JP7236886B2 - Anomaly detection device, anomaly detection method, and anomaly detection system - Google Patents

Description

The present invention relates to an anomaly detection device, an anomaly detection method, and an anomaly detection system, and more particularly to an anomaly detection technology for manufacturing/processing equipment.

As an anomaly detection system for equipment, Patent Document 1 describes "(1) focusing on the similarity between data, generating compact learning sensor data consisting of normal cases, (2) generating new data based on the similarity and the presence or absence of anomalies. Add to learning sensor data, (3) Delete the alarm occurrence section of the equipment from learning sensor data, (4) Model the learning sensor data updated from time to time with the subspace method, based on the distance relationship between the observation data and the subspace, Detecting anomaly candidates, (5) Detecting anomalies from anomaly candidates by combining analysis targeting event information, (6) Obtaining the degree of divergence of observation data based on the utilization frequency distribution of learning sensor data, and observing data Identify anomalous elements of (summary excerpt)" configuration is disclosed.

JP 2013-218725 A

When predicting the state of a device that performs anomaly detection (hereinafter referred to as "target device"), if the sensor data contains noise or outliers, the characteristics of the noise or outliers will appear as anomalies in the state analysis results. , unable to correctly analyze the state. Therefore, noise and outliers are generally removed by data preprocessing, but for this purpose, it is necessary to set an appropriate threshold as a removal criterion.

However, when the state of the target device changes from time to time depending on the operating time and operating conditions, such as the polishing plate of a polishing device or the cutting blade of a cutting device, the state of the target device, which should be used as the removal determination criterion, changes moment by moment. Therefore, there is a problem that it is difficult to appropriately set the removal determination criteria in accordance with the change state of the target device.

In Patent Document 1, anomaly detection is performed based on "(1) Compact learning sensor data consisting of normal cases, focusing on similarity between data", but when the state of normal cases changes, the change is not taken into consideration, and the above problem cannot be solved.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a technique for improving the abnormality detection accuracy of an apparatus whose normal state changes sequentially.

In order to achieve the above object, the present invention has the configuration described in the claims. For example, an anomaly detection device that detects anomalies in a target device capable of performing processing operations multiple times, wherein observation sensor data output by an observation sensor that observes the state of the target device , and the observation sensor data is a data acquisition unit that acquires preconditions including at least one of processing conditions and work details of the target device at the time of output, and a cumulative number of times of processing performed under the preconditions; Observation sensor data output by the observation sensor during operation in a state, preconditions when the observation sensor data was output, and the cumulative number of processing performed under the preconditions are associated with each other and stored as learning sensor data. a sensor data storage unit, and the learning sensor data output under the same preconditions, the number of times of processing of which is less than the number of times of observation of the target device as a reference , is recorded in the learning sensor data storage unit; a prediction sensor data selection unit that selects from a learning sensor data group and generates prediction sensor data that predicts the state of the target device at the observation target time based on the selected learning sensor data; and an observation sensor at the observation target time. an anomaly detection unit that detects occurrence of an anomaly in the target device during the observation target time based on a comparison result of the data and the predicted sensor data.

According to the above invention, it is possible to provide a technique for improving the accuracy of abnormality detection of a device whose normal state changes successively. Objects, configurations, and effects other than those described above will be clarified by the following embodiments.

Diagram showing the schematic configuration of an anomaly detection system for manufacturing/processing equipment Functional block diagram of anomaly detection device Flowchart showing the processing flow of the anomaly detection system Diagram showing an example of the data structure acquired by the data acquisition unit The figure which shows the data structure example of the learning sensor data group memorize|stored in a learning sensor data memory|storage part. Diagram showing the results of principal component analysis of resistance value sensor data features Diagram showing an example of abnormality determination Flowchart showing the flow of predictive sensor data selection processing

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a schematic configuration of an abnormality detection system 1 for manufacturing/processing equipment.

The anomaly detection system 1 is configured by connecting a target device 10 to be an anomaly detection target and an anomaly detection device 20 for communication via a network 30 .

The target device 10 may be of any type, such as a product manufacturing device or a product processing device. More specifically, it may be a manufacturing/processing device such as a wafer manufacturing device, a polishing device, or a cutting device, or may be an IoT device.

The target device 10 is a device capable of performing processing operations multiple times. In this embodiment, an apparatus that completes one machining process by performing machining operations a plurality of times is taken as an example. For example, the target device 10 is a wafer polishing device. For the processing work performed by the polishing machine, the processing work conditions such as the type of wafer, the finishing state, and the number of times of polishing per unit time (adjusted by setting the rotation speed (rpm) of the plate) are set. Polishing work is performed a plurality of times during the polishing work. In this embodiment, the condition of the processing work is called "precondition", and the process from the start to the end of the polishing work for one wafer is called "one processing step". For example, 100 polishing operations are performed in one processing step. Preconditions include, for example, sensor data showing the finished state of the wafer (specifically, there are height, surface roughness, flatness, etc. as processing target values), type of wafer, cumulative use of polishing plates used for polishing. There is part position information specifying the part of the polishing plate to be used for polishing, such as the number of times and cumulative use time.

If the polishing work is continued with the same polishing plate while exchanging wafers in the polishing apparatus, the polishing plate will wear out. Then, even if the same preconditions and the same number of polishing operations are performed for the same type of wafer, the wear amount of the polishing plate changes. For example, the amount of wear (corresponding to the amount of change in state of the polishing plate) when polishing a wafer once with a new polishing plate is greater than that when polishing a wafer once with a polishing plate that has been polished many times. The larger the value, the greater the amount of wear.

When trying to detect an abnormality in the wear state of the polishing plate, the observation sensor data of the polishing plate in the normal state is used as reference data, and this is compared with the observation sensor data of the polishing plate at the time of observation. At that time, as described above, even if the same preconditions and the same number of polishing operations are performed, the wear amount (state change amount) of the polishing plate is different, so the reference data should be appropriately changed accordingly.

Therefore, in the anomaly detection device 20, instead of fixed reference data, sequentially changing reference data is selected, and an anomaly detection determination is performed by comparing the degree of similarity with the observation sensor data of the observation target time.

The target device 10 includes at least one or more observation sensors 11 that observe the device state of the target device 10 . In this embodiment, the observation sensor data output by the observation sensor 11 is multi-dimensional data. Therefore, statistical processing including PCA analysis (principal component analysis) is performed on observed sensor data, and abnormality determination is performed using the results. The details will be described later. A resistance value sensor and a pressure sensor may be used as examples of observation sensors. Since the device state to be observed differs depending on the type of the target device 10, such as a temperature sensor, an atmospheric pressure sensor, a humidity sensor, a vibration sensor, and an acceleration sensor, various observation sensors corresponding to the type of the target device 10 may be used. PCA analysis, which will be described later, is a useful analysis method when there are multiple dimensions of observed sensor data. The "multiple-dimensional observation sensor data" here means that the same type of observation sensors (for example, resistance value sensors) are attached at multiple locations, and the outputs from each observation sensor are combined to form multi-dimensional observation sensor data. A case, a case in which a plurality of observation sensors 11 with different collection contents such as resistance values and pressure values are attached, and the output from each observation sensor is combined to form multi-dimensional observation sensor data; When FFT (Fast Fourier Transform) is performed on the time-series sensor data series obtained, basically 2^n results are obtained according to the number of target data of FFT, so the output of one observation sensor 11 There are cases of forming multi-dimensional observed sensor data from .

The anomaly detection device 20 acquires observation sensor data output by the observation sensor 11 attached to the target device 10, and performs anomaly detection of the target device 10 based on this. The values of observation sensor data are various such as current value, pressure, temperature, and resistance value. The anomaly detection device 20 automatically predicts the state of the target device 10 using the waveforms of these various observation sensor data, and detects an anomaly.

The abnormality detection device 20 is composed of a computer having a CPU 201, a ROM 202, a RAM 203, an HDD 204, a communication I/F 205, an input/output I/F 206, and a bus 207 connecting them together. The input/output I/F 206 is connected to a display 70 and an input device 71 including a keyboard, mouse, and touch panel.

FIG. 2 is a functional block diagram of the abnormality detection device 20. As shown in FIG.

The abnormality detection device 20 includes a data acquisition unit 21, a prediction sensor data selection unit 22, a learning sensor data storage unit 23, a preprocessing unit 24, a statistical processing unit 25, an abnormality detection unit 26, a determination result information output unit 27, and a learning sensor. A data update unit 28 is included. The preprocessing unit 24 includes a threshold excess determination unit 24a and a feature amount extraction unit 24b. The statistical processing section 25 includes a PCA section 25a and a clustering section 25b. The abnormality detection unit 26 includes a cluster comparison unit 26a, an abnormality candidate determination unit 26b, and an abnormality candidate level determination unit 26c. The processing contents of each functional block will be described later. The learning sensor data storage unit 23 is configured by a storage area of the HDD 204 . Other functional blocks are formed by the CPU 201 executing software that implements the functions of each functional block.

FIG. 3 is a flowchart showing the flow of processing of the anomaly detection system 1. As shown in FIG.

The operator uses the GUI of the anomaly detection device 20 to set the file paths of the prediction parameters and learning/observation sensor data (S101). Here, "setting a file path" means specifying (setting) a setting file for processing preconditions and a list file for learning/observation sensor data. The file path indicates the storage location of each file and the storage location of the prediction result output file.

The data acquisition unit 21 acquires observation sensor data and preconditions for outputting the observation data (S102). FIG. 4 is a diagram showing an example of the data structure acquired by the data acquisition unit 21. As shown in FIG. In this example, observation sensor data (sensor values) associated with an observation sensor ID that uniquely identifies the observation sensor 11, a precondition for outputting the observation sensor data, and the cumulative number of times of processing performed under the precondition are targeted. The device 10 generates it and transmits it to the anomaly detection device 20 . Observation sensor data including the cumulative number of times of polishing and the observation sensor data (sensor values) observed at that time for a plurality of times, preferably for one processing step, constitutes time-series sensor data. When the observation sensor data in FIG. 4 is observation sensor data including sensor values for each observation, the abnormality detection device 20 side may stack the data to form time-series sensor data, or the target device 10 side may perform one processing. By stacking the sensor values from the first time to the last time in the process and putting them together, observation sensor data composed of time-series sensor data may be generated and transmitted to the anomaly detection device 20 .

The data acquisition unit 21 outputs preconditions from the received data to the prediction sensor data selection unit 22 . The predicted sensor data selection unit 22 reads the learned sensor data associated with the same precondition as the acquired precondition from the learned sensor data storage unit 23, and selects the predicted sensor data (S103).

FIG. 5 is a diagram showing an example data structure of a learned sensor data group stored in the learned sensor data storage unit 23. As shown in FIG. The learning sensor data group is the observation sensor data collected in advance under the same processing conditions (same preconditions) in the past normal state (the observation sensor data collected in advance will be referred to as (referred to as learning sensor data in the explanation).

The predicted sensor data selection unit 22 obtains the deviation (state change amount) of machining efficiency fluctuations due to part deterioration from learning sensor data associated with the same precondition as the precondition added to the observed sensor data. The predicted sensor data selection unit 22 refers to the previous observation sensor data as reference information, and selects the learned sensor data whose number of executions of the processing process is one less than the number of executions of the processing process as prediction sensor data. For example, when predicting the state of the fifth polishing process, learning data composed of time-series sensor data obtained in the fourth polishing process is selected. The details of the process of selecting predicted sensor data executed in this step will be described later.

Furthermore, the predicted sensor data selection unit 22 adds the standard deviation or variance of ±Nσ to the selected predicted sensor data to set the upper limit threshold and the lower limit threshold as outliers and noise determination thresholds (S104).

Based on the aforementioned threshold, the threshold excess determination unit 24a treats sensor data exceeding the threshold as noise or an outlier, and excludes the sensor data from use in the subsequent PCA analysis (S105). This is to avoid the problem that features such as noise and outliers are extracted as high contribution rates as a result of summarizing the data by reducing the dimension of the feature amount data in the PCA analysis in the statistical processing unit 25, which will be described later. .

In statistically analyzing and comparing waveforms of time-series observed sensor data, the feature amount extraction unit 24b performs processing for extracting feature amounts for statistical processing, such as filtering and Fourier transform (S106). ).

In this step, the data exceeding the threshold is excluded from the target of statistical analysis processing in the latter stage, and the above-mentioned prediction and observation sensor data are subjected to filtering processing (cutting of DC components by differentiation) and Hanning window function as feature extraction processing. Application and FFT (Fast Fourier Transform) are performed, and the feature quantities are statistically analyzed and compared.

The statistical processing unit 25 performs statistical processing for subsequent abnormality determination processing on each of the sensor data and predicted sensor data (S107).

FIG. 6 is a diagram showing the principal component analysis result of the resistance value sensor data feature amount. In the graph of FIG. 6, the horizontal axis indicates the number of times of processing, and the vertical axis indicates the sensor value. Also, the graph of FIG. 5 is obtained by superimposing the graphs of the observed sensor data when performing one machining process after performing the machining work a plurality of times for a plurality of machining processes. In step S106, since the observed sensor data of times when noise/outlier appears is below the lower limit, it is excluded from feature quantity analysis. Then, in step S106, by performing statistical processing only on the observed sensor data within the threshold, signs of component deterioration can be observed without being affected by noise due to the occurrence of a short circuit.

The PCA unit 25a performs PCA analysis on the observed sensor data and predicted sensor data, and reduces the dimension of the data (=abstract). From the result, N principal components are selected in descending order of the cumulative contribution rate, and the principal component score of each of the sensor data and predicted sensor data is obtained.

The clustering unit 25b performs clustering based on the principal component scores of each of the observed sensor data and predicted sensor data. Predicted sensor data and observed sensor data after FFT are subjected to PCA analysis, and the principal component scores of principal components with high cumulative contribution rates are plotted on a scatter diagram for clustering.

The abnormality detection unit 26 performs abnormality determination by comparing the predicted sensor data and the observed sensor data preprocessed in S103 (S108).

The cluster comparison unit 26a compares the distribution of principal component scores between the sensor data cluster and the predicted sensor data cluster.

If there is a distribution that deviates from the predicted sensor data cluster in the observed sensor data cluster, the abnormality candidate determination unit 26b determines that the sensor data of that time behaves differently from normal. The deviation determination range is determined by the maximum value in the positive direction and the minimum value in the negative direction for each axis of the distribution coordinates of the predicted sensor data. However, a function for adjusting the defined range is provided separately.

FIG. 7 is a diagram showing an example of abnormality determination. In FIG. 7, since the top two cumulative contribution ratios are used, they are represented two-dimensionally. The abnormality candidate determination unit 26b classifies data deviating from the distribution range of predicted sensor data as abnormality candidates.

The abnormality candidate level determination unit 26c classifies the deviation axial direction and distance of the observation sensor data of the time when the abnormality candidate determination unit 26b determines that the behavior is different from normal. As for level classification of anomaly candidates, there is a method of classifying them according to variations in deviation distance and deviation direction from the distribution of predicted sensor data.

If the abnormality candidate determination unit 26b determines that there is an abnormality, the determination result information output unit 27 outputs to the GUI of the display 70 whether or not there is behavior suspected to be an abnormality candidate or exceeds the threshold value in the observed sensor data based on the analysis described above ( S109). If neither the threshold value excess nor the abnormality candidate is detected, it is determined to be normal. Further, the determination result information output unit 27 may display the determination result for each processing on a history screen (GUI).

When the learning sensor data updating unit 28 determines that the observation sensor data is normal as a result of the analysis, it determines that the characteristics of the observation sensor data are adopted as learning sensor data for the next time. Then, it is stored in the learning sensor data storage unit 23 . As a result, the existing learned sensor data group stored in the learned sensor data storage unit 23 is updated.

FIG. 8 is a flowchart showing the flow of predictive sensor data selection processing.

When the predictive sensor data selection unit 22 determines that the acquired precondition is unknown, that is, it is a precondition that has not yet been recorded in the learned sensor data storage unit 23 (S201/No), the learned sensor data storage unit 23 (S202). The process of recording new preconditions and the observation sensor data acquired under the preconditions corresponds to the learning process.

The predicted sensor data selection unit 22 determines that the precondition is known, that is, the precondition recorded in the learning sensor data storage unit 23 (S201 / Yes), and the processing process of the observation target time is known. Although it is a prerequisite, if it is determined that it is the first machining process after the start of state prediction or that the prerequisite has been changed (S203 / Yes), from the learning sensor data group recorded in the learning sensor data storage unit 23 Only learning sensor data having the same preconditions as the observation target times are selected (S204).

When the prediction sensor data selection unit 22 determines that the precondition of the observation target time is the first processing step (S205/Yes), it selects learning sensor data for the first processing step (S206), and selects prediction sensor data. End the process. It is assumed that learning sensor data for the machining process for the first time is stored in advance in the learning sensor data storage unit 23 .

If the prediction sensor data selection unit 22 determines that the precondition of the observation target time is not the first processing step and the precondition is not changed (S203/No), or there is a change in the precondition of the observation target time (S203/Yes, S204, S205/No), the prediction sensor data selection unit 22 selects the prediction sensor used in the previous processing if the learning sensor data of the previous processing has not been updated (S207/No). The learned sensor data, in which the amount of state change is advanced by the difference Δ times from the data, is selected as the predicted sensor data (S208). After that, the predictive sensor data selection process ends. The difference Δ times is the number of times of processing in the prediction sensor data used in the previous processing from the observation target time. For example, the observation target time is the fifth processing work, but if the predicted sensor data used in the previous processing is the learning sensor data obtained in the third processing work, two state changes occur. Assuming that it is not, the state change amount is advanced twice to generate predicted sensor data.

When the prediction sensor data selection unit 22 determines that the previously processed learning sensor data has been updated (S207/Yes), the prediction sensor data selection unit 22 compares the latest learning sensor data with the feature amount of the learning sensor data group, and selects the closest learning sensor. Learning sensor data advanced by one state change amount is calculated from the data, and the data is selected as predicted sensor data (S209). After that, the predictive sensor data selection process ends.

The effects of this embodiment are as follows. When comparing the processing settings of the observed processing itself with the learning sensor data of the same processing settings in the past in the manufacturing and processing equipment that is the target of anomaly detection, the wear and deterioration of the parts of the device itself can be detected from the observation time. Since it depends on the contents of the previous processing and differs, there are cases where the shapes of the learned sensor data waveforms differ even among normal learned sensor data with the same processing settings. In the conventional technology, in such a case where the shape of normal sensor data itself is not fixed, by simply selecting learning sensor data with the same past processing settings and setting a fixed threshold, excessive abnormal detection due to threshold error , and there are cases where an abnormality that is originally intended to be found cannot be detected.

On the other hand, the anomaly detection system 1 according to the present embodiment learns by associating preconditions that define the manufacturing/processing state of the target device 10 with observed sensor data when operating normally under the preconditions. This learning sensor data is used to generate prediction sensor data that predicts the state of the observation target time. An abnormality is detected by comparing the predicted sensor data and the observed sensor data.

As a result, even if the shape of the normal learning sensor data waveform is not determined, it can be compared with the observation sensor data of the observation target time.

In addition, when data is summarized (dimension reduction) by principal component analysis, the result shows that the characteristic components such as noise and outliers are less likely to appear in the principal components that are ranked at the top of the cumulative contribution rate, and the anomaly that we want to find is difficult to find. By placing the feature component at the top, it is possible to detect a significant sign (a state in which the behavior is different from normal within a range that does not exceed the threshold).

The above-described embodiments do not limit the present invention, and modifications without departing from the gist of the present invention belong to the technical scope of the present invention. For example, although a polishing apparatus has been described above as an example, the application field of the abnormality detection system 1 of the present embodiment is not limited to that. For example, as case 1, the present invention can be applied to preventive maintenance of a wire electric discharge cutting machine (a device that cuts an object by electric discharge from a wire). In this case, learning/observation sensor data is used to detect a sign of disconnection of the wire of the electrical discharge processing unit and perform maintenance. It is believed that wire wear affects electrical observation sensor data during discharge.

Further, as case 2, the present embodiment may be applied to failure prediction of a spinning machine. Using learning and observation sensor data, detect and prevent signs such as thread breakage in synthetic fiber winders and processing machines. Fibers are believed to be susceptible to environmental factors.

In this way, in addition to wear and tear due to use, deterioration over time and environmental factors such as temperature and humidity affect the performance of equipment, the state of parts and workpieces, and the observed sensor data during normal operation This embodiment can be applied to a case where the waveform and the threshold value for determining abnormality and normality fluctuate.

1: Anomaly detection system 10: Target device 11: Observation sensor 20: Anomaly detection device 21: Data acquisition unit 22: Predicted sensor data selection unit 23: Learning sensor data storage unit 24: Preprocessing unit 24a: Threshold excess determination unit 24b: Feature amount extraction unit 25: Statistical processing unit 25a: PCA unit 25b: Clustering unit 26: Abnormality detection unit 26a: Cluster comparison unit 26b: Abnormality candidate determination unit 26c: Abnormality candidate level determination unit 27: Determination result information output unit 28: Learning Sensor data update unit 30: network 70: display 71: input device

Claims (8)

An abnormality detection device that detects an abnormality in a target device capable of performing processing operations multiple times,
Observation sensor data output by an observation sensor that observes the state of the target device , preconditions including at least one of processing conditions and work details of the target device when the observation sensor data is output, and the preconditions a data acquisition unit that acquires the cumulative number of times of processing performed below ;
Observation sensor data output by the observation sensor while the target device was operating in a normal state in the past, preconditions when the observation sensor data was output, and the cumulative number of processing performed under the preconditions are associated. a learning sensor data storage unit for storing as learning sensor data;
Selecting from the group of learned sensor data recorded in the learned sensor data storage section the learned sensor data output under the same preconditions and processed less than the observation target times of the target device as a reference. a prediction sensor data selection unit that generates prediction sensor data that predicts the state of the target device at the observation target time based on the selected learning sensor data;
an anomaly detection unit that detects occurrence of an anomaly in the target device at the observation target time based on a comparison result of the observed sensor data at the observation target time and the predicted sensor data;
An anomaly detection device comprising: In the abnormality detection device according to claim 1,
The predicted sensor data selection unit calculates, based on the learned sensor data, a state change amount when the target device performs the processing operation once under the same preconditions,
reading from the learning sensor data storage unit the learning sensor data with the number of times of processing that is less than the number of times to be observed and output under the same preconditions, the learning sensor data with the number of times of processing closest to the number of times to be observed; Predicting the state of the target device at the observation target time by advancing the state change amount for the difference count between the closest number and the observation target time to the read learning sensor data, and obtaining the predicted data as the predicted sensor data;
An anomaly detection device characterized by: In the abnormality detection device according to claim 1,
The predictive sensor data selection unit records the learned sensor data output under the same precondition as the known precondition in the case of the first machining process, although the precondition is a known precondition, in the learned sensor data storage unit. selecting from the learning sensor data group, selecting the selected learning sensor data as the predicted sensor data,
The abnormality detection unit detects the occurrence of an abnormality in the target device during the observation target time in the first processing process based on a comparison result of the observed sensor data in the first processing process and the predicted sensor data.
An anomaly detection device comprising: In the abnormality detection device according to claim 2 or 3,
Further comprising a threshold excess determination unit for removing observed sensor data containing noise or outliers,
The predicted sensor data selection unit obtains a variance or standard deviation of the predicted sensor data,
The threshold excess determination unit sets an upper threshold value and a lower threshold value of variations that can occur in the predicted sensor data according to a normal distribution based on the variance value or the standard deviation, and sets the observed sensor data of the observation target time. If the value of exceeds the upper limit threshold or falls below the lower limit threshold, the observation sensor data is excluded from anomaly detection processing,
An anomaly detection device characterized by: In the abnormality detection device according to claim 4,
Statistical features including at least one or more principal components included in the observed sensor data determined by the threshold excess determination unit to fall within the lower limit threshold and the upper limit threshold, and included in the predicted sensor data a feature quantity extraction unit for extracting statistical feature quantities containing at least one or more principal components,
A statistical processing unit that performs clustering of the statistical feature amount included in the observed sensor data and the statistical feature amount included in the predicted sensor data;
an abnormality detection unit that determines the degree of similarity between the observed sensor data and the predicted sensor data based on the clustering result of the statistical processing unit, and determines whether or not there is an abnormality in the target device based on the similarity;
a determination result information output unit that outputs a determination result of the presence or absence of the abnormality;
An anomaly detection device, further comprising: In the abnormality detection device according to claim 5,
Further comprising a learned sensor data update unit that records the observation sensor data used by the abnormality detection unit for determining that the target device is not abnormal in the learned sensor data storage unit,
An anomaly detection device characterized by: An abnormality detection method for detecting an abnormality in a target device capable of performing processing operations multiple times,
Observation sensor data output by an observation sensor that observes the state of the target device , preconditions including at least one of processing conditions and work details of the target device when the observation sensor data is output, and the preconditions a step of obtaining the cumulative number of times of machining performed below ;
Observation sensor data output by the observation sensor while the target device was operating in a normal state in the past, preconditions when the observation sensor data was output, and the cumulative number of processing performed under the preconditions are associated. storing as learning sensor data;
Based on the observation target times of the target device, the number of times of processing is less than that and the learning sensor data output under the same preconditions is read, and based on the read learning sensor data, at the observation target times generating predictive sensor data that predicts the state of the target device;
detecting the occurrence of an abnormality in the target device at the observation target time based on a comparison result of the observed sensor data at the observation target time and the predicted sensor data;
An anomaly detection method, comprising: The abnormality detection device according to claim 1 ;
An anomaly detection system configured by connecting a target device whose anomaly is to be detected by the anomaly detection device via a network.

Images