value V generated in plant 2. Then, the process values ​​V acquired by these sensors 3 are collected in the data input computer 4.

[0035] The data input computer 4 associates each acquired process value V with a sensor management number, sensor name, etc., and outputs it as a process quantity P. In other words, each of the multiple process values ​​V related to plant 2 is converted into multiple process quantities P, which are used in the machine learning format of the anomaly prediction detection model M. These multiple process quantities P are sent to the learning computer 5.

[0036] The learning computer 5 preprocesses the learning input data (input signals) to be input to the neural network. For example, the learning computer 5 acquires multiple process quantities P sent from the data input computer 4. Each of these process quantities P is then input to the data classification unit 12.

[0037] The data classification unit 12 classifies each process quantity P into correlated data, which allows the system to learn the correlation between these process quantities, and uncorrelated data, which does not allow the system to learn the correlation between these process quantities, based on at least one condition. Here, the data classification unit 12 outputs a physical correlation signal determination flag F1 when the process quantity P is classified as correlated data, and outputs a standalone monitoring signal determination flag F2 when the process quantity P is classified as uncorrelated data. The physical correlation signal determination flag F1 and the standalone monitoring signal determination flag F2 are input to the classification result processing unit 13.

[0038] The physical correlation signal is a process quantity P that does not have the potential to induce false correlations. On the other hand, the standalone monitoring signal is a process quantity P that may induce false correlations.

[0039] The classification result processing unit 13 associates a physical correlation signal determination flag F1 with process quantities P classified as correlated data, and associates a standalone monitoring signal determination flag F2 with process quantities P classified as uncorrelated data. The process quantities P with the physical correlation classification flag and the process quantities P with the standalone monitoring classification flag output from the classification result processing unit 13 are input to the learning model generation unit 14. In other words, the classification result processing unit 13 generates learning input data in which each process quantity P is associated as either correlated data or uncorrelated data, according to the classification by the data classification unit 12. This learning input data is input to the learning model generation unit 14.

[0040] Furthermore, the multiple process quantities P used as input data for training are divided into multiple groups. For example, machine learning is performed by dividing them into system groups of Plant 2, such as the water supply system, condensate system, power distribution system, and control system. Here, at least one process quantity P linked as correlated data changes in response to changes in other process quantities P within the same group. On the other hand, at least one process quantity P linked as uncorrelated data does not change in response to changes in other process quantities P within the same group. In this way, appropriate machine learning can be performed even when multiple process quantities P linked to correlated and uncorrelated data belong to the same group.

[0041] The learning model generation unit 14 inputs the learning input data into the anomaly prediction detection model M and performs machine learning. Here, process quantities P with a physical correlation classification flag are treated as physical correlation signals, and machine learning is performed on process quantities P that take into account their correlation with other process quantities P. On the other hand, process quantities P with a standalone monitoring classification flag are treated as standalone monitoring signals, and machine learning is performed on standalone process quantities P that do not take into account their correlation with other process quantities P. The anomaly prediction detection model M generated by the learning model generation unit 14 is set in the detection computer 6.

[0042] The detection computer 6 performs anomaly prediction detection using an autoencoder. For example, the detection computer 6 acquires multiple process quantities P sent from the data input computer 4. The detection computer 6 inputs these process quantities P as judgment input data to the input layer of the trained anomaly prediction detection model M. Then, the detection computer 6 acquires judgment output data, which is output from the output layer of the anomaly prediction detection model M in response to the input judgment data, and in which the normal state of the multiple process quantities P has been restored. Based on the difference between the judgment input data and the judgment output data, the detection computer 6 determines whether or not there is an anomaly or an anomaly prediction in plant 2. In this embodiment, even when anomaly prediction detection is performed using an autoencoder, false detections can be suppressed and accuracy can be improved.

[0043] This anomaly prediction detection model M comprises an input layer, a hidden layer, and an output layer. The input layer receives either training input data or judgment input data. The output layer outputs judgment output data in response to the judgment input data. The hidden layer's parameters are machine-learned using the training input data. The anomaly prediction detection model M then causes the detection computer 6 to function in such a way that it determines whether or not there is an anomaly or an anomaly precursor in plant 2 based on the difference between the judgment input data and the judgment output data.

[0044] Next, with reference to Figure 3, the classification method of the data classification unit 12 will be described.

[0045] As shown in Figure 3, the data classification unit 12 classifies the process quantity P input from the data input computer 4 according to multiple classification conditions. For example, classification conditions 1 through 5 are pre-set. Depending on these classification results, a physical correlation signal determination flag F1 or a standalone monitoring signal determination flag F2 is input to the classification result processing unit 13.

[0046] For example, if any one of the classification conditions from the first to the fifth is met (if the result is YES in any one step), the single monitoring signal determination flag F2 is input to the classification result processing unit 13. On the other hand, if none of the classification conditions from the first to the fifth are met (if the result is NO in all steps), the physical correlation signal determination flag F1 is input to the classification result processing unit 13.

[0047] Next, referring to Figures 4 through 13, we will explain the classification criteria from the first to the fifth.

[0048] In the graphs described below (Figures 5, 7, 9, 11, and 13), the process quantity P obtained during the learning period becomes the learning input data, and the process quantity P obtained during the monitoring period becomes the judgment input data. In the graphs, the solid line represents the measured value, and the dashed line represents the predicted value generated by the anomaly prediction detection model M. Here, the measured value (solid line) during the learning period becomes the learning input data. Furthermore, the measured value (solid line) during the monitoring period becomes the judgment input data, and the predicted value (dashed line) during the monitoring period becomes the judgment output data.

[0049] First, let's explain the first classification criterion. The first classification criterion involves separating process quantities P (Figure 3) that are expected to be fixed values ​​in operation into individual monitoring signals. This suppresses false correlations between process quantities that are expected to be fixed values.

[0050] As shown in Figure 4, let's assume, for example, that there is a first system piping 21 and a second system piping 22. The first system piping 21 is equipped with a flow control valve 23. The second system piping 22 is equipped with a flow meter 24 for measuring its flow rate. Here, the graph in Figure 5(A) is obtained based on the opening signal (process amount P) of the flow control valve 23. The graph in Figure 5(B) is obtained based on the measured value (process amount P) of the flow meter 24.

[0051] Here, assuming that the equipment that obtains the process quantity P (Figure 3) is operated with fixed values ​​such as the flow control valve 23, there is no physical correlation between the opening signal indicating the open / closed state of the flow control valve 23 and the flow rate of the second system piping 22. However, if the trends during the learning period are similar, a correlation (spurious correlation) between the process quantities will be learned as existing.

[0052] For example, if, during a certain portion of the learning period 25, the opening signal of the flow control valve 23 and the flow rate of the second system piping 22 happen to show the same tendency, such as becoming a constant value, they will be learned as being correlated with each other.

[0053] As a result, the predicted flow rate of the second piping system 22 becomes misaligned. For example, during a certain portion 26 of the monitoring period, the opening signal of the flow control valve 23 may fluctuate due to noise. In this case, even though the flow control valve 23 fluctuates due to a single factor, the predicted flow rate of the second piping system 22, which has been learned to correlate with this fluctuation, may fluctuate simultaneously.

[0054] In other words, the predicted flow rate of the second system piping 22 (output data for judgment) will differ from the measured value (input data for judgment). As a result, even though there is no actual abnormality in the flow rate of the second system piping 22, it will be mistakenly judged as being abnormal.

[0055] Therefore, process quantities P that take a fixed value in operation, such as the opening signal of the flow control valve 23, are assigned a standalone monitoring signal determination flag F2 (Figure 3) and are classified as process quantities P for which correlation is not learned.

[0056] In this embodiment, the data classification unit 12 (Figure 3) classifies at least one process quantity P as uncorrelated data if it is operated at a fixed value. In this way, the classification of process quantities P is performed based on whether or not it is a fixed value, thus enabling appropriate classification.

[0057] Next, the second classification criterion will be explained. The second classification criterion involves classifying process quantity P (Figure 3), which causes mis-interlocking due to changes in correlation caused by switching operating conditions in a system with redundant equipment, as a single monitoring signal. This suppresses mis-interlocking caused by switching operating conditions of redundant equipment.

[0058] As shown in Figure 6, for example, suppose there is a system in which a first pipe 31 and a second pipe 32 merge to form a third pipe 33. The first pipe 31 is equipped with a first pump 34 and a first pressure gauge 35 for measuring its discharge pressure. The second pipe 32 is equipped with a second pump 36 and a second pressure gauge 37 for measuring its discharge pressure. The third pipe 33 is equipped with a flow meter 38 for measuring its flow rate. Here, the graph in Figure 7(A) is obtained based on the discharge pressure (process amount P) obtained by the first pressure gauge 35. The graph in Figure 7(B) is obtained based on the discharge pressure (process amount P) obtained by the second pressure gauge 37. The graph in Figure 7(C) is obtained based on the flow rate (process amount P) of the pipe obtained by the flow meter 38.

[0059] The system is designed to be redundant with the first pump 34 and the second pump 36. The operator can select the conditions for redundant operation of the first pump 34 and the second pump 36. For example, during the learning period, only the first pump 34 is operated at all times. During the learning period, the other pump, the second pump 36, remains stopped.

[0060] If only the first pump 34 was operating during the learning period, the correlation between the discharge pressure of the first pump 34, which indicates the operating state, and the flow rate of the third pipe 33, which indicates the movement of the entire system, will be learned. On the other hand, since the discharge pressure of the second pump 36, which was stopped during the learning period, did not fluctuate, the correlation between the discharge pressure of the first pump 34 and the flow rate of the third pipe 33 will not be learned.

[0061] Here, for example, during the monitoring period, the conditions for redundant operation may change due to the operator's selection, resulting in a state where the first pump 34 is stopped and only the second pump 36 is operating. In this case, the predicted value of the discharge pressure of the first pump 34, which has been learned to correlate with the flow rate of the third pipe 33 that is linked to the discharge pressure of the second pump 36, will be mis-operated. The predicted value of the flow rate of the third pipe 33 will also be affected, leading to mis-operation.

[0062] In other words, the predicted values ​​(output data for judgment) of the discharge pressure of the first pump 34 and the flow rate of the third pipe 33 will differ from the measured values ​​(input data for judgment). As a result, even though there is no actual abnormality in the discharge pressure of the first pump 34 and the flow rate of the third pipe 33, the system will mistakenly judge them as abnormal.

[0063] Therefore, a single monitoring signal determination flag F2 (Figure 3) is assigned to the process volume P of the discharge pressure of the first pump 34 and the second pump 36, whose correlation changes when the operating conditions are switched, and they are classified as process volume P for which the correlation is not learned.

[0064] In this embodiment, the data classification unit 12 (Figure 3) detects when some of the operating conditions of the plant 2 are switched, and the correlation between at least one process quantity P and other process quantities P changes. do In this case, at least one of these process quantities P is classified as uncorrelated data. In this way, the classification of process quantities P is performed based on whether or not there is a change when some of the operating conditions of plant 2 are switched, thus enabling appropriate classification.

[0065] Next, the third classification criterion will be explained. The third classification criterion involves classifying process quantities P (Figure 3) that cause mis-interlocking due to spurious correlation in systems with equipment that operates intermittently by operator control as individual monitoring signals. This suppresses mis-interlocking due to spurious correlation between equipment that operates intermittently.

[0066] As shown in Figure 8, let's assume, for example, that there is a first system piping 41 and a second system piping 42. The first system piping 41 is equipped with a first pump 43 and a first pressure gauge 44. The second system piping 42 is equipped with a second pump 45, a second pressure gauge 46, a tank 47, and a water level gauge 48. Here, the graph in Figure 9(A) is obtained based on the discharge pressure (process amount P) obtained by the first pressure gauge 44. The graph in Figure 9(B) is obtained based on the discharge pressure (process amount P) obtained by the second pressure gauge 46. The graph in Figure 9(C) is obtained based on the water level (process amount P) in the tank 47 obtained by the water level gauge 48.

[0067] For example, the integrity of the equipment is confirmed by periodic operation by an operator. Here, there is a physical correlation between the discharge pressure of the second pump 45 located in the second system piping 42 and the water level in the tank 47. On the other hand, there is no physical correlation between the discharge pressure of the first pump 43 located in the first system piping 41 and the water level in the tank 47 located in the second system piping 42. However, because the discharge pressure of the first pump 43 and the water level in the tank 47 coincide for a long period during the learning period, they are learned to have a correlation (spurious correlation).

[0068] For example, suppose the second pump 45 is operated during a certain portion 49 of the learning period. During this operation, the discharge pressure, which indicates the operating state of the second pump 45, and the water level in the tank 47 fluctuate in conjunction. If, by chance, the first pump 43 is operated at the same time that the water level in the tank 47 is fluctuating due to the operation of the second pump 45, then the fluctuations in the discharge pressure, which indicates the operating state of the first pump 43, and the water level in the tank 47 will be incorrectly learned as being correlated.

[0069] As a result, during a certain portion 50 of the monitoring period, when the first pump 43 is operated, the predicted value of the water level in tank 47, which was incorrectly learned as being correlated, malfunctions in conjunction with the discharge pressure of the first pump 43.

[0070] In other words, the predicted value of the water level in tank 47 (output data for judgment) will differ from the actual measured value (input data for judgment). As a result, even though there is no abnormality in the actual water level of tank 47, it will be mistakenly judged as being abnormal.

[0071] Furthermore, in such cases, the discharge pressure of the second pump 45 may be mistakenly learned as being correlated with fluctuations in other intermittently moving process quantities P that have no physical correlation.

[0072] Therefore, process quantities P that operate intermittently, such as the discharge pressure of the first pump 43 or the discharge pressure of the second pump 45, are assigned a single monitoring signal determination flag F2 (Figure 3) and are classified as process quantities P for which correlation is not learned.

[0073] In this embodiment, the data classification unit 12 (Figure 3) classifies at least one process quantity P as uncorrelated data if it changes due to intermittent operation of a part of the plant 2 (Figure 2). In this way, the classification of process quantity P is performed based on whether or not intermittent operation is expected, thus enabling appropriate classification.

[0074] Next, we will explain the fourth classification criterion. The fourth classification criterion involves separating process quantity P (Figure 3), which is affected by external factors such as climate and causes mis-interlocking, into individual monitoring signals. This suppresses mis-interlocking due to spurious correlations in equipment that is affected and fluctuates due to external factors.

[0075] As shown in Figure 10, let's assume there is a predetermined building 51. A first pressure gauge 54 is provided to measure the differential pressure between a first piece of equipment 52 located inside the building 51 and a second piece of equipment 53 located outside the building 51. Furthermore, a second pressure gauge 57 is provided to measure the differential pressure between a third piece of equipment 55 and a fourth piece of equipment 56, both located inside the building 51. Note that the fourth piece of equipment 56 is located in a different room 58 from the third piece of equipment 55. Here, the graph in Figure 11(A) is obtained based on the differential pressure inside and outside the building 51 (process amount P), which is obtained by the first pressure gauge 54. The graph in Figure 11(B) is obtained based on the differential pressure inside the building 51 (process amount P), which is obtained by the second pressure gauge 57. Note that since the second piece of equipment 53 is located outside the building 51, it is susceptible to the effects of external environmental factors such as changes in atmospheric pressure.

[0076] For example, during the learning period, fluctuations in atmospheric pressure cause the readings of the first pressure gauge 54 and the second pressure gauge 57 to fluctuate. If the first pressure gauge 54 and the second pressure gauge 57, which have no physical correlation with each other, show similar fluctuations, the system may learn that there is a correlation (spurious correlation) between the process quantities because the trends during the learning period are similar.

[0077] As a result, predicted values ​​may become misaligned. For example, during a certain portion of the monitoring period 59, the differential pressure inside and outside the building (process amount P) of the first pressure gauge 54 may fluctuate due to external environmental factors such as weather or wind. In this case, the predicted value of the indoor differential pressure (process amount P) of the second pressure gauge 57, which has been learned to be correlated, may become misaligned.

[0078] In other words, the predicted value of the indoor differential pressure from the second pressure gauge 57 (output data for judgment) will differ from the measured value (input data for judgment). As a result, even though there is no actual abnormality in the indoor differential pressure, it will be mistakenly judged as being abnormal.

[0079] Since there are countless patterns of fluctuations caused by such external environmental factors, comprehensive learning is generally difficult. For this reason, process quantities P that fluctuate due to external environmental factors such as differential pressure inside and outside a building are assigned a single monitoring signal judgment flag F2 (Figure 3) and are classified as process quantities P for which correlation is not learned.

[0080] In this embodiment, the data classification unit 12 (Figure 3) classifies at least one process quantity P as uncorrelated data if that quantity changes due to the external environment of plant 2 (Figure 2). In this way, the classification of process quantities P is performed based on whether or not they change due to the external environment of plant 2, thus enabling appropriate classification.

[0081] Next, let's explain the fifth classification criterion. The fifth classification criterion involves classifying process quantities P (Figure 3) that include sudden fluctuations (abrupt changes) that cannot be fully covered during the learning period as individual monitoring signals. This suppresses false correlations between process quantities that cause sudden fluctuations.

[0082] Here, a sudden change refers to a sudden or abrupt change in which at least one process quantity P undergoes a change compared to a predetermined threshold.

[0083] For example, let's define sudden changes as fluctuations that occur at a shorter interval than the sampling period of the process quantity P. If the fluctuation occurs at a shorter interval than the sampling period, the value will change infinitely many times if the sampling timing is slightly off. It is impossible to learn all patterns of such fluctuations.

[0084] As shown in Figure 12, let's assume, for example, that there is a first system piping 61 and a second system piping 62. The first system piping 61 is equipped with a first pump 63, a first flow control valve 64, and a first flow meter 65. The second system piping 62 is equipped with a second pump 66, a second flow control valve 67, and a second flow meter 68. Here, the graph in Figure 13(A) is obtained based on the flow rate (process amount P) of the first system piping 61, which is the measurement value of the first flow meter 65. The graph in Figure 13(B) is obtained based on the flow rate (process amount P) of the second system piping 62, which is the measurement value of the second flow meter 68.

[0085] During the learning period, the fluctuation trends, including sudden changes in the flow rates of the first system piping 61 and the second system piping 62, may become similar depending on the operating conditions of the first pump 63 and the second pump 66, or the driving conditions of the first flow control valve 64 and the second flow control valve 67. In this case, the system may learn that there is a correlation (spurious correlation) between the process quantities.

[0086] As a result, if the flow rate of the first piping system 61 undergoes a sudden change of magnitude that was not present during the learning period during a certain portion of the monitoring period 69, the predicted flow rate of the second piping system 62, which was learned to be correlated, will be misinterpreted.

[0087] In other words, the predicted flow rate of the second system piping 62 (output data for judgment) will differ from the measured value (input data for judgment). As a result, even though there is no actual abnormality in the flow rate of the second system piping 62, it will be mistakenly judged as being abnormal.

[0088] Therefore, a single monitoring signal determination flag F2 (Figure 3) is assigned to process quantities P that cause sudden changes, and they are classified as process quantities P for which correlation is not learned.

[0089] In this embodiment, the data classification unit 12 (Figure 3) classifies at least one process quantity P as uncorrelated data if it undergoes a sudden change. In this way, the classification of process quantities P is performed based on whether or not they undergo a sudden change, thus enabling appropriate classification.

[0090] As described above, process quantities P are classified according to at least one classification criterion. Then, physical correlation signals that learn the correlation between process quantities and process quantities P that may induce false correlations due to spurious correlations are classified as individual monitoring signals, making it possible to generate input data for training. In this way, when generating a learning model using machine learning that handles a large number of sensors 3 and may induce false detections due to false correlations, it is possible to improve the detection performance.

[0091] Next, a modified example of the anomaly prediction detection system 1 will be described using Figure 14. Note that components identical to those shown in the previously described embodiment are denoted by the same reference numerals, and redundant descriptions are omitted. The configuration applied in this modified example may also be applied to the previously described embodiment, or combined as appropriate.

[0092] In the embodiment described above, the data classification unit 12 of the learning computer 5 automatically classifies the process amount P. In this modified example, however, by adding a process amount classification flag setting unit 15, the process amount classification flag F3 is output outside of the learning computer 5.

[0093] For example, the process quantity classification flag F3 is input from the process quantity classification flag setting unit 15 to the data input computer 4. This input process for the process quantity classification flag F3 may be performed based on user operation, or it may be performed automatically by the process quantity classification flag setting unit 15.

[0094] The data input computer 4 outputs a process quantity classification flag F3 associated with the process quantity P. The data classification unit 12 of the learning computer 5 reads the process quantity classification flag F3 and sets either a physical correlation signal determination flag F1 or a standalone monitoring signal determination flag F2 for the process quantity P.

[0095] For example, as in the first classification condition mentioned above, there are cases where it is difficult to determine from the data whether a process quantity P that should be classified as an eigenvalue is an eigenvalue due to conditions such as noise. Even in this case, the process quantity classification flag F3 is set based on external information such as design information, so that it is appropriately classified as a standalone monitoring signal.

[0096] Furthermore, as with the second classification criterion mentioned above, changes in correlation due to switching operating conditions may not be fully covered due to reasons such as missing or abnormal data. Also, if only training data is available, automatic classification becomes difficult. Even in such cases, the process quantity classification flag F3 is set based on external information such as design information, allowing for appropriate classification into individual monitoring signals.

[0097] Furthermore, as with the third classification criterion mentioned above, even for equipment that operates intermittently, it may be possible to obtain only training data from periods when intermittent operation does not occur, due to reasons such as missing data or the inclusion of abnormal data. In this case, it becomes difficult to automatically classify the process quantity P, but by setting the process quantity classification flag F3 based on external information such as design information, it is appropriately classified into a single monitoring signal.

[0098] Furthermore, as mentioned in the fourth classification criterion above, some process quantities P are affected by external factors. In this case, if the process quantity P does not include external factors that should be excluded, such as the atmospheric pressure outside building 51 (Figure 10), automatic classification becomes difficult. Even in this case, the process quantity classification flag F3 is set based on external information such as design information, allowing for appropriate classification into individual monitoring signals.

[0099] Furthermore, as mentioned in the fifth classification criterion above, some processes exhibit rapid fluctuations in the process quantity P. Here, it can be difficult to determine whether a rapid fluctuation occurs at a shorter time than the sampling period. For example, if the set value temporarily increases or decreases due to interlock conditions, it may not be classified as a sudden change. Even in this case, the process quantity classification flag F3 is set based on external information such as design information, ensuring that the signal is appropriately classified as a standalone monitoring signal.

[0100] In this way, process quantities P that may induce false correlations due to spurious correlations based on external information such as design information not included in the process quantity P can be classified as individual monitoring signals. These can then be provided as training input data (input information) for machine learning.

[0101] The modified learning computer 5 sets process quantity classification flags F3 to classify each process quantity P as either correlated or uncorrelated data based on external information. Here, the external information is information that is not expected from the process quantity P generated from plant 2, such as design information, or information that is not included in the process quantity P generated from plant 2. In this way, the classification of process quantity P is performed based on external information such as design information, so an appropriate classification can be made.

[0102] The anomaly prediction detection system 1 of the above-described embodiment comprises a control device with highly integrated processors such as an FPGA (Field Programmable Gate Array), GPU (Graphics Processing Unit), CPU (Central Processing Unit), and dedicated chips; storage devices such as ROM (Read Only Memory) and RAM (Random Access Memory); external storage devices such as HDD (Hard Disk Drive) and SSD (Solid State Drive); a display device such as a display; input devices such as a mouse and keyboard; and a communication interface. This anomaly prediction detection system 1 can be implemented with a hardware configuration using a normal computer.

[0103] The program executed by the anomaly prediction detection system 1 of the above-described embodiment is provided pre-installed in ROM or the like. Additionally or alternatively, this program is provided as an installable or executable file stored on a computer-readable non-temporary storage medium such as a CD-ROM, CD-R, memory card, DVD, or flexible disk (FD).

[0104] Furthermore, the program executed by this anomaly prediction detection system 1 may be stored on a computer connected to a network such as the Internet and provided for download via the network. Alternatively, this anomaly prediction detection system 1 can be configured by combining separate modules, each independently performing its respective function, which are interconnected via a network or dedicated line.

[0105] According to the embodiment described above, by classifying each process quantity P into correlated data that learns the correlation between multiple process quantities and uncorrelated data that does not learn the correlation between multiple process quantities, false detections caused by spurious correlations can be suppressed.

[0106] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be carried out in a variety of other forms, and various omissions, substitutions, modifications, and combinations are possible without departing from the spirit of the invention. These embodiments or their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0107] 1...Anomaly prediction detection system, 2...Plant, 3...Sensor, 4...Data input computer, 5...Learning computer, 6...Detection computer, 7...Input unit, 8...Output unit, 9...Communication unit, 10...Storage unit, 11...Processing circuit, 12...Data classification unit, 13...Classification result processing unit, 14...Learning model generation unit, 15...Process volume classification flag setting unit, 21...First system piping, 22...Second system piping, 23...Flow rate control valve, 24...Flow meter, 25...Part of the learning period, 26...Part of the monitoring period, 31...First piping, 32...Second piping, 33...Third piping, 34...First pump, 35...First pressure gauge, 36...Second pump, 37...Second pressure gauge, 38...Flow meter, 41...First system piping, 42...Second system piping, 43...First pump, 44...First pressure gauge, 45...Second pump, 46...Second pressure gauge, 47...Tank, 48...Water level gauge, 49...Part of the learning period, 50...Part of the monitoring period, 51...Building, 52...First equipment, 53...Second equipment, 54...First pressure gauge, 55...Third equipment, 56...Fourth equipment, 57...Second pressure gauge, 58...Room, 59...Part of the monitoring period, 61...First system piping, 62...Second system piping, 63...First pump, 64...First flow control valve, 65...First flow meter, 66...Second pump, 67...Second flow control valve, 68...Second flow meter, 69...Part of the monitoring period, F1...Physical correlation signal judgment flag, F2...Standalone monitoring signal judgment flag, F3...Process amount classification flag, M...Anomaly prediction detection model, P...Process amount, V...Process value.

Claims

1. The system includes at least one computer that detects abnormalities or signs of abnormalities in the monitored facility using an abnormality prediction detection model, The aforementioned at least one computer is The quantities of multiple processes occurring at the aforementioned facility are acquired, The aforementioned multiple process quantities are classified into a predetermined number of groups, and according to predetermined classification conditions related to switching the operating conditions of the target facility, when one process quantity in the same group correlates with other process quantities, it is classified as correlated data, and when one process quantity in the same group does not correlate with other process quantities, it is classified as uncorrelated data. The aforementioned multiple process quantities generate training input data linked as either the correlated data or the uncorrelated data for each of the aforementioned multiple groups. The aforementioned training input data is input to the anomaly prediction detection model to perform machine learning. An anomaly prediction detection system configured to determine the anomaly or an anomaly prediction of the target facility based on the difference between the input data and output data of the anomaly prediction detection model on which the machine learning has been performed.

2. The aforementioned at least one computer is An anomaly prediction detection system according to claim 1, configured to set process quantity classification flags that classify the plurality of process quantities into correlated data or uncorrelated data according to the predetermined classification conditions.

3. An anomaly prediction detection system according to claim 1, wherein, according to the predetermined classification conditions, if one of the multiple process quantities is operated at a fixed value, the one process quantity is classified as uncorrelated data.

4. An anomaly prediction detection system according to claim 1, wherein if the correlation between one process quantity and the other process quantities among the plurality of process quantities changes according to the predetermined classification conditions, the one process quantity is classified as uncorrelated data.

5. An anomaly prediction detection system according to claim 1, wherein, according to the predetermined classification conditions, if one of the multiple process quantities changes due to intermittent operation of a part of the target facility, the one process quantity is classified as uncorrelated data.

6. An anomaly prediction detection system according to claim 1, wherein, according to the predetermined classification conditions, if one of the multiple process quantities changes due to the external environment of the target facility, the one process quantity is classified as uncorrelated data.

7. An anomaly prediction detection system according to claim 1, wherein if one of the multiple process quantities undergoes a sudden change according to the predetermined classification conditions, the one process quantity is classified as uncorrelated data.

8. An anomaly prediction and detection method that uses at least one computer to detect anomalies or signs of anomalies in a target facility under monitoring using an anomaly prediction detection model, The quantities of multiple processes occurring at the aforementioned facility are acquired, The aforementioned multiple process quantities are classified into a predetermined number of groups, and according to predetermined classification conditions related to switching the operating conditions of the target facility, when one process quantity in the same group correlates with other process quantities, it is classified as correlated data, and when one process quantity in the same group does not correlate with other process quantities, it is classified as uncorrelated data. The aforementioned multiple process quantities generate training input data linked as either the correlated data or the uncorrelated data for each of the aforementioned multiple groups. The aforementioned training input data is input to the anomaly prediction detection model to perform machine learning. An anomaly prediction detection method for determining the anomaly or an anomaly prediction of the target facility based on the difference between the input data and output data of the anomaly prediction detection model on which the machine learning has been performed.

Images