Abnormality diagnosis device, abnormality diagnosis method, and program

Abstract

The abnormality diagnosis device includes a diagnosis unit that diagnoses the type of abnormality occurring in the abnormality diagnosis object based on the difference between abnormality simulation results of each type of abnormality simulating multiple types of abnormalities of the abnormality diagnosis object and multiple time-series observation results of the abnormality diagnosis object observed in time by multiple detection units.

Description

Technical Field

[0001] The present invention relates to an abnormality diagnosis device, an abnormality diagnosis method, and a program.

[0002] This application claims priority based on Japanese Patent Application No. 2018-204502 filed on October 30, 2018, the contents of which are incorporated herein by reference. Background Art

[0003] Conventionally, as a method for diagnosing an abnormality occurring in an object, there is known a technique for accumulating information on malfunctions of the object occurring in the past and diagnosing the abnormality based on the tendency of the accumulated information (for example, Patent Document 1).

[0004] Prior Art Literature

[0005] Patent Literature

[0006] Patent Document 1: Japanese Patent Application Publication No. 2010-126130 Summary of the invention

[0007] Technical problem to be solved by the invention

[0008] However, in the prior art, there is little information about the object, and when there is no information about the malfunction or the information about the malfunction is little, it is sometimes difficult to diagnose the abnormality of the object.

[0009] Conventionally, when the number of devices (for example, sensors) that acquire information related to abnormality diagnosis targets is large, it may be necessary to store a large amount of information when diagnosing the abnormality.

[0010] One aspect of the present invention is made in consideration of such a situation, and one object of the present invention is to provide an abnormality diagnosis device, an abnormality diagnosis method, and a program capable of diagnosing an abnormality of an object without using information on malfunctions that occurred in the past.

[0011] Solutions for solving technical problems

[0012] One embodiment of the present invention relates to an abnormality diagnosis device including a diagnosis unit, which diagnoses the type of abnormality occurring in an abnormality diagnosis object based on the difference between abnormality simulation results of various types of abnormalities obtained by simulating multiple types of abnormalities of the abnormality diagnosis object and multiple time series observation results obtained by observing the abnormality diagnosis object based on time series through multiple detection units.

[0013] Effects of the Invention

[0014] According to the present invention, it is possible to diagnose an abnormality of an abnormality diagnosis target without using information on a malfunction that occurred in the past. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a diagram showing an example of the structure of the abnormality diagnosis device according to the embodiment.

[0016] Figure 2 This is a diagram showing an example of the content of the observation result information.

[0017] Figure 3 This is a diagram showing an example of the contents of the valve pressure loss abnormality simulation result.

[0018] Figure 4 This is a diagram showing an example of the content of normal simulation information.

[0019] Figure 5 This is a diagram showing an example of the content of the correlation information.

[0020] Figure 6 This is a graph showing an example of the temporal change of the observation result of the sensor in a normal state.

[0021] Figure 7 This is a diagram showing an example of the shape of a locus on a phase plane in a normal state.

[0022] Figure 8 This is a tree diagram showing an example of the classification result based on the complex power cepstrum distance of the diagnosis part classification.

[0023] Fig. 9 This is a tree diagram showing an example of a result of classification based on the degree of difference evaluated by the diagnosis unit using the dynamic time warping method.

[0024] Fig.10 This is a flowchart showing a series of processes of the abnormality diagnosis device.

[0025] Fig.11 Yes means Fig.10 Flow chart of a series of processes of the processing of step S106 in .

[0026] Fig.12 Yes means Fig.10 A flowchart of a series of processes of the processing of step S108 in .

[0027] Fig.13 It is a diagram showing an example of the structure of an abnormality diagnosis device according to a modification.

[0028] Fig.14 This is a diagram showing an example of contribution rate. DETAILED DESCRIPTION

[0029] <Implementation Method>

[0030] Hereinafter, embodiments of the present invention will be described with reference to the drawings.

[0031] [Overall structure]

[0032] Figure 1 1 is a diagram showing an example of the structure of an abnormality diagnosis device 1 according to an embodiment. The abnormality diagnosis device 1 diagnoses an abnormality that has occurred or may occur in a device or system that is an abnormality diagnosis object. The abnormality diagnosis device 1 includes, for example, a control unit 100 and a storage unit 200. The control unit 100 implements the functions of the acquisition unit 102, the diagnosis unit 104, the trajectory derivation unit 106, the first difference calculation unit 108, the second difference calculation unit 110, and the output unit 112 by executing the program (software) stored in the storage unit 200 through a hardware processor such as a CPU (Central Processing Unit). Some or all of these components can be implemented by hardware (including circuitry) such as LSI (Large Scale Integration) or ASIC (Application Specific Integrated Circuit), FPGA (Field-Programmable Gate Array), GPU (Graphics Processing Unit), or by the collaboration of software and hardware.

[0033] The storage unit 200 may be implemented, for example, by a storage device such as a HDD (Hard Disk Drive) or a flash memory (including a storage device of a non-temporary storage medium), or by a removable storage medium such as a DVD or CD-ROM (a non-temporary storage medium), or by a storage medium mounted on a drive device. A part or all of the storage unit 200 may be an external device accessible to the abnormality diagnosis device 1, such as a NAS (Network Attached Storage) or an external storage server. The storage unit 200 stores, for example, abnormal simulation information 202, normal simulation information 204, correlation information 206, and complex autoregression model information 208. Details of various information will be described later.

[0034] The acquisition unit 102 acquires information (hereinafter referred to as observation result information 300) indicating multiple observation results of the state of the observation object observed by multiple sensors (not shown) configured in the device or system of the abnormal diagnosis object at each predetermined time. The observation object is, for example, a physical quantity that changes in the device or system of the abnormal diagnosis object, and the state of the observation object is, for example, the value of the physical quantity. The physical quantity can be, for example, a physical quantity representing at least one of mechanical properties, electromagnetic properties, thermal properties, sound properties, chemical properties, etc., or spatial information or time information represented by them. The physical quantity includes, for example, pressure, flow rate, valve opening, atmospheric pressure, outside temperature, water temperature, blood pressure, pulse, etc. The device or system of the abnormal diagnosis object is, for example, a rocket (for example, a reused rocket, a rocket engine), a transportation system, a power generation device, a life support device, etc. In the case where the device of the abnormal diagnosis object is a rocket, multiple sensors are mounted on the rocket to observe the observation object (physical quantity) involved in the control of the rocket. The device or system of the abnormal diagnosis object is not limited to this, as long as it is a device that can obtain the observation object.

[0035] The acquisition unit 102 may acquire the observation result information 300 from a plurality of sensors connected to the abnormality diagnosis device 1, or may acquire the observation result information 300 from a sensor connected to a location where information can be transmitted and received via a network, or may acquire (extract) the observation history of the sensor included in the action history (e.g., log information) of the abnormality diagnosis object that has completed the action as the observation result information 300. In the following description, the case where the acquisition unit 102 acquires the observation result information 300 from the log information stored in the data recorder carried by the reused rocket is described. The sensor is an example of a "detection unit".

[0036] Figure 2 300 is a diagram showing an example of the content of the observation result information 300. In the observation result information 300, for example, information that can identify a sensor (sensor ID shown in the figure), a sensor name, and information that establishes a correspondence between observation results of the observation object observed by the sensor at a predetermined time are stored separately for each sensor. Figure 2 In FIG. 1 , the sensor whose sensor ID is “0001” and whose sensor name is “temperature sensor A” observes the temperature of the abnormality diagnosis target, and the observation result shows the temperature observed by the sensor (for example, the device temperature).

[0037] return Figure 1 The diagnosis unit 104 diagnoses the type of abnormality that occurs or may occur in the abnormality diagnosis target based on the plurality of observation results acquired by the acquisition unit 102. The abnormality is, for example, an abnormality in a device or system of the abnormality diagnosis target (hereinafter referred to as a system abnormality).

[0038] The trajectory derivation unit 106, the first difference calculation unit 108, and the second difference calculation unit 110 respectively generate indices used by the diagnosis unit 104 when diagnosing an abnormality. Details of the trajectory derivation unit 106, the first difference calculation unit 108, and the second difference calculation unit 110 will be described later.

[0039] The output unit 112 outputs information related to the type of abnormality diagnosed by the diagnosis unit 104. The output unit 112 may, for example, display an image indicating the type of abnormality on a display device (not shown) connected to the abnormality diagnosis device 1, or may output information indicating the type of abnormality to another device connected via a network, or may cause the storage unit 200 to store information indicating the type of abnormality.

[0040] [Processing of Diagnosis Unit 104]

[0041] The following describes the details of the processing of the diagnosis unit 104. First, the diagnosis unit 104 determines a certain observation result from the observation results obtained by the acquisition unit 102. The diagnosis unit 104 can determine the observation result of the diagnosis object based on an instruction input by the user via an input device (not shown) connected to the abnormal diagnosis device 1, for example, or can determine the observation result based on a predetermined order. The diagnosis unit 104 diagnoses the type of abnormality that occurs in the determined abnormal diagnosis object or that may occur in the abnormal diagnosis object based on the determined observation result and the simulation result simulated using a model that simulates (models) the device or system of the abnormal diagnosis object. In the subsequent description, the situation in which the diagnosis unit 104 diagnoses the type of abnormality that occurs in the abnormal diagnosis object or that may occur in the abnormal diagnosis object is sometimes referred to as the diagnosis unit 104 diagnosing the type of abnormality of the abnormal diagnosis object.

[0042] The abnormal diagnosis object of this embodiment is, for example, a flow control valve driven by an electric motor, and is a device including an electric motor, a mechanism component that transmits the power generated by the electric motor to a transmission, an actuator that converts the rotational motion generated by the electric motor into a telescopic motion, and a valve core driven by the actuator. The diagnostic unit 104 diagnoses four types of abnormalities that may occur in the abnormal diagnosis object, namely, "valve pressure loss abnormality" or "fluid torque abnormality" of the valve core, "motor coil abnormality" of the electric motor, and "mechanical torque abnormality" of the actuator. The abnormalities that may occur in the abnormal diagnosis object are not limited to these four abnormalities.

[0043] The following describes the details of the simulation results. The abnormal simulation information 202 is information indicating the results of the simulation using a model that simulates the abnormal state of the abnormal diagnosis object (hereinafter referred to as the abnormal simulation result). The abnormal simulation information 202 includes a valve pressure loss abnormal simulation result 202a, a fluid torque abnormal simulation result 202b, a motor coil abnormal simulation result 202c, and a mechanism torque abnormal simulation result 202d.

[0044] The valve pressure loss abnormality simulation result 202a is an abnormality simulation result of a model that simulates a state where a "valve pressure loss abnormality" occurs in the abnormality diagnosis object. The fluid torque abnormality simulation result 202b is an abnormality simulation result of a model that simulates a state where a "fluid torque abnormality" occurs in the abnormality diagnosis object. The motor coil abnormality simulation result 202c is an abnormality simulation result of a model that simulates a state where a "motor coil abnormality" occurs in the abnormality diagnosis object. The mechanism torque abnormality simulation result 202d is an abnormality simulation result of a model that simulates a state where a "mechanism torque abnormality" occurs in the abnormality diagnosis object.

[0045] Although the models used for simulation and the simulation results are different, the abnormal simulation results (202a to 202d) include the same contents. Therefore, the valve pressure loss abnormal simulation result 202a will be described below as an example. Figure 3 202a. The valve pressure loss abnormality simulation result 202a is information showing the simulation results of the observation object generated by multiple sensors configured for the abnormality diagnosis object for each sensor. The following describes a case where the sensor whose observation result is shown in the observation result information 300 is consistent with the sensor whose abnormality simulation result is shown in the abnormality simulation information 202.

[0046] In the valve pressure loss abnormality simulation result 202a, the abnormality simulation result when multiple inputs with different interference patterns are given to the model emulating the "valve pressure loss abnormality" is represented according to each interference pattern. Interference is, for example, the difference in input given to the model due to the action timing, action status or individual differences between the abnormal diagnosis objects. Figure 3 In the “valve pressure loss abnormality” model, the input disturbance patterns are eight disturbance patterns No. 1 to No. 8. Therefore, in the valve pressure loss abnormality simulation result 202a, the abnormality simulation results when the model is given inputs that generate disturbances of each pattern are shown for each disturbance pattern.

[0047] Figure 42 is a diagram showing an example of the content of the normal simulation information 204. The normal simulation information 204 is information showing the result of simulation using a model simulating a case where the abnormal diagnosis object is in a normal state (hereinafter referred to as a normal simulation result) for each sensor. The normal simulation information 204 shows the normal simulation result when an input including a disturbance is given to the model simulating the normal state of the abnormal diagnosis object.

[0048] Based on the observation results and simulation results, the diagnosis unit 104 diagnoses the type of abnormality of the abnormality diagnosis object using (1) a method using a dissimilarity distance evaluated by a known method such as complex power cepstrum and (2) a dissimilarity distance evaluated by a known method such as DTW (Dynamic Time Warping). The details of each method are described below.

[0049] [(1) About the method using the dissimilarity distance evaluated by the complex power cepstrum etc.]

[0050] The following describes the content of the process in which the diagnosis unit 104 diagnoses the type of abnormality using (1) a method using a dissimilarity distance evaluated by a complex power cepstrum or the like. First, the diagnosis unit 104 determines the observation object (hereinafter referred to as the first observation object) observed by a certain sensor (hereinafter referred to as the first sensor) from the observation result information 300 obtained by the acquisition unit 102. Then, based on the correlation information 206, the diagnosis unit 104 determines the observation result (hereinafter referred to as the second observation result) of the sensor (hereinafter referred to as the second sensor) that observes the observation object (hereinafter referred to as the second observation object) that is in a correlation relationship with the first sensor. With respect to the second observation object, only one sensor can be selected from a plurality of sensors, or a plurality of sensors can be selected as the second observation object by aggregating them. Moreover, with respect to the second observation object, a sensor that is not in a correlation relationship with the first sensor can also be selected. Selecting a sensor that is in a correlation relationship allows for rapid analysis, so the following describes the case in which a sensor that is in a correlation relationship is selected.

[0051] Figure 5 2 is a diagram showing an example of the content of the correlation information 206. The correlation information 206 is information showing sensors whose observation results are correlated with each other, for example. Figure 5In the example, the sensors whose observation results are correlated with the sensor whose sensor ID is "0001" and whose sensor name is "temperature sensor A" are sensors whose sensor IDs are "0002" to "0006" and whose sensor names are "temperature sensor B" to "temperature sensor F". The sensors whose observation results are correlated with a certain sensor are, for example, a sensor that observes an observation object of the same type as the observation object of the certain sensor, or a sensor (for example, a "current sensor") that observes an observation object of a different type (for example, "current") that changes according to a change in an observation object (for example, "voltage") of a certain sensor (for example, a "voltage sensor").

[0052] Although the observation object is the same as the observation object of the first sensor or the type of the observation object changes according to the change of the first observation object, the sensor whose observation results are not correlated due to the installation location or installation environment may not be included in the correlation information 206.

[0053] For example, when the first sensor is "temperature sensor A", the diagnosis unit 104 refers to the correlation information 206 and determines "temperature sensor B" to "temperature sensor F" as the second sensor that is correlated with the first sensor, "temperature sensor A". When there are multiple sensors that are correlated with the first sensor, "sensor A", the diagnosis unit 104 may first process one of the sensors as the second sensor and then process the other sensors as the second sensor, but may also process multiple sensors in a group. The following describes a case where the diagnosis unit 104 performs diagnosis using "sensor B" as the second sensor when there are two correlated sensors.

[0054] The trajectory derivation unit 106 performs principal component analysis (hereinafter referred to as PCA; Principal Component Analysis) processing on the first observation result and the second observation result determined by the diagnosis unit 104 using the principal component analysis technique in multivariate time series data. For example, the trajectory derivation unit 106 reconstructs a set of two-variable data having a time index as a two-dimensional vector representing the first observation result as the first element and the second observation result as the second element. For this data (in this example, it is converted into a set of two-variable data having the first observation result as the first element and the second observation result as the second element according to each time when the value is obtained), the PCA processing is performed to extract features. At this time, in the k-variable PCA processing, the principal components from the first principal component to the k-th principal component can be extracted. Therefore, when the trajectory derivation unit 106 performs the two-variable PCA processing on the first observation result and the second observation result, the first principal component score and the second principal component score can be derived at the same time. In this way, the trajectory derivation unit 106 reconstructs the observation results obtained from multiple sensors into multivariate data, thereby being able to derive the principal component scores of multiple sensor information at one time through a single PCA processing.

[0055] Instead of reconstructing the first observation result and the second observation result into two-variable data and performing the PCA processing on them in aggregate, the trajectory deriving unit 106 may reconstruct the single sensor information into multivariate data by introducing a hysteresis structure using a sliding window method for the first observation result and the second observation result, and perform the PCA processing. The trajectory deriving unit 106 may reconstruct the first observation result in which the single sensor information is reconstructed into multivariate data using a sliding window method and the PCA processing is performed, and the second observation result in which the single sensor information is reconstructed into multivariate data using a sliding window method and the PCA processing is performed as two-variable data and perform the PCA processing on them in aggregate.

[0056] In the following, the first observation result and the second observation result respectively include m (m is a natural number) observation results obtained at the same time every specified time. Therefore, for the above-mentioned first observation result and second observation result, when PCA processing is performed using the PCA processing method in multivariate time series data, m groups of first principal component scores and second principal component scores derived by the trajectory derivation unit 106 are obtained. In the following description, the first principal component scores of "Sensor A" and "Sensor B" derived by the trajectory derivation unit 106 are recorded as first principal component scores x, and the second principal component scores of "Sensor A" and "Sensor B" are recorded as second principal component scores y.

[0057] In the k-variable PCA processing, the principal components from the first principal component to the k-th principal component can be extracted. Therefore, the trajectory export unit 106 is not limited to the first principal component score and the second principal component score used in subsequent analysis, and any score among the third principal component score, the fourth principal component score, ..., the k-th principal component score (k is a natural number) can be used in subsequent analysis.

[0058] The trajectory derivation unit 106 can convert the time series information of a single variable into a set of vectors with nearby measurement values ​​as elements, which is called a local time series, by using a sliding window method or the like, which is a well-known technique. The trajectory derivation unit 106 reconstructs the detection results into multivariate data with nearby measurement values ​​as variables. For the detection results reconstructed by the sliding window method, when multivariate analysis is performed using the principal component analysis method that can obtain multimodal information, the trajectory derivation unit 106 can not only import the time structure (lag structure) into the principal component analysis result, but also obtain multimodal information of the number of window widths from the detection results. Details of the sliding window method will be described later. The multimodal information obtained in this way respectively represents abnormalities of different properties (periodic changes, mixing of irregular noise, vibration of measurement values, etc.).

[0059] In the following description, the process performed by the trajectory derivation unit 106, i.e., the process of reconstructing the time series information of a single sensor into multivariate data by vectorizing using a lag structure and obtaining the mode information by using a multivariate analysis method (e.g., principal component analysis, core principal component analysis, fuzzy principal component analysis, sparse principal component analysis, probabilistic principal component analysis, robust principal component analysis, etc.) capable of obtaining multi-mode information, is described as "principal component score acquisition process for each sensor". In contrast, the process performed by the trajectory derivation unit 106, i.e., the process of reconstructing the time series information of two or more sensors into multivariate data by vectorizing them according to each time concentration and obtaining the mode information by using a multivariate analysis method in a general time region by processing them as a plurality of independent multivariate data, is described as "principal component score acquisition process for each sensor group".

[0060] The trajectory deriving unit 106 can obtain mode information from a different viewpoint from that of the "principal component score acquisition process for each sensor group" when two or more sensors are aggregated for multivariate analysis through the "principal component score acquisition process for each sensor group". When the trajectory deriving unit 106 performs the "principal component score acquisition process for each sensor group" using principal component analysis, unlike the case where the mode information of the number of window widths is extracted when the sliding window method is used in the "principal component score acquisition process for each sensor group" using principal component analysis, it can obtain multi-mode information of the number of sensors used for reconstruction of multivariate data. In the "principal component score acquisition process for each sensor group", the trajectory deriving unit 106 obtains mode information based on the detection results of multiple sensors at one time, so compared with the "principal component score acquisition process for each sensor" that obtains mode information for each sensor, when the number of sensors included in the abnormality detection target is large, it is possible to save calculation effort.

[0061] The trajectory derivation unit 106, for example, uses the detection result of the detection result of a single sensor represented by the observation result information 300 before the mode information acquisition process (hereinafter referred to as the 0th order mode information) and the data after the detection result of the sensor is standardized (hereinafter referred to as the standardized 0th order mode information) as the object data of the "principal component score acquisition process for each sensor". The trajectory derivation unit 106, for example, uses the "0th order mode information", "standardized 0th order mode information", "the result of obtaining the principal component score of each sensor, i.e., the mode information from the first order to the higher order", "the result of obtaining the principal component score of each sensor after standardization, i.e., the mode information from the first order to the higher order", etc. as the object data of the "principal component score acquisition process for each sensor group".

[0062] The trajectory derivation unit 106 uses the data that has been subjected to the "principal component score acquisition process for each sensor group" as described above to perform the "principal component score acquisition process for each sensor group", thereby enabling the import of the hysteresis structure lost in the "principal component score acquisition process for each sensor group". This is based on the following reasons. The data is reconstructed using the sliding window method, and the hysteresis structure is imported into the "principal component score acquisition process for each sensor" data that has been subjected to multivariate analysis. On the other hand, in the case of the "principal component score acquisition process for each sensor group", the trajectory derivation unit 106 reconstructs the time series information of two or more sensors into multivariate data by vectorizing them according to each time concentration, so the time correlation between the data is lost. The trajectory derivation unit 106 uses the result of the "principal component score acquisition process for each sensor" into which the hysteresis structure is introduced as the time series information of the sensor in the "principal component score acquisition process for each sensor group", thereby enabling the restoration of the time correlation lost when the model information is acquired.

[0063] The trajectory export unit 106 performs "principal component score acquisition processing for each sensor" according to each sensor or "principal component score acquisition processing for each sensor group" for a combination of more than two sensors based on the observation result information 300 obtained by the acquisition unit 102, and extracts the characteristics of various anomalies such as anomalies for which no threshold is set or anomalies for which threshold is difficult to set (periodic changes or mixing of irregular noise, etc.) through multi-mode information.

[0064] The trajectory derivation unit 106 derives a trajectory in which the coordinates (coordinates (x1, y1), ..., coordinates (xm, ym) in the figure) representing the elements {x1, x2, ..., xm} included in the first principal component score x and the elements {y1, y2, ..., ym} included in the second principal component score y, which are represented by straight lines and have the same observation timing, on a phase plane with the first principal component score x as the first axis (e.g., horizontal axis) and the second principal component score y as the second axis (e.g., vertical axis).

[0065] Figure 6 This is a graph showing an example of temporal changes in a sensor in a normal state. Figure 6 The waveform W1 shown is a waveform indicating the temporal change of the detection result of “sensor A”, and the waveform W2 is a waveform indicating the temporal change of the detection result of “sensor B”.

[0066] Figure 7 : is a diagram showing an example of the shape of a trajectory on a phase plane in a normal state. The trajectory derivation unit 106 generates a trajectory (illustrated trajectory Orb1) in which coordinates (illustrated coordinates (x1, y1), ..., coordinates (xm, ym)) represented by elements of the first principal component score x and elements of the second principal component score y are connected in time sequence by straight lines in a phase plane with the first principal component score x derived based on the principal component score acquisition processing results of each sensor group as the horizontal axis and the second principal component score y as the vertical axis.

[0067] The trajectory deriving unit 106 may derive the first principal component score x and the second principal component score y after performing processing to make the average and the dispersion complete, respectively, when the first observation object of the first sensor "sensor A" providing the raw data and the second observation object of the second sensor "sensor B" are different from each other (for example, temperature and pressure, etc.) when deriving the trajectory. In this case, for example, the trajectory deriving unit 106 performs PCA processing on the standardized first observation object and the second observation object to derive the first principal component score x and the second principal component score y.

[0068] Although the case where the trajectory deriving unit 106 derives the first principal component score x and the second principal component score y based on the two-variable analysis performed by the principal component score acquisition process of each sensor group is described, it is not limited to this. As long as the trajectory deriving unit 106 can determine the shape of the trajectory in the normal state and the shape of the trajectory in the abnormal state, when processing k sensor information, it can derive the trajectory by combining different principal component analysis results (for example, obtaining the third principal component score from the third principal component, and obtaining the kth principal component score from the kth principal component, etc.) based on other principal components (for example, the first principal component, the second principal component, ... the kth principal component). The trajectory deriving unit 106 can derive the trajectory using the result of the principal component score acquisition process of each sensor.

[0069] The trajectory derivation unit 106 determines the abnormal simulation results of the first sensor (in this example, "sensor A") and the abnormal simulation results of the second sensor (in this example, "sensor B") based on the abnormal simulation information 202. Here, as described above, the abnormal simulation information 202 includes the abnormal simulation results when the interference inputs of interference patterns No. 1 to No. 8 are respectively given to the model that imitates the state of "valve pressure loss abnormality", "fluid torque abnormality", "motor coil abnormality" and "mechanism torque abnormality" of the abnormal diagnosis object. Therefore, the trajectory derivation unit 106 determines eight patterns (i.e., 32) of abnormal simulation results for the four abnormal states as the respective abnormal simulation results of the first sensor and the second sensor.

[0070] The trajectory deriving unit 106 performs PCA processing on each of the determined abnormal simulation results to derive a trajectory (hereinafter referred to as trajectory Orb2). In this example, the trajectory deriving unit 106 derives 32 trajectories Orb2 based on 32 abnormal simulation results. The process of deriving trajectory Orb2 by the trajectory deriving unit 106 is the same as the process of deriving trajectory Orb1 described above, so the description is omitted. Trajectory Orb1 is an example of a "first trajectory", and trajectory Orb2 is an example of a "second trajectory".

[0071] The first difference calculation unit 108 extracts the shapes of all the trajectories derived by the trajectory derivation unit 106, and converts the shapes of the extracted trajectories into complex autoregressive models. The trajectory derivation unit 106, for example, converts the trajectory shown on the phase plane into an image of a predetermined size (or a predetermined number of pixels), and extracts the shape of the trajectory (contour line (edge)) based on the brightness (pixel value) of each pixel when the converted image is formed into grayscale. Then, the first difference calculation unit 108 derives an m-order complex autoregressive model when the point sequence obtained by tracing the contour line of the extracted trajectory is set as equation (1) and its complex expression is set as equation (2). The m-order complex autoregressive model is defined as a model that approximates the contour points by a linear combination of the contour points up to the first m as shown in equation (3).

[0072] [Mathematical formula 1]

[0073] (x j ,y j )(j=0,1,…,N-1)···(1)

[0074] [Mathematical formula 2]

[0075] z j =x j +iy j ···(2)

[0076] [Mathematical formula 3]

[0077]

[0078] Complex autoregressive coefficient

[0079] The first difference calculation unit 108 converts the shape of the trajectory into a complex autoregressive model according to each trajectory derived by the trajectory derivation unit 106. The first difference calculation unit 108 derives the difference (distance) between the complex autoregressive model z1 of a certain trajectory after conversion and the complex autoregressive model z2 of other trajectories other than the trajectory, namely, the complex power cepstrum distance Dc (z1, z2). The complex power cepstrum distance Dc (z1, z2) is the mean square distance of the logarithm of the spectral envelope of the complex autoregressive model z1 and the complex autoregressive model z2, and is expressed by formula (4).

[0080] [Formula 4]

[0081]

[0082] Complex power cepstrum coefficients of complex autoregressive model z1

[0083] Complex power cepstrum coefficients of complex autoregressive model z2

[0084] The first difference calculation unit 108 derives complex power cepstrum distances Dc of all combinations of two selected trajectories from among the trajectory Orb1 and a plurality of (32 in this example) trajectories Orb2. The diagnosis unit 104 classifies the complex power cepstrum distances Dc derived by the first difference calculation unit 108 by, for example, hierarchical clustering. Figure 8 : is a tree diagram showing an example of a result of classification based on the complex power cepstrum distance Dc classified by the diagnosis unit 104. Figure 8 In the figure, the complex power cepstrum distance Dc is represented by the length of the branch of the dendrogram. Generally, there is a tendency that the larger the value of the complex power cepstrum distance Dc, the greater the difference between the two complex autoregressive models (i.e., the types of anomalies are different), and the smaller the value of the complex power cepstrum distance, the smaller the difference between the two complex autoregressive models (i.e., the types of anomalies are the same). Figure 8 Indicates such a tendency. Figure 8 , the complex power cepstrum distance Dc of the complex autoregressive model based on the observation results of the first sensor and the observation results of the second sensor is close to the complex power cepstrum distance Dc of the complex autoregressive model based on the abnormal simulation results of the model that produces the "mechanism torque abnormality". In this case, the diagnosis unit 104 diagnoses the type of abnormality of the abnormality diagnosis object as "mechanism torque abnormality".

[0085] In the above, the first difference calculation unit 108 may be configured to derive Euclidean distance, likelihood ratio distance, or complex power Mel-cepstral distance based on a complex autoregressive model after transforming the shape of the trajectory. Euclidean distance, likelihood ratio distance, and complex power Mel-cepstral distance are examples of “difference”.

[0086] In the above description, the first difference calculation unit 108 is not limited to using a complex autoregressive model in the digitization of the shape of the trajectory. The first difference calculation unit 108 is a method of digitizing the shape of the phase plane trajectory and calculating the degree of difference, in which the complex autoregressive model only extracts the outer shape of the figure. Therefore, when more detailed judgment is required about the shape of the trajectory, the first difference calculation unit 108 calculates the degree of difference using a known method such as a neural network that can extract the internal shape, in addition to a method such as DTW that calculates the degree of difference of the shape of the phase plane trajectory in consideration of the time structure.

[0087] In the above, the trajectory derivation unit 106 and the first difference calculation unit 108 derive the trajectory Orb2 based on the abnormal simulation results of each of the first sensor and the second sensor according to the situation determined by the diagnosis unit 104, and convert it into a complex autoregressive model based on the derived trajectory Orb2, but it is not limited to this. For example, it can be a structure in which a complex autoregressive model is pre-derived for all combinations of two sensors taken out from the sensors set as the first sensor and the second sensor. For example, the storage unit 200 can pre-store information representing the complex autoregressive model of each combination, that is, complex autoregressive model information 208. In this case, the diagnosis unit 104 can determine the complex autoregressive model corresponding to the complex autoregressive model of the trajectory Orb1 derived by the first difference calculation unit 108 based on the complex autoregressive model information 208, and calculate the complex power cepstrum distance Dc. The following describes the case where the complex autoregressive model information 208 is stored in the storage unit 200.

[0088] Thus, the diagnosis unit 104 diagnoses the type of abnormality involved in the complex autoregressive model closest (smaller) to the complex autoregressive model based on the observation result among the complex power cepstrum distances Dc between the complex autoregressive models derived by the first difference calculation unit 108 as the type of abnormality of the abnormal diagnosis object, and can diagnose the abnormality of the abnormal diagnosis object without using information on past adverse events. The diagnosis unit 104 diagnoses the abnormal diagnosis object based on the abnormal simulation results obtained as the output of the model, so that a robust diagnosis can be made for changes in the response of the abnormal diagnosis object caused by bias or time lag of the device or system of the abnormal diagnosis object.

[0089] The diagnosis unit 104 performs classification by a classification method such as hierarchical clustering, and diagnoses the type of abnormality involved in the complex autoregressive model with the closest (smallest) complex power cepstrum distance Dc among the classified clusters as the type of abnormality to be diagnosed as the abnormality, thereby being able to diagnose the type of abnormality by a simpler process. The diagnosis unit 104 performs PCA processing to perform diagnosis based on the observation results of the sensor with reduced noise (in this example, the trajectory represented by the first principal component score x and the second principal component score y), so that the type of abnormality can be diagnosed with higher accuracy.

[0090] [(2) Regarding the method using the dissimilarity distance evaluated by DTW etc.]

[0091] The following describes the content of the process of the diagnosis unit 104 diagnosing the type of abnormality using the method (2) using the dissimilarity distance evaluated by DTW or the like. The diagnosis unit 104 diagnoses the type of abnormality based on the dissimilarity distance evaluated by DTW or the like between the observation result and the abnormal simulation result. The dissimilarity distance evaluated by DTW or the like is a distance scale that can compare the problem point in the Euclidean distance, that is, time series data with different lengths, and can also cope with the deviation in the time axis direction by comparing the best points. First, the diagnosis unit 104 determines the observation object observed by a certain sensor (hereinafter referred to as the first sensor) from the observation result information 300 obtained by the acquisition unit 102. In addition, the diagnosis unit 104 determines the abnormal simulation result corresponding to the determined first observation object in the abnormal simulation information 202. The second difference calculation unit 110 calculates the dissimilarity distance by DTW or the like based on the temporal change of the observation result determined by the diagnosis unit 104 and the temporal change of the abnormal simulation result. In order to make the abnormal state feature more prominent, it is preferable to perform pre-processing such as obtaining the difference between the abnormal simulation result and the observation result and the normal simulation result and performing time differentiation on the difference before calculating the difference distance based on DTW or the like.

[0092] As described above, the abnormal simulation information 202 includes abnormal simulation results when the interference of interference mode No. 1 to interference mode No. 8 is input, which are respectively assigned to the model that imitates the state of "valve pressure loss abnormality", "fluid torque abnormality", "motor coil abnormality" and "mechanism torque abnormality" of the abnormal diagnosis object. Therefore, the second difference calculation unit 110 determines 32 abnormal simulation results as abnormal simulation results corresponding to the observed object. The second difference calculation unit 110 calculates the difference distance between the waveforms involved in all combinations of two selected waveforms based on each waveform representing the time-varying of the determined abnormal simulation result and the waveform representing the time-varying of the determined observation result. The waveform representing the time-varying of the determined observation result is an example of the "first waveform", and the waveform representing the change of the cause of the determined abnormal simulation result is an example of the "second waveform".

[0093] The diagnosis unit 104 classifies the difference distances evaluated by DTW or the like calculated by the first difference calculation unit 108 using a classification method such as hierarchical clustering. Fig. 9 is a tree diagram showing an example of the result of classification by the diagnosis unit 104 based on the degree of difference evaluated by the dynamic time warping method. Fig. 9 In the above diagram, the dissimilarity distance is represented by the length of the branch of the dendrogram. Generally, there is a tendency that the larger the dissimilarity distance value, the greater the difference between the two waveforms (i.e., the types of abnormalities are different), and the smaller the dissimilarity distance value, the smaller the difference between the two waveforms (i.e., the types of abnormalities are the same). Fig. 9Indicates such a tendency. Fig. 9 In the example, the DTW distance between the waveform of the observation result of the first sensor and the waveform of the abnormality simulation result of the model that produces the “mechanism torque abnormality” is close (small). In this case, the diagnosis unit 104 diagnoses the type of abnormality of the abnormality diagnosis target as “mechanism torque abnormality”.

[0094] Thus, the diagnosis unit 104 diagnoses the type of abnormality involving the waveform whose dissimilarity distance among the dissimilarity distances between the waveforms derived by the second difference calculation unit 110 is closest (smaller) to the waveform based on the observation result as the type of abnormality of the abnormal diagnosis object, and is able to diagnose the abnormality of the abnormal diagnosis object without using information on adverse conditions that occurred in the past.

[0095] The diagnosis unit 104 classifies the differences by a classification method such as hierarchical clustering, and diagnoses the type of abnormality associated with the waveform having the closest (smallest) difference distance among the classified clusters as the type of abnormality to be diagnosed, thereby making it possible to diagnose the type of abnormality by simpler processing.

[0096] [Processing Flow]

[0097] Next, the contents of the operation of the abnormality diagnosis device 1 will be described. Fig.10 1 is a flowchart showing a series of processes of the abnormality diagnosis device 1. First, the acquisition unit 102 acquires the observation result information 300 from the log information of the abnormality diagnosis target (step S100). Next, the diagnosis unit 104 specifies a certain observation result in the observation result information 300 acquired by the acquisition unit 102 (step S102).

[0098] The diagnosis unit 104 determines whether the determined observation result can be analyzed by a single variable (step S104). Single variable analysis is an analysis method for diagnosing the type of abnormality of the abnormal diagnosis object based on the observation result involved in a certain type (for example, a certain sensor). In contrast, multivariate analysis is an analysis method for diagnosing the type of abnormality of the abnormal diagnosis object based on multiple observation results. Observation results that can be analyzed by a single variable are, for example, observation results that read changes or abnormalities related to the action of the abnormal diagnosis object through a certain type of observation result, and observation results that produce a sharp change in the observation result within a specified time. Observation results that cannot be analyzed by a single variable (i.e., observation results that require multivariate analysis) are, for example, observation results that read changes or abnormalities related to the action of the abnormal diagnosis object based on the correlation with other observation results, and observation results that have a small change in the observation result within a specified time (i.e., a large time constant).

[0099] When the determined observation result is an observation result that cannot be analyzed by a single variable, the diagnosis unit 104 uses "(1) a method using a dissimilarity distance evaluated by a complex power cepstrum, etc." to diagnose the type of abnormality of the abnormal diagnosis object (step S106). When the determined observation result is an observation result that can be analyzed by a single variable, the diagnosis unit 104 uses "(2) a method using a dissimilarity distance evaluated by DTW, etc." to diagnose the type of abnormality of the abnormal diagnosis object (step S108). The details of the processing of step S106 and step S108 are described later. While the diagnosis unit 104 is diagnosing the abnormal diagnosis object based on all the observation results included in the observation result information 300, the processing of steps S102 to S108 is repeated (step S110). The output unit 112 outputs information indicating the diagnosis result after diagnosis by the diagnosis unit 104 (step S112). The diagnosis result includes, for example, the type of abnormality diagnosed as abnormal in step S106 or step S108. When the abnormality diagnosed as the abnormality diagnosis target does not correspond to any abnormality type, the diagnosis unit 104 may diagnose that the abnormality diagnosis target has no abnormality. In this case, the diagnosis result may indicate that the abnormality diagnosis target has no abnormality.

[0100] [Processing flow: (1) Using a method that uses the dissimilarity distance evaluated by the complex power cepstrum, etc.]

[0101] The following describes in detail the process in which the diagnosis unit 104 diagnoses the type of abnormality of the abnormality diagnosis target using (1) a method using a degree of difference distance evaluated by a complex power cepstrum or the like. Fig.11 Yes means Fig.10Flowchart of a series of processes of the processing of step S106 in . First, the diagnosis unit 104 determines the second observation result that is correlated with the first observation result determined in step S102 based on the correlation information 206 (step S202). When there are multiple second observation results, the diagnosis unit 104 can also analyze them in a summary by reconstructing them into multidimensional data with each observation result as an element. If necessary, one observation result can also be determined as the second observation result, and the other observation results can be processed later. Moreover, if the sensor used as the second observation result is optimized by the method described later, the subsequent analysis can be performed by the best two sensors, so the calculation load can be reduced compared to the case where the phase plane trajectory is created by a combination of more than three sensors. The trajectory derivation unit 106 can perform PCA processing in a summary by reconstructing the two observation results determined by the diagnosis unit 104 into multidimensional data with each observation result as an element (step S204). The trajectory derivation unit 106 can perform PCA processing according to each sensor instead of the structure of performing PCA processing in a summary. Next, the trajectory derivation unit 106 performs PCA processing in a summary according to each sensor group. The trajectory derivation unit 106 derives the trajectory on the phase plane based on the derived multiple principal components (scores) (step S206). The phase plane is, for example, a phase plane with the first principal component score x as the first axis (for example, the horizontal axis) and the second principal component score y as the second axis (for example, the vertical axis) in the case of two variables. The trajectory derivation unit 106 connects the coordinates (coordinates (x1, y1), ..., coordinates (xm, ym) shown in the figure) of the elements with the same detection timing among the elements {x1, x2, ..., xm} included in the first principal component score x and the elements {y1, y2, ..., ym} included in the second principal component score y through a straight line, thereby deriving the trajectory.

[0102] The first difference calculation unit 108 extracts the shapes of all the trajectories derived by the trajectory derivation unit 106, and converts the shapes of the extracted trajectories into complex autoregressive models respectively (step S108). The first difference calculation unit 108 determines a plurality of complex autoregressive models involved in the abnormal simulation results corresponding to the first observation result and the second observation result based on the complex autoregressive model information 208 (step S210). The first difference calculation unit 108 selects two complex autoregressive models from the complex autoregressive models converted in step S208 and the complex autoregressive models determined in step S210, and calculates the complex power cepstrum distance Dc representing the difference between the selected complex autoregressive models (step S212). The first difference calculation unit 108 repeats step S212 until the complex power cepstrum distance Dc of all the combinations of the selected two complex autoregressive models is calculated (step S214).

[0103] The diagnosis unit 104 classifies the complex power cepstrum distance Dc calculated in step S212 by hierarchical clustering (step S216). The result of the hierarchical clustering performed by the diagnosis unit 104 is to determine a complex autoregressive model whose complex power cepstrum distance Dc is smaller than a prescribed value (that is, the type of abnormality is the same) as the complex autoregressive model involved in the observation result (step S218). The diagnosis unit 104 diagnoses the type of abnormality involved in the complex autoregressive model whose complex power cepstrum distance Dc is smaller than a prescribed value as the type of abnormality of the abnormal diagnosis object (step S220). When there is no complex autoregressive model whose complex power cepstrum distance Dc is smaller than a prescribed value than the complex autoregressive model involved in the observation result, the diagnosis unit 104 diagnoses that the abnormal diagnosis object has no abnormality (step S222). The diagnosis unit 104 may determine the complex autoregressive model having the smallest complex power cepstrum distance Dc with the complex autoregressive model related to the observation result, and diagnose the type of abnormality related to the complex autoregressive model as the abnormality to be diagnosed.

[0104] [Processing flow: (2) Method using dissimilarity distance evaluated by DTW etc.]

[0105] The following describes in detail the process in which the diagnosis unit 104 diagnoses the type of abnormality of the abnormality diagnosis target using (2) the method using the degree of difference distance evaluated by DTW or the like. Fig.12 Yes means Fig.10 The flowchart of a series of processes of the processing of step S108. First, the diagnosis unit 104 determines multiple abnormal simulation results of various types of abnormalities corresponding to the first observation result determined in step S102 based on the abnormal simulation information 202 (step S302). The second difference calculation unit 110 selects the determined first observation result and two of the multiple abnormal simulation results, and calculates the difference between the selected observation result or the abnormal simulation result and the time-varying waveforms of the normal state simulation result evaluated by DTW (step S304). The second difference calculation unit 110 repeatedly performs the processing of step S304 until the difference distance of all combinations of the selected first observation result and two of the multiple abnormal simulation results is calculated (step S306).

[0106] The diagnosis unit 104 classifies the complex power cepstrum distance Dc derived in step S304 by hierarchical clustering (step S308). The result of the hierarchical clustering performed by the diagnosis unit 104 is to determine whether there is an abnormal simulation result of a waveform whose complex power cepstrum distance Dc of the waveform involved in the observation result is smaller than a prescribed value (that is, the type of abnormality is the same) (step S310). In the case where there is an abnormal simulation result of a waveform whose DTW distance to the waveform involved in the observation result is smaller than a prescribed value, the diagnosis unit 104 diagnoses the type of abnormality of the abnormal simulation result as the type of abnormality of the abnormal diagnosis object (step S312). In the case where there is no abnormal simulation result of a waveform whose DTW distance to the waveform involved in the observation result is smaller than a prescribed value, the diagnosis unit 104 diagnoses that there is no abnormality in the abnormal diagnosis object (step S314). The diagnosis unit 104 can determine the waveform with the smallest DTW distance involved in the observation result, and diagnose the type of abnormality involved in the abnormal simulation result of the waveform as the abnormality of the abnormal diagnosis object.

[0107] [When performing univariate analysis and multivariate analysis]

[0108] In the above, it is explained that Fig.10 In step S104, the diagnosis unit 104 determines whether the observation result can be analyzed by a single variable. If the single variable analysis cannot be performed, the abnormality is diagnosed by using "(1) a method using a dissimilarity distance evaluated by a complex power cepstrum, etc."; if the single variable analysis can be performed, the abnormality is diagnosed by using "(2) a method using a dissimilarity distance evaluated by DTW, etc.", but the present invention is not limited thereto. The diagnosis unit 104 may be configured to diagnose the abnormality by using "(1) a method using a dissimilarity distance evaluated by a complex power cepstrum, etc." (i.e., step S106) and "(2) a method using a dissimilarity distance evaluated by DTW, etc." (i.e., step S108) for the observation result determined in step S102. For example, when the diagnosis unit 104 diagnoses the abnormality by using both the method (1) and the method (2), the abnormality can be diagnosed even when it is not known in advance whether the observation result can be analyzed by a single variable.

[0109] <Modification>

[0110] A modified example of the present embodiment is described below. In the above-mentioned embodiment, a case is described in which the abnormal diagnosis device 1 diagnoses the abnormality of the abnormal diagnosis object based on all abnormal simulation results involving the first observation result and the second observation result. In the modified example, a case is described in which the abnormal diagnosis device 1a uses the phase plane trajectory formation contribution rate (hereinafter, abbreviated as "phase plane trajectory formation contribution rate") using the main component load to diagnose the abnormality of the abnormal diagnosis object. For the same structure as the above-mentioned embodiment, the same symbol is marked and the description is omitted. This method can also be used as a method for reducing the calculation load by reducing the sensor pairs used for abnormal diagnosis to a group of high-sensitivity sensor pairs (or, in the case of a single variable analysis in which a phase plane trajectory is not generated, a single sensor with high sensitivity) when there are multiple sensors representing abnormal values.

[0111] Fig.13 1 is a diagram showing an example of the structure of an abnormality diagnosis device 1a of a modified example. The abnormality diagnosis device 1a of the modified example includes a control unit 100a and a storage unit 200 instead of (or in addition to) the structure included in the abnormality diagnosis device 1. The control unit 100a includes an acquisition unit 102, a diagnosis unit 104, a trajectory derivation unit 106, a first difference calculation unit 108, a second difference calculation unit 110, an output unit 112, and a contribution rate calculation unit 114 instead of (or in addition to) the structure included in the control unit 100.

[0112] As described above, the diagnosis unit 104 determines, for example, all abnormal simulation results corresponding to the first observation result and the second observation result, and performs diagnosis on the abnormal diagnosis object based on the trajectory Orb2 derived from the determined abnormal simulation result or the dissimilarity distance evaluated by DTW, etc. Here, the abnormal simulation result sometimes does not highly sensitively represent the abnormality of the abnormal diagnosis object. The contribution rate calculation unit 114 calculates the degree of contribution of each sensor that forms the phase plane trajectory involved in the diagnosis of the abnormal diagnosis object for each abnormal simulation result (i.e., the phase plane trajectory forms a contribution rate). The diagnosis unit 104 of the modified example determines that the result is highly sensitive to represent the abnormality of the abnormal diagnosis object when the phase plane trajectory formation contribution rate calculated by the contribution rate calculation unit 114 is high, and performs diagnosis on the abnormal diagnosis object based on the trajectory Orb2 derived from the determined abnormal simulation result and the dissimilarity distance.

[0113] The following describes the content of the processing of the contribution rate calculation unit 114 when determining the best abnormal simulation result that highly sensitively represents the abnormality using the phase plane trajectory to form the contribution rate. First, the contribution rate calculation unit 114 uses the abnormal simulation results related to the k sensors for a certain abnormality to perform PCA processing by the aforementioned method to calculate the first principal component to the kth principal component. The contribution rate calculation unit 114 calculates the cumulative contribution rate of each principal component, for example, and selects the principal component with a high contribution rate (for example, the first principal component and the second principal component) based on the calculated cumulative contribution rate. The cumulative contribution rate of each principal component is the value obtained by adding the value obtained by dividing the eigenvalue of each principal component by the sum of the eigenvalues ​​of each principal component (the contribution rate of each principal component). The higher the contribution rate value, the less information is lost due to the dimensionality reduction effect of the PCA processing. In addition, the contribution rate calculation unit 114 calculates the contribution rate of the phase plane trajectory when the phase plane trajectory is formed using the principal component (for example, the first principal component and the second principal component) selected based on the cumulative contribution rate of each principal component. The principal component load is the correlation coefficient between the principal component score and each factor (each factor in this example is the observation value of each sensor). The sensor with a higher contribution rate to the formation of the phase plane trajectory is more contributed to the formation of the phase plane trajectory derived using the principal component score, and the more contributed to the extraction of features using the phase plane trajectory. If there are k sensors, the first principal component load, the second principal component load, ..., the kth principal component load are obtained according to each sensor. For example, when there are two sensors measuring abnormalities, the contribution rate calculation unit 114 can generate a phase plane trajectory using the first principal component and the second principal component, so that the value of the sum of the square of the first principal component load and the square of the second principal component load can be obtained as the phase plane trajectory formation contribution rate of each sensor involved in the formation of the phase plane trajectory.

[0114] Fig.14 This is a diagram showing an example of contribution rate. Fig.14 The vertical axis shows the value of the second principal component loading (the larger the correlation coefficient between the principal component score and each factor, the greater the contribution to the formation of the principal component score), and the horizontal axis shows the value of the first principal component loading. Fig.14 This is a graph showing the values ​​of the first principal component load and the second principal component load of each sensor involved in the abnormal simulation result when a certain abnormality occurs and the sensors A to C are the three sensors measuring the abnormality, respectively, as coordinates and expressed as points (points A to C). This abnormal simulation result is the abnormal simulation result when the principal component score of each sensor group is obtained using the three sensor information to form a phase plane trajectory. Fig.14Among the contributions calculated by the contribution rate calculation unit 114, sensor C has the highest contribution and sensor A has the lowest contribution. Sometimes, only based on the ratio of the contribution rates formed by the phase plane trajectories of the three sensors, the abnormal simulation result that highly sensitively represents the abnormal case can be uniquely determined. Moreover, in the case where multiple sensors are associated with the abnormality, the contribution rate is formed using the phase plane trajectory, and the number of sensors of interest is limited, thereby reducing the number of sensors that perform PCA processing, thereby reducing the overall calculation load. Moreover, the result can also be used in data assimilation and parameter estimation when constructing a simulation model for deriving abnormal simulation results. Moreover, the diagnosis unit 104, for example, in the abnormal case, focuses on the trajectory of sensor C, thereby being able to sense the abnormality with high sensitivity, so by using a combination including the sensor to perform the aforementioned analysis, it is possible to reduce misdiagnosis and efficiently perform abnormal diagnosis. It also serves as a reference for judging the accuracy or appropriateness of abnormal simulation results. For example, when the abnormal simulation result of the abnormality that highly sensitively indicates the abnormal case can be uniquely determined based on the ratio of the phase plane trajectory formation contribution rate, the abnormal simulation result with similar ratio of the phase plane trajectory formation contribution rate between sensors among multiple simulation results with different abnormal causes is selected as the abnormal simulation result used in the diagnosis of the abnormal diagnosis object, and is used for the diagnosis of the abnormal diagnosis object, thereby improving the reliability and diagnostic efficiency of the abnormal diagnosis. As described above, by optimizing the sensor used to derive the abnormal simulation result that highly sensitively indicates the abnormality using the phase plane trajectory formation contribution rate, the abnormal simulation result is optimized, thereby reducing the load involved in the processing of the diagnosis unit 104, the trajectory deriving unit 106, the first difference calculation unit 108, and the second difference calculation unit 110, and performing accurate diagnosis. Sensors with the same ratio of the phase plane trajectory formation contribution rate are estimated to have a strong correlation and causal relationship, so this result can also be used in the preparation of the correlation information 206.

[0115] [Detailed inspection of sensors installed in abnormality diagnosis targets]

[0116] In the above, it is explained that the phase plane trajectory formation contribution rate calculated by the contribution rate calculation unit 114 is used in the optimization of the abnormal simulation result for deriving the abnormal simulation result that highly sensitively represents the abnormality of the diagnosis unit 104, but the embodiment of the present invention is not limited to this. A sensor with a low phase plane trajectory formation contribution rate calculated by the contribution rate calculation unit 114 may be a sensor that directly intervenes in the cause of the abnormality even when the information related to the operation of the device or system of the abnormal diagnosis object is not observed with high sensitivity. The phase plane trajectory formation contribution rate of such a sensor can also be used in the case of determining whether the corresponding abnormal simulation result exists based on the sensor information. In addition, the output unit 112 outputs information representing the phase plane trajectory formation contribution rate calculated by the contribution rate calculation unit 114, which can promote the re-evaluation of the sensor configured for the abnormal diagnosis object for the designer of the device or system of the abnormal diagnosis object.

[0117] [Diagnosis of abnormal diagnosis object based on normal simulation results]

[0118] In the above, the case where the trajectory derivation unit 106, the first difference calculation unit 108, and the second difference calculation unit 110 calculate the trajectory Orb2 and the dissimilarity distance based on the abnormal simulation result shown in the abnormal simulation information 202 is described, but it is not limited to this. The trajectory derivation unit 106, the first difference calculation unit 108, and the second difference calculation unit 110 can calculate the trajectory Orb2 and the dissimilarity distance based on the normal simulation result shown in the normal simulation information 204. In this case, the diagnosis unit 104 can diagnose that the abnormal diagnosis object has no abnormality when the complex power cepstrum distance Dc between the complex autoregressive model based on the observation result and the complex autoregressive model based on the normal simulation result is smaller than a prescribed value (that is, the complex power cepstrum distance Dc is close). The diagnosis unit 104 can diagnose that the abnormal diagnosis object has no abnormality when the DTW distance between the waveform representing the time-varying observation result and the waveform representing the time-varying normal simulation result is smaller than a prescribed value (that is, the DTW distance is close). This method can be used when the purpose is to sense abnormality. However, here, when an abnormality is sensed, in order to diagnose the cause of the abnormality, it is sometimes necessary to compare with the above-mentioned abnormality simulation result.

[0119] As mentioned above, although the form for implementing this invention was demonstrated using embodiment, this invention is not limited to such embodiment at all, Various deformation|transformation and substitution can be added in the range which does not deviate from the summary of this invention.

[0120] Description of Reference Numerals

[0121] 1, 1a…abnormality diagnosis device, 100, 100a…control unit, 102…acquisition unit, 104…diagnosis unit, 106…trajectory derivation unit, 108…first difference calculation unit, 110…second difference calculation unit, 112…output unit, 114…contribution rate calculation unit, 200…storage unit, 202…abnormality simulation information, 204…normal simulation information, 206…correlation information, 208…complex autoregression model information, 300…observation result information.

Claims

1. An abnormality diagnosis device, wherein: The abnormality diagnosis device comprises: a diagnostic unit that diagnoses the type of abnormality occurring in the abnormality diagnosis object based on a difference between abnormality simulation results of the various types of abnormalities obtained by simulating the abnormalities of the abnormality diagnosis object and a plurality of time series observation results obtained by time series observation of the abnormality diagnosis object by a plurality of detection units; and a trajectory deriving unit, which reconstructs multidimensional data at each observation time according to each sensor or each sensor group for the plurality of time series observation results, and derives, based on two analysis results among the analysis results obtained by performing principal component analysis on the reconstructed multidimensional data, a first trajectory on a phase plane in which a principal component involved in one of the time series observation results is set as a first axis, a principal component analysis result involved in another time series observation result is set as a second axis, and corresponding elements are set as coordinates, according to each combination of the two analysis results; The diagnosis unit diagnoses the type and cause of the abnormality based on the abnormal simulation result of the combination that is consistent with the combination of the observation results involved in the first trajectory among the plurality of abnormal simulation results, by setting the principal component of the abnormal simulation result corresponding to the first axis involved in the first trajectory as the first axis, setting the principal component of the abnormal simulation result corresponding to the second axis involved in the observation results involved in the first trajectory as the second axis, and setting the corresponding element as the difference between the shape of the second trajectory and the shape of the first trajectory on the phase plane of the coordinates, The trajectory deriving unit derives the first trajectory for the time series observation results having a correlation among the plurality of time series observation results. The time series observation results having the correlation relationship are observation results of the same type of observation object, or a combination of observation results of a certain observation object and observation results of a different type of observation object that changes according to changes in the certain observation object. The abnormality diagnosis device further includes a first difference calculation unit, which converts the shape of the second trajectory and the shape of the first trajectory of each type of the abnormality through a complex autoregressive model, and calculates a dissimilarity distance between the converted complex autoregressive models, wherein the complex autoregressive model is defined as a model that approximates a point sequence obtained by respectively tracking the shape of the second trajectory and the shape of the first trajectory by linear combination, and the dissimilarity distance is a complex power cepstrum distance, The diagnosis unit specifies the shape of the second trajectory having the smallest shape difference distance from the first trajectory among the difference distances calculated by the first difference calculation unit, and diagnoses the type of abnormality related to the specified second trajectory as the type of abnormality to be diagnosed.

2. The abnormality diagnosis device according to claim 1, wherein: The abnormality diagnosis device further includes a second difference calculation unit, which calculates, by a dynamic time warping method, the difference distances between a first waveform representing the time variation of the observation result set as the object of abnormality diagnosis and a plurality of second waveforms representing the time variation of the abnormality simulation results of the plurality of time series observation results according to the various types of the abnormalities, among the plurality of time series observation results, respectively. The diagnosis unit identifies the second waveform having the smallest difference distance with the first waveform among the difference distances calculated by the second difference calculation unit among the plurality of second waveforms, and diagnoses the type of abnormality related to the identified second waveform as the type of abnormality to be diagnosed.

3. The abnormality diagnosis device according to claim 1, wherein: The diagnosis unit classifies the dissimilarity distances using a classification method, and diagnoses the type of abnormality related to the abnormality simulation result of the cluster having the smallest information dissimilarity distance with the time series observation result among the classified clusters as the type of abnormality to be diagnosed as the abnormality.

4. The abnormality diagnosis device according to claim 1, wherein: The abnormality diagnosis device further includes a contribution rate calculation unit, which calculates the contribution rate involved in the formation of the phase plane trajectory of each sensor included in the time series observation result and the abnormality simulation result based on the time series observation result and the principal component load obtained by multivariate time series principal component analysis of the abnormality simulation result. The diagnosis unit diagnoses the type of the abnormality using the abnormality simulation result related to the contribution rate when the contribution rate calculated by the contribution rate calculation unit is higher than a prescribed value, and does not diagnose the type of the abnormality using the abnormality simulation result related to the contribution rate when the contribution rate is less than the prescribed value.

5. A method for abnormality diagnosis, wherein: include: The computer diagnoses the type of abnormality generated in the abnormality diagnosis object based on the difference between the abnormality simulation results of the various types of abnormalities obtained by simulating the abnormalities of the abnormality diagnosis object and the multiple time-series observation results obtained by time-series observation of the abnormality diagnosis object by multiple detection units; The computer reconstructs the multidimensional data at each observation time according to each sensor or each sensor group for the plurality of time series observation results, and based on two analysis results among the analysis results obtained by performing principal component analysis on the reconstructed multidimensional data, derives a first trajectory on a phase plane in which a principal component involved in one of the time series observation results is set as a first axis, a principal component analysis result involved in another time series observation result is set as a second axis, and corresponding elements are set as coordinates according to each combination of the two analysis results; and The computer diagnoses the type and cause of the abnormality based on the abnormality simulation result of the combination that is consistent with the combination of the observation results involved in the first trajectory among the plurality of abnormality simulation results, by setting the principal component of the abnormality simulation result corresponding to the first axis involved in the first trajectory as the first axis, setting the principal component of the abnormality simulation result corresponding to the second axis of the observation results involved in the first trajectory as the second axis, and setting the corresponding elements as the difference between the shape of the second trajectory and the shape of the first trajectory on the phase plane of the coordinate. deriving the first trajectory for the time series observation results having a correlation among the plurality of the time series observation results, The time series observation results having the correlation relationship are observation results of the same type of observation object, or a combination of observation results of a certain observation object and observation results of a different type of observation object that changes according to changes in the certain observation object. The method further includes: transforming the shapes of the second trajectories and the shapes of the first trajectories of the various types of the abnormalities respectively through a complex autoregressive model, and calculating the dissimilarity distance between the transformed complex autoregressive models; and The shape of the second trajectory having the smallest shape difference distance from the first trajectory among the calculated difference distances is determined, and the type of abnormality related to the determined second trajectory is diagnosed as the type of abnormality to be diagnosed. The complex autoregressive model is defined as a model that approximates a point sequence obtained by respectively tracking the shape of the second trajectory and the shape of the first trajectory by linear combination, and the dissimilarity distance is a complex power cepstrum distance.

6. A storage medium storing the following program, wherein: The procedures include: causing a computer to diagnose the type of abnormality occurring in the abnormality diagnosis object based on a difference between abnormality simulation results of the various types of abnormalities obtained by simulating the abnormalities of the abnormality diagnosis object and a plurality of time-series observation results obtained by time-series observation of the abnormality diagnosis object by a plurality of detection units; The computer is caused to reconstruct the multidimensional data at each observation time according to each sensor or each sensor group for the plurality of the time series observation results, and based on two of the analysis results obtained by performing principal component analysis on the reconstructed multidimensional data, a first trajectory on a phase plane in which a principal component involved in one of the time series observation results is set as a first axis, a principal component analysis result involved in another time series observation result is set as a second axis, and corresponding elements are set as coordinates is derived according to each combination of the two analysis results; The computer is caused to diagnose the type and cause of the abnormality based on the abnormal simulation result of the combination that is consistent with the combination of the observation results involved in the first trajectory among the plurality of abnormal simulation results, by setting the principal component of the abnormal simulation result corresponding to the first axis involved in the first trajectory as the first axis, setting the principal component of the abnormal simulation result corresponding to the second axis involved in the observation results involved in the first trajectory as the second axis, and setting the corresponding element as the difference between the shape of the second trajectory and the shape of the first trajectory on the phase plane of the coordinate system, The program derives the first trajectory for the time series observation results having a correlation among the plurality of the time series observation results, The time series observation results having the correlation relationship are observation results of the same type of observation object, or a combination of observation results of a certain observation object and observation results of a different type of observation object that changes according to changes in the certain observation object. Further, the shapes of the second trajectories and the shapes of the first trajectories of the various types of the abnormalities are respectively transformed by complex autoregressive models, and the dissimilarity distances between the transformed complex autoregressive models are calculated. The shape of the second trajectory having the smallest shape difference distance from the first trajectory among the calculated difference distances is determined, and the type of abnormality related to the determined second trajectory is diagnosed as the type of abnormality to be diagnosed. The complex autoregressive model is defined as a model that approximates a point sequence obtained by respectively tracking the shape of the second trajectory and the shape of the first trajectory by linear combination, and the dissimilarity distance is a complex power cepstrum distance.

Images