Procedure for diagnosing anomalies, facility for diagnosing anomalies, program for diagnosing anomalies, and system for diagnosing anomalies

Abstract

Method for diagnosing anomalies to diagnose whether an anomaly exists in a plant to be diagnosed, wherein the method for diagnosing anomalies comprises: obtaining from the plant data that exhibits a state variable of a single evaluation element or a plurality of evaluation elements; classifying the data obtained from the plant for each operating state of the plant (S610); evaluating the sufficiency of the number of data units for each classified data group (S320); calculating a plurality of parameter values ​​that configure a trained model according to the sufficiency of the number of data units and storing the parameter values ​​in association with the operating state (S340);Obtaining the multitude of parameter values ​​that configure the trained model, without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage state of the parameter values ​​associated with each operating state (S630); creating a trained model for diagnosing data to be diagnosed using the obtained parameter values ​​(S640); calculating an anomaly grade to determine with respect to an anomaly based on the trained model (S650); and determining whether an anomaly is present in a diagnostic target based on the anomaly grade (S670);where obtaining the multitude of parameter values ​​that configure the trained model, without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage state of the parameter values ​​associated with each operating state, comprises: obtaining all parameter values ​​that configure the trained model, based on the fact that all parameter values ​​are stored in an operating state identical to the operating state of data to be diagnosed; and obtaining, based on the fact that only some of the parameter values ​​that configure the trained model are stored in the operating state identical to the operating state of data to be diagnosed, only some of the parameter values, and obtaining the remaining parameter values ​​from parameter values ​​of an identical type stored in another operating state.

Description

TECHNICAL AREA

[0001] The present disclosure relates to a method for diagnosing anomalies, a facility for diagnosing anomalies, and a program for diagnosing anomalies. STATE OF THE ART

[0002] Monitoring the condition and diagnosing anomalies in a plant or device are generally performed based on data representing the state variable of a single or multiple evaluation element obtained from a diagnostic target. This data, representing the state variable of the evaluation element(s), includes various types of operational data from the plant or device being diagnosed, as well as measurement data on temperature, vibration, and other parameters generated by the diagnostic target and measured by various sensors. This operational and measurement data is time-series data obtained through continuous observation of changes over time.

[0003] In one approach to anomaly diagnosis, a trained model is created using time-series data (normal data) obtained during normal operation as training data. Based on the data to be diagnosed and the trained model, a determination is made as to whether or not an anomaly exists in the plant or device. In such anomaly diagnosis, for example, a change in time-series data that is not due to an anomaly, such as a change after maintenance or a change due to a change in operating mode, will incorrectly identify such a normal change in condition as an anomaly if the plant or device is diagnosed based on the trained model, which was created using training data obtained before the maintenance or change in operating mode.Therefore, a procedure for diagnosing the plant or device without false identification when such a normal change in operating condition occurs was required.

[0004] Japanese patent application no. 2017-102826 (PTL 1) discloses a technique "to enable highly accurate diagnosis of anomalies according to a status change of a mechanical system" (see "Subject Matter" in "Abstract"). According to this technique, "when diagnosing anomalies using a pattern recognition approach for diagnostic target data, training data to be used for training information is extracted for each diagnostic target data unit based on the diagnostic target data and extraction condition information to determine an extraction condition stored by a training information creator 11, and then the training information is created according to the training data and an anomaly diagnosis is performed" (see "Means of Solving the Problem" in "Abstract").

[0005] Japanese Publication No. 2014-32657 (PTL 2) discloses "a method for monitoring a plant condition with a method for highly sensitive and rapid anomaly detection and a device for this purpose" (see paragraph 0009). According to this method, a procedure for "storing a cluster center and data belonging to a cluster by pre-clustering training data, selecting data near new observation data from the training data belonging to the cluster near the new observation data, generating a normal model from these selected data, and calculating an anomalous measure" is disclosed (see "Abstract").

[0006] In Japanese Publication No. 2015-181072 (PTL 3), a technique is disclosed “in a method for monitoring the condition of a plant based on a time series signal emitted by the plant, to provide an operating pattern identifier for each specified period based on the time series signal, to select training data for each specified period based on the operating pattern identifier, to create a normal model based on these selected training data, to calculate an abnormal measure based on the time series signal and the normal model, and to identify, based on this calculated abnormal measure, whether the condition of the plant is abnormal or normal” (see “Summary”).It further describes a procedure based on a macro-characteristic value (an average, a variance, a maximum value, a minimum value or the like) as a method for searching for an operational trampling indicator that closely approximates the state. LIST OF CAPLETS PATENT LITERATURE PTL 1: Japanese Disclosure Document No. 2017-102826 PTL 2: Japanese Disclosure Document No. 2014-32657 PTL 3: Japanese Disclosure Document No. 2015-181072 BRIEF DESCRIPTION OF THE INVENTIONAL PROBLEM

[0007] Some systems to be diagnosed operate only in certain predetermined operating states, while others operate based on an optimal operating state determined each time by a peripheral device or environment. For the latter, it is difficult to define in advance the number of operating states the system can be in, or even any specific state within that state.In an example where a diagnosis is performed using a trained model created from training data determined based on a condition (an extraction condition or a cluster) for classifying data to be diagnosed for each operating state, as in the technique disclosed in PTL 1 or PTL 2, if diagnostic target data that do not satisfy the pre-generated classification condition are input, the training data may be selected inappropriately, and the determination of an anomaly may be incorrect. To perform a proper diagnosis of such diagnostic target data, training data should be re-collected in the same operating state as the diagnostic target data; however, the diagnosis may not be possible until the data has been collected.

[0008] To address such a problem, according to the technique disclosed in PTL 3, the macro characteristic value is used as the method for searching for an operating tread pattern that closely approximates the operating state, operating state data with a similar macro characteristic value are selected as training data, a threshold for individual determination of an anomaly is set for the selected training data, and the plant is diagnosed.

[0009] According to this technique, even if the state of the system changes, training data with a similar state can be selected to create the trained model, enabling diagnosis. However, diagnosing the data requires a series of processes, including selecting training data, creating a trained model, and determining a threshold for anomaly detection, thus increasing the computational scope of these processes. The anomaly detection threshold should be appropriately determined based on the similarity between the training data and the diagnostic target data. However, PTL 3 is silent on a specific procedure for determining anomaly detection threshold for diagnosis using the trained model created from the training data determined based on the macro-feature value.If the threshold for determining an anomaly is set incorrectly, the anomaly determination may be incorrect.

[0010] A technique is needed to diagnose whether an anomaly exists in a diagnostic target system without incorrectly identifying a change in time-series data as an anomaly when the system's state changes, even if a change in time-series data is present when the system's state is not caused by an anomaly, such as a change following maintenance or a change due to a change in operating mode. Additionally, a technique is needed to reduce the computational scope required for these processes.

[0011] The present disclosure, made in light of the background described above, discloses a technique for improving the accuracy of anomaly detection such that the anomaly is essentially detected without falsely identifying a change in time-series data, even when a change in time-series data is not caused by an anomaly, such as a change following maintenance or a change due to a change in operating mode. The present disclosure further discloses a technique for reducing the time required to collect training data necessary for diagnosis and for performing a diagnosis without false identification, even when a change in the operating state of the plant is not predictable in advance. Furthermore, the present disclosure discloses a technique for reducing the computational scope required for these processes. SOLUTION TO THE PROBLEM

[0012] According to one form of the present disclosure, a method for diagnosing anomalies is provided for diagnosing whether an anomaly exists in a plant under diagnosis. This method for diagnosing anomalies includes obtaining from the plant data that exhibits a state variable of a single evaluation element or a plurality of evaluation elements; classifying the data obtained from the plant for each operating state of the plant; evaluating the sufficiency of the number of data units for each classified data group; calculating a plurality of parameter values ​​that configure a trained model according to the sufficiency of the number of data units and storing the parameter values ​​in association with the operating state; and obtaining the plurality of parameter values ​​that configure the trained model without overlaps for each parameter.Based on the operating state of the data to be diagnosed and a storage status of the parameter values ​​associated with each operating state, a trained model is created to diagnose the data to be diagnosed using the obtained parameter values. An anomaly level is then calculated to determine whether an anomaly exists based on the trained model, and whether an anomaly is present in a diagnostic target is determined based on the anomaly level.

[0013] According to another embodiment, an anomaly diagnosis device is provided to diagnose whether an anomaly exists in a system under diagnosis. This anomaly diagnosis device includes a data reader for obtaining data from the system that exhibits a state variable of a single evaluation element or a plurality of evaluation elements; a data classifier for classifying the data obtained from the system for each operating state of the system; a parameter calculator and storer for evaluating the sufficiency of the number of data units for each classified data group, for calculating a plurality of parameter values ​​that configure a trained model according to the sufficiency of the number of data units, and for storing the parameter values ​​in association with the operating state; and a parameter acquisition unit for obtaining the plurality of parameter values.that configure the trained model without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage state of the parameter values ​​associated with each operating state; a trained model creator for creating a trained model to diagnose the data to be diagnosed with the obtained parameter values; an anomaly grade calculator for calculating an anomaly grade to determine with respect to an anomaly based on the trained model; and an anomaly determination unit for determining whether an anomaly is present in a diagnostic target, based on the anomaly grade.

[0014] According to another embodiment, an anomaly diagnosis program is provided to cause a computer to perform processing to diagnose whether an anomaly exists in a plant under diagnosis. This anomaly diagnosis program causes the computer to perform the following: obtain data from the plant that includes a state variable of a single rating element or a plurality of rating elements; classify the data obtained from the plant for each operating state of the plant; assess the sufficiency of the number of data units for each classified data group; calculate a plurality of parameter values ​​that configure a trained model according to the sufficiency of the number of data units and store the parameter values ​​associated with the operating state; and obtain the plurality of parameter values ​​that configure the trained model.without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage state of the parameter values ​​associated with each operating state, creation of a trained model for diagnosing the data to be diagnosed with the obtained parameter values, calculation of an anomaly level to determine with regard to an anomaly based on the trained model and determination of whether an anomaly is present in a diagnostic target, based on the anomaly level.

[0015] According to yet another embodiment, an anomaly diagnosis system is provided to diagnose whether an anomaly exists in a system under diagnosis. This anomaly diagnosis system includes a data reader for obtaining data from the system that represents a state variable of a single evaluation element or a plurality of evaluation elements; a data classifier for classifying the data obtained from the system for each operating state of the system; a parameter calculator and storer for evaluating the sufficiency of the number of data units for each classified data group, for calculating a plurality of parameter values ​​that configure a trained model according to the sufficiency of the number of data units, and for storing the parameter values ​​in association with the operating state; and a parameter acquisition unit for obtaining the plurality of parameter values.that configure the trained model without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage state of the parameter values ​​associated with each operating state; a trained model creator for creating a trained model to diagnose the data to be diagnosed with the obtained parameter values; an anomaly grade calculator for calculating an anomaly grade to determine with respect to an anomaly based on the trained model; and an anomaly determination unit for determining whether an anomaly is present in a diagnostic target, based on the anomaly grade. ADVANTAGEOUS EFFECTS OF THE INVENTION

[0016] According to the present disclosure, a normal change in the operating state due to maintenance or a change in the operating mode is not incorrectly determined, and the accuracy of the diagnosis can be improved. Even if a change in the operating state of the system is not predictable in advance, the period for re-collecting training data necessary for the diagnosis can be reduced, and a diagnosis can be performed without incorrect determination. In addition, the computational scope required for these calculation processes can be reduced.

[0017] The foregoing and other tasks, features, aspects and advantages of this invention will become more apparent through the following detailed description of this invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS Fig. Figure 1 is a block diagram showing a functional configuration of a facility for diagnosing anomalies 100. Fig. Figure 2 is a representation that exemplifies a procedure for classifying into an operating state based on processing in a data classifier 104 to determine whether a change in time series data is present in at least one evaluation element or not. Fig. Figure 3 is a flowchart showing part of the processing carried out by a parameter calculator and storer 105. Fig. Figure 4 is a flowchart showing part of a processing operation that is carried out to obtain a parameter in a parameter acquisition unit 106. Fig. Figure 5 is a representation showing an exemplary hardware configuration of a computer system 500, which is switched on as an anomaly diagnosis device 100. Fig. Figure 6 is a flowchart showing part of the processing carried out by the computer system 500, which implements the device 100 for diagnosing anomalies according to a second embodiment. DESCRIPTION OF EXECUTION FORMS

[0018] An embodiment of the present invention is described below with reference to the drawings. In the following description, identical components are assigned the same reference numerals, and their designations and functions are also the same. Therefore, a detailed description of these is not repeated. First embodiment.

[0019] A configuration of a device for diagnosing anomalies 100 according to an embodiment is described with reference to Fig. 1 described. Fig. Figure 1 is a block diagram showing a functional configuration of the anomaly diagnostic device 100 for implementing an anomaly diagnostic method of the present invention. The anomaly diagnostic device 100 determines whether a condition of a system or device to be diagnosed (hereinafter also collectively referred to as a "system") is normal or not by analyzing time-series data obtained from the system. Additionally, the anomaly diagnostic device 100 is configured to output a result of this determination.

[0020] The plant or device to be diagnosed is, for example, a facility or plant, such as a generator, a manufacturing automation device (MAD), a power receiving and distribution device, an elevator, and an electrical device for railways. Examples of a plant include a power station, a wastewater treatment plant, and an industrial plant for petrochemicals, metals, and food and pharmaceuticals. Depending on the type of plant, it may consist of multiple plants or devices. A single plant or device can be diagnosed as described above, or the plant or device may consist of multiple plants or devices, as in the case of a plant.The following is an example configuration in which the anomaly diagnosis facility 100 is used to monitor a generator's condition and diagnose an anomaly.

[0021] As in Fig. As shown in Figure 1, the anomaly diagnosis device 100 includes a data reader 102, a data display unit 103, a data classifier 104, a parameter calculator and storer 105, a parameter acquisition unit 106, a trained model creator 107, an anomaly grade calculator 108, an anomaly determination unit 109, and a determination result output unit 110. The anomaly diagnosis device 100 accepts input data acquired from a system or device 10. [Reading data]

[0022] The data reader 102 reads time-series data obtained from the equipment under diagnosis. This time-series data is obtained by continuously observing changes in a state variable of a single evaluation element or a multitude of evaluation elements over time. The time-series data can be any data output by the equipment and data output by a sensor (not shown) provided within the equipment. The following is a description in the case of the generator. Evaluation elements include evaluation elements related to operational data, such as an output, speed, voltage, or armature current of the generator, and evaluation elements related to measurement data obtained by measuring temperature and vibration, as measured by the sensor attached to a device or component within the generator.The evaluation elements can include elements relating to the environment in which the generator is deployed, such as ambient air temperature, as well as plant operating and measurement data relating to the generator's operating state. An evaluation element can consist of a single element or a multitude of elements, and the state variable of the evaluation element represents a physical quantity obtained as time-series data from each element. These evaluation elements are examples, and the number of elements is unlimited. The time-series data can include timestamp information. When obtaining time-series data for a multitude of evaluation elements, data measurement cycles (sampling) should ideally coincide between / among the evaluation elements.If the cycles do not coincide, processing such as omitting or interpolating data for at least one evaluation element after the data has been obtained determines a form such as an assignment of state variables to the multitude of evaluation elements at a single point in time (sampling). The data reader 102 transmits the read data to the data display unit 103 and the data classifier 104. [Data display]

[0023] The data display unit 103 displays data received from the data reader 102 in a graph as time-series data. In one aspect, the data display unit 103 can display each data unit in a time series according to the time specified by the timestamp. The data display unit 103 can be implemented as a monitoring device included in a computer system or as an external monitoring device connected to the computer system. [Data classification]

[0024] The data classifier 104 classifies data received by the data reader 102 for each operating state. The operating state represents the state of a system being diagnosed. The state variable of the time-series data obtained from a diagnostic target changes when the system's operating conditions or control settings change. The operating state is defined as differing before and after this change, and data before and after the change are classified. In other words, the data classifier 104 classifies the obtained data for each similar operating state (state variable). The system's state can change, and the state variable of the obtained time-series data can change, not only when the operating conditions or control settings change, but also when the system is shut down or undergoes maintenance.These changes must also be classified. The number of operating states into which classification can be made is unlimited, and the number is determined according to the characteristics of a diagnostic target system or a change in a state variable of the data during a period in which the system, operating in a normal state, was already known. Classification into operating states is carried out such that one operating state is assigned for each measurement period of data (for each sampling period), regardless of the number of evaluation elements in the obtained data.

[0025] The data classifier 104 classifies data received by the data reader 102 for each operating state based on cluster processing or processing to determine whether or not there has been a change in data for at least one evaluation element.

[0026] Cluster processing can be a non-hierarchical clustering approach, such as k-means, or a hierarchical clustering approach. In another aspect, the data classifier 104 can learn a classification rule during processing for classification into the operational state, apply the classification rule to data newly received by the data reader 102 after learning, and determine the operational state classification to be performed. For example, a centroid of each cluster can be calculated from data belonging to each cluster and determined during cluster processing; a distance between the centroid of each cluster and classification target data can be calculated as a determination index; and the operational state classification to be performed can be determined based on a predefined threshold.In another example, data classified into each cluster and its operational state can be linked together as training data, and the training data can be applied to machine learning, such as a Support Vector Machine, Random Forest and k-Nearest Neighbor, and deep learning, such as a neural network, to create a classifier that can be used to classify into the operational state.

[0027] A procedure for classifying data into its operational state based on processing to determine whether or not there has been a change in time series data for at least one evaluation element is now described.

[0028] The Data Classifier 104 can select any evaluation element and determine the operating state into which data is to be classified based on whether or not a predefined threshold for the state variable is exceeded. If the threshold is exceeded, it is determined that the time series data has changed, and the operating state of data exceeding the threshold is distinguished and classified. The number or value of the thresholds used for this determination can be freely defined.

[0029] Furthermore, another example is given with reference to Fig. 2 described. Fig. Figure 2 is an exemplary representation of a procedure for classifying time series data into an operational state based on processing to determine whether or not a change has occurred in at least one evaluation element. Data classifier 104 splits the time series data for the selected evaluation element at arbitrary time intervals and calculates a statistical measure, such as a mean, variance, standard deviation, maximum, minimum, kurtosis, and skewness for each data group resulting from the split. At least one or a combination of these statistical measures is evaluated based on a predetermined threshold. If the threshold is exceeded, it is determined that the time series data has changed, and the operational state into which data in a split range are to be classified is determined.As in . Fig. As shown in Figure 2, the time series data (obtained data) for any evaluation element are split at any time intervals, and a mean and variance are calculated based on the split data set and arranged chronologically. Thresholds (m1 to m2 and v1 to m2) are determined in advance for the mean and variance, and the operational status of data in the split time range is determined based on these thresholds. Fig. Figure 2 shows an example of a classification into three operating states, from first to third, based on these threshold values. If the average is less than m1 and the variance within a range is not less than v1 and not less than v2, it is classified as the first operating state; if the average is not less than m2 and the variance is not less than v2, it is classified as the second operating state; and if the average within a range is not less than m1 and not less than m2, and the variance within a range is not less than v1 and not less than v2, it is classified as the third operating state.

[0030] At least one evaluation element for calculating the statistical measurement can be defined, and an evaluation can be performed using a variety of evaluation elements. This method allows for the classification of a change in an operating state while considering variations in data, which is difficult when classifying based solely on the threshold value of the state variable.

[0031] When processing to classify data based on cluster processing or processing to determine whether or not there has been a change in data relating to at least one evaluation element, the number of operating states into which the classification is to be performed and various thresholds used as criteria can be freely determined or updated after they have been set.

[0032] If it is clear that data of interest has changed due to maintenance of the plant or a change in a condition controlling the operation of the plant, the data classifier 104 can classify it into the operating state at the time of the maintenance of the plant or a change in a condition controlling the operation. [Calculation and saving of parameter values]

[0033] With renewed reference to Fig. 1. The parameter calculator and storer 105 calculates and stores a value of a parameter necessary for creating a trained model for each data set classified by the data classifier 104. According to one embodiment, the trained model refers to a mathematical expression or data processing operation for calculating an analysis output value (an anomaly grade) for anomaly detection. The parameter represents part of a constant that configures the trained model (the mathematical expression or data processing operation) and refers to a value that changes according to the training data (the data used for learning).

[0034] The following description provides an example where a Mahalanobis-Taguchi (MT) method (a unit space as defined therein) is used as the trained model. The unit space represents a normal data set that serves as a reference for diagnosis. When the Mahalanobis-Taguchi (MT) method (the unit space as defined therein) is used as the trained model, a mean, a standard deviation, and an inverse matrix of a correlation matrix of data defining the unit space are included as necessary parameters for constructing the trained model.

[0035] The trained model is not limited to the MT method in the present disclosure, and it is also applicable to a method for using a statistical model, a known anomaly diagnosis model used in the field of machine learning or deep learning, and a data analysis approach.

[0036] Only processing for calculating and storing a parameter value with reference to Fig. 3 described. Fig. Figure 3 is a flowchart showing part of the processing performed by a parameter calculator and storer 105. Subsequent processing is implemented by executing an instruction to carry out the processing by a CPU with a sufficiently known configuration or by another processor. Alternatively, such processing can also be implemented by a circuit element configured to perform the processing.

[0037] In step S310, the parameter calculator and storer 105 receives an input of data that has been classified by the data classifier 104.

[0038] In step S320, the parameter calculator and storer 105 performs a sufficiency assessment of the number of data units for each classified data group. The parameter calculator and storer 105 determines in advance, as a sufficiency assessment, the number of data units necessary to calculate a parameter value and allow a diagnosis without incorrect determination, for example, based on past diagnostic records and the creation of a trained model, and evaluates the sufficiency based on the number of data units. In other words, in the present embodiment, the sufficiency assessment determines whether the number of data units in the data group classified into each operating state is equal to or greater than the number that allows a diagnosis without incorrect determination.In addition to this method, the parameter calculator and storer 105 can perform an assessment of sufficiency based on a correlation coefficient between / among a multitude of evaluation elements or on the degree of change of a basic statistical measure (an average, a variance, a maximum value, a minimum value, a kurtosis, or a skewness) of any evaluation element. Specifically, the parameter calculator and storer 105 first divides a data group classified into a specific operating state into arbitrary time ranges (any number of data units) and calculates the correlation coefficient or the basic statistical measure of a data group in the smallest divided range, which is the oldest in chronological order.The parameter calculator and storer 105 then calculates the correlation coefficient or basic statistical measure based on the data group obtained by combining the second oldest data group in chronological order with the oldest data group. The parameter calculator and storer 105 further calculates the correlation coefficient or basic statistical measure based on a data group obtained by combining the third oldest data group in chronological order with the second oldest data group and the oldest data group. The parameter calculator and storer 105 thus incrementally adds a data group that is newer in chronological order to calculate the correlation coefficient or basic statistical measure until the calculation is complete for all subdivided data groups.The parameter calculator and storer 105 calculates the magnitude of change of the correlation coefficient or the basic statistical measure, which is calculated stepwise, and determines the number of data units to be used for the calculation when the magnitude of change is less than the predetermined threshold. This number is sufficient for calculating the parameter for diagnostic purposes without incorrect determination. Using this method, even when data is added for training, the parameter calculator and storer 105 can determine the minimum number of data units beyond which the correlation coefficient or the basic statistical measure will no longer change, and determine the sufficiency rating as OK when the magnitude of change is less than the predetermined threshold.Alternatively, the parameter calculator and storer 105 can calculate the correlation coefficient or basic statistical measure for each of data groups split at any time interval (the number of data units), evaluate a variation thereof, and take the number of split data units whose variation is not greater than the threshold as the number of data units determined to be OK in the sufficiency assessment. If there are multiple parameters configuring the trained model, the parameter calculator and storer 105 can individually select and set an evaluation procedure for the sufficiency assessment of the number of data units for each parameter and a threshold thereof from the above.

[0039] In step S330, the parameter calculator and storer 105 determines whether a sufficient number of data units are present. If the parameter calculator and storer 105 determines that a sufficient number of data units are present, that is, if the result of a sufficiency assessment of the number of data units indicates OK (YES in step S330), it switches the control to step S340. Otherwise, that is, if the result of the sufficiency assessment indicates NG (NO in step S330), the parameter calculator and storer 105 switches the control to step S350.

[0040] In step S340, the parameter calculator and storer 105 calculates a value of a parameter (referred to as the parameter value in the description below) based on data (the classified data group) that was entered in step S310.

[0041] In step S350, the parameter calculator and storer 105 does not calculate the parameter value.

[0042] In step S360, the parameter calculator and storer 105 saves the calculated parameter value (step S340) in conjunction with the operating state. Conversely, if the parameter calculator and storer 105 does not calculate the parameter value (step S350), it saves the absence of the parameter value in conjunction with the operating state.

[0043] A similar processing procedure is performed for all parameters that configure the trained model. The parameter calculator and storer 105 performs a similar processing procedure for each operating state classification performed in the data classifier 104, calculates the parameter value that configures the trained model for all operating states, and stores the value or absence of the value according to a result of the sufficiency evaluation.

[0044] The processing to calculate and store the parameter can be seen as part of a step in the "learning" process to create the trained model for diagnosing anomalies. [Obtaining parameter values]

[0045] The parameter acquisition unit 106 obtains, without overlap for each parameter, a multitude of parameter values ​​that configure the trained model, according to the operating state of the data to be diagnosed and a storage status of the parameter values ​​associated with each operating state. In the present embodiment, the storage status of the parameter values ​​associated with each operating state refers to a result of processing in the parameter computer and storer 105 and represents the linking of information about the parameter value that configures the trained model, or the absence of the parameter value for each operating state and one state thereof.

[0046] Obtaining the parameter refers to part of the "learning" step to create the trained model for anomaly diagnosis and can be viewed as a configuration for creating the trained model that is suitable for diagnosing data without false determination.

[0047] A process for obtaining a parameter value is now described in the parameter acquisition unit 106.

[0048] First, with reference to Fig. 1. A process is described up to which diagnostic data reaches the parameter acquisition unit 106. Data to be diagnosed is read by the data reader 102 and then subjected to classification processing in the data classifier 104. This processing can be carried out based on the classification rule described in the data classifier's description. Subsequently, processing is performed in the parameter acquisition unit 106.

[0049] The processing to obtain the parameter value in the parameter acquisition unit 106 is now carried out with reference to Fig. 4 described. Fig. Figure 4 is a flowchart showing part of the processing carried out to obtain a parameter in the parameter acquisition unit 106. Subsequent processing is carried out by the CPU with a sufficiently known configuration or by another processor, through the execution of an instruction to implement the processing. Alternatively, such processing can also be implemented by a circuit element configured to carry out the processing.

[0050] In step S410, the parameter acquisition unit 106 refers to the storage status of the parameter, which is linked to the operating state that corresponds to the operating state of data to be diagnosed, the operating state being determined in the data classifier 104.

[0051] In step S420, the parameter acquisition unit 106 determines whether all parameter values ​​that configure the trained model can be obtained based on the referenced operating state. In other words, the parameter acquisition unit determines whether all parameter values ​​in the operating state are stored as the presence of a value in the parameter computer and storer 105. If the parameter acquisition unit 106 determines that all parameter values ​​can be obtained (are stored as the presence of a value) (YES in step S420), it switches to step S430. Otherwise (NO in step S420), the parameter acquisition unit 106 switches to step S440.

[0052] In step S430, the parameter acquisition unit 106 obtains all parameter values ​​stored in the operating state.

[0053] In step S440, the parameter acquisition unit 106 determines whether certain parameter values ​​that configure the trained model can be obtained based on the referenced operating state. In other words, the parameter acquisition unit determines whether certain parameters are stored as having a value in the parameter computer and storer 105 in the operating state. If the parameter acquisition unit 106 determines that some of the parameter values ​​that configure the trained model can be obtained (are stored as having a value) (YES in step S440), it switches control to step S450. Otherwise (NO in step S440), the parameter acquisition unit 106 switches control to step S460.

[0054] In step S450, the parameter acquisition unit 106 obtains only some of the parameter values ​​that can be obtained, and obtains remaining parameter values ​​based on parameter values ​​of the same type that are stored in another operating state.

[0055] In step S460, the parameter acquisition unit 106 does not obtain any of the parameter values.

[0056] The processing operations performed by the data classifier 104, the parameter calculator and storer 105, and the parameter acquisition unit 106 are described below with reference to specific examples. If data from any known period, during which the normal operation of the plant has been confirmed in advance, is used as training data, the data is applied to a processing unit for classifying data and calculating and storing parameters. The following description assumes that the operating state of the plant is classified into data groups with three operating states A, B, and C as a result of the processing in the data classifier 104.Additionally, an example is considered in which unit space information (the mean, the standard deviation and the inverse matrix of the correlation matrix) is generated as the trained model for performing anomaly diagnosis using the MT method.

[0057] In such a case, the parameter calculator and storer 105 performs a data sufficiency assessment for the calculation of each parameter for each of the operating states A, B and C.

[0058] For example, if in operating state A the result of the sufficiency assessment of all parameters indicates OK, the parameter calculator and storer 105 consequently calculates in operating state A an average a, a standard deviation a and an inverse matrix a of the correlation matrix with the training data classified in operating state A, and stores each calculated value in association with operating state A.

[0059] If, on the other hand, in operating state B the result of the sufficiency assessment of data only specifies the inverse matrix of the correlation matrix NG, the parameter calculator and storer 105 calculates only an average b and a standard deviation b based on the training data classified in operating state B and stores these calculated values ​​in association with operating state B. The inverse matrix of the correlation matrix is ​​stored in association with operating state B as the absence of the value.

[0060] If, in operating state C, the result of the sufficiency assessment of data of all parameters indicates NG, the parameter calculator and storer 105 does not calculate any of the parameters from the training data classified in operating state C and stores information indicating the absence of a value of the parameters in association with operating state C.

[0061] An exemplary processing procedure in the parameter acquisition unit 106 is now described. For illustrative purposes, an example is described in which data to be diagnosed is entered into the data reader 102 and classified by the data classifier 104, and subsequently the operating state of the data is determined as operating state A. The parameter acquisition unit 106 obtains a parameter value (the mean a, the standard deviation a, and the inverse matrix a of the correlation matrix) which is stored in operating state A, which is the operating state corresponding to the diagnostic target data.

[0062] An example is now described in which the operating state of the diagnostic target data is defined as operating state B. The parameter acquisition unit 106 obtains the parameter value stored in operating state B, which corresponds to the diagnostic target data. Since only the mean b and the standard deviation b are stored as the presence of a value in operating state B, the parameter acquisition unit 106 obtains these values. The value of the inverse matrix of the correlation matrix, however, is not stored in operating state B, and the parameter acquisition unit 106 is unable to obtain this value. The parameter acquisition unit 106 then obtains the parameter value that has not yet been obtained by using the parameter value stored in another operating state (for example, operating state A).Therefore, in operating state B, the parameter acquisition unit 106 obtains the average b, the standard deviation b, and the inverse matrix a of the correlation matrix.

[0063] When parameter acquisition unit 106 acquires a parameter stored in a different operating state, it can freely choose which operating state to select. However, it is desirable that the parameter value acquired from a different operating state be similar across operating states. For example, if the inverse matrix of the correlation matrix is ​​acquired from a different operating state, parameter acquisition unit 106 acquires the parameter value stored in the operating state whose correlation between evaluation elements appears to be high. A similar approach can also be applied to a parameter of a different type. Depending on the type of trained model or parameter, a process such as pre-setting a prohibition on redirecting the parameter value in a different operating state can be applied.

[0064] Thus, even if some of the parameter values ​​stored in the operating state corresponding to the diagnostic target data are not saved, the time required to re-collect training data in the operating state corresponding to the diagnostic target data (in order to calculate the parameter) can be reduced by obtaining the calculated and saved parameter value using training data in a different operating state. The parameter whose value is not saved (whose sufficiency rating is NG) requires a larger number of training data units necessary for an accurate diagnosis than another parameter whose sufficiency rating is OK, and the collection of this training data takes some time.Therefore, the period for collecting data to obtain this parameter value is not necessary, and accordingly, the period for collecting data before the start of the diagnosis can be reduced.

[0065] Finally, an example is considered where the operating state of the diagnostic target data is defined as operating state C. Although the parameter acquisition unit 106 obtains the parameter value stored in operating state C, which corresponds to the diagnostic target data, there is no value for the parameter in operating state C, and thus the parameter acquisition unit 106 does not obtain any of the parameter values. [Creation of a trained model]

[0066] The trained model creator 107 creates the trained model using the parameter value obtained by the parameter acquisition unit 106. As described above, the MT procedure or any other approach to diagnosing anomalies can be used as the trained model. In the case of the MT procedure, the obtained parameter value (the mean, standard deviation, and inverse matrix of the correlation matrix) is used to define the unit space. If the parameter acquisition unit 106 does not obtain any of the parameter values ​​associated with the operating state corresponding to the operating state of the data to be diagnosed, the trained model creator 107 does not create the trained model but temporarily pauses the diagnosis.Pausing a diagnosis means that subsequent anomaly grade calculation processing in the anomaly grade calculator 108 and anomaly determination processing in the anomaly determination unit 109 are temporarily suspended. If sufficient training data necessary for the diagnosis has not been collected (adequacy assessment is NG), the anomaly diagnosis unit 100 postpones the diagnosis to prevent an incorrect determination. Furthermore, as with the conventional technique, processing to select training data necessary for creating the trained model is not required for each diagnostic target data unit, and the parameter value and the trained model are automatically determined when the operational state of the diagnostic target data is established. Therefore, computational capacity can be reduced. [Calculation of the degree of anomaly]

[0067] The anomaly grade calculator 108 calculates the anomaly grade to determine whether an anomaly of the facility is present, based on data to be diagnosed and the trained model created by the trained model creator 107. In an example according to one aspect, the anomaly grade calculator 108 calculates a Mahalanobis distance as the anomaly grade in the MT method example, based on the data to be diagnosed and the unit space, which is the created trained model.

[0068] The anomaly detection unit 109 determines, on the basis of the degree of anomaly calculated by the degree of anomaly calculator 108 and a threshold value that is predetermined to determine whether an anomaly is present or not, whether an anomaly is present in the system or device to be diagnosed or not.

[0069] For example, the anomaly detection unit 109 predetermines a threshold for the degree of anomaly. If the degree of anomaly is less than the threshold, the anomaly detection unit considers a determination normal, and if the degree of anomaly is equal to or greater than the threshold, it considers a determination abnormal. The parameter calculator and storer 105 is configured to perform a sufficiency assessment of the number of data units. If the parameter acquisition unit 106 can acquire at least one parameter that configures the trained model, and a diagnosis is performed, the threshold for determining that the degree of anomaly is abnormal is therefore set to a uniform value.Specifically, the fact that the number of data units is sufficient for the sufficiency assessment in parameter calculator and storer 105 (the parameter values ​​are stored) means that the parameter values ​​necessary for creating the trained model, which performs the diagnosis with a certain accuracy, can be calculated. Even if the training data continues to increase, the parameter value and the accuracy of the diagnosis will not change. Therefore, in a diagnosis where the trained model is applied, which was created using the parameter value calculated from the number of data units that was rated as OK in the sufficiency assessment, the threshold for determining normality can be set to a constant value.In other words, processing to determine an anomaly detection threshold for each diagnostic target data unit is not necessary, and diagnostic target data can be determined using a predetermined anomaly detection threshold that corresponds to the operating state.

[0070] In conventional techniques, a diagnosis without false positives requires processing to individually determine the threshold for anomaly detection according to the diagnostic target data and the trained model for each diagnostic target data point. In contrast, the present invention automatically determines, based on the result of the sufficiency assessment of the training data in the parameter calculator and storer 105 (whether the parameter value is stored as the presence of a value or not), whether the trained model should be created or not, i.e., whether a diagnosis should be performed or not, or whether the parameter acquisition unit 106 can acquire the parameter value or not.Since the anomaly detection unit 109 can determine a value as the threshold for anomaly detection, as described above, when a diagnosis is performed, processing to determine the threshold for anomaly detection for each diagnostic target data unit is not necessary and the computational scope for the diagnosis can be reduced. [Output of the result of the determination]

[0071] The determination result output unit 110 outputs a change in the anomaly grade over time, calculated by the anomaly grade calculator 108, and a result of the determination in the anomaly determination unit 109. The determination result output unit 110 can transmit these determination results to a monitoring device (not shown) connected to the anomaly diagnosis device 100, or to an external device (for example, a server computer provided in a central monitoring device) via a communication line (not shown). [Hardware configuration]

[0072] With reference to Fig. Section 5 describes a hardware configuration of the facility for diagnosing anomalies 100. Fig. Figure 5 is a representation showing an exemplary hardware configuration of a computer system 500, which is switched on as an anomaly diagnosis device 100.

[0073] The Computer System 500 includes a central processing unit (CPU) 510, a read-only memory (ROM) 520, a random access memory (RAM) 530, a hard disk drive (HDD) 540, a communication interface (I / E) 550, and an input / output interface (I / O interface) 560. The input / output interface 560 can be connected to an input unit 570 and a display unit 580.

[0074] The CPU 510 executes every instruction contained in a program. When the CPU 510 executes the program, processing occurs to implement the instructions in the program. Fig. The function shown in 1 was performed according to each instruction.

[0075] A pre-created program or pre-created data is permanently (in a non-volatile manner) stored in the ROM 520.

[0076] Data to be used by the CPU 510 during the execution of the program is temporarily stored in RAM 530, and RAM 530 acts as a temporary data storage area to be used as a workspace.

[0077] The HDD 540 is a non-volatile storage device, and data generated by the CPU 510, data received via the communication interface 550, or data input into the input / output interface 560 are stored in the HDD 540. The non-volatile storage device is not limited to the HDD 540, and a semiconductor storage device, such as flash memory, can be used instead of or in addition to the HDD 540.

[0078] The communication interface 550 communicates with another device that is communicatively connected to the computer system 500. For example, when the computer system 500 is communicatively connected to an external device, including a system to be diagnosed, the communication interface 550, acting as a data reader 102, reads (receives) data transmitted by the external device.

[0079] The I / O interface 560 accepts a signal input from the external device to the computer system 500 or outputs a signal to the external device. The external device could be, for example, a central management unit or another server computer, a wirelessly connected smartphone, a tablet device, and the like.

[0080] The input unit 570 accepts an instruction or signal input into the computer system 500. In one aspect, the input unit 570 is implemented as an input device for accepting an action from a user of the computer system 500, such as a keyboard, a mouse, a touch panel, or other devices. In another aspect, the input unit 570 can also be implemented as a device that outputs a signal, such as a camera or various sensors.

[0081] The display unit 580 implements a representation based on a signal input via the I / O interface 560. The display unit 580 is implemented, for example, by a monitoring device, an indicator, a lamp, or the like. In one aspect, the display unit 580 corresponds to an example of the data display unit 103 and the determination result output unit 110. The display unit 580 can display time-series data obtained from the system or device being diagnosed, and the result of the determination by the anomaly detection unit 109, indicating whether or not an anomaly is present in the system or device.

[0082] The processing in the Computer System 500 is implemented by each hardware and software unit executed by the CPU 510. Such software can be pre-stored on the HDD 540. Software can be distributed as a program product stored on a Compact Disc Read-Only Memory (CD-ROM) or other storage medium. Alternatively, software can be provided as a program product that can be downloaded from an information provider connected to what is known as the Internet. Such software is read from the storage medium by a device with an optical disk drive (not shown) or other reader, or downloaded via the communication interface 550, and then stored once on the HDD 540. This software is read from the HDD 540 by the CPU 510 and stored in RAM 530 as an executable program. The CPU 510 then executes the program.

[0083] Each component that is in the Fig. The computer system 500 shown in Figure 5 is generally defined. Accordingly, a significant part of the present invention can be software stored in the ROM 520, RAM 530, HDD 540, or another storage medium, or other software that can be downloaded via a network. Since the operations of each hardware unit of the computer system 500 are sufficiently known, a detailed description is not repeated.

[0084] A recording medium is not limited to a CD-ROM, a flexible disk (FD), or a hard disk drive 540, but can be a medium that transports a program in a fixed manner, such as a magnetic tape, a cassette tape, an optical disc (a magnetic optical disc (MO) / a mini-disc (MD) / a digital versatile disc (DVD)), an integrated circuit (IC) card (including a memory card), an optical card, or a semiconductor memory such as a mask ROM, an electronically programmable read-only memory (EPROM), an electronically erasable programmable read-only memory (EEPROM), a flash ROM, or a secure socket disk (SSD).

[0085] The present program includes not only a program that can be executed directly by the CPU 510, but also a program in source code form, a compressed program, an encrypted program, and the like. [Effect of the first embodiment]

[0086] The inventors discovered that the number of training data units required for a diagnosis without misdiagnosis varies for each parameter that configures the trained model. Furthermore, the inventors found that a diagnosis without misdiagnosis can also be performed with a trained model created from a selection of parameter values ​​that configure the model. These parameters are obtained from a single value of an identical parameter derived from a data set that differs in its operating state. In other words, the trained model can be generated by redirecting certain parameter values ​​between different operating states.Additionally, it was found that data necessary for calculating the parameter value in the operating state do not need to be collected, since the parameter value obtained from training data in a different operating state, where the parameter value has already been calculated and stored, can be redirected as the parameter, which requires a relatively large number of training data units (which takes time to collect data).

[0087] Then, with such a configuration as described in the first embodiment, where processing is carried out to calculate and store the parameter value according to the result of the sufficiency assessment of training data for each parameter, and if some of the parameter values ​​are not stored in the operating state corresponding to the diagnostic target data (the number of training data units is insufficient), the parameter value of the identical parameter in another operating state is redirected to generate the trained model, and a diagnosis is performed even if there is a normal change in the operating state, a diagnosis without incorrect determination can be carried out with the minimum number of training data units, that is, in a training data collection period that is shorter than in the conventional example.Since the trained model is created based on the parameter value for which a sufficiency assessment of training data is OK, a value can be determined as the threshold for anomaly detection, a computational processing to determine the threshold for anomaly detection does not need to be performed for each diagnostic target data unit, and the computational scope can be reduced.

[0088] Furthermore, with such a configuration, if as a result of a sufficiency assessment none of the parameter values ​​that configure the trained model are calculated (the trained model that enables a diagnosis without erroneous determination cannot be created), the trained model is not created and the diagnosis is postponed, thus reducing the risk of incorrect determination. Second embodiment.

[0089] A second embodiment is described below. An anomaly diagnosis device according to the second embodiment differs from the anomaly diagnosis device 100 according to the first embodiment in that the anomaly diagnosis device 100 according to the first embodiment performs a series of diagnostic processes, including data classification, calculation and storage of a parameter, obtaining the parameter, creating a trained model, calculating an anomaly level, and determining it with training data for any pre-set period, whereas the anomaly diagnosis device according to the second embodiment performs a diagnosis over a long period, such that the parameter value stored in each operating state is updated and the diagnosis is performed whenever additional training data can be obtained.

[0090] The essential functional configuration and hardware configuration of the anomaly diagnostic device according to the second embodiment are identical to the essential functional configuration and hardware configuration of the anomaly diagnostic device 100 according to the first embodiment. Therefore, the description thereof is not repeated. A description of the second embodiment follows, optionally with reference to the configuration of the anomaly diagnostic device 100 according to the first embodiment.

[0091] The data classifier 104 according to the second embodiment classifies additional training data for each operating state at a time when the additional training data is available. The classification procedure can be carried out as in the first embodiment.

[0092] The parameter calculator and storer 105 performs a sufficiency assessment of classified data using a classified data group. At this point, the parameter calculator and storer 105 can perform a sufficiency assessment based solely on the additional training data, or a sufficiency assessment of a data group obtained by combining training data already used to calculate and store the parameter with the additional training data, or a data group obtained by selecting some from that data group.

[0093] The parameter calculator and storer 105 calculates a parameter value to be stored (or makes a determination as "absence of value") based on the result of a data sufficiency assessment for each parameter and overwrites the parameter value stored in each operating state for use in the current build of the trained model with these values. The number of data units used to calculate the parameter value can be similar to the number of data units used for the sufficiency assessment.

[0094] The parameter calculator and storage unit 105 can use as additional training data data newly acquired from the plant or data determined as normal in the anomaly detection unit 109.

[0095] Now, with reference to Fig. 6 a control structure of the device for diagnosing anomalies 100 according to the second embodiment is described. Fig. Figure 6 is a flowchart showing part of the processing carried out by the CPU 510 of the computer system 500, which implements the anomaly diagnosis device 100 according to the second embodiment.

[0096] In step S610, the CPU 510 classifies data received from the system or device under diagnosis for each operating state. This processing is performed in the data classifier 104. Afterwards, a stream of training data for parameter calculation and storage passes to S620 (parameter calculator and storer 105), and a stream of data to be diagnosed passes to S630 (parameter acquisition unit 106).

[0097] In step S620, the CPU 510, acting as parameter calculator and storer 105, calculates the parameter value based on the training data and stores the value. As in the first embodiment, the CPU 510 performs a sufficiency evaluation of each parameter from the training data in all operating states and calculates the parameter value, storing it in association with each operating state according to an evaluation result. Alternatively, the CPU 510 stores the absence of the parameter value.

[0098] In step S630, the CPU 510, acting as parameter acquisition unit 106, obtains the parameter value associated with the operating state that corresponds to the operating state of the diagnostic target data. If, as in the first embodiment, only some of the parameters are stored, the CPU 510 obtains the remaining parameters based on parameter values ​​of the same type in a different operating state. If none of the parameter values ​​are stored, none of the parameter values ​​are obtained.

[0099] In step S640, CPU 510, acting as the creator of trained models, creates the trained model based on each obtained parameter value. If CPU 510 does not obtain any parameter values, it does not create the trained model and temporarily pauses the diagnosis of the diagnostic target data.

[0100] In step S650, the CPU 510, acting as an anomaly level calculator 108, calculates the anomaly level of the plant or device based on the created trained model and the diagnostic target data.

[0101] In step S660, the CPU 510, as an anomaly detection unit 109, compares the calculated anomaly level and the predefined threshold for anomaly detection and performs an anomaly detection to diagnose a condition of the plant or device.

[0102] In step S670, if the result of the anomaly detection in step S660 is normal (YES in step S670), the CPU 510 switches control to step S680. If the result of the anomaly detection in step S660 is abnormal (NO in step S670), the CPU 510 switches control to step S690.

[0103] In step S680, the CPU 510 uses the diagnostic target data, which has been determined to be normal, as training data. The controller then returns to step S620. In other words, the data designated as training data undergoes a sufficiency evaluation or processing to calculate and store the parameter in the parameter calculator and storer 105.

[0104] In step S690, the CPU 510 discards the diagnostic target data whose diagnostic result was determined to be abnormal and excludes the diagnostic target data as candidates for training data. [Effect of the second embodiment]

[0105] As explained above, the anomaly diagnostic device 100 according to the second embodiment regularly calculates a latest value of the parameter based on a data set containing the latest data about the plant and updates a stored value, thus enabling a diagnosis that reflects the current state of the plant.

[0106] If the result of the sufficiency assessment of the training data indicates NG, none of the parameter values ​​that the trained model configures can be calculated or obtained, and the diagnosis is temporarily paused, as exemplified in the first embodiment, the anomaly diagnostic device 100 according to the second embodiment performs the aforementioned function to carry out the sufficiency assessment when it obtains additional training data. If the result of the assessment indicates OK, it can calculate and store all or some of the parameter values. Then, at the time when the parameter value associated with the operating state corresponding to the diagnostic target data can be obtained, the trained model can be created that allows for a diagnosis without false determination, and diagnostic accuracy can be ensured, a paused diagnosis can be resumed.By performing a system diagnosis over an extended period, data for training is accumulated, and a diagnosis can be carried out in various operating states without incorrect determination of diagnostic target data. Since the diagnosis can be resumed at the point when at least some parameter values, requiring a relatively small number of training data units, can be calculated, as in the first embodiment, the training data collection period can be minimized while maintaining diagnostic accuracy. Third embodiment.

[0107] An anomaly diagnosis according to a third embodiment is now described. The hardware configuration of an anomaly diagnosis device according to the third embodiment is the same as the hardware configuration of the anomaly diagnosis device 100 according to the first or second embodiment. Except for functions and a control structure specific to the anomaly diagnosis device according to the third embodiment, the functions and control structure are the same as those of the anomaly diagnosis device 100 according to the first or second embodiment. Therefore, the description of the identical hardware configuration, functions, and control structure is not repeated.The anomaly diagnostic device 100 according to the third embodiment is described below with reference to the hardware configuration, functions and control structure of the anomaly diagnostic device 100 according to the first or second embodiment.

[0108] The anomaly diagnosis device 100 according to the first and second embodiments performs anomaly diagnosis when the operating state in which the diagnostic target data is classified falls under any operating state that is predetermined during the classification of training data into the operating state. Conversely, the operating state of the diagnostic target data may not fall under any of the operating states predetermined during the classification of training data into the operating state when determined by the data classifier 104. The anomaly diagnosis device according to the third embodiment differs from the anomaly diagnosis device 100 according to the first or second embodiment in that it also performs anomaly diagnosis in such a case.

[0109] The operating state class of the training data under which the operating state of the diagnostic target data falls is first determined in the data classifier 104, shown in the first embodiment, based on each classification approach and a determination threshold. An example in which classification into the operating state is performed based on the classification rules determined during cluster processing is described by way of example. The data classifier 104 calculates the centroid of each cluster from data belonging to each cluster, which are determined during cluster processing, and obtains a distance between the centroid of each cluster and the classification target data.If there is no such cluster such that a distance between the center of gravity of each cluster and classification target data is equal to or shorter than a predetermined determination threshold in comparison between this distance between the center of gravity of each cluster and the classification target data and the threshold for determination with respect to classification into each predetermined cluster, the data classifier 104 determines that the operating state of the diagnostic target data does not fall under any of the operating states that were predetermined in advance when classifying the training data into the operating state.As with a procedure for classifying into the operating state based on processing to determine whether or not there is a change in time series data at at least one evaluation element, the data classifier 104 similarly determines, if none of the conditions (including the evaluation element or the statistical measure thereof and the threshold for determination) for classification into each operating state are met, that the operating state of the diagnostic target data does not fall under any of the operating states that were determined in advance when classifying the training data into the operating state.

[0110] A method for processing an anomaly diagnosis according to the third embodiment can be selected from two methods below, depending on the operating characteristics of the plant or device to be diagnosed. In a first method, if the operating state of the diagnostic target data does not fall under any of the classes of operating states of the training data, the anomaly diagnosis device 100 according to the third embodiment does not perform any subsequent processing and determines that the diagnostic target data is anomalous. This first method is effective if the plant or device to be diagnosed operates only in certain predetermined operating states.

[0111] In a second method, the anomaly diagnostic device 100 according to the third embodiment adds the operating state of the diagnostic target data as the operating state on which a series of anomaly diagnostic processes are to be performed, even if the operating state of the diagnostic target data does not fall under any of the operating states determined in advance during the classification of the training data into the operating state. In other words, the parameter calculator and storer 105, using the diagnostic target data and the data determined to be in the operating state corresponding to the diagnostic target data as the training data, performs the sufficiency assessment and carries out processing to calculate and store the parameter value.If the data, which has been determined to be in the operating state corresponding to the diagnostic target data and which are defined as the diagnostic target data, are further processed by the anomaly diagnostic device 100 to obtain the parameter value in the parameter acquisition unit 106, processed to create the trained model in the trained model creator 107, processed to calculate the anomaly grade in the anomaly grade calculator 108, and processed to determine an anomaly in the anomaly determination unit 109. This second procedure is effective when it is not possible to predict the number of operating states in which the equipment to be diagnosed can be, or to predict this state to determine the optimal operating state each time according to a peripheral device or an environment for operations.

[0112] In the second procedure, the data classifier 104 inputs the diagnostic target data as the training data into the parameter calculator and storer 105 if the operating state of the diagnostic target data does not fall under any of the operating states that are determined in advance during the classification of the training data into operating states, that is, if it does not fall under any of the operating states that are to be processed in the parameter calculator and storer 105. The parameter calculator and storer 105 calculates the parameter value to be stored in this operating state and stores it, adding the operating state to which the diagnostic target data belongs as a new operating state to be learned.

[0113] The parameter calculator and storer 105 then performs processing on the input data as in the first embodiment. Specifically, the parameter calculator and storer 105 performs a sufficiency evaluation of the number of data units for a multitude of parameters that configure the trained model. If the sufficiency evaluation is OK, the parameter calculator and storer 105 calculates the parameter value and stores the parameter value in association with the newly learned operating state. If the sufficiency evaluation is NG, the parameter calculator and storer 105 stores the parameter value as the absence of the value. The parameter acquisition unit 106 then acquires the parameter value associated with the newly learned operating state.If all of the multitude of parameters that configure the trained model can be obtained, as in the first embodiment, the parameter acquisition unit obtains all parameter values. If some of the parameter values ​​can be obtained, the parameter acquisition unit 106 obtains some of the obtainable parameter values ​​and obtains the remaining parameter values ​​based on parameter values ​​of the same type stored in another operating state. If all parameter values ​​fall under the absence of the value, the parameter acquisition unit 106 obtains none of the parameter values. Thus, if the parameter value can be obtained after a determination of its presence has been made, the trained model creator 107 creates the trained model, and the anomaly grade calculator 108 and the anomaly detection unit 109 each perform processing to determine whether an anomaly is present or not.If none of the parameter values ​​can be obtained, the trained model creator 107 does not create the trained model and the anomaly diagnosis facility 100 temporarily pauses the diagnosis.

[0114] A specific sequence of the previous processing operations is shown. The parameter calculator and storer 105 collects new data necessary for calculating parameters in the newly added operating state. Therefore, in an initial phase of training data collection, the sufficiency assessment of all parameter values ​​that configure the trained model is negative, and the parameter calculator and storer 105 does not calculate the parameter value. Consequently, the trained model creator 107 does not create the trained model. During this processing, the anomaly diagnostic facility 100 can temporarily pause the diagnostic process and avoid an incorrect determination if the number of data units entered into the parameter calculator and storer 105 is less than the number of data units required to calculate the parameter and enable a diagnosis without false determination.

[0115] As more time passes, the amount of input training data increases. Then, the sufficiency assessment of the parameter requiring the smallest number of data units for calculating the parameter values ​​is successful, and this parameter value is calculated. At this stage, the anomaly diagnostic facility 100 can resume the diagnostic process, as the parameter acquisition unit 106 and the trained model creator 107 can obtain remaining parameter values ​​based on parameter values ​​of the same type that were calculated and stored in another operating state, and create the trained model. A parameter that is redirected from the other operating state requires a larger number of data units than other parameters for diagnosis without incorrect determination, and data collection takes some time.According to the configuration of the present invention, the period required for this data collection in the conventional example can be reduced, and accordingly, the anomaly diagnosis device 100 can perform a diagnosis without false determination in a short data collection period.

[0116] As explained above, the anomaly diagnosis device disclosed above can create the trained model in the shortest data collection period without false determination and can perform anomaly diagnosis even when it is not possible to predict in advance the number of operating states in which the plant to be diagnosed may be, or to predict this state in advance for determining the optimal operating state each time according to a peripheral device or an environment for operations.

[0117] Furthermore, the anomaly diagnostic facility 100 is configured such that the parameter calculator and storer 105 performs the sufficiency assessment of the number of data units, thus eliminating the need to calculate the anomaly detection threshold for each diagnostic target data unit. Therefore, since the computational overhead is reduced by the anomaly diagnostic facility 100, an increase in computing resources can be avoided.

[0118] Some of the technical features disclosed above can be summarized as follows.

[0119] [Configuration 1] According to one embodiment, a method for diagnosing anomalies, performed by a computer (for example, the computer system 500), is provided. This method for diagnosing anomalies includes: obtaining, by the CPU 510 as a data reader 102, data containing a state variable of a single evaluation element or a plurality of evaluation elements from a diagnostic target via a communication interface 550; classifying, by the CPU 510 as a data classifier 104, the data containing the state variable of the single evaluation element or the plurality of evaluation elements for each operating state of the system or device to be diagnosed; evaluating, by the CPU 510 as a parameter calculator and storer 105, the sufficiency of the number of data units for each classified data group; and calculating (deriving), by the CPU 510 as a parameter calculator and storer 105.a multitude of parameter values ​​that configure a trained model, according to the sufficiency of the number of data units and saving these parameter values ​​in association with the operating state, by the CPU 510 as parameter acquisition unit 106, the multitude of parameter values ​​that configure the trained model, without overlaps for each parameter, according to the operating state of diagnostic target data and a saving state of the parameter values ​​that are associated with each operating state, creation, by the CPU 510 as creator of trained models 107, of a trained model with the acquired parameter values, calculation, by the CPU 510 as anomaly grade calculator 108, of an anomaly grade to determine with respect to an anomaly based on the trained model and determination, by the CPU 510 as anomaly determination unit 109, of whether an anomaly is present in a diagnostic target, based on the calculated anomaly grade.

[0120] According to the configuration described above, a multitude of parameter values, which configure the trained model, are obtained without overlap for each parameter, based on the operating state of the data to be diagnosed and the storage status of the parameter values ​​associated with each operating state. The trained model is then created using these parameter values. The anomaly threshold for determining whether an anomaly exists in a diagnostic target is calculated based on the created trained model. This anomaly threshold determines whether an anomaly is present. Thus, a normal change in the operating state due to maintenance or a change in operating mode will not be incorrectly identified as an anomaly, but the accuracy of the diagnosis can be improved.Furthermore, even if a change in the system's operating state is unpredictable, the period for re-collecting training data necessary for diagnostics can be reduced. Additionally, the computational scope required for such processing can be reduced.

[0121] [Configuration 2] In addition to the above configuration, the procedure for diagnosing anomalies according to one aspect includes classifying the data exhibiting the state size of the single evaluation element or the multitude of evaluation elements for each operating state of the plant. Performing the classification based on cluster processing or processing to determine whether there is a change in the data with respect to at least one evaluation element.

[0122] [Configuration 3] In addition to the above configuration, the procedure for diagnosing anomalies according to one aspect includes determining whether there is a change in the data, where at least one of a mean, variance, standard deviation, maximum, minimum, kurtosis and skewness is calculated from any time range of the data, which are used as the index of determination.

[0123] [Configuration 4] In addition to the above configuration, the procedure for diagnosing anomalies according to one aspect includes: Assessing the adequacy of the number of data units; performing the assessment using at least one of a procedure for determining in advance the number of data units that can be diagnosed without error from past diagnostic records; and performing an assessment based on this number of data units, or a procedure for performing an assessment based on a correlation coefficient between any assessment elements calculated from any time range of each data group, or a measure of change of a basic statistical measure of any assessment element.

[0124] [Configuration 5] In addition to the configuration above, the procedure for diagnosing anomalies according to one aspect involves obtaining the multitude of parameter values ​​that configure the trained model, without overlaps for each parameter, according to the operating state of the diagnostic target data and a storage state of the parameter values ​​associated with each operating state; obtaining all parameter values ​​that configure the trained model, based on the fact that all parameter values ​​are stored in an operating state identical to the operating state of the data to be diagnosed; and obtaining, based on the fact that only some of the parameter values ​​that configure the trained model are stored in the operating state identical to the operating state of the data to be diagnosed.only some of the parameter values ​​and obtaining the remaining parameter values ​​from parameter values ​​of an identical type that are stored in another operating state.

[0125] [Configuration 6] In addition to the configuration described above, the procedure for diagnosing anomalies according to one aspect involves calculating a large number of parameter values ​​that configure a trained model, according to the sufficiency of the number of data units, and storing the parameter values ​​in association with the operating state. It also involves regularly calculating the latest value of a parameter from a data set containing the most recent data obtained from the plant and updating the stored parameter values. "Regular" refers to a predetermined time interval and refers, for example, to a time on a day that is predetermined, a specific time each week, a specific time and date in each month, or the like.

[0126] [Configuration 7] In addition to the configuration above, in the procedure for diagnosing anomalies according to one aspect, the trained model is a unit space according to a Mahalanobis-Taguchi method. The set of parameter values ​​is a mean, a standard deviation, and an inverse matrix of a correlation matrix of data that define the unit space.

[0127] [Configuration 8] In addition to the configuration described above, the procedure for diagnosing anomalies according to one aspect further includes output, by the CPU 510 as the determination result output unit 110, of a result of a determination with respect to the diagnostic target. An output target is a monitoring device or other display unit 580, or a server computer or other information management device connected via a communication interface and located remotely.

[0128] [Configuration 9] According to another embodiment, the anomaly diagnostic device 100 (or the computer system 500) is provided to diagnose whether an anomaly exists in a diagnostic target. This anomaly diagnostic device 100 includes a memory (for example, EOM 520, RAM 530, or HDD 540) in which a program is stored, and a processor (for example, CPU 510) coupled to the memory to execute the program. The program causes the processor to: obtain, from the diagnostic target, data that has a state size of a single rating element or a plurality of rating elements; classify the data that has the state size of the single rating element or the plurality of rating elements for each operating state of the diagnostic target; and evaluate the adequacy of the number of data units for each classified data group.Calculating (deriving) a multitude of parameter values ​​that configure a trained model, according to the sufficiency of the number of data units, and storing these parameter values ​​in association with the operating state; obtaining the multitude of parameter values ​​that configure the trained model without overlaps for each parameter, according to the operating state of diagnostic target data and a storage state of the parameter values ​​associated with each operating state; creating a trained model with the obtained parameter values; calculating an anomaly level to determine with respect to an anomaly based on the trained model; and determining whether an anomaly exists in the diagnostic target based on the anomaly level.

[0129] [Configuration 10] In addition to the above configuration, the program in the facility for diagnosing anomalies according to one aspect causes the processor to further perform an output of a result of determining the diagnostic target.

[0130] [Configuration 11] According to yet another embodiment, an anomaly diagnosis program is provided to diagnose whether an anomaly exists in a diagnostic target. This anomaly diagnosis program causes a processor (for example, CPU 510) to perform the following: obtain, from the diagnostic target, data containing a state variable of a single rating element or a plurality of rating elements; classify the data containing the state variable of the single rating element or the plurality of rating elements for each operating state of the diagnostic target; assess the sufficiency of the number of data units for each classified data group; calculate (derive) a plurality of parameter values ​​that configure a trained model, according to the sufficiency of the number of data units; and store these parameter values ​​in association with the operating state.Obtaining the multitude of parameter values ​​that configure the trained model, without overlaps for each parameter, according to the operating state of the diagnostic target data and a storage state of the parameter values ​​associated with each operating state; creating a trained model with the obtained parameter values; calculating an anomaly level to determine whether an anomaly exists in the diagnostic target based on the trained model; and determining whether an anomaly is present in the diagnostic target based on the anomaly level.

[0131] [Configuration 12] In addition to the above configuration, the program for diagnosing anomalies according to one aspect causes the processor to further perform an output of a result of determining the diagnostic target.

[0132] [Configuration 13] According to yet another embodiment, an anomaly diagnosis system is provided to diagnose whether or not an anomaly exists in a diagnostic target. This anomaly diagnosis system causes a processor (for example, CPU 510) to perform the following: obtain, from the diagnostic target, data containing a state variable of a single rating element or a plurality of rating elements; classify the data containing the state variable of the single rating element or the plurality of rating elements for each operating state of the diagnostic target; assess the sufficiency of the number of data units for each classified data group; calculate (derive) a plurality of parameter values ​​that configure a trained model, according to the sufficiency of the number of data units; and store these parameter values ​​in association with the operating state.Obtaining the multitude of parameter values ​​that configure the trained model, without overlaps for each parameter, according to the operating state of the diagnostic target data and a storage state of the parameter values ​​associated with each operating state; creating a trained model with the obtained parameter values; calculating an anomaly level to determine whether an anomaly exists in the diagnostic target based on the trained model; and determining whether an anomaly is present in the diagnostic target based on the anomaly level.

[0133] [Configuration 14] In addition to the above configuration, the system for diagnosing anomalies according to one aspect causes the processor to further perform an output of a result of determining the diagnostic target.

[0134] It is understood that the embodiments disclosed in this document are in every respect illustrative and not limiting. The scope of the present invention is defined by the content of the claims and not by the foregoing description and is intended to include any modification that lies within the scope of the claims and has an equivalent meaning to their content. REFERENCE MARK LIST

[0135] 10 Plant or device; 100 Anomaly diagnostic device; 102 Data reader; 103 Data display unit; 104 Data classifier; 105 Parameter calculator and storer; 106 Parameter retrieval unit; 107 Trained model creator; 108 Anomaly grade calculator; 109 Anomaly determination unit; 110 Determination result output unit; 500 Computer system; 510 CPU; 520 ROM; 530 RAM; 540 Hard disk drive; 550 Communication interface; 560 Input and output interface; 570 Input unit; 580 Display unit. QUOTES INCLUDED IN THE DESCRIPTION

[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited patent literature

[0000] 2017-102826 [0004, 0006] 2014-32657 [0005, 0006] 2015-181072

[0006]

Claims

[1] An anomaly diagnosis method for diagnosing whether an anomaly exists in an installation to be diagnosed, the anomaly diagnosis method comprising: Obtaining, from the system, data that comprise a state variable of a single evaluation element or a plurality of evaluation elements; Classifying the data obtained from the plant for each operating state of the plant; Assessing the sufficiency of the number of data units for each classified data group; Calculating a plurality of parameter values configuring a trained model according to the sufficiency of the number of data units and storing the parameter values in association with the operating state; Obtaining the plurality of parameter values that configure the trained model, without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage status of the parameter values associated with each operating state; Creating a trained model to diagnose data to be diagnosed with the obtained parameter values; Calculating an anomaly level to determine an anomaly based on the trained model; and Determine whether an abnormality exists in a diagnostic target based on the abnormality level. [2] The anomaly diagnosis method according to claim 1, wherein classifying the data for each operating state of the plant comprises performing the classification based on cluster processing or processing for determining whether there is a change in the data with respect to at least one evaluation item. [3] The method for diagnosing abnormalities according to claim 2, wherein determining whether there is a change in the data comprises determining whether there is a change by calculating at least one of an average, a variance, a standard deviation, a maximum, a minimum, a kurtosis, and a skewness from any temporal range of the data used as a determination index. [4] A method for diagnosing anomalies according to any one of claims 1 to 3, wherein evaluating the sufficiency of the number of data units comprises performing the evaluation with at least one of the following: a method for determining in advance the number of data units that can be diagnosed without misidentification from past diagnostic records and making an assessment based on the number of data units; or a method for making an evaluation based on a correlation coefficient between any evaluation items calculated from any time range of each data group or an amount of change in the basic statistical measure of any evaluation item. [5] A method for diagnosing anomalies according to any one of claims 1 to 4, wherein obtaining the plurality of parameter values configuring the trained model without overlaps for each parameter according to the operating state of the data to be diagnosed and a storage status of the parameter values associated with each operating state comprises: Obtaining all parameter values that configure the trained model based on the fact that all parameter values are stored in an operating state that is identical to the operating state of data to be diagnosed; and Obtaining, based on a fact that only some of the parameter values configuring the trained model are stored in the operating state identical to the operating state of the data to be diagnosed, only some of the parameter values and obtaining the remaining parameter values from parameter values of an identical type stored in another operating state. [6] The anomaly diagnosis method according to any one of claims 1 to 5, wherein calculating a plurality of parameter values configuring a trained model according to the sufficiency of the number of data units and storing the parameter values in association with the operating state comprises regularly calculating a latest value of a parameter from a data group including the latest data acquired from the plant and updating the stored parameter values. [7] The anomaly diagnosis method according to any one of claims 1 to 6, wherein the trained model is a unit space according to a Mahalanobis-Taguchi method, and the plurality of parameter values is an average, a standard deviation, and an inverse matrix of a correlation matrix of data defining the unit space. [8] Anomaly diagnosis device for diagnosing whether an anomaly exists in a system to be diagnosed, the anomaly diagnosis device comprising: a data reader for obtaining, from the system, data comprising a state variable of a single evaluation element or a plurality of evaluation elements; a data classifier for classifying the data obtained from the system for each operating state of the system; a parameter calculator and storer for evaluating the sufficiency of the number of data units for each classified data group, calculating a plurality of parameter values configuring a trained model according to the sufficiency of the number of data units, and storing the parameter values in association with the operating state; a parameter acquisition unit for acquiring the plurality of parameter values configuring the trained model without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage status of the parameter values associated with each operating state; a trained model creator for creating a trained model for diagnosing the data to be diagnosed with the obtained parameter values; an anomaly degree calculator for calculating an anomaly degree for determining an anomaly based on the trained model; and an abnormality determination unit for determining whether an abnormality exists in a diagnosis target based on the abnormality degree. [9] An anomaly diagnosis program for causing a computer to perform processing to diagnose whether an anomaly exists in a system to be diagnosed, the anomaly diagnosis program causing the computer to perform the following: Obtaining, from the system, data that comprise a state variable of a single evaluation element or a plurality of evaluation elements; Classifying the data obtained from the plant for each operating state of the plant; Assessing the sufficiency of the number of data units for each classified data group; Calculating a plurality of parameter values configuring a trained model according to the sufficiency of the number of data units and storing the parameter values in association with the operating state; Obtaining the plurality of parameter values that configure the trained model, without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage status of the parameter values associated with each operating state; Creating a trained model to diagnose the data to be diagnosed with the obtained parameter values; Calculating an anomaly level to determine an anomaly based on the trained model; and Determine whether an abnormality exists in a diagnostic target based on the abnormality level. [10] An anomaly diagnosis system for diagnosing whether an anomaly exists in a system to be diagnosed, comprising: a data reader for obtaining, from the system, data comprising a state variable of a single evaluation element or a plurality of evaluation elements; a data classifier for classifying the data obtained from the system for each operating state of the system; a parameter calculator and storer for evaluating the sufficiency of the number of data units for each classified data group, calculating a plurality of parameter values configuring a trained model according to the sufficiency of the number of data units, and storing the parameter values in association with the operating state; a parameter acquisition unit for acquiring the plurality of parameter values configuring the trained model without overlaps for each parameter, according to the operating state of the data to be diagnosed and a storage status of the parameter values associated with each operating state; a trained model creator for creating a trained model for diagnosing the data to be diagnosed with the obtained parameter values; an anomaly degree calculator for calculating an anomaly degree for determining an anomaly based on the trained model; and an abnormality determination unit for determining whether an abnormality exists in a diagnosis target based on the abnormality degree.

Images