Equipment abnormality diagnosis system and equipment abnormality diagnosis method

Abstract

To reduce a processing load and improve diagnostic accuracy by allowing efficient and accurate detection of abnormalities and estimating causes using a minimum number of sensors.SOLUTION: An equipment abnormality diagnosis system 1 comprises: a data preprocessing unit 21 that adjusts the number of dimensions, which is the number of individual data included in target data to be diagnosed, to a specific number of dimensions; a data analysis unit 23 that performs topological data analysis on the target data adjusted to the specific number of dimensions by using a mapper, which is an algorithm that handles data with the specific number of dimensions; and a result display control unit 26 that displays analysis results using the mapper in a cluster display or time series display in a manner that makes it possible to identify the presence or absence of an abnormality, wherein the data preprocessing unit 21 has a function of determining a correlation between multiple sensors and selecting the data to be diagnosed based on the strength of the correlation.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] FIELD An embodiment of the present invention relates to a device abnormality diagnosis technique. [Background technology]

[0002] If equipment installed in a plant stops or breaks down unplanned, it will affect the operation of the plant and cause significant damage. Therefore, it is necessary to monitor the condition of the equipment and properly evaluate its health. Furthermore, if the evaluation reveals signs of an abnormality, it is necessary to clarify the cause and take measures as soon as possible.

[0003] Plant equipment includes rotating machinery such as motors and pumps, as well as piping and valves, all of which are controlled by control signals from the control system via cables. The health of these devices is assessed by measuring vibration, temperature, current, pressure, flow rate, sound, and other values ​​depending on the type of device, and monitoring changes in these values. However, due to the large number of devices and monitoring items within a plant, it is difficult for humans to accurately diagnose each device. In addition, some devices have significant noise or individual differences, requiring the experience and skill of an engineer to properly diagnose them.

[0004] For this reason, methods for efficient diagnosis that do not depend on the skill level of the engineer have been considered, and in recent years, progress has been made in developing efficient and effective abnormality diagnosis technologies using diagnostic methods that use AI (Artificial Intelligence).For example, a technology is known that collects and analyzes large amounts of time-series data from numerous sensors installed in large-scale, complex systems, and uses AI to determine whether or not an abnormality exists. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 2017-021702 Summary of the Invention [Problem to be solved by the invention]

[0006] For measurement data without knowledge of abnormalities, it is difficult to determine whether something is normal or abnormal using only one signal due to the influence of noise and operating conditions, and it is also difficult to identify features related to abnormalities. Therefore, it is possible to effectively determine abnormalities by removing the influence of noise and operating conditions from the relationship between two signals.

[0007] The method in Patent Document 1 detects anomalies by monitoring changes in the relationships between signals acquired by each sensor. However, because the relationships between multiple sensors are monitored by brute-force investigation, the number of data combinations becomes enormous, increasing the cost and processing time of the processing device.

[0008] Furthermore, when adding new sensors to enable monitoring in order to reduce inspection labor, the cost required to install the sensors and processing equipment becomes an issue, so optimization is required. However, when installing new sensors, past data is required to evaluate the relationship between the data.

[0009] Furthermore, for events where no abnormality has been experienced in the past, it is difficult to identify the source of the abnormality (location of the abnormality) from changes in each piece of data and its relationship.

[0010] The embodiments of the present invention have been made in consideration of these circumstances, and aim to enable efficient and accurate detection of abnormalities and estimation of their causes using a minimum number of sensors, thereby reducing the processing load and improving diagnostic accuracy. [Means for solving the problem]

[0011] An equipment abnormality diagnosis system according to an embodiment of the present invention comprises: a data pre-processing unit that adjusts the number of dimensions, which is the number of individual data items included in target data to be diagnosed, of at least one of time-series data and frequency spectrum data that indicate the relationship between multiple sensors provided in at least one piece of equipment, to a specific number of dimensions; a data analysis unit that performs topological data analysis on the target data adjusted to the specific number of dimensions using a mapper, which is an algorithm that handles data with the specific number of dimensions; and a result display control unit that displays the analysis results using the mapper in a cluster display or a time-series display that allows the presence or absence of an abnormality to be identified, and the data pre-processing unit has a function of determining the correlation between the multiple sensors and selecting data to be diagnosed based on the strength of the correlation. [Effects of the Invention]

[0012] According to the embodiment of the present invention, it is possible to efficiently and accurately detect an abnormality and estimate its cause using a minimum number of sensors, thereby reducing the processing load and improving the accuracy of diagnosis. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a block diagram showing a device abnormality diagnosis system according to an embodiment of the present invention; [Figure 2] FIG. 1 is a schematic diagram showing an example of monitoring data. [Figure 3] FIG. 10 is an explanatory diagram showing a first example of a coupling installation defect as an example of an abnormality model. [Figure 4] FIG. 10 is an explanatory diagram showing a second example of a coupling installation defect as an example of an abnormality model. [Figure 5] FIG. 10 is an explanatory diagram showing an example of a change in data when an abnormality occurs. [Figure 6] FIG. 3 is a block diagram showing a physical model unit. [Figure 7] FIG. 2 is a block diagram showing an analysis processing unit. [Figure 8] FIG. 2 is a block diagram showing a data preprocessing unit. [Figure 9] FIG. 4 is a block diagram showing a processing setting unit. [Figure 10] FIG. 4 is a block diagram showing a parameter tuning unit. [Figure 11] FIG. 4 is a block diagram showing a result display control unit. [Figure 12] FIG. 4 is a block diagram showing a display adjustment unit. [Figure 13] FIG. 4 is a block diagram showing an abnormality cause analysis unit. [Figure 14] FIG. 10 is an explanatory diagram showing an example of pre-processed acceleration amplitude data. [Figure 15] FIG. 10 is an explanatory diagram showing an example of transfer function data after preprocessing. [Figure 16] FIG. 10 is a screen diagram showing an example of a cluster display. [Figure 17] FIG. 10 is a screen diagram showing an example of a time series display. [Figure 18] 10 is a table showing an example of the correlation between parameters and analysis results. [Figure 19] An illustration of how clustering can be used to infer the cause of an anomaly. [Figure 20] FIG. 10 is an explanatory diagram showing the analysis results using hierarchical clustering. [Figure 21] FIG. 10 is an explanatory diagram showing the analysis results using hierarchical clustering. [Figure 22] FIG. 10 is an explanatory diagram showing the analysis results using hierarchical clustering. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, embodiments of a device abnormality diagnosis system and a device abnormality diagnosis method will be described in detail with reference to the drawings.

[0015] 1, reference numeral 1 denotes an equipment abnormality diagnosis system of this embodiment. This equipment abnormality diagnosis system 1 diagnoses whether or not there is an abnormality in equipment installed in a facility such as a power plant.

[0016] For example, as shown in Figure 2, assume there is a given tank 60 and a controller 61 that controls the water level in the tank. This tank 60 is equipped with a water level meter 62 as a sensor. Suppose a malfunction of a flow control valve 63, which adjusts the flow rate of water supplied to the tank 60, causes an abnormality, such as a drop in the water level in the tank, or there are signs of such an abnormality. In such a case, the relationship between the opening signal from the controller 61 and the flow rate discharged from the flow control valve 63 tends to differ from that under normal conditions. If the difference between normal and abnormal conditions is large, it is easy to detect the abnormality, but if the difference is small, it is difficult to detect the abnormality. Furthermore, the amount of data related to phenomena that occur during equipment operation is enormous, making it difficult to manually evaluate such a large amount of data.

[0017] Furthermore, while it is easy to identify abnormalities in a simple system like the one shown in Figure 2, it is difficult to do so in a large-scale plant, requiring a large number of sensors. For example, suppose there is a branch in the pipe between the tank 60 and the flow control valve 63, which is connected to another system. In this case, it is impossible to determine whether a change in the relationship between the opening signal from the controller 61 and the flow rate discharged from the flow control valve 63 is due to an abnormality in the flow control valve 63 or an abnormality in another system. Identifying such an abnormality requires adding sensors to acquire data, which is extremely costly. Furthermore, depending on the location of the sensors, the correlation between the acquired data may be weak, and there is a risk of overlooking a serious abnormality.

[0018] In this embodiment, a function between two signals is calculated for data indicating the status of equipment in a facility, and a combination of two signals with a strong correlation is selected to evaluate the characteristics between two signals that are physically related, thereby enabling efficient and effective improvement in diagnostic accuracy.

[0019] The characteristics between the two signals may be evaluated by a physical model simulation.By evaluating the behavior of the equipment when an abnormality occurs in advance through simulation, a diagnostic technology is provided that can extract data feature amounts acquired when an abnormality occurs for equipment that has not experienced an abnormality in the past, and evaluate the presence or absence of abnormality symptoms or the cause of the abnormality.

[0020] In this case, the feature quantities of the data evaluated in the simulation may also include those related to the relationship between the data of the sensors. In this way, it is possible to grasp in advance the changes in the relationship between the data acquired by each sensor in response to a predicted abnormality, and to identify the cause from the data at the time of the abnormality occurrence.

[0021] In this way, by understanding changes in the relationships between data in advance, it is possible to select sensors that can detect anomalies and identify their causes, making it possible to operate at the lowest possible cost.

[0022] In recent years, advances in machine learning and artificial intelligence technologies have made it possible to extract useful information from large amounts of data. Examples of such data analysis algorithms include topological data analysis (TDA) and clustering. The equipment abnormality diagnosis system 1 of this embodiment diagnoses equipment abnormalities using an algorithm that can handle large amounts of data.

[0023] Conventional analytical techniques require dimensionality reduction when dealing with large-scale, multidimensional data. For example, dimension reduction using statistical methods can result in the loss of information in the data. Furthermore, when it comes to anomaly diagnosis using supervised learning, building a learning model can be difficult if there are only a few anomaly data labels.

[0024] Therefore, in this embodiment, data analysis can be performed without a teacher, and abnormalities are diagnosed from within a data group using a data analysis technique that focuses on the topological characteristics of the data group. Note that with regard to supervised data analysis, there is a method that uses a learning model of a neural network and deep learning, using the results of a simulation of a certain abnormality factor as teaching data.

[0025] For example, conventional topological data analysis requires manual judgment to set analytical parameters or interpret analytical results. In contrast, this embodiment sweeps analytical parameters used in topological data analysis and analyzes the corresponding output results (e.g., number of data, number of edges, number of node groups), thereby extracting characteristic data (abnormal data). It is also possible to estimate which parameters contribute to an abnormal event.

[0026] As shown in FIG. 1, a user of the equipment abnormality diagnosis system 1 attaches at least one inspection sensor 2 to equipment installed in a plant to acquire data. This inspection sensor 2 monitors the status of the equipment to acquire monitoring data, which is then input into the equipment abnormality diagnosis system 1. The monitoring data includes data on flow rate, current, pressure, temperature, vibration, voltage, etc. The inspection sensor 2 may be permanently installed or may be installed each time a measurement is made. The equipment to be diagnosed may be one or multiple.

[0027] The monitoring data is at least one of time waveform data and frequency spectrum data of a signal generated from an inspection sensor 2 attached to the equipment. The monitoring data may be the absolute value of the time waveform or frequency spectrum of the signal, or a quantity representing the characteristics of the signal, or a combination thereof. Transfer functions, correlation functions (coherence), phases, etc., which are used to express the relationship between two data, may also be used as monitoring data.

[0028] The user can select any measurement conditions for acquiring the state quantity data, such as the number of sampling points or frequency range. In this way, equipment abnormalities can be diagnosed based on data that can be easily acquired by the inspection sensor 2.

[0029] For example, in a system for controlling the water level in a tank as shown in Figure 2, monitoring data includes a correlation function between the signal data of the opening signal from the controller 61 and the flow rate discharged from the flow control valve 63. Note that the monitoring data includes not only continuous data obtained by constantly monitoring the state of the equipment, but also intermittent data measured periodically. Furthermore, although this embodiment uses opening signal data and flow rate data as an example, a combination of two other signal data may also be used.

[0030] The monitoring data includes primary data indicating the main objective of the device's operation, as well as secondary data that is generated as a result of the device's operation. For example, in the case of the flow control valve 63, the discharge flow rate is the primary data. Data other than the main objective of the device's operation is the secondary data. In this embodiment, the primary data may be used to diagnose an abnormality, or the secondary data, or a combination of these may be used.

[0031] This monitoring data is then input to the equipment abnormality diagnosis system 1 and analyzed. For example, the monitoring data is input to the equipment abnormality diagnosis system 1 via a user terminal 3 (for example, a PC) connected to the Internet. Note that the monitoring data may also be input to the equipment abnormality diagnosis system 1 directly.

[0032] This equipment abnormality diagnosis system 1 may be implemented on a cloud server outside the plant, or on a personal computer or server within the plant. Furthermore, after monitoring data is collected in an electric device such as a server or in a predetermined memory, the collected data may be transferred to the equipment abnormality diagnosis system 1.

[0033] The device abnormality diagnosis system 1 of this embodiment is configured as a computer having hardware resources such as a CPU (Central Processing Unit), GPU (Graphics Processing Unit), ROM (Read Only Memory), RAM (Random Access Memory), HDD (Hard Disk Drive), and SSD (Solid State Drive), and the CPU executes various programs to realize software-based information processing using the hardware resources. Furthermore, the device abnormality diagnosis method of this embodiment is realized by having the computer execute the various programs.

[0034] Next, the system configuration of the device abnormality diagnosis system 1 will be described with reference to the block diagrams shown in Fig. 1 and Fig. 6 to Fig. 13. Note that the arrows in the block diagrams are examples showing the flow of processes, and there may be other flow of processes besides the arrows. Furthermore, the order of processes is not necessarily fixed, and the order of some processes may be interchanged. Furthermore, some processes may be executed in parallel with other processes.

[0035] As shown in FIG. 1, the equipment abnormality diagnosis system 1 includes a main control unit 4, a communication unit 5, an information input unit 6, a physical model unit 7, an analysis processing unit 8, and a display output unit 9.

[0036] At least some of the functions of the main control unit 4, the physical model unit 7, and the analysis processing unit 8 are realized by the CPU executing programs stored in the memory, HDD, or SSD.

[0037] Furthermore, the equipment abnormality diagnosis system 1 includes a monitoring database 10, a process database 11, a specification database 12, a simulation database 13, and an abnormality cause database 14. These are collections of information stored in memory, HDD, or SSD, and organized so that they can be searched or accumulated.

[0038] The main control unit 4 controls the device abnormality diagnosis system 1 in an overall manner.

[0039] Predetermined information is input to the information input unit 6 in response to operations by a user of the equipment abnormality diagnosis system 1. This information input unit 6 includes input devices such as a mouse, keyboard, and touch panel. That is, predetermined information is input to the equipment abnormality diagnosis system 1 in response to operations of these input devices. Predetermined information such as monitoring data and model parameters is input to this information input unit 6.

[0040] The communication unit 5 communicates with electronic devices such as the user terminal 3 via a communication line such as the Internet. The communication unit 5 may be mounted on a wireless communication device. The communication unit 5 may also be mounted on a predetermined network device, for example, a wireless LAN access point or antenna. The communication unit 5 may also communicate with the user terminal 3 via a WAN (Wide Area Network) or a mobile communication network.

[0041] The display output unit 9 displays various information showing the analysis results. For example, a screen such as a cluster display (see FIG. 16) or a time series display (see FIG. 17) is displayed. In other words, the display output unit 9 is a graphic user interface (GUI) for displaying the analysis results.

[0042] The equipment abnormality diagnosis system 1 of this embodiment includes a display device such as a display that outputs analysis results. In other words, the display output unit 9 controls the screen displayed on the display. The display may be separate from the computer main body or may be integrated with the computer main body. Additionally or alternatively, the display output unit 9 may control images displayed on a display provided in another computer connected via a network.

[0043] In this embodiment, a display is exemplified as a device for displaying an image, but other modes are also possible. For example, an image may be displayed using a head-mounted display or a projector. Furthermore, a printer that prints information on a paper medium may be used instead of a display. In other words, the objects controlled by the display output unit 9 may include a head-mounted display, a projector, or a printer.

[0044] The monitoring database 10 stores monitoring data that a user acquires from equipment using an inspection sensor 2.

[0045] The process database 11 stores process data indicating processes related to equipment. The process data is collected separately from the monitoring data and includes, for example, operating conditions such as rotation speed, flow rate, pressure, temperature, and current, and usage conditions such as operating period and usage period. This process data may be stored by the user or may be stored in advance.

[0046] Specification data indicating the specifications of the equipment is stored in the specification database 12. For example, in the case of a rotating machine, the specification data includes information such as the model, manufacturer, size, head, flow rate type, bearing type, and number of blades. Note that this specification data may be stored by the user or may be stored in advance.

[0047] The analysis processing unit 8 performs operations required for calculating the Fourier transform (Fast Fourier Transform: FFT), envelope processing, averaging, transfer function, correlation function, phase, and the like.

[0048] Furthermore, the analysis processing unit 8 analyzes target data or past data contained in the monitoring data using an algorithm that handles data with a specific number of dimensions (see FIG. 6). Note that in this embodiment, monitoring data is exemplified as data indicating the state of an equipment during operation, but a data set created by combining monitoring data and process data may also be data indicating the state of an equipment during operation. In the following description, the term "monitoring data" may also include process data.

[0049] Here, the specific number of dimensions in this embodiment is four or more. For example, the analysis processing unit 8 performs analysis using at least one of topological data analysis and cluster analysis. In this way, analysis can be performed using data with a specific number of dimensions.

[0050] The target data is the data included in the monitoring data that is the target of diagnosis. Note that the target data may also include process data acquired at the same time as the monitoring data or around the same time.

[0051] Historical data is data in which the presence or absence of an abnormality contained in the monitoring data is known. For example, the presence or absence of an abnormality in monitoring data acquired from equipment currently in operation is unknown. Furthermore, historical data may also include information regarding the cause of the abnormality.

[0052] This monitoring data is used as target data to diagnose whether there is an abnormality and its cause. Furthermore, monitoring data acquired from equipment several months or years ago allows us to know whether the equipment actually operated normally or whether an abnormality occurred after the data was acquired. Such monitoring data is treated as past data. Note that process data acquired around the same time as past monitoring data or around the same time may also be included in past data.

[0053] Furthermore, simulation data may be used as past data in addition to monitoring data. In this way, it is possible to evaluate the presence or absence of an abnormality and its cause even for equipment that has not previously experienced an abnormality.

[0054] The physical model unit 7 models the operation of equipment made up of one or more devices, and reproduces the state of the devices that occurs in relation to the operation through simulation.

[0055] For example, the physical model unit 7 constructs a physical model that reproduces the operation of the device as a piece of equipment and reproduces the state of the device that occurs in relation to the operation, and evaluates the behavior of the device through simulation. Note that a physical model is a virtual model that can simulate physical quantities related to the device. By using the physical model, the physical model unit 7 can simulate not only the device operating normally, but also the device operating abnormally.

[0056] The physical model reproduces the behavior of the device by a combination of parameters, and includes, for example, at least one of a mechanical model, a control model, an electrical model, a magnetic model, a thermal model, a fluid model, a physical property model, and a structural model.

[0057] Here, the mechanical model represents the behavior of the mechanical system related to at least one of the rotational and translational motions of each piece of equipment, such as an electric motor, a rotating object, or a group of load-side machines. The control model is used by a control panel, which serves as a control unit constituting at least one piece of equipment, to control the behavior of the entire device. The electrical model represents the behavior of the device's electrical system. The magnetic model represents the behavior of the device's magnetic system. The thermal model represents the thermal behavior of the device. The fluid model represents the behavior of at least one of the flow rate and pressure of a fluid flowing through piping connected to a pump constituting at least one piece of equipment. The physical property model represents the material properties of each element. The structural model has information on at least one of the dimensions and shape of each piece of equipment. In this way, the behavior of the device can be precisely reproduced according to various models.

[0058] In other words, a physical model is a model that models and formulates the structure or function of a device, and a set of equations corresponding to these are formulated.

[0059] The physical model unit 7 may construct a model based on the information about the device stored in the specification database 12.

[0060] The physical model also includes an abnormality model that indicates an abnormality in each device that constitutes the device. For example, as shown in Figures 3 and 4, a coupling of a rotating body 70 can be considered. The eccentricity e from the center of rotation caused by improper installation of the coupling can be calculated. c or declination d c The effect of eccentricity e is formulated as a model. c or declination d c Parameters that indicate the degree of abnormality, such as the size of the

[0061] There are various abnormality models for the rotating body 70. For example, there are bearing misalignment, breakage of the rotor bars or short circuit of the motor, protrusion of the rotor bars, and protrusion of the coils in the stator or slots. There are also bends or cracks in the rotating body 70, imbalance (missing parts, residual imbalance), short circuit or burnout of the stator winding, unbalanced power supply voltage, irregular waves in the inverter, and damage to the rolling bearings (inner ring, outer ring, rolling elements). Parameters are defined for each abnormality.

[0062] When vibrations occur in a rotating shaft due to such an abnormality, the effects propagate to structures such as bearings and casings that support the rotating body, as well as to the floor, making it difficult to identify the source of oscillation in a plant with multiple rotating bodies. Therefore, by evaluating the correlation between data acquired by installed sensors through simulations that take into account the rigidity and damping characteristics of the structure in advance, it is possible to optimize the installation location of the sensors and make it easier to identify the cause.

[0063] Parameters related to abnormalities may be determined in advance through measurements. For example, eccentricity occurring in a flange coupling can be measured from the displacement of the flanges on the drive and load sides. This reduces the number of parameters in the physical model, improving diagnostic accuracy and reducing calculation costs.

[0064] The physical model includes not only the individual devices that make up the facility, but also a system model constructed by combining these devices. As an example, a system model corresponding to Figure 2 is shown in Figure 5. For example, it is possible to evaluate the change in each data from its normal state when a seat leak occurs in the flow control valve 63 using a simulation. In this case, if the amount of seat leak in the flow control valve 63 is set as a parameter related to the abnormality, each data will change from its normal state in response to changes in the amount of leak. This makes it possible to link the degree of the abnormality with the resulting change in data.

[0065] The simulation database 13 stores the simulation data executed by the physical model unit 7 .

[0066] The anomaly cause database 14 stores feature quantities of the target data extracted based on the analysis results of the analysis processing unit 8 and the analysis results of the physical model section 7 stored in the simulation database 13. Furthermore, the anomaly cause database 14 stores cause information (anomaly cause data) indicating a correspondence relationship with the cause of an abnormality of the equipment. Here, the cause information indicates a correlation between the feature quantities of an abnormality and the cause of the abnormality. In other words, the anomaly cause database 14 stores in advance the correspondence relationship between the feature quantities and the cause of an abnormality.

[0067] The cause information may be a correlation input or set by the user, may be a correlation evaluated in advance by simulation, or may be pre-stored. In this way, the cause of an abnormality can be easily estimated based on the correspondence between the feature amount and the abnormality cause.

[0068] As shown in FIG. 6, the physical model unit 7 includes a model construction unit 15, a model information acquisition unit 16, a model analysis unit 17, a model determination unit 18, and a model update unit 19.

[0069] The model construction unit 15 constructs a physical model that reproduces the operation of the device as a device and reproduces the state of the device that occurs in relation to the operation.

[0070] The model information acquisition unit 16 acquires various data from the model construction unit 15 and the information input unit 6. In addition, the model information acquisition unit 16 may acquire various data from the monitoring database 10, the process database 11, and the specification database 12, as necessary.

[0071] The model analysis unit 17 performs calculations by reflecting parameters in equations (group of equations) formulated based on a physical model, and generates simulation data. This calculation includes structural analysis using the finite element method (FEM) or the like.

[0072] The model determination unit 18 determines the degree of reproducibility of the physical model from the simulation data, past data, and the abnormality factor determination results.

[0073] The model determining unit 18 may evaluate the degree of reproducibility from the variation in the results of analyzing a plurality of data.

[0074] If the model determination unit 18 determines that the degree of reproducibility of the physical model is low, the model update unit 19 updates the physical model to reproduce the processed data.

[0075] For example, the physical model unit 7 determines the reproducibility of the physical model based on the data reproduced by the simulation, past data for which the presence or absence of anomalies is known, and the results of anomaly cause determination based on the past data. If the reproducibility of the physical model is determined to be low, the physical model unit 7 updates the physical model to reproduce the processed data.

[0076] 7, the analysis processing unit 8 includes a data acquisition section 20, a data preprocessing section 21, a processing setting section 22, a data analysis section 23, a parameter tuning section 24, and a result organization section 25. The analysis processing unit 8 further includes a result display control section 26, a display adjustment section 27, an abnormality cause analysis section 28, a cause information acquisition section 29, and a cause display control section 30.

[0077] The data acquisition unit 20 acquires various data from the monitoring database 10, the process database 11, the specification database 12, the simulation database 13, and the abnormality cause database 14. In the following description, the term "monitoring data" may include process data, specification data, simulation data, or abnormality cause data acquired by the data acquisition unit 20.

[0078] The data preprocessing unit 21 extracts target data to be diagnosed from at least one of time-series data and frequency spectrum data that indicate the relationships between multiple sensors installed in at least one device.The data preprocessing unit 21 then adjusts the number of dimensions, which is the number of individual data items included in the target data, to a specific number of dimensions.The data preprocessing unit 21 also has the function of determining correlations between multiple sensors and selecting data to be diagnosed based on the strength of the correlation.

[0079] For example, the data preprocessing unit 21 extracts data to be used for analysis, such as target data or past data, from the monitoring data, and preprocesses this data before it is analyzed by the data analysis unit 23 (see FIG. 7). Here, the number of dimensions of the target data or past data is adjusted to a specific number of dimensions. Note that the target data and past data are adjusted to have the same specific number of dimensions.

[0080] The data preprocessing unit 21 calculates the correlation of the data acquired by each sensor. For simulation data, the relationship between each piece of data may be found by processing such as sensitivity analysis (see FIG. 8).

[0081] The process setting unit 22 sets various settings or analysis algorithms used when the data analysis unit 23 performs analysis (see FIG. 9).

[0082] The data analysis unit 23 performs topological data analysis on the target data adjusted to the specific number of dimensions using a mapper, which is an algorithm for handling data with a specific number of dimensions, as will be described later.

[0083] For example, the data analysis unit 23 analyzes the target data or past data adjusted to a specific number of dimensions using at least one of topological data analysis and cluster analysis, which are algorithms that handle data with a specific number of dimensions. This data analysis unit 23 can detect target data that indicates an abnormal state from the target data included in the monitoring data.

[0084] The parameter tuning unit 24 tunes various settings used when the data analysis unit 23 performs analysis or the settings of an algorithm for analysis (see FIG. 10).

[0085] The result organizing unit 25 organizes whether or not the analysis results of the data obtained by the data analysis unit 23 are optimal. In other words, the result organizing unit 25 organizes the correspondence between the analysis parameters and the analysis results for the results analyzed by the data analysis unit 23.

[0086] If the data analysis result is not optimal, the settings are tuned by parameter tuning unit 24. That is, tuning by parameter tuning unit 24 and data analysis by data analysis unit 23 are repeated until an analysis result is obtained that enables the target data to be displayed in a manner that makes it possible to identify the presence or absence of an abnormality in the device, based on the data analysis by data analysis unit 23.

[0087] The result display control unit 26 displays the analysis results using a mapper, which will be described later, in a cluster display or a time series display, which are formats that allow the presence or absence of abnormalities to be identified.

[0088] For example, the result display control unit 26 controls the display output unit 9 to display a screen showing the analysis results of the data obtained by the data analysis unit 23, in a manner that makes it possible to identify the presence or absence of an abnormality in the device (see FIG. 11). For example, the result display control unit 26 controls the display output unit 9 to display a screen such as a cluster display (see FIG. 16) or a time series display (see FIG. 17).

[0089] Here, in this embodiment, a mode in which the presence or absence of an abnormality can be identified is, for example, a graph of three dimensions or less. For example, assume that the cluster display (see FIG. 16) is a two-dimensional graph. This cluster display may also be a three-dimensional graph. In this way, target data of four or more dimensions that cannot be visualized using ordinary display technology can be visualized as a graph of three dimensions or less, and can be used to determine the presence or absence of an abnormality in equipment. Note that a time series display (see FIG. 17) is included in the mode in which the presence or absence of an abnormality can be identified.

[0090] When using topological data analysis, the result display control unit 26 organizes the anomaly detection results, including the number of nodes, the number of edges, the total number of input data, the number of unique data contained in a node, the number of node groups, etc. In the following description, node groups may be referred to as clusters. The result display control unit 26 also organizes the analysis parameters of the topological data analysis, including distance metrics, filter functions, intervals, overlaps, clustering methods, etc.

[0091] The display adjustment unit 27 controls the screen to be displayed on the display output unit 9 in response to a user operation (see FIG. 12).

[0092] If there is an abnormality in the target data, the abnormality cause analysis unit 28 analyzes the cause of the abnormality based on the analysis result of the data obtained by the data analysis unit 23 (see FIG. 13).

[0093] The cause information acquisition unit 29 acquires cause information from the abnormality cause database 14. Based on this cause information, the abnormality cause analysis unit 28 performs analysis.

[0094] The cause display control unit 30 controls the display output unit 9 to display a screen showing the analysis results of the data obtained by the abnormality cause analysis unit 28, in a format that allows the user to diagnose the cause of the abnormality in the equipment.

[0095] 8, the data preprocessing unit 21 includes a target data adjustment unit 31, a past data adjustment unit 32, a missing information interpolation unit 33, a data label assignment unit 34, and a correlation analysis unit 35. The data preprocessing unit 21 further includes a data scaling unit 36, a spectrum analysis unit 37, an amplitude value calculation unit 38, a data format conversion unit 39, and a signal selection unit 40.

[0096] The target data adjustment unit 31 adjusts the target data to be diagnosed, which is included in the monitoring data or the process data, so that the number of dimensions of the target data becomes a specific number of dimensions.

[0097] The past data adjustment unit 32 adjusts the past data so that the number of dimensions of the past data, which is known to have an abnormality included in the monitoring data or the process data, becomes a specific number of dimensions.

[0098] In this embodiment, "adjusting data to have a specific number of dimensions" may include a user thinning out or compressing the data to have any number of dimensions. Furthermore, it may also include extracting only data that is likely to contribute to anomalies from the data to be analyzed based on physical knowledge about the equipment. For example, if there is frequency data with 2000 dimensions, preprocessing may be performed to extract only 48-dimensional data that is likely to contribute to anomalies (to adjust the dimensions).

[0099] When target data or past data has a specific number of dimensions, if information for some of the information for each number of dimensions is missing, the missing information interpolation unit 33 interpolates the missing information. In this way, the information for each number of dimensions of the target data or past data is compiled, and target data or past data that can be analyzed by an algorithm that handles data with a specific number of dimensions can be created.

[0100] The data labeling unit 34 assigns a predetermined label to the target data or past data whose specific number of dimensions has been adjusted. For example, with regard to data that has been analyzed in the past, or data in which significant features were observed during data collection, it is known whether the data is normal or abnormal. Such past data is assigned a label L as information that can identify the presence or absence of an abnormality (see FIGS. 14 and 15). In other words, the past data is associated with a label L that can identify the presence or absence of an abnormality. Note that the label L of the target data is left blank.

[0101] In the case of data in which an abnormality exists, the data labeling unit 34 may set a label relating to the cause of the abnormality.

[0102] The data scaling unit 36 ​​performs a process to normalize or standardize the monitoring data. This process makes it possible to align the scales of data that differ in scale for each dimension. For example, normalization causes the range of feature values ​​to fall within a certain range. This is often the case, primarily within the range of [0, 1] or [-1, 1]. The standardization process also includes a process of converting the feature values ​​so that their mean is 0 and their variance is 1.

[0103] The spectrum analysis unit 37 extracts any frequency component from the waveform data. Based on this analysis, vibration spectrum data can be generated for target data or past data. The spectrum analysis unit 37, for example, uses FFT to extract specific frequency components from the monitoring data, or performs envelope processing to extract specific periodic frequency components. When dealing with two signals, the spectrum analysis unit 37 also analyzes the transfer function, correlation function, and phase between the signals. For example, the frequency spectrum data includes the transfer function, correlation function, and phase between the two signals.

[0104] The correlation analysis unit 35 performs correlation analysis between each piece of data, such as correlation coefficients or scatter diagrams.

[0105] The signal selection unit 40 selects two signals that have a physical relationship based on the correlation evaluation result between the data obtained by the correlation analysis unit 35 or the sensitivity analysis result by simulation. In this way, changes in the relationship between the data due to an abnormality can be detected efficiently, and the accuracy of diagnosis can be improved.

[0106] The correlation can be evaluated using a correlation coefficient or a scatter diagram, etc. For example, it is determined that there is a correlation when the absolute value of the correlation coefficient exceeds a certain threshold.

[0107] The signal selection unit 40 may also select two signals to be evaluated based on the correlation between any two signals. In this case, the signal selection unit 40 may select multiple combinations of signals depending on the strength of the correlation. For example, a correlation function (coherence) can be used to measure the correlation. The closer the coherence is to 1, the more likely it is that there is a physical causal relationship between the behavior of the two signals, or that they share a common influencing factor. An example is a process that extracts the features of a transfer function and estimates the transfer path from the data of each sensor. This process estimates the coherence between two signals corresponding to the estimated transfer path and examines the strength of the correlation. Since a strong correlation is a necessary condition for a fluctuation source, it is evaluated whether this condition is met.

[0108] The Fourier transforms of the two time series data are defined as X(f) and Y(f), respectively, and the cross spectrum is defined as follows:

number

[0109] The normalized cross spectrum is called coherence, which is defined as follows:

number

[0110] Here, the values ​​in "<>" in the formula represent a certain averaging process. Coherence takes a value between 0 and 1 and represents the correlation between two signals in frequency space. In other words, when coherence is close to 1, the two signals contain that frequency component.

[0111] The averaging process may be performed on time series data or on frequency spectra, and examples of the averaging process include moving average and triangular moving average.

[0112] The phase between the frequency components contained in the two signals is defined as follows:

number

[0113] Similarly, coherence may be determined for the simulation results, or sensitivity analysis may be used.

[0114] This reduces the time required for diagnosis and improves the accuracy of the diagnosis, while also reducing the cost of installing additional sensors.

[0115] The amplitude value calculation unit 38 analyzes the amplitude values ​​contained in the monitoring data. The analyzed values ​​may be overall values, if necessary. Based on this analysis, amplitude value data that is the target data or past data can be created.

[0116] That is, in this embodiment, it is possible to obtain, based on the monitoring data, information relating to the time-series waveform of the equipment, such as amplitude value data or vibration spectrum data, a transfer function between two signals, a correlation function, and a phase.

[0117] The data format conversion unit 39 converts the data format of the target data or past data so that the data format matches the algorithm used by the data analysis unit 23. Note that converting the data format means adjusting the shape of the data in order to perform anomaly analysis.

[0118] As shown in FIG. 9, the process setting section 22 includes a preprocessing condition setting section 41 and an algorithm setting section .

[0119] The preprocessing condition setting unit 41 sets the preprocessing conditions, for example, the settings of the above-mentioned interpolation or normalization process are changed.

[0120] The algorithm setting unit 42 sets the algorithms to be used by the data analysis unit 23. For example, the algorithm setting unit 42 sets whether to use topological data analysis, cluster analysis, or both. The algorithm setting unit 42 also sets the hyperparameters of these algorithms. For example, when using topological analysis, the filter function, overlap rate, etc. are changed by setting these hyperparameters.

[0121] As shown in FIG. 10, the parameter tuning unit 24 includes a specific dimension number setting unit 43 and a parameter optimization unit 44.

[0122] The specific number of dimensions setting unit 43 sets the specific number of dimensions to be used when analyzing target data or past data. In this way, the specific number of dimensions can be set so that the presence or absence of an abnormality can be identified. Note that the specific number of dimensions may be set by the preprocessing condition setting unit 41 or the algorithm setting unit 42.

[0123] Furthermore, the setting of the specific number of dimensions is repeated so that the display of the target data or past data obtained after the analysis is in a form that makes it possible to identify the presence or absence of an abnormality. In this way, the setting of the specific number of dimensions can be repeated until the form that makes it possible to identify the presence or absence of an abnormality is reached. This setting may be performed manually or automatically.

[0124] The parameter optimization unit 44 tunes the settings of the parameters used for analysis in the data analysis unit 23 based on the correspondence between the analysis parameters and the analysis results organized by the result organization unit 25. In this tuning, the parameters are repeatedly changed and optimized so that the display of the target data or past data obtained after the analysis is in a form that makes it possible to identify the presence or absence of an abnormality. At this time, the parameters may be tuned in advance using the simulation results regarding an abnormality calculated by the model analysis unit 17 as abnormal data.

[0125] As shown in FIG. 11, the result display control unit 26 includes a cluster display control unit 45 and a time-series display control unit 46.

[0126] The cluster display control unit 45 controls the display output unit 9 to display a cluster display screen (see FIG. 16) based on the data analysis results obtained by the data analysis unit 23. The cluster display is a screen that displays target data as nodes N. In this way, the target data can be visualized as a cluster display, making it easier to identify whether or not there is an abnormality in the equipment.

[0127] The time series display control unit 46 controls the display output unit 9 to display a time series display screen (see FIG. 17) showing the analysis results of the data obtained by the data analysis unit 23. The time series display is a screen that displays the target data in the chronological order in which it was monitored, and also displays the node N containing the target data in an identifiable manner. In this way, the target data can be displayed in chronological order, making it easier to identify whether or not there is an abnormality in the equipment.

[0128] As shown in FIG. 12, the display adjustment unit 27 includes a node selection receiving unit 47 and a target information display unit .

[0129] The node selection receiving unit 47 performs processing to receive the selection of a node N or a cluster C on the screen where the cluster is displayed.

[0130] The target information display unit 48 displays the breakdown information B included in the selected node N received by the node selection receiving unit 47 (see FIG. 16). This breakdown information B includes the target information of the target data included in the selected node N. In this way, it is possible to easily obtain information on the target data corresponding to the node N.

[0131] As shown in FIG. 13, the abnormality cause analysis unit 28 includes a feature extraction unit 49 and an abnormality cause estimation unit 50.

[0132] The feature extraction unit 49 extracts the feature of the target data after the analysis by the data analysis unit 23 .

[0133] Furthermore, the feature extraction unit 49 compares the calculation results in a normal state with those in an abnormal state, and selects a feature of the target abnormal state from the amount of change. The calculation results to be compared may be time waveforms, FFT, envelope processing, averaging, transfer functions, correlation functions, phases, or a combination of these. For example, in the case of frequency spectra, the feature is the spectrum at a frequency where the spectrum changes significantly between normal and abnormal. For two signals, the feature is, for example, the gain of the transfer function where the gain changes significantly between normal and abnormal. Note that multiple feature values ​​may be selected, not just the one with the largest change.

[0134] The anomaly cause estimation unit 50 estimates the cause of the anomaly based on the feature quantities extracted by the feature quantity extraction unit 49. In other words, the anomaly cause estimation unit 50 organizes the data so that the target data can be displayed in a manner that enables the cause of the anomaly to be diagnosed, based on the feature quantities extracted by the feature quantity extraction unit 49. In this way, when an anomaly is discovered in a device, the cause can be estimated.

[0135] Next, the target data or past data analyzed by the data analysis unit 23 will be described with reference to FIGS.

[0136] FIG. 14 is a table containing an example of data after preprocessing by the data preprocessing unit 21, taking acceleration amplitude as an example. In this table, acceleration amplitudes of a specific number of dimensions are registered in association with the monitoring dates (measurement dates, data collection dates and times). Note that any number can be selected as the specific number of dimensions for acceleration amplitude. Here, the data of the most recent date is the target data T that is the target of diagnosis for the presence or absence of an abnormality.

[0137] Note that multiple columns S of acceleration amplitudes with a specific number of dimensions are provided. The number of these columns S is the number of dimensions of the data. In the table of FIG. 14, acceleration amplitudes of five or more dimensions, for example, 50 dimensions, are registered in association with a predetermined date (measurement date, data collection date and time). Also, a label L that can identify the presence or absence of an abnormality is registered in association with past data P. For example, a label L of "A" is registered for past data P that was normal, and a label L of "B" is registered for past data P that was abnormal.

[0138] The data recorded in the table of FIG. 14 may be a single file or multiple files.

[0139] FIG. 15 is a table containing an example of frequency data calculated by the data preprocessing unit 21. In this table, vibration values ​​for each specific number of dimensions are registered in association with the monitoring date. Any number can be selected as the specific number of frequency dimensions. Here, the data of the most recent date is the target data T that is the target of diagnosis for the presence or absence of an abnormality. The frequency data includes a frequency spectrum, a transfer function, a phase, and a coherence.

[0140] Note that multiple columns F of frequencies with a specific number of dimensions are provided. The number of these columns F is the number of dimensions of the data. In the table of FIG. 15, a transfer function with five or more dimensions, for example, 50 dimensions, is registered in association with a specific date. Also, a label L that can identify the presence or absence of an abnormality is registered in association with past data P. For example, a label L of "A" is registered for normal past data P, and a label L of "B" is registered for abnormal past data P.

[0141] Similarly to the transfer function, the phase and coherence are also calculated. That is, the table in FIG. 15 contains data related to the transfer function, coherence, and phase. The target data may be any one of these, or a combination of these may be used. When the transfer function, coherence, and phase are all used, the table in FIG. 15 has the number of dimensions F (number of frequency sequences) × 3 (transfer function, coherence, phase) for each specified date.

[0142] In Figures 14 and 15, the specific number of dimensions of the target data and past data is the same. However, there are cases where data corresponding to each number of dimensions cannot be extracted in association with each date. For example, among 50-dimensional data, data for several dimensions may be missing. This missing data is interpolated by the missing information interpolation unit 33. The value of the interpolated data may be zero, the average value of the data included in the same row as the missing item, or a value interpolated or extrapolated based on data before and after this average value. Furthermore, the value of the interpolated data may be the average value of the data included in the same column as the missing item, or a value interpolated or extrapolated based on data before and after this average value, or any other arbitrarily determined value.

[0143] In this embodiment, the past data is data acquired from a device that is the target of an abnormality diagnosis, but other configurations are also possible. For example, if there is a device with the same manufacturer, specifications, and model, data acquired from this other device may be used as the past data of the device that is the target of an abnormality diagnosis.

[0144] The specific number of dimensions is interpolated and format converted in the data preprocessing unit 21 so that the number of dimensions is the same for all dates. The processing in the data preprocessing unit 21 may be performed automatically or may be set manually.

[0145] Conventional statistical methods using machine learning are based on statistical analysis and assume that the data being handled follows some kind of distribution, such as a normal distribution. Therefore, it is necessary to develop an analytical method that corresponds to the characteristics of each piece of data. Furthermore, when handling large-scale, multidimensional data, processing such as reducing the dimensions of the data is necessary. However, dimension reduction using such statistical methods has the problem of losing information in the data. This embodiment can solve this problem.

[0146] Next, we will explain the topological data analysis used in this embodiment, so-called topological data analysis (hereinafter referred to as "TDA"). TDA is a method for analyzing the characteristics of large-scale multidimensional data by focusing on the topological information (shape of the data) of the data rather than on statistical characteristics. There are two types of TDA methods: persistent homology and mapper.

[0147] Persistent homology is a method that can capture geometric invariants and quantitatively evaluate the details of data shapes, while mappers are a method for representing high-dimensional data as simple graphs that can capture geometric information.

[0148] This embodiment illustrates an example in which TDA Mapper is used, although other analysis techniques may also be used.

[0149] TDA Mapper makes it possible to visualize high-dimensional (specific number of dimensions) data as a graph of three or fewer dimensions. It also makes it possible to represent parts of the data that are important for understanding them as a single node N. For example, nodes N that have consecutive data are connected by an edge E, and the set of data is converted into a graph (see Figure 16).

[0150] In this embodiment, "visualization" refers to a form in which a human can recognize a shape. For example, a three-dimensional graph displayed on a display is a two-dimensional image, but a human can recognize its three-dimensional shape.

[0151] TDA Mapper's analysis results are less dependent on the coordinate system of the target data, and it is possible to compare data in different coordinate systems without depending on the selected coordinate system. TDA Mapper is also invariant to small displacements. For example, it is less susceptible to noise.

[0152] TDA Mapper also has the ability to compress multidimensional data into lower dimensions and represent the shape of the data. For example, if a set of data containing several thousand points is input, it can be represented as a network with 13 vertices and 12 edges.

[0153] TDA Mapper accepts multidimensional data as input and sets a distance metric for the input data. Next, it sets a filter function to map the data with the distance set to a lower dimension. One or more filter functions can be set. For example, a function that projects data with Euclidean distance as a metric onto the Y-axis can be used as a filter. Next, intervals are determined for the projected data, and the data is divided into bins. The intervals and the overlap between intervals are important for the relationship between nodes and edges, which will be described later. Next, the data contained within each interval is clustered. An example of a clustering algorithm is the k-means method. In this case, the number of clusters, k, is also a parameter. The clusters for each interval become nodes when visualized. If data within a cluster overlaps, a line is drawn between the nodes. This line is called an edge and indicates a connection between the nodes.

[0154] Mapper hyperparameters that can (and must) be set by the user during the above processing process include, for example, metric, filter function, interval, overlap, and clustering method. Here, metric is the definition of distance. Filter functions include density, eccentricity, graph Laplacian, projection onto each axis, SVD (singular value decomposition), and isolation forest. Regarding intervals, for example, increasing the number of intervals will increase the number of clusters after clustering processing. However, there is a possibility that many empty clusters will be created. In other words, the amount of data per cluster may be small. Regarding overlap, increasing the number of overlaps will increase the connections (edges) between nodes. Clustering methods include hierarchical clustering, k-means, and other methods.

[0155] The parameter tuning unit 24 and the result organizing unit 25 organize the correlation between these control parameters and the analysis results (see FIG. 18). Then, the parameter tuning unit 24 sweeps the analysis parameters until the data point of interest is plotted at a node of the graph.

[0156] This parameter tuning ensures that the data point of interest is always represented on a node. Next, it is checked whether the node containing the data point of interest also contains data points from other measurement dates, and whether it has a relationship with other nodes, i.e., whether it is connected by an edge. In this case, if the data from other measurement dates has a normal label L, it is determined that the data of interest is also likely to be normal. On the other hand, if there is no connection with normal data, it is determined that the data of interest is not normal, that is, that it is likely to be abnormal.

[0157] By following the above procedure, it becomes possible to automatically determine whether or not an abnormality exists, without the user having to set analysis parameters.

[0158] 16, a cluster display (graph) that is generated by the data analysis unit 23 and is the result of TDA Mapper is displayed on the display. The cluster display is controlled by a cluster display control unit 45 of the result display control unit 26.

[0159] This cluster display includes a graph display area G where the node N and cluster C are displayed, a breakdown information display area D where the breakdown information B contained in the node N is displayed, and a color label contour R that indicates the color-coding manner of the node N.

[0160] In the example of FIG. 16, nodes N are displayed as circles in the graph display area G. For example, the greater the amount of target data or past data included in one node N, the larger the circle displayed. Furthermore, multiple nodes N are connected by edges E. The group of nodes connected by these edges E is cluster C. Cluster C1 of nodes N containing normal data and clusters C2, C3, and C4 of nodes N containing abnormal data are displayed separately.

[0161] For example, one cluster C1 containing normal data is displayed, and three clusters C2 containing abnormal data are displayed. In principle, a cluster C in this embodiment refers to a plurality of nodes N connected by edges E, but even if there is a single node N, if it is separated from other clusters C, it will be treated as one cluster C.

[0162] The cluster display shows a graph generated based on both the target data and past data. Note that a given node N may contain both the target data and past data.

[0163] By selecting a specific node N displayed in the graph display area G with the mouse cursor M, the user can display the breakdown information B contained in the selected node N in the breakdown information display area D.

[0164] For example, if the selected node N includes multiple target data, the dates (measurement date, data collection date and time) corresponding to each target data are displayed. Attribute information such as the number of data points and data labels of the target data included in the selected node N may be displayed in the breakdown information display area D. Statistical quantities included in the node N may also be displayed in the breakdown information display area D.

[0165] Note that the item selected with the mouse cursor M may be cluster C. In this case, the breakdown information B included in the selected cluster C can be displayed in the breakdown information display area D.

[0166] The color label contour R displays the display color setting of the node N. For example, a node N including past data in which a label L of "A" indicating normality is registered is displayed in yellow, and a node N including past data in which a label L of "B" indicating abnormality is registered is displayed in blue. In this way, past data can be displayed in a manner in which the associated label L can be identified based on an analysis of the past data.

[0167] Note that the data may be color-coded for each node N that the user wishes to check, or may be color-coded by the frequency corresponding to each node N. Also, for example, data may be color-coded by labeling whether the equipment has been overhauled or not. Also, the nodes N may be color-coded for each parameter that was affected during the analysis.

[0168] Furthermore, in the cluster display, if it is determined that the data of interest is highly likely to be abnormal, the function for highlighting unusual nodes can be used to highlight the node N containing the target data with an abnormality on the GUI (displaying the possibility of an abnormality).

[0169] 17, the time series display generated by the data analysis unit 23 and resulting from the TDA Mapper is displayed on the display. The time series display is controlled by the time series display control unit 46 of the result display control unit 26.

[0170] In this time series display, the data that was graphed in the cluster display is displayed in chronological order. For example, the time series display has a column K1 for the monitoring dates (measurement dates, data collection dates and times) and a column K2 for the nodes N corresponding to each date. Furthermore, the time series display has a column K3 for the clusters C corresponding to each node N and a column K4 showing the possibility of an anomaly in each node N or cluster C.

[0171] By referring to this time series display, the user can view whether or not there is a possibility of an abnormality in each node N or cluster C. For example, "Normal" is displayed in correspondence with a node N that includes past data in which a label L of "A" indicating normality is registered, and "Abnormal" is displayed in correspondence with a node N that includes past data in which a label L of "B" indicating an abnormality is registered.

[0172] That is, the result display control unit 26 displays the target data in a manner (for example, a chronological display) that allows the presence or absence of an abnormality to be identified based on the analysis of the target data. Furthermore, based on the analysis of the past data, the result display control unit 26 displays the past data in a manner that allows the associated label L to be identified. In this way, the presence or absence of an abnormality in the target data can be identified by referring to the display manner of the past data.

[0173] The cluster display (see FIG. 16) and the time series display (see FIG. 17) can be switched between at will by the user's operation.

[0174] The abnormality cause analysis unit 28 uses a function for extracting the feature amount of abnormal data to extract the feature amount of data that is determined to be possibly abnormal.

[0175] Here, we will explain the case where the analysis target and analyzed data are frequency data. For example, for data determined to be abnormal and data labeled as normal, the difference between each frequency component is calculated, and frequencies with large differences in each frequency component are extracted. The difference between this extracted frequency and frequency component is considered to be the feature of the abnormal data. The definition of a frequency component with a large difference can be any extraction method, such as one that exceeds a preset threshold or extracting differences in the top few percent. The extracted feature (frequency, difference in frequency component) is then input to the abnormality cause analysis unit 28.

[0176] Furthermore, as shown in FIG. 19, when past data for which the cause of an anomaly is known is displayed in a cluster format together with the target data, the linking relationship between the target data and the past data is input to the anomaly cause analysis unit 28, allowing the cause to be identified.

[0177] The results of a simulation of an abnormal state in advance may be used as past data, which makes it possible to identify the cause of an abnormality even for equipment that has never experienced an abnormality in the past.

[0178] Although the present embodiment exemplifies a cluster display using topological data analysis as a mode capable of identifying the presence or absence of an abnormality, other modes may be used. For example, the mode capable of identifying the presence or absence of an abnormality may be a hierarchical cluster display based on a hierarchical cluster analysis, or a non-hierarchical cluster display based on a non-hierarchical cluster analysis. For example, a method such as the K-means method may be used for the non-hierarchical cluster analysis.

[0179] An embodiment using hierarchical clustering will be described below with reference to FIGS.

[0180] For example, hierarchical clustering takes N pieces of unclustered input data and fuses them in descending order of similarity to create larger and larger clusters. Finally, N pieces of data are integrated into one cluster, and the process of integrating the data or clusters is represented in the form of a tree called a dendrogram.

[0181] Here, the parameters of hierarchical clustering include a scale and a method for measuring the similarity of data. Scales for measuring similarity include Euclidean distance and Minkowski distance. Methods for measuring the distance between clusters include Ward's method and group average method. These parameters are optimized by the parameter tuning unit 24.

[0182] Furthermore, an index for parameter tuning and optimization is, for example, the correlation coefficient between the cluster distance and the Cohen matrix (Cohen correlation coefficient).

[0183] FIG. 19 shows the display results using a hierarchical cluster. For example, the data to be analyzed processed by the data preprocessing unit 21 is input to the data analysis unit 23, analyzed using a hierarchical clustering method, and the analysis results are displayed on the display output unit 9. FIG. 19 shows the cluster display results of the results. Here, 57 amplitude value data consisting of k dimensions are input. Of the 57 amplitude value data, some are within the normal range and some indicate abnormalities.

[0184] 19, if each individual data is a node N, then multiple nodes N or a group of nodes including at least one node N can be defined as a cluster C. When using this hierarchical clustering, as with the TDA described above, it is possible to display clusters C5 and C6 of nodes N that contain normal data and cluster C7 of nodes N that contain abnormal data.

[0185] As shown in Figure 20, in hierarchical clustering, normal data will be in an average branching pattern as various data and clusters are combined. In contrast, as shown in Figure 21, abnormal data will be displayed branching in the upper layers. Note that while the above explanation exemplifies a case where the number of normal data is greater than the number of abnormal data, there may also be cases where the number of normal data is smaller than the number of abnormal data.

[0186] The equipment abnormality diagnosis system 1 described above includes a control device, a storage device, an output device, an input device, and a communication interface. Here, the control device includes a highly integrated processor such as a central processing unit (CPU), a graphics processing unit (GPU), a field programmable gate array (FPGA), or a dedicated chip. The storage device includes a read-only memory (ROM), a random access memory (RAM), a hard disk drive (HDD), or a solid state drive (SSD). The output device includes a display panel, a head-mounted display, a projector, a printer, etc. The input device includes a mouse, a keyboard, a touch panel, etc. This equipment abnormality diagnosis system 1 can be realized with a hardware configuration using a normal computer.

[0187] The program executed by the above-described equipment abnormality diagnosis system 1 is provided by being pre-installed in a ROM or the like. Additionally or alternatively, the program is provided by being stored in a computer-readable non-transitory storage medium as a file in an installable or executable format. This storage medium includes a CD-ROM, a CD-R, a memory card, a DVD, a flexible disk (FD), and the like.

[0188] The programs executed by this equipment abnormality diagnosis system 1 may also be stored on a computer connected to a network such as the Internet and provided by downloading them via the network. In other words, the programs may be provided from cloud computing resources. Alternatively, a server on the cloud may execute the programs, and only the processing results may be provided via the cloud. This equipment abnormality diagnosis system 1 may also be configured by combining separate modules that independently perform the functions of the components and interconnect them via a network or dedicated lines.

[0189] According to the embodiment described above, the data preprocessing unit 21 determines the correlations between multiple sensors and selects the data to be diagnosed based on the strength of the correlations. This makes it possible to efficiently and accurately detect anomalies and estimate their causes using a minimum number of sensors. Furthermore, by limiting the diagnosis to the relationships between data acquired by sensors installed in the facility that are highly correlated, it is possible to reduce the processing load and improve the accuracy of the diagnosis.

[0190] Furthermore, by inputting the feature quantities into the anomaly cause analysis unit 28, the degree of influence of the feature quantities on the cause of the anomaly can be determined, making it possible to estimate the cause of the anomaly. Furthermore, by automating the parameter tuning required for anomaly detection, it becomes possible to estimate the cause of the anomaly without manual tuning of the anomaly detection. Furthermore, it becomes possible to estimate the cause of the anomaly from the results of the anomaly detection.

[0191] Furthermore, by analyzing target data adjusted to a specific number of dimensions using an algorithm that handles data with a specific number of dimensions, diagnosis can be performed without relying on the skill of the engineer.

[0192] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, modifications, and combinations can be made without departing from the spirit of the invention. These embodiments and modifications thereof are within the scope and spirit of the invention, as well as the inventions described in the claims and their equivalents. Note that the term "a" or "an" does not necessarily mean to be limited to one thing, and the term "a" or "an" may mean plural. [Explanation of symbols]

[0193] 1...equipment abnormality diagnosis system, 2...inspection sensor, 3...user terminal, 4...main control unit, 5...communication unit, 6...information input unit, 7...physical model unit, 8...analysis processing unit, 9...display output unit, 10...monitoring database, 11...process database, 12...specification database, 13...simulation database, 14...abnormality cause database, 15...model construction unit, 16...model information acquisition unit, 17...model analysis unit, 18...model judgment unit, 19...model update unit, 20...data acquisition unit, 21...data preprocessing unit, 22...processing setting unit, 23...data analysis unit, 24...parameter tuning unit, 25...result organization unit, 26...result display control unit, 27...display adjustment unit, 28...abnormality cause analysis unit, 29...cause information acquisition unit, 30...cause display control unit, 31...target data adjustment unit, 32...past data adjustment unit, 33...missing information interpolation unit, 3 4...Data label assignment unit, 35...Correlation analysis unit, 36...Data scaling unit, 37...Spectral analysis unit, 38...Amplitude value calculation unit, 39...Data format conversion unit, 40...Signal selection unit, 41...Preprocessing condition setting unit, 42...Algorithm setting unit, 43...Specific dimension number setting unit, 44...Parameter optimization unit, 45...Cluster display control unit, 46...Time series display control unit, 47...Node selection reception unit, 48...Target information display unit, 4 9...Feature extraction section, 50...Anomaly cause estimation section, 60...Tank, 61...Controller, 62...Water level meter, 63...Flow control valve, 70...Rotating body, B...Detailed information, C...Cluster, D...Detailed information display area, E...Edge, F...Frequency column, G...Graph display area, K1 to K4...Time series display columns, L...Label, M...Mouse cursor, N...Node, P...Past data, R...Color label contour, S...Acceleration amplitude column, T...Target data.

Claims

1. a data preprocessing unit that extracts time-series data and / or frequency spectrum data indicating a relationship between a plurality of sensors provided in at least one device, and adjusts the number of dimensions, which is the number of individual data included in target data to be diagnosed, of the at least one data, to a specific number of dimensions; a data analysis unit that performs topological data analysis on the target data adjusted to the specific number of dimensions using a mapper that is an algorithm that handles data of the specific number of dimensions; a result display control unit that displays the analysis results using the mapper in a cluster display or a time series display in a manner that allows the presence or absence of an abnormality to be identified; Equipped with the data preprocessing unit has a function of determining a correlation between the plurality of sensors and selecting data to be diagnosed based on the strength of the correlation; Equipment abnormality diagnosis system.

2. The frequency spectrum data includes a transfer function, a correlation function, and a phase between two signals. The device abnormality diagnosis system according to claim 1 .

3. a physical model unit that models the operation of a facility configured with one or more of the devices and reproduces the state of the devices that occurs in relation to the operation through simulation; 3. The device abnormality diagnosis system according to claim 1 or 2.

4. the physical model unit determines a degree of reproduction of the physical model from the data reproduced by the simulation, past data for which the presence or absence of the abnormality is known, and an abnormality cause determination result based on the past data, and if it is determined that the degree of reproduction of the physical model is low, updates the physical model to reproduce the processed data. The device abnormality diagnosis system according to claim 3 .

5. the specific number of dimensions is four or more, The aspect in which the presence or absence of the abnormality can be identified is a three-dimensional or less graph as the cluster display.

3. The device abnormality diagnosis system according to claim 1 or 2.

6. the data preprocessing unit adjusts the past data, which is included in the data indicating the state of the device and for which the presence or absence of the abnormality is known, so that the number of dimensions of the past data becomes the specific number of dimensions, and associates a label capable of identifying the presence or absence of the abnormality with the past data adjusted to the specific number of dimensions; the data analysis unit analyzes the past data adjusted to the specific number of dimensions using the mapper; the result display control unit displays the analysis result using the mapper in a manner that allows the associated label to be identified.

3. The device abnormality diagnosis system according to claim 1 or 2.

7. a parameter tuning unit that sets the specific number of dimensions, 3. The device abnormality diagnosis system according to claim 1 or 2.

8. a parameter tuning unit that repeatedly sets the specific number of dimensions so that the display of the target data is in a form that makes it possible to identify the presence or absence of the abnormality; 3. The device abnormality diagnosis system according to claim 1 or 2.

9. the data preprocessing unit interpolates missing information when information on some of the numbers of dimensions is missing when the target data has the specific number of dimensions; 3. The device abnormality diagnosis system according to claim 1 or 2.

10. an anomaly cause analysis unit that extracts feature quantities of the target data and organizes the target data to display the target data in a manner that enables the cause of the anomaly to be diagnosed based on the extracted feature quantities; 3. The device abnormality diagnosis system according to claim 1 or 2.

11. an anomaly cause database that stores in advance a correspondence relationship between the feature amount and the cause of the anomaly; 3. The device abnormality diagnosis system according to claim 1 or 2.

12. the cluster display is a display in which the target data is a node, Accepting a selection of the nodes displayed in the cluster; displaying information about the target data included in the selected node; A display adjustment unit is provided.

3. The device abnormality diagnosis system according to claim 1 or 2.

13. the result display control unit displays the target data in chronological order of monitoring, and displays the node including the target data in an identifiable manner.

3. The device abnormality diagnosis system according to claim 1 or 2.

14. An inspection sensor attached to the equipment monitors the state of the equipment to acquire monitoring data, and the monitoring data is input.

3. The device abnormality diagnosis system according to claim 1 or 2.

15. a data preprocessing unit that extracts time-series data and / or frequency spectrum data indicating a relationship between a plurality of sensors provided in at least one device, and adjusts the number of dimensions, which is the number of individual data included in target data to be diagnosed, of the at least one data, to a specific number of dimensions; a data analysis unit that performs topological data analysis on the target data adjusted to the specific number of dimensions using a mapper that is an algorithm that handles data of the specific number of dimensions; a result display control unit that displays the analysis results using the mapper in a cluster display or a time series display in a manner that allows the presence or absence of an abnormality to be identified; This is a method using the data preprocessing unit determines correlations between the plurality of sensors and selects data to be diagnosed based on the strength of the correlations; Equipment abnormality diagnosis method.

Images