Abnormality detection device, abnormality detection method, and abnormality detection program

Abstract

The present invention improves the performance of identifying the cause of abnormality and irregularity in a production facility. An abnormality detection device comprises: a data acquisition unit that reads image data and process data from a storage device in which image data that is continuously output by an imaging device in a production facility, and process data that is continuously output by a sensor in the production facility are stored; an abnormality determination unit that calculates an image abnormality degree using a model that has learned the trend of feature amounts of image data in a normal state, and a feature amount of the image data acquired by the data acquisition unit, and calculates a sensor value abnormality degree indicating the degree of irregularity of the process data read by the data acquisition unit; and a cause diagnosis unit that outputs a cause calculated by the abnormality determination unit, on the basis of causal relationship information defining association of a causal relationship between a cause and a combination of the sensor value abnormality degree and the image abnormality degree.

Description

Anomaly detection device, anomaly detection method, and anomaly detection program

[0001] The present disclosure relates to an anomaly detection device, an anomaly detection method, and an anomaly detection program.

[0002] Conventionally, techniques have been proposed for estimating the cause of an abnormality using operation data received from a plant. For example, a technique has been proposed in which the probability of occurrence of a first abnormal event that has occurred in the past is weighted more heavily than the probability of occurrence of a second abnormal event that has not yet occurred, thereby estimating the cause of a sign of an abnormality (Patent Document 1).

[0003] JP 2018-109851 A

[0004] In general, it is desirable to prevent abnormalities in production facilities and suppress their impact on safety, stability, product quality, costs, etc. This technology aims to improve the ability to identify the cause of abnormalities in production facilities.

[0005] An example of an anomaly detection device according to the present disclosure is as follows: An anomaly detection device including: a data acquisition unit that reads image data and process data continuously output by an imaging device provided in production equipment from a storage device that stores the image data and the process data continuously output by a sensor provided in the production equipment, an anomaly determination unit that calculates a degree of image abnormality using a model that has learned the tendency of feature quantities of the image data in a normal state and the feature quantities of the image data acquired by the data acquisition unit, and calculates a sensor value abnormality degree that represents the degree of modulation of the process data read by the data acquisition unit, and a cause diagnosis unit that outputs the cause corresponding to the sensor value abnormality degree and the image abnormality degree calculated by the anomaly determination unit, based on causal relationship information that defines the association between a cause and a combination of the sensor value abnormality degree and the image abnormality degree.

[0006] The anomaly detection device according to the above, wherein the production facility is a plant for manufacturing a granular product, and the imaging device is provided in a tank of the manufacturing plant.

[0007] The anomaly determination unit uses the image data to output feature quantities at multiple levels from local features to global features using a convolutional deep learning model that outputs feature quantities at multiple levels from local features to global features, and calculates, as the image anomaly degree, a corrected anomaly degree obtained by correcting a deviation based on a difference between the feature quantities at multiple levels and the average of the corresponding levels using a model that has learned an average of the feature quantities in a normal state for each of the multiple levels using multiple image data in a normal state, based on smoothness of the texture of the image data read out by the data acquisition unit.

[0008] The anomaly detection device as described above, wherein the convolutional deep learning model is an EfficientNet, the deviation is calculated using Gaussian AD, and the corrected anomaly degree is calculated by multiplying the deviation by a value based on a Gray-Level Co-Occurrence Matrix (GLCM).

[0009] The anomaly detection device as described above, wherein the model is created using a pseudo-abnormal image created by smoothing the image data in a normal state.

[0010] The model is generated by learning the characteristics of the area where the product is imaged on an agitator blade installed in the tank, and estimates the area of ​​the image data where the product is imaged by segmentation.The abnormality determination unit calculates the maximum value of the size of the area where the product is imaged for each rotation cycle of the agitator blade, and determines that the smaller the maximum value, the higher the degree of image abnormality.

[0011] The abnormality determination unit calculates a degree of abnormality of the sensor value depending on whether a direction of deviation from a predetermined standard is positive or negative.

[0012] The causal relationship information defines the process data used to calculate the degree of abnormality of the sensor value by a timing, period, or interval in a process performed by the production equipment for each piece of process data output by the sensor, and the abnormality determination unit calculates the degree of abnormality of the sensor value using a value extracted from the process data read by the data acquisition unit based on the timing, period, or interval defined by the causal relationship information.

[0013] An anomaly detection method in which a computer executes the following processes: reading image data from a storage device that stores image data that is continuously output by an imaging device equipped in production equipment; reading the process data from a storage device that stores process data that is continuously output by a sensor equipped in the production equipment; calculating a degree of image abnormality using a model that has learned the tendency of feature amounts of the image data in a normal state and the feature amounts of the read image data, and calculating a sensor value abnormality degree that represents the degree of modulation of the read process data; and outputting the calculated sensor value abnormality degree and the cause corresponding to the image abnormality degree based on causal relationship information that defines the correspondence between a cause and a combination of the sensor value abnormality degree and the image abnormality degree.

[0014] An anomaly detection program for causing a computer to execute the following processes: reading image data from a storage device that stores image data that is continuously output by an imaging device equipped in production equipment; reading the process data from a storage device that stores process data that is continuously output by a sensor equipped in the production equipment; calculating a degree of image abnormality using a model that has learned the tendency of feature amounts of the image data in a normal state and the feature amounts of the read image data, and calculating a sensor value abnormality degree that indicates the degree of modulation of the read process data; and outputting the calculated sensor value abnormality degree and the cause corresponding to the image abnormality degree based on causal relationship information that defines the correspondence between a cause and a combination of the sensor value abnormality degree and the image abnormality degree.

[0015] The contents of the means for solving the problem can be combined as much as possible without departing from the problem and technical idea of ​​the present disclosure. The contents of the means for solving the problem can be provided as a device such as a computer or a system including multiple devices, a method executed by a computer, or a program executed by a computer. A recording medium storing the program may also be provided.

[0016] According to the disclosed technology, it is possible to improve the accuracy of identifying the cause of abnormal modulation in production equipment.

[0017] FIG. 1 is a diagram illustrating an example of a system according to the present embodiment. FIG. 2 is a schematic diagram illustrating an example of a process performed by equipment included in a plant. FIG. 3 is a diagram illustrating an example of equipment included in a plant. FIG. 4 is a diagram illustrating an example of process data in a batch process. FIG. 5 is a diagram illustrating an example of a pre-defined process line definition table. FIG. 6 is a diagram illustrating an example of a pre-defined tag definition table. FIG. 7 is a diagram illustrating an example of process data in a continuous process. FIG. 8 is a diagram illustrating an example of traceability information. FIG. 9 is a diagram illustrating the association between process data in a continuous process and a serial number in a batch process. FIG. 10 is a diagram illustrating an example of information pre-registered in a knowledge base. FIG. 11 is a diagram illustrating an example of a logic tree representing the relationship between modulation and its cause. FIG. 12 is a diagram illustrating synchronization processing of process data. FIG. 13 is a diagram illustrating an example of calculating an anomaly level for time-series data based on the distance from a reference. FIG. 14 is a diagram illustrating an example of calculating an anomaly level for time-series data based on the distance from a reference, taking into account the positive and negative directions of the time-series data. FIG. 15 is a diagram illustrating anomaly detection using an autoencoder. FIG. 16 is a diagram for explaining EfficientNet-B0. FIG. 17 is a diagram showing the relationship between GLCM features and the above-mentioned image abnormality degree. FIG. 18 is an example of an image in a retention box under normal and abnormal conditions. FIG. 19 is a block diagram showing an example of the configuration of an abnormal modulation cause identifying device. FIG. 20 is a process flow diagram showing an example of a learning process executed by the abnormal modulation cause identifying device. FIG. 21 is a diagram showing an example of an action table. FIG. 22 is a process flow diagram showing an example of an abnormality detection process executed by the abnormal modulation cause identifying device. FIG. 23 is a diagram showing an example of a screen output to an input / output device. FIG. 24 is a diagram showing another example of a screen output to an input / output device.

[0018] Hereinafter, an embodiment of an abnormal modulation cause identifying device (an abnormality detecting device) will be described with reference to the drawings.

[0019] <Embodiment> Fig. 1 is a diagram illustrating an example of a system according to this embodiment. The system 100 includes an abnormal modulation cause identification device 1, a control station 2, and a plant 3. The system 100 is, for example, a distributed control system (DCS) and includes multiple control stations 2. That is, the control system of the plant 3 is divided into multiple sections, and each control section is controlled in a distributed manner by the control station 2. The control station 2 is an existing facility in the DCS, and receives status signals output from sensors and the like included in the plant 3 and outputs control signals to the plant 3. Then, actuators such as valves and other devices included in the plant 3 are controlled based on the control signals.

[0020] The abnormal modulation cause identification device 1 acquires status signals (process data) from the plant 3 and image data obtained from an imaging device such as an ITV (Industrial Television) via a control station 2. The process data includes the temperature, pressure, flow rate, etc. of the processing target, such as raw materials and products, as well as setting values ​​that determine the operating conditions of equipment equipped in the plant 3. The image data corresponds to, for example, each of multiple frames continuously output from the imaging device. The abnormal modulation cause identification device 1 may perform feature extraction processing on the image data to obtain predetermined feature quantities. The abnormal modulation cause identification device 1 also creates an abnormality detection model based on a knowledge base that stores correspondences between assumed causes and, for example, effects that appear as an abnormality. For example, a model for identifying abnormal modulation, its precursors, and its causes is created based on a method for detecting deviations of process data changes from an acceptable range, created based on the knowledge base. The model makes judgments using feature quantities obtained from the image data in addition to changes in the process data. The abnormal modulation cause identification device 1 can then detect the occurrence or precursors of abnormal modulation using the model, process data, and image data. Furthermore, the abnormal modulation cause identification device 1 may determine, for example, candidate operating conditions for suppressing abnormal modulation based on a table that stores the causes of abnormal modulation and actions to address them, and the identified cause, and present these to the user.

[0021] FIG. 2 is a schematic diagram showing an example of a process performed by equipment provided in a plant. In this embodiment, a manufacturing plant in which a resin such as cellulose acetate is produced as a product will be described as an example. The manufacturing process may include a batch process 31 and a continuous process 32. In the batch process 31, the processing target is sequentially processed for each predetermined processing unit, and processing such as receiving, holding, and discharging raw materials into each device is performed in sequence. In the continuous process 32, the processing target that is continuously introduced is continuously processed, and processing such as receiving, holding, and discharging raw materials is performed in parallel. The process may also include multiple lines 33 that perform the same processing in parallel.

[0022] FIG. 3 is a schematic diagram showing an example of a plant. The devices performing each process include, for example, reactors, distillation apparatuses, heat exchangers, compressors, pumps, tanks, etc., which are connected via piping. Sensors, valves, etc. are provided at predetermined positions in the devices and piping. The sensors may include thermometers, flow meters, pressure gauges, level gauges, concentration meters, etc. The sensors monitor the operating status of each device and output status signals. The sensors in the plant 3 are each assigned a "tag," which is identification information for identifying the sensor. The type of process data can be identified based on the tag. The abnormal modulation cause identification device 1 and the control station 2 manage input and output signals to each device based on the tag.

[0023] 3 is a diagram showing an example of equipment included in a plant. For example, a plant 3 for producing resin includes a shredder 301, a cyclone 302, a pretreatment machine 303, a precooler 304, a reactor 305, an aging tank 306, a retention tank (RT) 307, a precipitator 308, a hardening tank 309, a solid-liquid separator 310, a washer 311, a dehydrator 312, a dryer 313, and a product silo 314.

[0024] The crusher 301 receives raw materials such as cellulose (pulp) and performs a deflocculation process to break them into flocculent material. The deflocculated raw materials are then sent to the pre-treatment machine 303 via the cyclone 302. The pre-treatment machine 303 also performs a pre-treatment process.

[0025] The pre-cooler 304 performs a pre-cooling step of cooling a predetermined mixed solution to be reacted with the raw material. The reactor 305 receives the raw material treated in the pre-treatment step and the mixed solution cooled in the pre-cooling step, and performs an acetylation step.

[0026] In addition, the aging tank 306 introduces steam into the solution (dope) after the acetylation reaction to raise the temperature and perform an aging process. The above-mentioned process corresponds to the batch process 31 shown in Fig. 2. In the batch process 31, processes belonging to a plurality of series 33 are sequentially performed in the above-mentioned equipment. After that, the intermediate product is accumulated in the RT 307 and corresponds to the continuous process 32 in which the intermediate product is continuously treated.

[0027] The settling machine 308 extrudes the aging-treated dope through a perforated plate to perform a precipitation process in which a resin is precipitated. In this process, for example, cellulose acetate is precipitated. The retention box 309 retains the precipitated resin and performs a hardening process in which the resin is solidified. In this embodiment, image data of the resin in the retention box 309 is acquired and used to identify the cause of abnormal modulation.

[0028] The solid-liquid separator 310 performs a solid-liquid separation (acid removal) step to separate the precipitated resin from the aqueous solution. The washer 311 performs a washing step to wash the acid-containing resin with pure water. The acid is removed by the washing step.

[0029] The dehydrator 312 performs a dehydration process to remove water from the resin after washing. This dehydration process reduces the load on the subsequent dryer. The dryer 313 performs a drying process to dry the resin after dehydration. The dried resin is stored in a product silo 314 and then shipped.

[0030] <Batch Process> FIG. 4 is a diagram illustrating an example of process data in a batch process. The left column of FIG. 4 shows a portion of the batch process 31 shown in FIG. 2. Specifically, the process includes a crusher 301, a cyclone 302, a pre-treatment device 303, a pre-cooler 304, and a reactor 305. These processes are further classified into a pre-treatment process, a pre-cooling process, and a reaction process. The right column of FIG. 4 shows an example of process data acquired in each process. In the pre-treatment process, time-series data is acquired from sensors or imaging devices with tags 001 and 002. In the pre-cooling process, time-series data is acquired from sensors with tags 003 and 004. In the reaction process, time-series data is acquired from sensors with tags 005, 006, and 007. In the batch process, processing objects associated with serial numbers (also referred to as "production numbers," "batch numbers," or "control numbers") are intermittently processed. That is, the serial number is identification information for identifying processing targets that are processed collectively in a batch process. As shown in FIG. 4 , time-series data for processing targets associated with subsequent serial numbers is obtained over time. In this embodiment, the control station 2 manages serial numbers and steps that indicate processing stages in the subdivided processes that make up the batch process. Note that, if the steps are reset by a programmable logic controller (PLC) in the plant 3 connected to the control station 2, the serial number of the process data output from the control station 2 may be used appropriately depending on the timing of communication between the control station 2 and the plant 3 (e.g., after a step change in the PLC or after a set time has elapsed). The set time may also be set for each production line and each subdivided process.

[0031] 5 is a diagram showing an example of a pre-set process line definition table. The process line definition table registers, for each series and process, the serial number, a definition of the step indicating the processing stage in each process, and the type of product to be processed in each process. The process line definition table may be a so-called database table, or may be a file in a predetermined format such as CSV. The process line definition table is also created in advance by the user and read out by the abnormal modulation cause identification device 1.

[0032] The process line definition table includes attributes such as series, process, serial number, step, and product type. The series field stores identification information for specifying the process series. The process field stores identification information for indicating the subdivided processes in a batch process. The serial number field stores a serial number, which is identification information for identifying the processing objects that are processed collectively in a batch process. The step field stores a definition of the timing of multiple steps that indicate the processing stages in the process. The product type field stores the type of processing object.

[0033] 6 is a diagram showing an example of a pre-set tag definition table. The tag definition table defines the acquisition timing of process data or image data obtained from a sensor or imaging device corresponding to each tag. The tag definition table may be a so-called database table or a file in a predetermined format such as CSV (Comma Separated Values). The tag definition table is created in advance by the user and read out by the abnormal modulation cause identification device 1.

[0034] The tag definition table includes the attributes of tag, series, process, and collection interval. The tag field registers a tag that is identification information for a sensor or imaging device. The series field registers identification information for specifying a process series. The process field registers identification information that indicates a subdivided process in a batch process. The collection interval field registers information that indicates the interval at which the sensor output value or image data is acquired.

[0035] <Continuous Process> FIG. 7 is a diagram illustrating an example of process data in a continuous process. The left column of FIG. 7 shows part of the process of the continuous process 32 shown in FIG. 2. Specifically, the process includes the RT 307 and the settling machine 308. The right column of FIG. 7 shows an example of process data acquired in each process. In the continuous process 32, time-series data associated with tags but not with serial numbers is continuously acquired from a sensor or an imaging device. In the continuous process, time-series data is acquired from each sensor or imaging device with tags 102 and 103. In the continuous process, equipment continuously receives objects to be processed and continuously processes them.

[0036] When a continuous process is performed after a batch process, traceability information set in advance by a user is used in this embodiment to link the processing target in the batch process with the processing target in the continuous process. FIG. 8 is a diagram showing an example of traceability information. The traceability information includes attributes of a sampling interval and a residence time. The sampling interval field is registered with the interval at which sampling is performed in the continuous process for process inspection using, for example, a reduction method. The residence time field is registered with the time the processing target remains in the process from the completion of the batch process until it reaches the process included in the continuous process.

[0037] FIG. 9 is a diagram illustrating the association of process data or image data in a continuous process with serial numbers in a batch process. The process data or image data is acquired, for example, at intervals set in the traceability information. Furthermore, when a continuous process is performed after a batch process, the product of the batch process completed within a predetermined period is introduced into a tank or the like as the processing target for the continuous process. Therefore, the process data in the continuous process can be associated with a group of serial numbers whose completion time of the batch process falls within the predetermined period by tracing back the residence time of the processing target from the completion of the batch process to the time of measurement by a sensor or the output of image data by an imaging device. This association improves the accuracy of identifying the cause of an abnormality using the process data in the batch process when a batch process and a continuous process are performed consecutively.

[0038] As described above, by associating the serial number in batch processing with the timing of measurement in continuous processes, it is possible to improve the accuracy of identifying the cause of an abnormality.

[0039] FIG. 10 is a diagram illustrating an example of information pre-registered in the knowledge base. The knowledge base is pre-stored in the storage device of the abnormal modulation cause identification device 1. The table in FIG. 10 includes an "influence" column corresponding to each sensor or imaging device (tag) and rows indicating "assumed causes" of modulation. That is, the direction of value fluctuation is registered in the column corresponding to the sensor or image data affected by causes such as "Cause 1" and "Cause 2" shown in each row. In the knowledge base, the direction of fluctuation is represented by "up" indicating an increase (rise) in the sensor output value or "down" indicating a decrease (decrease). Note that, as shown in FIG. 10, the combination of cause and effect is not necessarily one-to-one. In addition, the process data calculation method, extraction timing, threshold values ​​used for abnormality determination, etc. are defined for each sensor. Information indicating the calculation to be performed on the output value of each sensor is registered in the calculation method row. In this embodiment, calculations are performed using machine learning techniques such as the Hotelling algorithm, k-nearest neighbor algorithm, DTW Barycenter Averaging, Autoencoder, Graphical Lasso, and EfficientNet, anomaly detection techniques such as Gaussian AD, and texture analysis techniques such as GLCM (Gray-Level Co-Occurrence Matrix). The extraction timing row contains information indicating the timing at which a value to be used for anomaly detection is extracted from the output values ​​of each sensor. For example, in batch processing, the timing may be defined by a step indicating a processing stage in each process, a specific period, or a specific time point. Furthermore, in continuous processing, the timing may be defined by a sampling interval, as shown in FIG. 8 . The threshold row contains a threshold value, which is the standard for determining an anomaly in each anomaly detection technique. The threshold value may include, for example, an upper limit and a lower limit. As described above, the knowledge base defines combinations of causal relationships between causal events and the resulting effects, which are the modulations in process data. In addition, combinations of causal relationships can be represented in a tree format, with the modulation that appears as an effect as the root and its assumed causes as the leaves, and the events that appear in the process from the cause to the modulation connected hierarchically in chronological order.

[0040] The knowledge base is prepared in advance by a user based on, for example, HAZOP (Hazard and Operability Study). HAZOP is a method for comprehensively listing, in association with each other, the following: detection means at monitoring points using instrumentation devices constituting a plant; control ranges (upper and lower limit thresholds and alarm setpoints); deviations from the control ranges (abnormalities, modulations); a list of assumed causes of deviations from the control ranges; logic (detection means) for determining which assumed cause caused the deviation; the impact of the deviation; measures to be taken when the deviation occurs; and actions for the measures. The knowledge base may be prepared based on methods other than HAZOP, such as Fault Tree Analysis (FTA), Failure Mode and Effect Analysis (FMEA), Event Tree Analysis (ETA), or methods based on these or similar methods; content extracted from interviews with operators; and content extracted from work standards and technical standards. In this embodiment, anomaly detection is performed based on parameters that are considered to have a causal relationship in the knowledge base.

[0041] Based on the information set in the tables, the abnormal modulation cause identification device 1 extracts data at a predetermined timing from the process data and image data acquired from the plant 3 and performs an abnormality determination using a predetermined method. FIG. 11 shows an example of a logic tree representing the relationship between modulation and its causes. The logic tree can be created based on the knowledge base shown in FIG. 10 . The logic tree in FIG. 11 places events occurring upstream in the production process and earlier in the time series on the left side, and events occurring downstream in the production process and later in the time series on the right side, with arrows connecting the assumed causes hierarchically to the modulation that appears as an effect. If multiple assumed causes exist for a single modulation in the knowledge base table, the logic tree branches and connects them, and displays a group of events that commonly appear in the process from the assumed cause to the modulation. The thick solid-line rectangles at the upstream end of each branch correspond to the assumed causes in the knowledge base table, and correspond to the numbers in parentheses in FIGS. 10 and 11 . The thin solid-line rectangles correspond to the effects in the knowledge base table and represent events that can be observed from the process data. For each of these effects, calculations are performed according to the calculation method defined in the knowledge base table. In addition, for each assumed cause, a model containing a formula for performing the above calculations is defined, and the model can be used to detect abnormalities or their precursors and to assist in identifying their causes.

[0042] <Calculation Method> The above calculation may include, for example, the following methods: Furthermore, the abnormal modulation cause identifying device 1 may be configured to display the results of these calculations.

[0043] ・Hotelling method (T ２(Hotelling method) For example, assuming that multiple process data obtained from one sensor follow a predetermined probability density function, the population mean and standard deviation are estimated from the sample mean and sample standard deviation calculated using the process data. The predetermined probability density function is, for example, a normal distribution. Then, the degree of anomaly (sensor value anomaly degree) obtained from the sensor value is calculated based on the distance from the population mean to the process data to be verified. For example, the sensor value anomaly degree is determined based on the square of the Mahalanobis distance. Note that the instantaneous value of the process data itself may be used, or the maximum value, minimum value, integral value, standard deviation, or differential coefficient (slope) of the process data over a predetermined period may be used to calculate the sensor value anomaly degree based on Hotelling's theory. The Hotelling method makes it possible to detect outliers from a predetermined standard.

[0044] - K-nearest neighbor method: For example, time-series process data obtained from one or more sensors is vectorized or converted into a matrix, and the distance between the data is calculated. The distance may be Euclidean distance, Mahalanobis distance, or Manhattan distance. The degree of abnormality of the sensor value is then determined based on the distance from the data to be verified to the kth closest data. In the k-nearest neighbor method, the determination is based on the relationship with other data. Therefore, for example, in cases where normal values ​​can be classified into multiple clusters, it is possible to detect outliers that are far from any of the multiple clusters.

[0045] Dynamic Time Wrapping (DTW) Barycenter Averaging: Average time series data can be calculated based on multiple time series data, such as process data from different batch processes. For example, the distance between each of process data with different serial numbers in a corresponding section of a batch process and the average time series data can be calculated. FIG. 12 is a diagram illustrating process data synchronization. For each element value included in time series data from batch processes with different serial numbers, the shortest distance between the values ​​included in the different time series data is calculated in a brute-force manner, and the time series data is aligned by sliding along the time axis so that the integrated value of the shortest distance is minimized. In other words, multiple time series data are synchronized based on the similarity of the time series data. This allows multiple process data to be displayed in an overlapping manner so that the steps in the processes performed in the plant 3 correspond in time series. The degree of sensor value anomaly is then calculated based on the integrated value of the distance between the synchronized time series data using k-nearest neighbors or Hotelling's theory. DTW Barycenter Averaging allows anomalies to be detected based on the degree of similarity between time series data.

[0046] The degree of sensor value anomaly may be calculated with a positive or negative sign for deviation from a reference such as the average. FIG. 13 is a diagram illustrating an example of calculating the degree of sensor value anomaly based on the magnitude of the distance from the reference for time-series data. FIG. 14 is a diagram illustrating an example of calculating the degree of sensor value anomaly based on the distance from the reference, taking into account the positive and negative directions for the same time-series data. In each diagram, the vertical axis represents, for example, the degree of deviation from the average. Although modulation actually occurs in the area indicated by the dashed rectangle, it is difficult to detect this based only on the values ​​shown in the example of FIG. 13. On the other hand, the example of FIG. 14 shows a tendency for deviations in the positive and negative directions to reverse, making it easy to detect modulation.

[0047] For example, in the Hotelling method described above, the degree of deviation from the reference is calculated as a positive or negative signed value without squaring the distance, thereby obtaining the degree of sensor value anomaly as shown in FIG. 14 . In DTW Barycenter Averaging and the like, a positive or negative sign is determined for a characteristic point, such as a maximum value, in the time-series data, for example, using the following formula, and the calculated value is multiplied by the magnitude of the distance. Sign Determination Formula = (μ - x) / |μ - x | where μ is the average value (reference value) of the training data, and x is the process data to be verified. In this way, using the sign determination formula, a sign indicating the direction of deviation from the reference at a given time point in the time-series data can be determined according to the magnitude relationship between the reference value at the time point and the process data to be verified at the corresponding time point. Furthermore, by using a signed value indicating the degree of deviation from the reference, the degree of sensor value anomaly as shown in FIG. 14 can be obtained, thereby reducing false detections. Furthermore, in addition to maximum values, minimum values, the difference between process data at a certain time point and process data at another time point, etc. may also be used as characteristic points in the time-series data.

[0048] Autoencoder (Autoencoder) FIG. 15 is a diagram illustrating anomaly detection using an autoencoder. This method performs anomaly detection based on the relationship between process data from multiple sensors. Specifically, a neural network is used to create a model that can compress (encode) and restore (decode) input data using, for example, continuous or batch process process data itself as training data. In the neural network, the number of nodes in the input and output layers corresponds to the number of sensors, and the number of nodes in the middle layer is less than the number of sensors. Information input to the input layer is compressed in the middle layer and restored in the output layer. Note that multiple middle layers may be present, and the connection structure between layers is not limited to full connections. Then, a learning process is performed using normal process data as training data, and a model is created with parameters adjusted to minimize the difference between the input layer value and the output layer value. Furthermore, in the anomaly detection process, the process data to be verified is input, and the sensor value anomaly degree is calculated based on the difference between the input layer value and the output layer value. In other words, when abnormal process data is input, the information compressed in the intermediate layer cannot be properly restored in the output layer, resulting in a large difference between the values ​​in the input layer and the output layer, and abnormalities can be detected based on this difference.By using an autoencoder, abnormalities can be detected based on the characteristics of the relationship between the output values ​​of multiple sensors.

[0049] Graphical Lasso: For example, the dependency between variables is quantified based on the covariance matrix of process data from multiple sensors in continuous processing or batch processing, and the quantified dependency is expressed as a sparse reference graph. Under normal conditions, it can be determined that the dependency between variables does not deviate significantly from the reference. Then, in the anomaly detection process, the dependency between variables is determined using the process data to be verified, and the degree of sensor value anomaly is calculated according to the magnitude of the difference from the above-mentioned reference. Graphical Lasso can quantify the correlation between process data, and the degree of sensor value anomaly can be detected based on the disruption of the relationship.

[0050] Furthermore, the following processing may be performed on the image data.

[0051] Gaussian AD: Normal image data is used as training data, and learning is performed using, for example, EfficientNet-B0. FIG. 16 is a diagram for explaining EfficientNet-B0. EfficientNet-B0 extracts feature vectors corresponding to multiple levels using convolutional deep learning. At multiple levels, the size of the image data gradually decreases, and the number of dimensions called channels increases. As shown in FIG. 16, nine levels of feature vectors with dimensionality ranging from 16 to 1280 are calculated. During the calculation process, feature vectors utilize GAP (Global Average Pooling) to average the spatial direction (width and height directions), compressing the dimensions only in the depth (channel) direction. Feature vectors at multiple levels can represent global features from local features of the image data. The average of the feature vectors at each level is then calculated and stored in a storage device as a normal model.

[0052] In the operation stage, a feature vector is similarly calculated for the input image data, and the difference M between the feature vectors at multiple levels and the average (normal model) at each level is calculated. ｉ is calculated by the Hotelling method described above. Also, the difference M ｉ The sum of these is calculated as the degree of abnormality (degree of image abnormality or deviation) obtained from the image.

[0053] Correction of Image Anomaly Degree Based on GLCM (Gray-Level Co-Occurrence Matrix) GLCM is a value that represents the smoothness of the texture of image data. Statistical values ​​(GLCM features) related to contrast, correlation, uniformity, etc. can be calculated based on the relationship with surrounding pixels. FIG. 17 shows the relationship between GLCM features and the image anomaly degree described above. In the graph of FIG. 17, the vertical axis represents the image anomaly degree, and the horizontal axis represents the GLCM feature. An image anomaly degree exceeding a predetermined threshold is determined to be abnormal. The area (1) indicated by the dashed line contains data that was erroneously determined to be abnormal even though there was no problem with the product state. The area (2) indicated by the dashed line contains data that was detected as abnormal and the product state was detected. The area (3) indicated by the dashed line contains data that was determined to be normal and not abnormal. The area (4) indicated by the dashed line contains data that was detected as abnormal and the product state was overlooked. As shown in FIG. 17, when the results of abnormality determination based on Gaussian AD are plotted based on GLCM features, the correctness or incorrectness of the determination can be roughly classified on the graph.

[0054] Therefore, the image abnormality degree may be corrected using GLCM. Specifically, the corrected abnormality degree a ｎｅｗ is the image abnormality degree a obtained by Gaussian AD and the GLCM feature (homogeneity) b ２ The result is obtained by multiplying ｎｅｗ = a × b ２ Such a correction abnormality degree a ｎｅｗ According to this, the accuracy of abnormality determination can be improved.

[0055] Machine Learning Using Pseudo-Abnormal Images Generally, image data of products in abnormal states in production plants is scarce, making supervised learning difficult. Therefore, image data of products in normal states is used to create pseudo-abnormal image data and use it as training data. The pseudo-abnormal image data is created by smoothing the normal image data. Smoothing can be performed, for example, using OpenCV's blur function to average a kernel size of 64, but any method is acceptable as long as it can blur the entire image. This pseudo-abnormal image is based on the operator's knowledge that when the image data is an image of a retention box 309 (Figure 4) that retains precipitated resin such as cellulose acetate, the texture displayed in the image becomes smooth in the event of an abnormality where the resin shrinks. Furthermore, normal and pseudo-abnormal images are used to perform learning using the VGG16 convolutional neural network model, and the output is the image anomaly level.

[0056] Segmentation: For example, semantic segmentation using DeepLab v3+ or similar may be used to estimate the amount of product lifted onto the agitator blades installed in the retention chamber 309 from the characteristics of the image data. Figure 18 shows example images of the retention chamber 309 under normal and abnormal conditions. The area highlighted by the solid white rectangle in Figure 18 captures the top of the agitator blade. Under normal conditions, the product is lifted to the top of the agitator blade, indicating sufficient precipitation. On the other hand, under abnormal conditions, the amount of product present on the top of the agitator blade is relatively small. This method can determine the size and coarseness of the precipitated product, so in addition to resin, compounds scraped up by the agitator can also be monitored.

[0057] In the segmentation method, multiple image data (frames) that are continuously output are divided into groups to be used for processing. For example, one rotation period of the impeller is used as the processing unit. Furthermore, annotation is performed by specifying pixels in the training data where the product on the impeller is imaged. Existing methods such as SEAM and AffinityNet can be used to select regions in the annotation. In this way, a model for estimating the amount of product from image data can be created. Furthermore, in the operational phase, the maximum estimated amount of product in one cycle is calculated, and the smaller the maximum value, the higher the degree of abnormality. Since the segmentation method calculates the degree of image abnormality based on the amount of product, the basis for the judgment is highly explainable, and it is expected to be applicable to other uses.

[0058] In addition, general anomaly detection methods or methods that apply these may also be used. Furthermore, the threshold values ​​used for anomaly detection in each method may be searched for using process data or image data actually obtained during operation of the plant 3, so as to minimize erroneous judgments under normal conditions and to quickly detect the occurrence of an anomaly or its sign under abnormal conditions, and the searched values ​​may be registered in the knowledge base shown in FIG.

[0059] <Device Configuration> FIG. 19 is a block diagram showing an example of the configuration of the abnormal modulation cause identification device 1. The abnormal modulation cause identification device 1 is a computer and includes a communication interface (I / F) 11, a storage device 12, an input / output device 13, and a processor 14. The communication I / F 11 may be, for example, a network card or a communication module, and communicates with other computers based on a predetermined protocol. The storage device 12 may be a main storage device such as a random access memory (RAM) or a read-only memory (ROM), or an auxiliary storage device (secondary storage device) such as a hard-disk drive (HDD), a solid-state drive (SSD), or a flash memory. The main storage device temporarily stores programs read by the processor 14 and information exchanged with other computers, and secures a working area for the processor 14. The auxiliary storage device stores programs executed by the processor 14 and information exchanged with other computers. The input / output device 13 is a user interface such as an input device such as a keyboard or a mouse, an output device such as a monitor, or an input / output device such as a touch panel. The processor 14 is an arithmetic processing device such as a CPU (Central Processing Unit), and performs each process according to this embodiment by executing a program. In the example of Fig. 19, functional blocks are shown within the processor 14. That is, by executing a predetermined program, the processor 14 functions as a data acquisition unit 141, a preprocessing unit 142, a learning processing unit 143, an abnormality determination unit 144, a cause diagnosis unit 145, and an output control unit 146.

[0060] The data acquisition unit 141 acquires process data from sensors provided in the plant 3 and image data from imaging devices, for example, via the communication I / F 11 and the control station 2, and stores the data in the storage device 12. As described above, the process data and image data are associated with the sensors and imaging devices by tags.

[0061] The preprocessing unit 142 processes the process data and image data when creating an anomaly detection model. For example, the preprocessing unit 142 links the process data and image data with a serial number. That is, based on the traceability information previously stored in the storage device 12, the preprocessing unit 142 links the process data and image data corresponding to a predetermined tag, system, and serial number in batch processing with the process data and image data corresponding to a predetermined tag in continuous processing and output at a predetermined timing. The preprocessing unit 142 also extracts data for a predetermined period to be used for anomaly detection based on settings in a table such as a knowledge base, and calculates feature values ​​corresponding to each method. Note that, in the learning process, the preprocessing unit 142 may perform data cleansing to extract training data by excluding data from non-steady operating periods, data when an anomaly occurs, and outliers such as noise. The preprocessing unit 142 may also calculate predetermined feature values ​​from the image data and perform predetermined image processing, such as smoothing, on the image data.

[0062] The learning processing unit 143 creates an anomaly detection model including one or more operations based on, for example, a knowledge base, and stores the model in the storage device 12. At this time, the learning processing unit 143 determines parameters that have learned the characteristics of training data, which is process data or image data. Note that when performing learning processing using output values ​​from multiple sensors, normalization may be performed as appropriate.

[0063] The anomaly determination unit 144 calculates the degree of sensor value anomaly using the process data and the anomaly detection model, and calculates the degree of image anomaly using the image data and the anomaly detection model. That is, in the learning process, the anomaly determination unit 144 calculates the degree of anomaly (i.e., the degree of sensor value anomaly or the degree of image anomaly) using test data for cross-validation and the anomaly detection model. In addition, in the anomaly determination process, the degree of anomaly is calculated using the process data or image data acquired from the plant 3.

[0064] The cause diagnosis unit 145 calculates the probability (likelihood) of each of the multiple assumed causes using the calculated abnormality degree. The probability is calculated, for example, using the abnormality degree calculated by the abnormality determination unit, based on the proportion and degree of influence that appears in the process data or image data among the influences associated with each assumed cause in the knowledge base. In addition, actions that indicate measures to be taken against the cause may be stored in the storage device 12 in association with each assumed cause, so that the actions can be presented to the user.

[0065] The output control unit 146 issues an alarm when an abnormality is detected, and outputs the probability of each assumed cause, for example, via the input / output device 13. The output control unit 146 is connected to the above-mentioned components via the bus 15 as appropriate in response to user operations. For convenience, one device shown in Fig. 19 includes the data acquisition unit 141, preprocessing unit 142, learning processing unit 143, abnormality determination unit 144, cause diagnosis unit 145, and output control unit 146, but at least some of the functions may be distributed and provided in different devices.

[0066] <Learning Process> FIG. 20 is a process flow diagram showing an example of the learning process executed by the abnormal modulation cause identification device 1. The processor 14 of the abnormal modulation cause identification device 1 executes a predetermined program to perform the process shown in FIG. 20 . The learning process is performed at any timing using process data or image data obtained through past operations of the plant 3. The learning process mainly includes preprocessing (FIG. 20: S1), model construction processing (S2), and verification processing (S3). That is, cross-validation may be performed using part of the process data or image data as training data and the rest as test data. Note that the above-mentioned tables and the like are assumed to be created by a user and stored in the storage device 12 in advance. For convenience, the preprocessing, learning processing, and verification processing are described in one process flow shown in FIG. 20 . However, at least part of the preprocessing, verification processing, etc. may be distributed and executed by different devices.

[0067] The data acquisition unit 141 of the abnormal modulation cause identification device 1 acquires process data or image data ( S11 in FIG. 20 ). In this step, data used for the anomaly detection model is extracted from the process data shown in FIGS. 4 and 7 , for example, image data output by an imaging device installed in the retention box 309. The process data is assumed to be stored in the storage device 12 in a file format such as OPC data, a so-called database table, or CSV. The image data is continuously acquired at a frame rate of, for example, 30 fps. The image data, which is each frame, is general raster data. The process data also includes attributes such as date and time and tags, and may further include attributes such as a serial number and step, particularly in batch processing process data. The image data is also assumed to be associated with information such as date and time and tags.

[0068] Furthermore, the pre-processing unit 142 of the abnormal modulation cause identifying device 1 associates the process data or image data of the continuous processing with the serial numbers (S12 in FIG. 20). In this step, as shown in FIG. 9, the process data or image data acquired in the continuous process is associated with the group of serial numbers of the process data or image data acquired in the batch processing, and the process data or image data used in calculating the degree of abnormality is associated. That is, in the knowledge base shown in FIG. 10 or the logic tree shown in FIG. 12, if a certain cause affects both the process data or image data of the batch processing and the process data or image data of the continuous processing, the degree of abnormality and the degree of occurrence are calculated based on the associated data in this step.

[0069] Then, the pre-processing unit 142 extracts and processes data to be used in the anomaly determination model (FIG. 20: S13). In this step, the pre-processing unit 142 extracts data for a predetermined period to be used for anomaly determination based on the setting values ​​of a table such as a knowledge base, and calculates feature quantities according to each method.

[0070] For example, when calculating the degree of anomaly using the Hotelling method, the preprocessing unit 142 extracts process data at a predetermined timing or period, calculates instantaneous values, which are the process data itself, maximum values, minimum values, integral values, or differences of the process data, integral values ​​of reaction rates, and differential coefficients at predetermined times, and stores these values ​​in the storage device 12. When calculating the degree of anomaly using the k-nearest neighbor method, the preprocessing unit 142 vectorizes or matrixes the time-series process data. When calculating the degree of anomaly using DTW Barycenter Averaging, the preprocessing unit 142 synchronizes multiple pieces of process data to obtain average time-series data. When calculating the degree of anomaly using an autoencoder or graphical lasso, the preprocessing unit 142 synchronizes multiple pieces of process data. When calculating the degree of anomaly based on image data, the preprocessing unit 142 performs predetermined image processing or calculates feature vectors.

[0071] The pre-processing unit 142 may perform predetermined data cleansing on the process data. The data cleansing process is a process for eliminating outliers, and various methods can be used. For example, a moving average value may be calculated using the most recent data. Alternatively, the difference between the moving average value and the actual measurement value may be calculated to obtain the standard deviation σ, which represents the variation in the difference. Then, values ​​that do not fall within a predetermined confidence interval, such as the interval from the mean value of the probability distribution -3σ to the mean value of the probability distribution +3σ (also called the 3σ interval), may be excluded. Similarly, values ​​that do not fall within the 3σ interval for the difference between the previous and next actual measurement values ​​may be excluded.

[0072] The learning processing unit 143 of the abnormal modulation cause identification device 1 then performs an anomaly detection model construction process ( S2 in FIG. 20 ). In this step, an anomaly detection model including an anomaly degree calculation is created based on the knowledge base shown in FIG. 10 . Specifically, for one or more “effects” associated with each of the “assumed causes” in FIG. 10 , an anomaly degree is calculated using the method registered in the “calculation method” field, and an anomaly detection model represented by a combination of the anomaly degrees is created. Furthermore, depending on the anomaly detection method, the learning processing unit 143 adjusts the model parameters using training data. For example, when calculating the sensor value anomaly degree using an autoencoder, the weight coefficients between layers are adjusted so that the input process data can be restored after being compressed. When calculating the sensor value anomaly degree using a graphical lasso, the dependency between variables is quantified based on the covariance matrix of process data from multiple sensors. Note that the image anomaly degree is calculated based solely on image data, but the calculated image anomaly degree can be combined with other anomaly degrees to identify the cause of the abnormal modulation. The learning processing unit 143 then stores the created anomaly detection model in the storage device 12.

[0073] The anomaly determination unit 144 of the abnormal modulation cause identification device 1 calculates the degree of anomaly using the created anomaly detection model and test data ( FIG. 20 : S31). In this step, the anomaly determination unit 144 calculates the degree of anomaly according to the method of calculating the degree of anomaly. For example, when calculating the degree of anomaly using the Hotelling method, the process data is used to estimate the sample mean and sample standard deviation of a population, and the degree of anomaly is calculated based on the distance from the population mean to the process data to be verified. When calculating the degree of anomaly using the k-nearest neighbor method, the distance between data is calculated, and the degree of anomaly is calculated based on the distance from the data to be verified to the kth closest data. When calculating the degree of anomaly using DTW Barycenter Averaging, the degree of anomaly is calculated using the k-nearest neighbor method or the Hotelling theory based on the integrated value of the distance between time-series data synchronized in preprocessing. When calculating the degree of anomaly using an autoencoder, the process data to be verified is input to the autoencoder, and the degree of anomaly is calculated based on the difference between the value of the input layer and the value of the output layer. When calculating the degree of anomaly using the Graphical Lasso method, the process data to be verified is used to determine the dependency relationship between variables, and the degree of anomaly is calculated based on the magnitude of the difference from the reference dependency relationship. When calculating the degree of image abnormality, for example, the method described above is used to calculate the degree of image abnormality using a model that has learned the tendency of the features of image data of products in a normal state and the features of the captured image data.

[0074] The cause diagnosis unit 145 of the abnormal modulation cause identification device 1 uses the calculated abnormality level to determine the probability of the assumed cause ( S32 in FIG. 20 ). In this step, the probability of each assumed cause in the knowledge base is calculated based on the proportion of modulations associated with the corresponding influence. For example, cause (1) in FIG. 10 is associated with three influences: an increase in flow rate in tag 001, a decrease in temperature 2 in tag 005, and the resin size in tag 009 being smaller than a predetermined standard. As described above, the sensor value abnormality level is calculated as a value corresponding to a positive or negative direction. In addition to the model for calculating the abnormality level when the resin size is smaller than the standard, a model for calculating the abnormality level when the resin size is larger than the standard may be created to enable detection of abnormalities in both positive and negative directions. In S32 in FIG. 20 , the abnormality levels calculated for each influence in S31 may be used to determine the probability of the proportion of the three influences whose abnormality levels exceed a threshold. For example, if the anomaly levels for two of the three influences exceed the threshold, the probability of validity can be 66.7%. Furthermore, in calculating the probability of validity, weighting may be further performed according to the type of influence (tag) or the magnitude of the anomaly level. For example, the probability of validity may be calculated by multiplying each influence by a weight and then calculating the sum of the weights.

[0075] The output control unit 146 also outputs the anomaly degree calculated in S31 and the validation degree calculated in S32 so that the user can evaluate the created model ( S33 in FIG. 20 ). In this step, cross-validation is performed using test data, which is different from the training data used to build the model, among process data collected during past operations of the plant 3. This step also verifies whether an anomaly is appropriately detected and whether an alarm or an action to address the anomaly is output, using process data from when an anomaly occurred in the past. The learning processing unit 143 also determines whether an anomaly can be detected with sufficient accuracy ( S4 in FIG. 20 ). If it is determined that the accuracy is insufficient ( S4: NO), the threshold value registered in the knowledge base (in other words, the normal range of the process data) is modified so that an anomaly can be appropriately detected, and the processing from S31 onward is repeated. If it is determined in S4 that an anomaly can be detected with sufficient accuracy ( S4: YES), operation is performed using the anomaly detection model and threshold value created in S2. Note that at least a portion of the determination in S4 may be made by the user.

[0076] It should be noted that, for example, actions to be taken by an operator of the plant 3 to deal with an assumed cause are associated with the assumed cause and are stored in advance in the storage device 12. FIG. 21 is a diagram showing an example of an action table. The table in FIG. 21 includes attributes of cause, action 1, and action 2. In the cause field, a cause corresponding to an assumed cause in the knowledge base is registered. In the action 1 and action 2 fields, information indicating measures to be taken by an operator of the plant 3 to resolve the corresponding cause is registered.

[0077] <Abnormality Detection Processing> FIG. 22 is a process flow diagram showing an example of the abnormality detection processing executed by the abnormality modulation cause identification device 1. The processor 14 of the abnormality modulation cause identification device 1 executes a predetermined program to perform the processing shown in FIG. 22 . The abnormality detection processing is performed almost in real time using process data or image data obtained by operation of the plant 3. The abnormality detection processing mainly includes preprocessing (FIG. 22: S10), model read processing (S20), and abnormality determination processing (S30). In FIG. 22 , steps corresponding to those in the learning processing shown in FIG. 20 are denoted by the same reference numerals, and the following description will focus on differences from the learning processing. For convenience, the processing will be described as being performed by the same device as the device that performs the learning processing; however, the device that performs the abnormality detection processing may be different from the device that performs the learning processing. In addition, it is assumed that tables such as the anomaly detection model, thresholds, and knowledge base created in the learning processing are stored in advance in the storage device 12.

[0078] The data acquisition unit 141 of the abnormal modulation cause identification device 1 acquires process data or image data ( S11 in FIG. 22 ). The process data is assumed to be stored in the storage device 12 in a file format such as OPC data, a so-called database table, or CSV. The image data is, for example, raster data output at a frame rate of 30 fps. This step is similar to S11 in FIG. 20 , except that data related to processes currently in operation in the plant 3 is acquired. Furthermore, the preprocessing unit 142 of the abnormal modulation cause identification device 1 associates the process data of the continuous processing with the serial number ( S12 in FIG. 22 ). This step is similar to S12 in FIG. 20 . The preprocessing unit 142 then extracts and processes data to be used in the abnormality determination model ( S13 in FIG. 22 ). This step is similar to S13 in FIG. 20 , except that data cleansing is not required.

[0079] Thereafter, the abnormality determination unit 144 of the abnormal modulation cause identification device 1 reads out the abnormality detection model created in the learning process from the storage device 12 (S20 in FIG. 22). The abnormality determination unit 144 also calculates the degree of abnormality using the created abnormality detection model and process data and image data obtained by operation of the plant 3 (S31 in FIG. 22). This step is the same as S31 in FIG. 20. The cause diagnosis unit 145 of the abnormal modulation cause identification device 1 also uses the calculated degree of abnormality to determine the probability of the assumed cause (S32 in FIG. 22). This step is the same as S32 in FIG. 20.

[0080] The output control unit 146 also outputs the degree of abnormality calculated in S31 and the degree of establishment calculated in S32, and issues an alarm if either degree of abnormality exceeds a predetermined threshold ( FIG. 22 : S303). In this step, the process data indicating the operating state of the plant 3, the degree of abnormality, and the degree of establishment of the assumed cause are presented to the user via the input / output device 13.

[0081] FIG. 23 is a diagram showing an example of a screen output to the input / output device 13. FIG. 23 is an example of a main control chart, showing the transition of individual process data using a line graph. An area 131 displayed on the input / output device 13 displays multiple combinations of identification information and the latest values ​​of process data acquired from the plant 3. The control chart in area 132 shows the transition of values ​​of specific process data using a line graph. The vertical axis represents the value of the process data, and the horizontal axis represents the time axis. In the example of FIG. 23 , the solid line represents the true value, and the dashed line represents the estimated value. The true value is the process data itself for which the degree of anomaly is to be calculated, and the estimated value may be an estimated value obtained by regression analysis of the process data for which the degree of anomaly is to be calculated. The thin dashed line represents the upper and lower limits of the normal range (in other words, a threshold value for anomaly detection). As shown by a balloon in FIG. 23 , when a user operates the input / output device 13, such as a pointing device, to move a pointer on the graph, the numerical value of the process data at the time indicated by the pointer may be displayed. The cause-and-effect diagram in area 133 displays causes of the process data modulation displayed in area 132 or tags that can identify them on the horizontal axis, and the vertical axis shows the degree of validity of each cause as a bar graph. A higher degree of validity indicates a higher possibility of the cause being the process data modulation. The cause-and-effect diagram is calculated by the cause diagnosis unit 145 based on the degree of abnormality calculated by the anomaly determination unit 144 for events that are assumed to be the cause of the process data modulation. Based on the degree of validity, the user can recognize potential causes of the modulation and their likelihood, thereby easily identifying the cause of the modulation. When the "Diagnose" button in area 134 is pressed, the anomaly determination unit 144 calculates the degree of abnormality at a specified time or the current time, and the cause-and-effect diagram is displayed by the output control unit 146. When the user operates the input / output device 13, such as a pointing device, to select one of the bar graphs in the cause-and-effect diagram, the cause of the modulation corresponding to the bar graph is highlighted in the logic tree.

[0082] FIG. 24 is a diagram showing another example of a screen output to the input / output device 13 by the output control unit 146. FIG. 24 is an example of a tree diagram, displaying a logic tree similar to that shown in FIG. 11. For example, when the bar graph for tag 004 is selected in FIG. 23, the effects corresponding to the process data for tag 004 are highlighted in the logic tree. The highlighting is achieved by changing the display mode, for example, by changing the color or line type. In FIG. 24, the corresponding rectangles are hatched. The thick rectangles connected upstream in the logic tree represent the assumed causes of the effects. In FIG. 24, each cause may be displayed as a speech bubble, or the effects on the process data other than the causes may be displayed. The probability of each cause calculated in S32 of FIG. 22 may also be displayed, or actions may also be displayed. The causes may also be displayed when the user moves the pointer over each rectangle.

[0083] 23 and 24, the process trend of each tag listed in the cause-and-effect diagram may be displayed, and in particular, the process trend of a tag for which the cause of modulation can be identified may be displayed. The process trend is calculated using the process data stored in the storage device 12, and values ​​for each period, such as for each predetermined hour, each predetermined number of days, each predetermined number of months, or each season, are plotted on a graph.

[0084] The output control unit 146 may also be configured to output a log of the degree of abnormality when, for example, the degree of abnormality calculated by each calculation method exceeds a predetermined threshold. Also, the output control unit 146 may be configured to output a log of the assumed cause and the degree of occurrence. Each log may be output in association with the date and time, serial number, calculation method, anomaly detection model, etc., thereby facilitating the analysis of abnormal modulation.

[0085] <Modifications> The configurations and combinations thereof in each embodiment are merely examples, and additions, omissions, substitutions, and other modifications of the configurations are possible as appropriate without departing from the spirit of the present disclosure. The present disclosure is not limited by the embodiments, but only by the scope of the claims. Furthermore, each aspect disclosed in this specification can be combined with any other feature disclosed in this specification.

[0086] Although the above-described success rate may be determined using only one or more image abnormality levels, the accuracy of abnormality determination can be improved by combining the image abnormality level and the sensor value abnormality level, as in the above-described embodiment. Furthermore, if abnormal variations in the amount of product, particularly in the retention box, can be detected early, operating conditions that result in optimal bulk density can be presented. For example, the size of the resin can be adjusted by the flow rate of the acid used, thereby preventing clogging of the device and deterioration of quality.

[0087] The products manufactured by the plant may be not only high molecular weight compounds such as resins, but also low molecular weight compounds. For example, the degree of abnormality can be calculated for granular compounds. Furthermore, although the above-described embodiment has been described using a chemical plant as an example, the present invention can be applied to manufacturing processes in general production facilities. For example, instead of the manufacturing number of the batch process in the embodiment, a lot number may be used as the processing unit, and processing similar to the batch process in the embodiment may be applied. Furthermore, the degree of image abnormality may be applied not only to the quantity or size of the product, but also to analysis of cell shape or fiber structure, analysis of wrinkles or sagging, detection of scratches on metal or concrete, analysis of crystal shape, detection of impurities, etc.

[0088] At least some of the functions of the abnormal modulation cause identification device 1 may be distributed among multiple devices, or multiple devices may provide the same functions in parallel. Also, at least some of the functions of the abnormal modulation cause identification device 1 may be provided on a so-called cloud.

[0089] The present disclosure also includes a method and a computer program for executing the above-described process, and a computer-readable recording medium having the program recorded thereon. The recording medium having the program recorded thereon enables the above-described process by causing a computer to execute the program.

[0090] Here, a computer-readable recording medium refers to a recording medium that stores information such as data and programs electrically, magnetically, optically, mechanically, or chemically and can be read by a computer. Among such recording media, those that can be removed from a computer include flexible disks, magneto-optical disks, optical disks, magnetic tapes, memory cards, etc. Furthermore, recording media that are fixed to a computer include HDDs, SSDs (Solid State Drives), ROMs, etc.

[0091] 1: Abnormal modulation cause identification device (abnormality detection device) 11: Communication I / F 12: Storage device 13: Input / output device 14: Processor 141: Data acquisition unit 142: Preprocessing unit 143: Learning processing unit 144: Abnormality determination unit 145: Cause diagnosis unit 146: Output control unit 2: Control station 3: Plant

Claims

1. a data acquisition unit that reads out image data and process data from a storage device that stores image data continuously output by an imaging device provided in the production equipment and process data continuously output by a sensor provided in the production equipment; an abnormality determination unit that calculates an image abnormality degree using a model that has learned the tendency of the feature values ​​of the image data in a normal state and the feature values ​​of the image data acquired by the data acquisition unit, and also calculates a sensor value abnormality degree that indicates the degree of modulation of the process data read by the data acquisition unit; a cause diagnosis unit that outputs the cause corresponding to the sensor value abnormality degree and the image abnormality degree calculated by the abnormality determination unit based on causal relationship information that defines a correspondence between a cause and a causal relationship between the sensor value abnormality degree and the image abnormality degree; An anomaly detection device comprising:

2. the production facility is a granular product manufacturing plant; The imaging device is installed in a tank of the manufacturing plant. The anomaly detection device according to claim 1 .

3. The abnormality determination unit Using the image data, a convolutional deep learning model that outputs feature amounts at multiple levels, from local features to global features, is used to output the feature amounts at multiple levels; Using a model that has learned the average of the feature amounts in the normal state for each of the plurality of levels using a plurality of image data in the normal state, a deviation degree based on the difference between the feature amounts of the plurality of levels and the average of the corresponding level is corrected based on the smoothness of the texture of the image data read by the data acquisition unit, and the image abnormality degree is calculated as the image abnormality degree. The anomaly detection device according to claim 1 or 2.

4. The convolutional deep learning model is EfficientNet, The deviation is calculated by Gaussian AD, The corrected anomaly degree is calculated based on the deviation degree using a GLCM (Gray-Level Co-Occurrence Matrix). It is calculated by multiplying the value based on The anomaly detection device according to claim 3 .

5. The model is created using a pseudo-abnormal image created by smoothing the image data of a normal state. The anomaly detection device according to claim 1 or 2.

6. The model is generated by learning features of an area where the product is imaged on an agitating blade provided in the tank, and the area where the product is imaged is estimated from the image data by segmentation; The abnormality determination unit calculates a maximum value of the size of the area in which the product is imaged for each rotation cycle of the agitating blade, and determines that the smaller the maximum value, the higher the degree of image abnormality. The anomaly detection device according to claim 2 .

7. The abnormality determination unit calculates the degree of abnormality of the sensor value depending on whether the direction of deviation from a predetermined standard is positive or negative. The anomaly detection device according to claim 1 , 2 , or 6 .

8. the causal relationship information defines the process data used to calculate the sensor value abnormality degree by a timing, a period, or an interval in a process performed by the production equipment for each of the process data output by the sensor; The abnormality determination unit calculates the sensor value abnormality degree using a value extracted from the process data read by the data acquisition unit based on the timing, period, or interval defined by the causal relationship information. The anomaly detection device according to claim 1 , 2 , or 6 .

9. reading out the image data from a storage device that stores image data continuously output by an imaging device provided in the production facility; reading the process data from a storage device that stores the process data continuously output by a sensor provided in the production equipment; calculating an image abnormality degree using a model that has learned the tendency of the feature amount of the image data in a normal state and the feature amount of the read image data, and also calculating a sensor value abnormality degree that indicates the degree of modulation of the read process data; Based on causal relationship information that defines the association between a cause and a combination of the sensor value abnormality degree and the image abnormality degree, the cause corresponding to the calculated sensor value abnormality degree and the image abnormality degree is output. An anomaly detection method in which processing is performed by a computer.

10. reading out the image data from a storage device that stores image data continuously output by an imaging device provided in the production facility; reading the process data from a storage device that stores the process data continuously output by a sensor provided in the production equipment; calculating an image abnormality degree using a model that has learned the tendency of the feature amount of the image data in a normal state and the feature amount of the read image data, and also calculating a sensor value abnormality degree that indicates the degree of modulation of the read process data; outputting the causes corresponding to the calculated sensor value abnormality degrees and image abnormality degrees based on causal relationship information that defines the association between the causes and the combinations of the sensor value abnormality degrees and image abnormality degrees; An anomaly detection program that causes a computer to execute a process.