Training data generation device and training data generation method

Abstract

One aspect of the present invention is a training data generation device 200 that comprises a training data adaptation unit 205 that generates training data for a target facility on the basis of training data for a reference facility by adaptation processing that is based on information about the difference between evaluation data composed of short-term sensor signals outputted from a sensor provided at the target facility and training data composed of long-term sensor signals outputted from a sensor provided at the reference facility.

Description

Learning Data Generation Device and Learning Data Generation Method

[0001] The present invention relates to a learning data generation device and a learning data generation method.

[0002] Power companies supply hot water for district heating using the waste heat of gas turbines or supply high-pressure steam or low-pressure steam to factories. Petrochemical companies operate gas turbines and the like as power facilities. In various plants and facilities using such gas turbines and the like, an abnormality detection device that detects equipment failures or signs of impending failures is extremely important for minimizing damage to society.

[0003] Facilities that require preventive maintenance as described above are not limited to gas turbines and steam turbines. Various things such as water wheels in hydroelectric power plants, nuclear reactors in nuclear power plants, wind turbines in wind power plants, engines in aircraft and heavy machinery, railway vehicle tracks, escalators, and elevators can be considered. Also, at the equipment or component level mounted on these devices, it is required to diagnose failures or their signs for the degradation and lifespan of the mounted batteries.

[0004] The abnormality detection device determines whether the target equipment or plant is normal or abnormal by analyzing sensor data from a plurality of sensors that acquire various physical information attached to the target equipment or plant. For example, Patent Document 1 discloses an equipment condition monitoring method in which data close to new observation data is selected from learning data belonging to a cluster close to the new observation data, a normal model is created from the selected data, an abnormality measure is obtained, a threshold value is determined, an abnormality measure is obtained from the new observation data and the normal model, and this is compared with the threshold value to detect equipment abnormalities. The abnormality measure is a value indicating the degree of deviation from the normal state.

[0005] Furthermore, Patent Document 2 discloses an anomaly analysis system comprising: a main factor extraction unit that extracts sensors that are the main factors influencing the measurement values ​​based on measurement values ​​measured by a plurality of sensors installed in the equipment; and a secondary factor correction unit that generates a model showing the normal state of the equipment using values ​​indicating secondary factors that affect the measurement values ​​in addition to the measurement values ​​measured by the main factor sensors.

[0006] Japanese Patent Publication No. 2014-32455, International Publication No. 2018 / 083720

[0007] The anomaly detection method described in Patent Document 1 learns from past normal data, and therefore requires the accumulation of sensor data over a long period of time when new equipment is put into operation. Even with equipment of the same model as the new equipment being put into operation, individual differences in sensor values ​​will occur due to individual differences in equipment and differences in installation or operating environment. For this reason, the learning data generated using the data of that equipment cannot be directly applied to the new equipment.

[0008] Furthermore, when operations are suspended for large-scale maintenance or other reasons and then resumed, the trends in sensor values ​​change. Therefore, in such cases, it is necessary to re-store the sensor data. For these reasons, there was a problem in that it took time to start or restart the operation of the anomaly detection device.

[0009] Similarly, the anomaly analysis method described in Patent Document 2 also requires the accumulation of training data, which means that it takes time before new equipment can be put into operation or restarted. Although a method for correcting the regression model used in the analysis using secondary factors is described, it is necessary to accumulate secondary factor data separately from sensor data. Furthermore, since the technology described in Patent Document 2 does not take into account individual differences in equipment, even when using the technology described in Patent Document 2, it is not possible to shorten the time from the start of operation of new equipment or the restart of operation of existing equipment to the start of anomaly detection processing.

[0010] This invention was made to solve the above problems, and its objective is to shorten the time from the start of operation of new equipment or the restart of operation of existing equipment to the start of anomaly detection processing.

[0011] A learning data generation device for an anomaly detection device according to one aspect of the present invention is a learning data generation device that generates learning data used for determining an anomaly in an anomaly detection device that determines whether or not an anomaly exists in a target facility that is subject to anomaly detection. The learning data generation device according to one aspect of the present invention includes a learning data adaptation unit that generates learning data for a target facility based on learning data for a reference facility by performing an adaptation process based on information of the difference between evaluation data consisting of sensor signals for a first period output from a sensor installed on the target facility and learning data consisting of sensor signals for a second period longer than the first period output from a sensor installed on a reference facility that is referenced in generating learning data for the target facility.

[0012] According to the present invention, the time from the start of operation of new equipment or the restart of operation of existing equipment to the start of anomaly detection processing can be shortened.

[0013] This is a block diagram showing an example configuration of an anomaly detection device according to one embodiment of the present invention. This is a block diagram showing an example configuration of a learning data generation device according to one embodiment of the present invention. This is a diagram showing an example configuration of a sensor signal storage unit or a sensor signal stored in a sensor signal storage unit according to one embodiment of the present invention. This is a flowchart showing an example of the procedure for anomaly measure calculation processing and anomaly determination processing by an anomaly detection device according to one embodiment of the present invention. This is a diagram showing an overview of the anomaly measure calculation processing when using the fast local subspace method according to one embodiment of the present invention. This is a flowchart showing an example of the procedure for generating target equipment learning data by a learning data generation device according to one embodiment of the present invention. This is a flowchart showing an example of the procedure for feature calculation processing by a feature calculation unit according to one embodiment of the present invention. This is a flowchart showing an example of the procedure for shift amount calculation processing according to one embodiment of the present invention. This is a diagram showing an overview of the individual shift amount search processing according to one embodiment of the present invention. Figure 9A is a diagram showing an example of learning data features, and Figure 9B is a diagram showing an example of evaluation data features. This is a flowchart showing an example of the procedure for learning data fitting processing according to one embodiment of the present invention. This is a diagram showing an example configuration of the learning data generation condition setting screen according to one embodiment of the present invention. This is a diagram showing an example configuration of the result confirmation screen according to one embodiment of the present invention. This is a diagram showing an example of the operation schedule of an anomaly detection device according to one embodiment of the present invention. This is a block diagram showing an example of the hardware configuration of an anomaly detection device and a learning data generation device according to one embodiment of the present invention.

[0014] One embodiment of the present invention will be described in detail below with reference to the drawings. In this specification and the drawings, elements having substantially the same function or configuration are denoted by the same reference numerals, and redundant descriptions are omitted.

[0015] <Configuration of the Anomaly Detection Device> First, the configuration of the anomaly detection device according to this embodiment will be described with reference to Figure 1. Figure 1 is a block diagram showing an example configuration of the anomaly detection device 100. The anomaly detection device according to this embodiment uses the learning data generated by the learning data generation device 200 shown in Figure 2 to detect anomalies or signs of anomalies occurring in equipment not shown.

[0016] As shown in Figure 1, the anomaly detection device 100 includes a sensor signal storage unit 103, a sensor signal input unit 104, a feature vector extraction unit 105, an anomaly measure calculation unit 106, a threshold calculation unit 107, an anomaly determination unit 108, and a learning data storage unit 109.

[0017] The sensor signal storage unit 103 stores sensor signals 102 output from sensors (not shown) attached to equipment 101, which is the equipment targeted for abnormality detection (hereinafter also simply referred to as "target equipment"). The target equipment 101 is, for example, equipment or plants such as gas turbines or steam turbines. In other words, the sensor signal 102 is a signal that represents the state of equipment 101. In the sensor signal storage unit 103, the sensor signal 102 is managed in association with the equipment ID, which is the identifier of the equipment, and the date and time information on when the sensor signal 102 was acquired. The sensor signal 102 is acquired at predetermined intervals (periodically) by the control unit 51 (see Figure 14) of the computer that constitutes the abnormality detection device 100, and stored in the sensor signal storage unit 103.

[0018] The sensor signal input unit 104 outputs the sensor signal 102 read from the sensor signal storage unit 103, or the sensor signal 102 directly input from a sensor attached to the equipment 101, to the feature vector extraction unit 105.

[0019] The feature vector extraction unit 105 extracts feature vectors from the sensor signal 102 at predetermined time intervals and sends the extracted feature vectors to the anomaly measure calculation unit 106. Hereinafter, the "determined time intervals" at which the feature vector extraction unit 105 extracts feature vectors from the sensor signal 102 may be referred to as "each time interval". The feature vector extraction unit 105 also stores the information necessary for feature vector extraction as training data in the training data storage unit 109. A feature vector is a representation of measurements from multiple sensors as a single vector value. Details of the "information necessary for feature vector extraction" will be described later.

[0020] The anomaly measure calculation unit 106 calculates the anomaly measure for each time point using feature vectors in a predetermined manner depending on whether it is during learning or anomaly detection, and sends the calculated anomaly measure to the threshold calculation unit 107 and the anomaly determination unit 108. "Learning" refers to the time when the "learning" process is executed, in which learning data is generated and saved using data stored in the sensor signal storage unit 103. "Anomaly detection" refers to the time when the anomaly determination unit 108 executes the process of detecting an anomaly in the equipment 101. Basically, "learning" is an offline process, and "anomaly detection" is an online process. However, it is also possible to process "anomaly detection" offline.

[0021] The anomaly measure is the amount of deviation (deviation) between the feature vector of interest and the reference feature vector. The reference feature vector is calculated using the feature vectors for a specified training period. The anomaly measure calculation unit 106 also stores the information used to calculate the anomaly measure as training data in the training data storage unit 109. Details of the "information used to calculate the anomaly measure" will be described later.

[0022] The threshold calculation unit 107 calculates a threshold based on the anomaly measure calculated by the anomaly measure calculation unit 106 during training, and sends the calculated threshold to the anomaly determination unit 108. The threshold calculated by the threshold calculation unit 107 is stored as training data in the training data storage unit 109. In other words, the training data includes information necessary for feature vector extraction, information necessary for calculating the anomaly measure, and the threshold.

[0023] The abnormality detection unit 108 detects abnormalities in the equipment 101 by comparing the abnormality measurement for each time point sent from the abnormality measurement calculation unit 106 with the threshold value calculated by the threshold value calculation unit 107. The abnormality detection unit 108 then outputs an abnormality detection result 110.

[0024] <Configuration of the Learning Data Generation Device> Next, the configuration of the learning data generation device 200 according to this embodiment will be described with reference to Figure 2. Figure 2 is a block diagram showing an example configuration of the learning data generation device 200.

[0025] As shown in Figure 2, the learning data generation device 200 includes a sensor signal storage unit 201, a reference equipment learning data storage unit 202, a feature calculation unit 203, a shift amount calculation unit 204, a learning data adaptation unit 205, and a target equipment learning data storage unit 206.

[0026] As explained in Figure 1, the sensor signal storage unit 103 stores the sensor signal 102 output from a sensor (not shown) attached to the equipment 101, which is the equipment targeted for abnormality detection. The sensor signal storage unit 201 shown in Figure 2 can be provided independently of the sensor signal storage unit 103 of the abnormality detection device 100, but it can also be used in conjunction with it.

[0027] The reference equipment learning data storage unit 202 stores learning data for reference equipment. Reference equipment refers to equipment that is referenced when generating learning data for target equipment, and includes other equipment of the same type (having the same specifications) as the target equipment, or equipment identical to the target equipment. Here, equipment identical to the target equipment refers to the target equipment before it was shut down for maintenance or other reasons.

[0028] In other words, the target equipment refers to newly installed equipment or equipment that resumes operation after maintenance, and is the equipment subject to anomaly detection. In the reference equipment learning data storage unit 202, the learning data of the reference equipment is managed in association with the learning data name and learning period. The reference equipment learning data stored in the reference equipment learning data storage unit 202 is learning data generated based on sensor signals 102 acquired over a long period of time, such as six months to one year or more (an example of the second period).

[0029] The feature calculation unit 203 reads the sensor signals of the reference equipment from the reference equipment learning data storage unit 202 and reads the short-term sensor signals of the target equipment from the sensor signal storage unit 201. "Short-term" refers to a short period of about 5 to 10 days (an example of the first period). In the following explanation, the short-term sensor signals of the target equipment will also be referred to as "evaluation data". The sensor signals used to generate the reference equipment learning data will also be referred to as "learning data". The feature calculation unit 203 then calculates features from the read learning data and sends the calculated features (hereinafter also referred to as "learning data features") to the shift amount calculation unit 204.

[0030] The shift amount calculation unit 204 calculates the shift amount of the training data by comparing the training data features, which are the characteristics of the reference equipment, with the evaluation data features, which are the characteristics of the target equipment, and sends the calculated shift amount to the training data fitting unit 205. The shift amount calculated by the shift amount calculation unit 204 is the difference between the training data features and the evaluation data features, and is the shift amount required to fit the training data features to the evaluation data.

[0031] The learning data adaptation unit 205 generates learning data suitable for the target equipment (hereinafter also referred to as "target equipment learning data") by shifting the learning data of the reference equipment by the shift amount calculated by the shift amount calculation unit 204. The target equipment learning data generated by the learning data adaptation unit 205 is stored in the target equipment learning data storage unit 206.

[0032] The target equipment learning data storage unit 206 stores the target equipment learning data, associated with the target equipment ID (not shown), which is the identifier of the target equipment, and the time information when the learning data was saved. The reference equipment learning data storage unit 202 and the target equipment learning data storage unit 206 may be shared with the learning data storage unit 109 of the anomaly detection device 100.

[0033] <Configuration of sensor signals stored in the sensor signal storage unit> Next, with reference to Figure 3, the configuration of sensor signals stored in the sensor signal storage unit 103 of the anomaly detection device 100 and the sensor signal storage unit 201 of the learning data generation device 200 will be described. Figure 3 is a diagram showing an example of the configuration of sensor signals stored in the sensor signal storage unit 103 or the sensor signal storage unit 201.

[0034] As shown in Figure 3, the sensor signal storage unit 103 or sensor signal storage unit 201 stores multiple sensor signals 102 (e.g., "27935" for "Sensor Sn1", "478" for "Sensor Sn2") in association with information about the date the sensor signal was acquired. The sensor signal 102 is a multidimensional time-series signal generated by acquiring multiple pieces of physical information with different physical characteristics at predetermined intervals. Depending on the equipment, the number of sensors may range from several hundred to several thousand.

[0035] Each of these sensors outputs sensor values, depending on its type, such as temperature of cylinders, oil, or coolant, pressure of oil or coolant, rotational speed of shafts, room temperature, and operating time. These sensor values ​​include not only outputs and states of equipment and plants, but also control signals used to control a state to a specific value (e.g., a target value). Furthermore, results calculated between sensor signals at the same time, or results obtained by applying statistical processing such as averaging, maximizing, and minimizing over a certain time-series range, may be treated similarly to sensor signals.

[0036] <Anomaly Measure Calculation Processing by Anomaly Detection Device> Next, with reference to Figure 4, the anomaly measure calculation processing and anomaly determination processing by the anomaly detection device 100 according to this embodiment will be described. Figure 4 is a flowchart showing an example of the procedure for the anomaly measure calculation processing and anomaly determination processing by the anomaly detection device 100. The anomaly detection device 100 according to this embodiment performs anomaly measure calculation processing during learning and anomaly determination processing when an anomaly is detected.

[0037] [Anomaly Measure Calculation Process During Learning] First, the anomaly measure calculation process of the anomaly detection device 100 will be explained. First, the sensor signal input unit 104 of the anomaly detection device 100 inputs the sensor signal input from the equipment 101, or the sensor signal during the learning period acquired from the sensor signal storage unit 103, to the feature vector extraction unit 105 (step S1). Next, the feature vector extraction unit 105 canonizes the input sensor signal and extracts a feature vector from the canonicalized sensor signal (step S2).

[0038] The feature vector extraction unit 105 canonically normalizes the sensor signals so that multiple sensor signals with different units and scales can be handled similarly. Specifically, the feature vector extraction unit 105 uses the mean μ and standard deviation σ of each sensor signal during the learning period to transform the value of each sensor signal x so that the mean of the solution to the equation (x - μ) / σ is "0" and the variance is "1". The feature vector extraction unit 105 also stores the mean μ and standard deviation σ of each sensor signal in the learning data storage unit 109 so that the same transformation can be performed even when an anomaly is detected.

[0039] Alternatively, the feature vector extraction unit 105 uses the maximum value mx and minimum value mn of each sensor signal during the learning period to transform the value of each sensor signal x such that the maximum solution to the equation (x - mn) / (mx - mn) is "1" and the minimum solution is "0". The feature vector extraction unit 105 may also use the top N percentile value and the bottom N percentile value instead of the maximum and minimum values. "N" is a parameter that is specified in advance.

[0040] The feature vector extraction unit 105 stores the maximum value mx and minimum value mn, or the top N percent value and bottom N percent value, of each sensor signal in the learning data storage unit 109 so that the same transformation can be performed even when an anomaly is detected. Hereinafter, these values ​​used for canonicalization will also be referred to as "canonicalization parameters".

[0041] The feature vector extraction unit 105 extracts a feature vector by arranging the canonicalized sensor signals as elements to form a vector. Alternatively, the feature vector extraction unit 105 may set windows such as ±1, ±2, ... for a given time and extract a feature vector consisting of elements equal to the window width (3, 5, ...) × the number of sensors. By extracting such a feature vector, the feature vector extraction unit 105 can also extract features representing the time change of the sensor signal as a feature vector.

[0042] After processing in step S2, the anomaly measure calculation unit 106 of the anomaly detection device 100 clusters the feature vectors of the learning period extracted by the feature vector extraction unit 105 (step S3). Then, the anomaly measure calculation unit 106 records the center position and cluster members of each cluster obtained as a result of the clustering in the learning data storage unit 109.

[0043] A member of a cluster is a feature vector belonging to the cluster. Any method may be used as the clustering method, but in this embodiment, an example of using the fast local subspace method is given. When using the fast local subspace method, the anomaly measure calculation unit 106 selects the initial positions of a specified number of cluster centers so as to be far from the feature vectors during the learning period. Then, after updating the cluster center positions and cluster members using the K-means method, the anomaly measure calculation unit 106 adjusts the number of clusters to a certain number by thinning out or adding cluster members.

[0044] Next, the anomaly measure calculation unit 106 calculates the anomaly measure of each feature vector extracted in step S12 (step S4). Here, referring to FIG. 5, the anomaly measure calculation process by the anomaly measure calculation unit 106 when using the fast local subspace method will be described. FIG. 5 is a diagram showing an overview of the anomaly measure calculation process when using the fast local subspace method. In the example shown in FIG. 5, the anomaly measure calculation unit 106 first searches for and selects one to k (k is a natural number of 2 or more) neighboring clusters of the target vector q indicated by a circle. In FIG. 5, the members of the cluster selected by the anomaly measure calculation unit 106 are shown by a solid rectangular frame.

[0045] Next, the anomaly measure calculation unit 106 measures the projection distance when the target vector q is projected onto the k - 1 dimensional affine subspace Sa spanned by the selected k neighboring vectors. FIG. 5 shows an example when k is 3, and the three neighboring vectors x1 to x3 are shown by a shaded rectangle. The three neighboring vectors x1 to x3 form an affine subspace Sa, that is, a plane shown by the shaded rectangle. And the point Xb (indicated by a black circle) on the affine subspace Sa closest to the target vector q becomes the projection point, that is, the reference vector. The anomaly measure calculation unit 106 calculates the distance from the target vector q to the reference vector Xb or its square as the anomaly measure. Note that the value of k may be any value as long as it is sufficiently smaller than the dimensionality of the feature vector.

[0046] Here, a specific method for calculating the anomaly measure will be described. First, the anomaly measure calculation unit 106 creates a matrix Q with k copies of the evaluation data q and a matrix X with the vectors xi (i = 1,..., k) arranged. Next, the anomaly measure calculation unit 106 obtains the correlation matrix C of the matrices Q and X using the following formula (1).

[0047] [Equation 1] C = (Q - X) Ｔ (Q - X) … Equation (1)

[0048] Next, the anomaly measure calculation unit 106 calculates a coefficient vector b representing the weighting of the neighboring vectors xi using the following formula (2). Then, the anomaly measure calculation unit 106 calculates the norm or its square of the vector (q - Xb) as the anomaly measure d.

[0049]

[0050] Returning to FIG. 4 again, the description continues. The threshold calculation unit 107 of the anomaly detection device 100 calculates a threshold (step S5). Here, the threshold calculated by the threshold calculation unit 107 is a threshold for not determining normal learning data as an anomaly. That is, in step S5, the threshold calculation unit 107 calculates the maximum value of the anomaly measure obtained from the normal learning data as the threshold. This threshold is compared with the anomaly measure in the anomaly determination unit 108, and the normal or abnormal state of the facility is determined based on the comparison result.

[0051] The anomaly detection device 100 that has performed the learning process stores the generated respective information as learning data in the learning data storage unit 109. The learning data includes at least information necessary for feature vector extraction, information necessary for anomaly measure calculation, and a threshold for anomaly determination. The information necessary for feature vector extraction is the average μ, standard deviation σ, maximum value mx, minimum value mn, etc. of each sensor signal calculated by the feature vector extraction unit 105 in the process of step S2 in FIG. The information necessary for anomaly measure calculation is the number of clusters calculated by the anomaly measure calculation unit 106 in step S3, and the ID of the center position and members of each cluster and all feature vector data that are members of any cluster.

[0052] The anomaly detection threshold is the threshold calculated by the threshold calculation unit 107 in step S5. In addition, the anomaly detection device 100 records analysis parameters that should be the same during training and anomaly detection in the training data storage unit 109. Analysis parameters that should be the same during training and anomaly detection include the number of cluster selections and the number of neighbors (number of neighbor vectors k). After completing the processing in step S5, the anomaly measurement calculation process during training by the anomaly detection device 100 is completed.

[0053] Figure 4 shows an example of anomaly measure calculation using the fast local subspace method, but the anomaly measurement calculation process may be performed using methods other than the fast local subspace method. Other methods include, for example, the local subspace method, vector quantization clustering, projected distance method, regression analysis, Gaussian process, and deep learning.

[0054] When using the local subspace method, the feature vector extraction unit 105 searches for and selects k neighboring vectors for the target vector q in Figure 5, using all feature vectors from the training period. Then, the anomaly measure calculation unit 106 uses the projection distance or its square when the target vector q is projected onto the k-1 dimensional affine subspace Sa spanned by the k neighboring vectors selected by the feature vector extraction unit 105 as the anomaly measure. When using the local subspace method, the clustering in step S3 of Figure 4 is omitted, and the information required for calculating the anomaly measure is the entire feature vector data from the training period.

[0055] When using the vector quantization clustering method, the feature vector extraction unit 105 searches for the nearest neighbor cluster of the vector q of interest. The anomaly measure calculation unit 106 then uses the distance to the center position of the searched cluster, or its square, as the anomaly measure. The information required for calculating the anomaly measure is the number of clusters and the center position of each cluster.

[0056] The projected distance method is a method for creating a subspace with its own origin for multiple feature vectors, i.e., an affine subspace Sa (the space with maximum variance). Therefore, when using the projected distance method, the feature vector extraction unit 105 calculates the affine subspace Sa for the feature vectors that are members of each cluster using the following method.

[0057] First, the feature vector extraction unit 105 finds the mean μ of the target feature vector and the covariance matrix Σ. Next, it solves the eigenvalue problem of Σ and arranges the eigenvectors corresponding to the r eigenvalues ​​specified in advance, in descending order of value, to form a matrix U. The feature vector extraction unit 105 then uses matrix U as an orthonormal basis for the affine subspace Sa. Here, the number of eigenvalues ​​r is set to be smaller than the dimension of the feature vector and smaller than the number of selected data points. Alternatively, the number of eigenvalues ​​r may not be a fixed number, but rather the value obtained when the cumulative contribution rate of the eigenvalues, starting from the largest, exceeds a predetermined proportion.

[0058] When the number of eigenvalues ​​r is set to such a value, the point on the affine subspace Sa closest to the vector of interest becomes the reference vector. The anomaly measure calculation unit 106 then uses the norm or square of the norm of the residual vector, which is obtained by subtracting the reference vector from the vector of interest, as the anomaly measure. The information required for calculating the anomaly measure is the number of clusters, the mean μ of the feature vectors of each cluster, and the orthonormal matrix U of the affine subspace Sa.

[0059] [Anomaly Determination Processing in Case of Anomaly Detection] Returning to Figure 4, we will now explain the anomaly determination processing in case of anomaly detection by the anomaly detection device 100.

[0060] First, the sensor signal input unit 104 of the anomaly detection device 100 receives a sensor signal 102 directly from the sensor signal storage unit 103 or from a sensor (not shown) attached to the equipment 101 (step S11). Next, the feature vector extraction unit 105 performs canonicalization of the input sensor signal and extracts feature vectors using the same method as in step S2 (step S12). In step S12, the feature vector extraction unit 105 performs canonicalization of the sensor signal using the mean μ of each sensor signal stored in the learning data storage unit 109, and either the standard deviation δ, the maximum value mx and minimum value mn, the upper N percent value and the lower N percent value.

[0061] Next, the anomaly measure calculation unit 106 uses the information necessary for calculating the anomaly measure stored in the learning data storage unit 109 to calculate the anomaly measure for the input feature vector in the same manner as in step S4 (step S13). Next, the anomaly determination unit 108 compares the calculated anomaly measure with the threshold value calculated in step S5 and stored in the learning data storage unit 109 to determine whether the target equipment is normal or abnormal (step S14). After completing the process in step S14, the anomaly determination process when an anomaly is detected by the anomaly detection device 100 is completed.

[0062] <Target Equipment Learning Data Generation Process by Learning Data Generation Device> Next, with reference to Figure 6, the target equipment learning data generation process by the learning data generation device 200 according to this embodiment will be described. Figure 6 is a flowchart showing an example of the procedure for target equipment learning data generation by the learning data generation device 200. The equipment targeted for learning data generation by the learning data generation device 200 is equipment of the same type as, or the same as, the equipment 101 targeted for anomaly detection by the anomaly detection device 100.

[0063] First, the feature calculation unit 203 of the learning data generation device 200 performs a feature calculation process (step S21). The feature calculation process by the feature calculation unit 203 will be described in detail with reference to Figure 7. Next, the shift amount calculation unit 204 performs a shift amount calculation process (step S22). The shift amount calculation process by the shift amount calculation unit 204 will be described in detail with reference to Figure 8. Next, the learning data fitting unit 205 performs a learning data fitting process (step S23). The learning data fitting process by the learning data fitting unit 205 will be described in detail with reference to Figure 10. Once the process in step S23 is completed, the processing of the feature calculation unit 203 of the learning data generation device 200 is completed.

[0064] [Feature Calculation Process] Figure 7 is a flowchart showing an example of the procedure for feature calculation processing by the feature calculation unit 203. First, the feature calculation unit 203 receives the sensor signals used to create the learning data for the reference equipment from the sensor signal storage unit 201 (step S31). Next, the feature calculation unit 203 receives the sensor signals of the target equipment that were in operation for a short period of time (step S32). Note that the processing order of steps S31 and S32 is not limited to this order and may be in the reverse order.

[0065] Furthermore, if the target equipment is newly operational and has been in operation for a short period of time, the sensor signals stored in the sensor signal storage unit 201 will be sensor signals output from other equipment of the same type as the target equipment, and will be sensor signals that have been stored for a long period of time. The reference equipment will be the other equipment from which the learning data was created by the anomaly detection device 100. On the other hand, if the target equipment was shut down for maintenance after the anomaly detection operation and has been restarted for a short period of time, the target equipment before the shutdown will be the reference equipment.

[0066] Next, the feature calculation unit 203 calculates feature calculation parameters (step S33). The feature calculation parameters are parameters that define the daily characteristics of the reference equipment calculated in the next step S34, and the daily characteristics of the target equipment calculated in step S35. If the daily characteristics of the reference equipment and the daily characteristics of the target equipment are shown, for example, as a daily histogram, the feature calculation parameters will be the number of divisions n of the daily histogram and the expansion coefficient m, etc.

[0067] Next, the feature calculation unit 203 calculates the daily features of the reference equipment based on the training data using the feature calculation parameters calculated in step S33 (step S34). The daily features of the reference equipment represent the features calculated for each day (training day) on which the reference equipment training data was created. As described above, the training data represents the sensor signals used to create the reference equipment training data. In the following description, the daily features of the reference equipment will also be referred to as "training data features". The process in step S34 is a loop process L2 that is performed for each training day of the reference equipment for which the training data features are to be calculated, that is, for all training days of the reference equipment.

[0068] After processing in loop L2, the feature calculation unit 203 uses the feature calculation parameters calculated in step S33 to calculate evaluation data features for each short-term operating day of the target equipment based on the evaluation data (step S35). As described above, the evaluation data is composed of sensor signals from the short-term operation of the target equipment, and the evaluation data features indicate the daily characteristics of the target equipment. The processing in step S35 is a loop processing L3 that is performed for each operating day of the target equipment, that is, for all operating days on which sensor signals that are the subject of the evaluation data features were acquired.

[0069] The feature calculation process ends after loop processing L3 is completed and loop processing L1, which is the loop processing for each sensor (applies to all sensors), is completed. Loop processing L1 is the process from step S33 to step S35.

[0070] [Example of Daily Histogram Generation] Here, we will explain an example of generating a daily histogram when the features of the training data and the evaluation data are shown in a daily histogram. The feature calculation parameters specified in advance for generating the daily histogram are the number of histogram divisions n and the expansion coefficient m.

[0071] In the feature calculation parameter calculation process in step S33 of Figure 7, the feature calculation unit 203 determines the maximum value MAX and minimum value MIN of the sensor signal value to be processed in the training data. Then, the feature calculation unit 203 calculates the step size St when dividing the range from the minimum value to the maximum value by the number of divisions n, using the formula St = (MAX - MIN) / n.

[0072] Next, the feature calculation unit 203 expands the range from the minimum value MIN to the maximum value MAX horizontally outward on both the minimum value MIN side and the maximum value MAX side. After expansion, the feature calculation unit 203 changes the minimum value MIN to "MIN - St × n × m" and the maximum value MAX to "MAX + St × n × m". Examples of the expansion of the minimum value MIN and maximum value MAX will be described in detail later with reference to Figure 9.

[0073] Next, the feature calculation unit 203 sets the number of bins in the histogram to "(2m+1)×n" and associates the range from the minimum value MIN of the sensor signal to the minimum value MIN+St×(2m+1)×n with the range from bin number 0 to (2m+1)×n-1.

[0074] In step S34, the feature calculation unit 203 first initializes all frequencies in the histogram region of bin number (2m+1) × n to "0". Then, for all data for the processing day, the feature calculation unit 203 calculates the bin number BNO from the signal value F of the processing sensor and counts up the frequency corresponding to BNO. The bin number BNO is calculated by the following formula (3).

[0075] [Mathematics 3] BNO = INT((F - MIN) / St) ... Equation (3)

[0076] In equation (3) above, the function INT(X) represents the integer part of X. In step S35, the feature calculation unit 203 creates a daily histogram of (2m+1) × n bins in the same manner as in step S34.

[0077] Note that setting the calculation unit for training data features and evaluation data features to "days" is merely an example, and the present invention is not limited to this. The calculation period for training data features and evaluation data features may be set at predetermined fixed intervals other than days, or at irregular intervals that match the operating cycle. Alternatively, no predetermined period may be set. For the sake of simplicity, the following explanation will continue to use the case where the calculation unit for features is "days" as an example.

[0078] [Shift Amount Calculation Process] Next, with reference to Figure 8, the shift amount calculation process performed in step S22 of Figure 6 will be described. Figure 8 is a flowchart showing an example of the procedure for the shift amount calculation process.

[0079] First, the shift amount calculation unit 204 of the learning data generation device 200 initializes the total shift amount to "0" (step S41). The total shift amount is a value calculated by the shift amount calculation unit 204 and output to the learning data fitting unit 205, and is used by the learning data fitting unit 205 to calculate the shift amount necessary to fit the learning data of the reference equipment to the learning data of the target equipment. Each process from step S41 to step S46 is a loop process (loop process L11) performed for each sensor (covering all sensors).

[0080] Next, the shift amount calculation unit 204 initializes the maximum similarity value to "0" (step S42). The maximum similarity value indicates the maximum similarity between the evaluation data features and the training data features. Similarity can be shown, for example, by histogram intersection. Histogram intersection is a value that indicates the sum of the overlaps of each bin in the histogram.

[0081] Next, the shift amount calculation unit 204 searches for individual shift amounts based on the maximization criterion for similarity (step S43). The individual shift amount represents the shift amount of one day's worth of evaluation data features relative to one day's worth of training data features. The process of searching for individual shift amounts based on the maximization criterion, performed in step S43, will be described in detail with reference to Figure 9. The processes in steps S42 to S45 are loop processes (loop process L12) performed on all days of evaluation data.

[0082] Next, the shift amount calculation unit 204 updates the maximum similarity value and the shift amount for each evaluation day (step S44). The shift amount for each evaluation day indicates the shift amount of one day's worth of evaluation data features relative to the entire set of training data features. That is, in step S44, if the maximum similarity value exceeds the previous maximum similarity value, the shift amount calculation unit 204 updates the existing maximum similarity value with the calculated maximum similarity value. The shift amount calculation unit 204 also changes (updates) the shift amount for each evaluation day when the maximum similarity value exceeds the previous value. Each process in steps S43 and S44 is a loop process (loop process L13) performed on all days of the training data.

[0083] After the completion of loop processing L13, that is, after processing for all days of training data, the shift amount calculation unit 204 adds the daily shift amount updated in step S44 to the total shift amount (step S45). Also, after the completion of loop processing L12, that is, after processing for all days of training data, the shift amount calculation unit 204 calculates the average shift amount by dividing the total shift amount by the number of days of evaluation data, and uses the calculated average shift amount as the shift amount of the evaluation data features relative to the training data features (step SS46). In the example shown in Figure 8, the average of the daily shift amounts is used as the shift amount, but the present invention is not limited to this. For example, the mode or median of the daily shift amounts may be used as the shift amount.

[0084] (Searching for individual shift amounts) Here, the process of searching for individual shift amounts using the maximization criterion, which is performed in step S43 of Figure 8, will be explained with reference to Figure 9. Figure 9 is a diagram showing an overview of the searching for individual shift amounts. Figure 9A is a diagram showing an example of training data features 801, and Figure 9B is a diagram showing an example of evaluation data features 802. Training data features 801 are daily histograms of a certain sensor (Sni) on a certain day (j), and evaluation data features 802 are daily histograms of a different day (l) from "j".

[0085] The training data feature 801 shown in Figure 9 is an expanded version of the histogram, with the range 801a from the minimum value MIN to the maximum value MAX extended horizontally outwards. In other words, the minimum value MIN in range 801a is changed to "MIN - S × n × m", and the maximum value MAX is changed to "MAX + S × n × m". The expanded range from the minimum value MIN to the maximum value MAX is then associated with bin numbers from "0" to "39".

[0086] In step S43 of Figure 8, the shift amount calculation unit 204 searches for the position where the similarity between the training data features and the evaluation data features is maximized, by shifting the evaluation data features (in the range 801a) in bin increments between the range 802a and range 802b shown in Figure 9B. In Figure 9, ranges 802a and 802b each have a size (=16) calculated by the histogram division number n (=8) × expansion coefficient m (=2), and are each indicated by dashed rectangular frames.

[0087] In this embodiment, the similarity between training data features and evaluation data features is shown by histogram intersection. Therefore, each feature is pre-normalized so that the sum of the frequencies of each histogram representing the training data features and evaluation data features is "1".

[0088] The maximum similarity value HImax, shown by the histogram intersection, can be calculated using the following formula (4).

[0089]

[0090] In equation (4) above, "ai" is the normalized value of the frequency of bin number Bni in the daily histogram of the training data (training data features 801), and "bi" is the normalized value of the frequency of bin number Bni in the daily histogram of the evaluation data (evaluation data features 802). The shift amount calculation unit 204 then determines the individual shift amount p when the maximum similarity value HImax is maximized.

[0091] Figure 9C shows an example of the position of range 802c when the maximum similarity value HImax is maximized. In the example shown in Figure 9C, the individual shift amount p of range 802c is "-2". In other words, the maximum similarity value HImax is maximized at the position where the range 801a of the histogram distribution in the training data feature 801 shown in Figure 9A is shifted by "2" (-2) in the "minimum" direction.

[0092] [Training Data Fitting Process] Next, with reference to Figure 10, the training data fitting process performed by the training data fitting unit 205 in step S23 of Figure 6 will be described. Figure 10 is a flowchart showing an example of the training data fitting process procedure.

[0093] First, the learning data fitting unit 205 receives reference equipment learning data from the reference equipment learning data storage unit 202 (see Figure 2) (step S51). Next, the learning data fitting unit 205 converts the shift amount calculated by the shift amount calculation unit 204 into a sensor signal value (step S52). The shift amount calculated by the shift amount calculation unit 204 in step S22 of Figure 6 is a value in bin units. Therefore, in step S52, the learning data fitting unit 205 calculates p × s using the step size S calculated as a feature calculation parameter in step S33 of Figure 7 to determine the shift amount in the sensor signal.

[0094] Next, the learning data fitting unit 205 uses the shift amount calculated in step S52 to shift the average of the canonicalization parameters used when canonicalizing the sensor signal in step S2 of Figure 4 (step S53). Since the sign of the shift amount represents the direction in which the evaluation data is moved horizontally, the learning data fitting unit 205 decreases the average to increase the evaluation data if the sign is positive. In other words, it subtracts the calculated shift amount from the average of the canonicalization parameters. On the other hand, if the sign is negative, it increases the average to decrease the evaluation data. In other words, it adds the shift amount to the average of the canonicalization parameters.

[0095] Furthermore, if the feature vector is canonicalized using the maximum and minimum values ​​of the sensor signal, or the upper and lower N percent values, the learning data fitting unit 205 adds or subtracts a shift amount to each of these canonicalization parameters. Steps S52 and S53 are loop processes (loop process L21) performed for each sensor (for all sensors).

[0096] After completing the loop process L21, the learning data fitting unit 205 saves the shifted canonicalization parameters and the learning data of the reference equipment as the learning data of the target equipment to the target equipment learning data storage unit 206 (step S54). After the processing in step S54, the learning data fitting process by the learning data fitting unit 205 is completed.

[0097] The process of converting the shift amount in step S52 of Figure 10 to a sensor signal value differs depending on the type of method employed by the anomaly measure calculation unit 106 when performing feature vector clustering. For example, if the method used by the anomaly measure calculation unit 106 is the fast local subspace method, the training data includes the center position of the cluster and all feature vector data that are members of any of the clusters.

[0098] Therefore, in step S52 of Figure 10, the learning data fitting unit 205 converts the shift amount into a feature vector shift amount using the standard deviation σ of the canonicalization parameter and the formula p × S / σ. Then, in step S53 of Figure 10, the learning data fitting unit 205 shifts the center position of all clusters and all feature vector data that become members of any cluster by the calculated shift amount, targeting the elements of the feature vector corresponding to the sensor being processed.

[0099] Furthermore, even when the method used by the anomaly measure calculation unit 106 is the local subspace method, the training data includes all feature vectors during the training period. Therefore, the training data fitting unit 205 performs the processes in steps S52 and S53 of Figure 10 in the same manner as when clustering is performed using the fast local subspace method.

[0100] Furthermore, if the method used by the anomaly measure calculation unit 106 is the vector quantization clustering method or the projection distance method, the training data fitting unit 205 shifts the center position of the clusters in the training data by the shift amount in step S53 of Figure 10.

[0101] Furthermore, while Figure 10 shows an example in which the learning data fitting unit 205 saves the shifted canonicalization parameters and the learning data of the reference equipment as the learning data of the target equipment, the present invention is not limited to this. The learning data fitting unit 205 may also save the shift amount converted to a shift signal value in step S52 of Figure 10 as the learning data of the target equipment. In this case, when an anomaly is detected in the target equipment, the feature vector extraction unit 105 of the anomaly detection device 100 shifts the sensor signal input from the sensor signal input unit 104 or the sensor signal storage unit 103 in step S1 of Figure 4 by the amount of the shift calculated in step S52. After that, the feature vector extraction unit 105 performs the processing from step S2 onwards in Figure 4.

[0102] When this method is used, it becomes unnecessary to generate target equipment learning data. Therefore, the learning data generation device 200 can use the reference equipment learning data stored in the reference equipment learning data storage unit 202 directly to perform anomaly detection processing. As a result, regardless of the method used for clustering processing in the anomaly measurement calculation process, the learning data generation device 200 can perform anomaly detection processing using the reference equipment learning data.

[0103] <Input screen for generating learning data> The conditions for generating learning data stored in the learning data storage unit 109 (an example of information necessary for generating learning data) can be set by the user, for example, via the learning data generation condition setting screen Sc1 shown in Figure 11. Figure 11 shows an example of the configuration of the learning data generation condition setting screen Sc1.

[0104] As shown in Figure 11, the learning data generation condition setting screen Sc1 includes a reference equipment ID selection unit 11, a learning data name selection unit 12, a target equipment ID selection unit 13, and a learning data name input unit 14. The learning data generation condition setting screen Sc1 also includes a data storage period display unit 15, an evaluation period setting unit 16, a histogram division number input unit 17, and an expansion coefficient input unit 18. Furthermore, the learning data generation condition setting screen Sc1 includes an execute button 19, a register button 20, and an exit button 21.

[0105] The reference equipment ID selection unit 11 is configured as a drop-down list. When the list display button, indicated by the downward-pointing triangle at the right end of the drop-down list, is pressed, the reference equipment ID selection unit 11 displays a list of reference equipment IDs associated with each reference equipment learning data stored in the reference equipment learning data storage unit 202. The user can select the reference equipment ID of the reference equipment to be targeted for generating target equipment learning data from the displayed list of reference equipment IDs.

[0106] The learning data name selection unit 12 is configured as a drop-down list, and when the list display button on the far right is pressed, it displays a list of learning data names for each reference equipment learning data stored in the reference equipment learning data storage unit 202. The user can select the learning data name of the learning data to be used for generating target equipment learning data from the displayed list of learning data names.

[0107] The target equipment ID selection unit 13 is configured as a drop-down list, and when the list display button on the far right is pressed, it displays a list of target equipment IDs stored in the sensor signal storage unit 201. The user can select the target equipment ID of the target equipment to be used for generating target equipment learning data from the displayed list of target equipment IDs.

[0108] The learning data name input unit 14 receives the learning data name of the target equipment learning data to be stored in the target equipment learning data storage unit 206. The user can set the learning name of the newly generated target equipment learning data by entering text into the learning data name input unit 14. Alternatively, if there are candidate names for the learning data, these candidates may be displayed in a list in the learning data name input unit 14, and the user may select one of the candidate names.

[0109] The data retention period display unit 15 shows the retention period information of the sensor signals acquired in the target equipment in the sensor signal storage unit 201, indicated by the start date and time and the end date and time of the retention period.

[0110] The evaluation period setting unit 16 receives input for the acquisition period of evaluation data used to generate target equipment learning data. The user can set the acquisition period of evaluation data via the evaluation period setting unit 16, within a range that does not exceed the storage range of the sensor signals displayed on the data storage period display unit 15. If the target equipment is equipment that will be restarted after being shut down, the user inputs the period after restart into the evaluation period setting unit 16.

[0111] The histogram division number input unit 17 receives the value of the histogram division number n used in step S33 of Figure 7. The histogram division number n is a value that defines the histogram step size St, and the resolution of the information shown in the histogram changes depending on the size of the step size St. The higher the resolution of the information shown in the histogram, the higher the accuracy of the shift amount calculation. Therefore, the user can set the histogram division number n via the histogram division number input unit 17 based on the desired accuracy of the shift amount calculation.

[0112] The expansion coefficient input unit 18 receives the value of the expansion coefficient m used in step S33 of Figure 7. Since the expansion coefficient m is a value that affects the magnitude of the shift amount, the user can set the expansion coefficient m according to the desired magnitude of the shift amount via the expansion coefficient input unit 18.

[0113] The execution button 19 is a button used to instruct the execution of the process for generating target equipment learning data. When the execution button 19 is detected to have been pressed by the user, the sensor signal corresponding to the learning data name of the reference equipment and the sensor signal of the target equipment during the specified evaluation period are read from the sensor signal storage unit 201 and input to the feature vector extraction unit 105, according to the feature calculation process flowchart shown in Figure 7. Next, the feature vector extraction unit 105 calculates the daily histogram of the reference equipment and the daily histogram of the target equipment for each sensor.

[0114] Subsequently, according to the flowchart of the shift amount calculation process shown in Figure 8, the shift amount calculation unit 204 calculates the shift amount based on the daily histogram of the reference equipment and the daily histogram of the target equipment. Next, according to the flowchart of the learning data fitting process shown in Figure 10, the learning data corresponding to the learning data name of the reference equipment is input from the reference equipment learning data storage unit 202 to the learning data fitting unit 205.

[0115] The learning data fitting unit 205 then uses the shift amount calculated by the shift amount calculation unit 204 to process the learning data of the reference equipment to fit the evaluation data, thereby generating the target equipment learning data. After the generation of the target equipment learning data is completed, the result confirmation screen Sc2 is displayed. The result confirmation screen Sc2 will be described in detail with reference to Figure 12.

[0116] The registration button 20 on the learning data generation condition setting screen Sc1 is a button for saving the target equipment learning data to the target equipment learning data storage unit 206. When the user presses the registration button 20, the target equipment learning data generated by the learning data generation device 200 is associated with the learning data name entered in the learning data name input unit 14 and saved to the target equipment learning data storage unit 206.

[0117] The exit button 21 is used to close the learning data generation condition setting screen Sc1. When the user presses the exit button 21, the learning data generation condition setting screen Sc1 is closed.

[0118] If the registration button 20 is detected to be pressed, the generated learning data is associated with the learning data name entered in the learning data name input unit 14 and saved in the target equipment learning data storage unit 206. If the exit button 21 is pressed, the learning data generation condition setting screen Sc1 is closed and the user returns to the learning data generation condition setting screen Sc1. If the exit button 21 is pressed without the registration button 20 being pressed, the generated learning data is discarded.

[0119] <Configuration of the Result Confirmation Screen> Next, the result confirmation screen Sc2 will be explained with reference to Figure 12. Figure 12 is a diagram showing an example configuration of the result confirmation screen Sc2.

[0120] As shown in Figure 12, the results confirmation screen Sc2 includes a reference equipment display unit 111, a learning period display unit 112, a target equipment display unit 113, an evaluation period display unit 114, a reference equipment sensor signal display unit 115, a target equipment sensor signal display unit 116, a reference equipment display period input unit 117, and a target equipment display period input unit 118. Furthermore, the results confirmation screen Sc2 includes a sensor selection unit 119 and an exit button 120.

[0121] The reference equipment display unit 111 displays the reference equipment ID selected in the reference equipment ID selection unit 11 in Figure 11. The learning period display unit 112 displays the information on the storage period in the sensor signal storage unit 201 of the sensor signals acquired by the target equipment, as displayed in the data storage period display unit 15 in Figure 11.

[0122] The target equipment display unit 113 displays the target equipment ID selected in the target equipment ID selection unit 13 in Figure 11. The evaluation period display unit 114 displays the evaluation period set in the evaluation period setting unit 16 in Figure 11.

[0123] The reference equipment sensor signal display unit 115 displays information on the time-series sensor signals 115a of the reference equipment with the reference equipment ID selected in the reference equipment ID selection unit 11 in Figure 11, for the display period set in the reference equipment display period input unit 117, which will be described later.

[0124] The target equipment sensor signal display unit 116 displays the time-series sensor signals 116a and 116b of the reference equipment for the target equipment ID selected in the target equipment ID selection unit 13 in Figure 11, for the display period set in the target equipment display period input unit 118 described later. The time-series sensor signal 116a is the time-series information of the sensor signal acquired from the sensor signal storage unit 201, and the time-series sensor signal 116b is the time-series information of the sensor signal after shifting by the learning data adaptation unit 205.

[0125] By comparing the time-series sensor signals 116a and 116b displayed on the target equipment sensor signal display unit 116, the user can confirm the change in the behavior of each sensor signal before and after the shift. Furthermore, if the behavior of time-series sensor signals 116a and 116b is almost the same, the user can confirm that a reasonable shift amount has been calculated by the shift amount calculation unit 204 of the learning data generation device 200.

[0126] The reference equipment display period input unit 117 receives the display period of the reference equipment's sensor signal (time-series sensor signal 115a). The user can input a period within the learning period range that they want to check in more detail into the reference equipment display period input unit 117. Initially, the reference equipment display period input unit 117 displays the same period as the learning period displayed in the learning period display unit 112. The user can change (set) the period set in the reference equipment display period input unit 117 to any period within the learning period range.

[0127] The Target Equipment Display Period Input Unit 118 receives the display period of the reference equipment's sensor signals (time-series sensor signal 116a and time-series sensor signal 116b). The user can input a period within the evaluation period range that they wish to check in more detail into the Target Equipment Display Period Input Unit 118. Initially, the Target Equipment Display Period Input Unit 118 displays the same period as the evaluation period displayed in the evaluation period display unit 114. The user can change (set) the period set in the Target Equipment Display Period Input Unit 118 to any period within the evaluation period range.

[0128] In the sensor selection unit 119, the sensor whose sensor signal behavior is to be checked is selected. The sensor selection unit 119 displays the name of the first sensor as an initial value. The exit button 120 is a button to end the display of the results confirmation screen Sc2. When the user presses the exit button 120, the results confirmation screen Sc2 is closed and the learning data generation condition setting screen Sc1 is displayed again.

[0129] <Example of an operation schedule for an anomaly detection device> Next, with reference to Figure 13, an example of an operation schedule for the anomaly detection device 100 using the target equipment learning data generated by the learning data generation device 200 will be described. Figure 13 is a diagram showing an example of an operation schedule for the anomaly detection device 100.

[0130] The horizontal axis of the operational schedule shown in Figure 13 represents the passage of time before and after the introduction of the target equipment. The introduction of the new equipment presupposes the existence of other equipment of the same type as the new equipment, from which sensor signals have been accumulated over a long period. A long period refers to, for example, a period of 6 to 12 months.

[0131] Prior to the time T1 when the new equipment is introduced, the anomaly detection device 100 generates learning data Da based on sensor signals from other equipment. Next, after the introduction of the new equipment, the anomaly detection device 100 accumulates sensor signals from the new equipment for a short period, for example, an evaluation period of 10 days or less. Then, the learning data generation device 200 generates learning data Db using the accumulated sensor signals (evaluation data) as input. Learning data Db is learning data adapted to the new equipment by shifting the learning data Da.

[0132] Next, at point T2, when the evaluation period ends, the anomaly detection device 100 starts anomaly detection using the training data Db generated by the training data generation device 200. In parallel, the anomaly detection device 100 stores the sensor signals of the new equipment necessary for training during a training period of 6 to 12 months. Then, the training data generation device 200 generates training data Dc using the stored sensor signals of the new equipment. From point T3 onward, when the generation of training data Dc is completed, the anomaly detection device 100 starts anomaly detection using the training data Dc.

[0133] Conventionally, it was not possible to start anomaly detection by the anomaly detection device 100 targeting the new equipment until the learning data Dc for the new equipment was generated, that is, until time point T3 arrived. In contrast, according to this embodiment, by using the learning data Db of other equipment that has been shifted to match the new equipment, it is possible to start anomaly detection by the anomaly detection device 100 from time point T2, which is less than 10 days after the introduction of the new equipment. In other words, according to this embodiment, the period required from the introduction of new equipment to the start of anomaly detection operation can be significantly shortened.

[0134] In a new piece of equipment where anomaly detection using training data Dc is performed, maintenance is carried out from time T4, and the equipment is restarted at time T5. In this case, the anomaly detection device 100 stores sensor signals from the restarted new equipment during the evaluation period from time T4 to time T5. The training data generation device 200 then generates training data Dd using the sensor signals stored during the evaluation period as input. Training data Dd is training data adapted to the restarted new equipment by shifting training data Dc. After time T5, when the evaluation period ends, the anomaly detection device 100 starts operating anomaly detection using the training data Dd.

[0135] Furthermore, the anomaly detection device 100 stores sensor signals from the newly restarted equipment, which are necessary for learning, for a learning period of 6 to 12 months from the time of restart T4. Then, the learning data generation device 200 generates learning data De using the stored sensor signals from the newly restarted equipment. After the generation of learning data De is completed at time T5, the anomaly detection device 100 starts operating anomaly detection using the learning data De.

[0136] According to this embodiment, similar to the introduction of new equipment, the period from restart to the start of anomaly detection operations after maintenance can be significantly shortened from approximately 6 to 12 months to approximately 10 days or less.

[0137] Furthermore, if the number of detected anomalies increases during operation of the anomaly detection device 100, the user shall investigate the status of the equipment within the facility and take countermeasures if there are any problems. On the other hand, if the user determines that changes in the behavior of sensor signals due to environmental changes or aging are the cause of the increase in anomalies, countermeasures against the anomalies may not be taken. In this case, it is conceivable that operation will continue as is, and anomaly detection will become the norm. If anomaly detection becomes the norm, the user may miss important anomalies.

[0138] To prevent such problems from occurring, if it is determined that no countermeasures are necessary for the increase in anomalies, the training data currently in operation may be processed to match the current state, i.e., the state in which anomalies are increasing. Specifically, the anomaly detection device 100 inputs evaluation data of the target equipment during the period from when anomaly detections begin to increase until the determination that no countermeasures are necessary is made, and the training data generation device 200 generates training data using the input evaluation data. By performing such processing, it is possible to prevent situations in which users miss important anomaly detections when it is determined that no countermeasures are necessary for anomalies.

[0139] <Example of Computer Hardware Configuration> Finally, with reference to Figure 14, the hardware configuration of the devices for realizing the functions of the anomaly detection device 100 and the learning data generation device 200 according to this embodiment will be described. Figure 14 is a block diagram showing an example of the hardware configuration of the anomaly detection device 100 and the learning data generation device 200. The computer 50 shown in Figure 14 is hardware used as a so-called computer.

[0140] As shown in Figure 14, the computer 50 includes a control unit 51 connected to bus B, a non-volatile storage 52, a display unit 53, an operation input unit 54, and a communication interface 55.

[0141] The control unit 51 includes a CPU (Central Processing Unit) 511, a ROM (Read Only Memory) 512, and a RAM (Random Access Memory) 513.

[0142] The CPU 511 reads the program code of the software that implements each function according to this embodiment from the ROM 512, loads it into the RAM 513, and executes it. Variables, parameters, etc. that occur during the calculation process are temporarily written to the RAM 513.

[0143] The control unit 51 may be equipped with a processing unit such as an MPU (Micro-Processing Unit) instead of a CPU 511. Alternatively, the control unit 51 may use both a CPU and an MPU.

[0144] As the non-volatile storage 52, for example, an HDD (Hard Disk Drive), SSD (Solid State Drive), flexible disk, optical disk, magneto-optical disk, CD-ROM, CD-R, non-volatile memory card, etc., can be used. In addition to the OS and various parameters, programs for operating the computer 50 are recorded in this non-volatile storage 52. The programs may also be stored in ROM 512. Furthermore, the non-volatile storage 52 comprises a sensor signal storage unit 103 of the anomaly detection device 100, a learning data storage unit 109, a sensor signal storage unit 201 of the learning data generation device 200, a reference equipment learning data storage unit 202, and a target equipment learning data storage unit 206.

[0145] The display unit 53 is a monitor, for example, composed of an LCD (Liquid Crystal Display), and displays the results of processing performed by the computer 50. The display unit 53 displays the learning data generation condition setting screen Sc1 shown in Figure 11, the result confirmation screen Sc2 shown in Figure 12, and the like.

[0146] The operation input unit 54 is composed of, for example, a keyboard, mouse, or touch sensor, and generates operation signals in response to user operations and supplies them to the CPU 511. The display unit 53 and the operation input unit 54 may be integrated as a single touch panel.

[0147] The program is stored in the form of computer-readable program code, and the CPU 511 sequentially executes operations according to the program code. In other words, the ROM 512 or non-volatile storage 52 is used as an example of a computer-readable, non-transient recording medium that stores a program executed by the computer.

[0148] Communication I / F55 uses, for example, a NIC (Network Interface Card), which enables the transmission and reception of various types of data with external devices via a network or communication line.

[0149] The embodiments described above are intended to explain the present invention in an easy-to-understand manner, and provide a detailed and specific description of the configurations of the device (anomaly detection device 100, learning data generation device 200). However, the invention is not necessarily limited to devices that include all of the configurations described.

[0150] Furthermore, the control lines or information lines indicated by solid arrows in Figures 1 and 2 are those deemed necessary for explanation and do not necessarily represent all control lines or information lines in the actual product. In reality, it can be assumed that almost all components are interconnected.

[0151] 100... Anomaly detection device, 103... Sensor signal storage unit, 104... Sensor signal input unit, 105... Feature vector extraction unit, 106... Anomaly measure calculation unit, 107... Threshold calculation unit, 108... Anomaly determination unit, 109... Learning data storage unit, 200... Learning data generation device, 201... Sensor signal storage unit, 202... Reference equipment learning data storage unit, 203... Feature calculation unit, 204... Shift amount calculation unit, 205... Learning data adaptation unit, 206... Target equipment learning data storage unit, Sc1... Learning data generation condition setting screen, Sc2... Result confirmation screen

Claims

1. A learning data generation device for an anomaly detection device that determines whether or not there is an anomaly in target equipment subject to anomaly detection, the learning data generation device comprising a learning data adaptation unit that generates learning data for the target equipment based on learning data for the reference equipment by performing an adaptation process based on information of the difference between evaluation data consisting of sensor signals for a first period output from a sensor installed on the target equipment and learning data consisting of sensor signals for a second period longer than the first period output from a sensor installed on reference equipment that is referenced in generating learning data for the target equipment.

2. The learning data generation device according to claim 1, wherein the target equipment is new equipment or equipment that has been restarted after being shut down, and in the case where the target equipment is new equipment, the reference equipment is other equipment having the same specifications as the target equipment, and in the case where the target equipment is equipment that has been restarted after being shut down, the reference equipment is the same equipment as the target equipment.

3. The learning data generation apparatus according to claim 2, wherein the first period is a short period of 10 days or less, and the second period is a long period of six months or more.

4. The learning data generation device according to claim 2, further comprising a feature calculation unit that calculates learning data features indicating the characteristics of the learning data and evaluation data features indicating the characteristics of the evaluation data, wherein the fitting process is a process of shifting the learning data features by the amount of the difference between the learning data features and the evaluation data features, with the difference being the shift amount.

5. The learning data generation device according to claim 4, wherein the feature calculation unit creates a first histogram of each sensor signal value constituting the evaluation data of the target equipment for a fixed period of time as the evaluation data features, and creates a second histogram of each sensor signal value constituting the learning data of the reference equipment for a fixed period of time as the learning data features.

6. The learning data generation device according to claim 5, further comprising a shift amount calculation unit that calculates the shift amount and outputs it to the learning data fitting unit, wherein the shift amount calculation unit calculates the similarity between the first histogram and the second histogram while shifting the first histogram horizontally, and calculates the shift amount to be used for the fitting process based on the shift amount of the first histogram when the similarity is maximized.

7. The learning data includes information necessary for calculating an abnormality measure indicating the degree of deviation of the equipment from a normal state, and a threshold value that is compared with the abnormality measure in determining the presence or absence of an abnormality, wherein the learning data and the evaluation data are information necessary for calculating the abnormality measure, and the threshold value is generated by the abnormality detection device based on the abnormality measure, according to any one of claims 1 to 6.

8. In the anomaly detection device, when the adjustment process based on the shift amount is performed, the learning data adjustment unit outputs the shift amount to the anomaly detection device instead of the learning data for the target equipment after the adjustment process is performed, according to claim 4.

9. A learning data generation method by a learning data generation device that generates learning data used for determining abnormalities in an abnormality detection device that determines whether or not there is an abnormality in target equipment subject to abnormality detection, the learning data generation method having a procedure in which a learning data matching unit generates learning data for the target equipment based on learning data for the reference equipment by matching information based on the difference between evaluation data consisting of sensor signals for a first period output from a sensor installed on the target equipment and learning data consisting of sensor signals for a second period longer than the first period output from a sensor installed on a reference equipment that is referenced in generating learning data for the target equipment.

Images