Learning device, learning method and fault prediction system

Abstract

A learning device readily acquires data clearly representing fault states used in supervised learning training data, and performs learning for high-precision determination of equipment fault states. The learning device includes: a pattern extraction unit that extracts temporal fluctuation patterns of characteristic frequency amplitudes associated with parts of the equipment from state observation signal data representing the operating state of the equipment up to a first time point; a training data generation unit that, based on the temporal fluctuation patterns of the characteristic frequency amplitudes, generates simulated state observation signal data representing temporal fluctuation patterns of the characteristic frequency amplitudes after the first time point, and generates training data including the simulated state observation signal data; and a learning unit that uses the training data to generate a classification model for determining the fault states of parts of the equipment.

Description

Technical Field

[0001] This disclosure relates to learning devices, learning methods, and fault prediction systems. Background Technology

[0002] Industrial equipment, machinery, and robots used in factories and other production facilities employ a large number of motors, gears, and other components. Besides sudden equipment failures, malfunctions caused by aging, wear, and deterioration can also lead to production line shutdowns, potentially reducing productivity or causing accidents.

[0003] Therefore, there is a growing demand for fault prediction systems that monitor the status of equipment containing these devices or machines and support effective periodic maintenance corresponding to the status of the equipment.

[0004] Related to this fault prediction system, Patent Document 1 discloses a technique in which conditions associated with faults in industrial machinery are learned based on a teaching dataset. This teaching dataset is a dataset created based on a combination of state variables containing sensor data reflecting the state of the industrial machinery and judgment data determining the degree of fault in the industrial machinery. Thus, by using supervised learning with teaching data (training data), the fault prediction accuracy of the equipment is improved compared to learning through unsupervised learning without using training data.

[0005] Existing technical documents

[0006] Patent documents

[0007] Patent Document 1: Japanese Patent Application Publication No. 2017-033526 Summary of the Invention

[0008] However, for individual motors, gears, and other components within devices with completely different operating conditions or structures, it is difficult to collect training data consisting of combinations of state variables and multiple fault states from operating equipment. For example, in reality, it is difficult to collect large amounts of data representing the actual states of motors, gears, etc., when they are in abnormal or fault states (hereinafter collectively referred to as "fault states"). Therefore, in practice, the following method is mostly used: learning the normal state through unsupervised learning and predicting the fault state by detecting the deviation from the normal state.

[0009] Thus, there is still room for research into the accuracy of prediction or determination in learning methods used to predict or determine the fault state of equipment through supervised learning.

[0010] The non-limiting embodiments of this disclosure help to provide learning devices, learning methods, and fault prediction systems that can easily obtain clearly represented fault states using supervised learning training data and perform learning for high-precision determination of device fault states.

[0011] One embodiment of the learning apparatus of this disclosure includes: a pattern extraction unit that extracts a time fluctuation pattern of the amplitude of a characteristic frequency associated with a part of the device from state observation signal data representing the operating state of the device up to a first time point; a training data generation unit that generates simulated state observation signal data representing a time fluctuation pattern of the amplitude of the characteristic frequency after the first time point based on the time fluctuation pattern of the amplitude of the characteristic frequency, and generates training data including the simulated state observation signal data; and a learning unit that uses the training data to generate a classification model for determining the fault state of a part of the device.

[0012] One embodiment of the learning method disclosed herein is executed by a learning device and includes the following steps performed by the learning device: extracting a time fluctuation pattern of the amplitude of a characteristic frequency associated with a part of the device from state observation signal data representing the operating state of the device up to a first time point; generating simulated state observation signal data representing a time fluctuation pattern of the amplitude of the characteristic frequency after the first time point based on the time fluctuation pattern of the amplitude of the characteristic frequency; generating training data including the simulated state observation signal data; and using the training data to generate a classification model for determining the fault state of a part of the device.

[0013] One embodiment of the fault prediction system disclosed herein includes: the learning device described above; and a state determination unit, which uses current state observation signal data representing the current operating state of the device and the classification model to determine the fault state of a part of the device.

[0014] Furthermore, these inclusive or specific methods can be implemented by systems, methods, integrated circuits, computer programs, or recording media, or by any combination of systems, methods, integrated circuits, computer programs, and recording media.

[0015] According to one embodiment of this disclosure, clearly represented fault states can be readily obtained for training data in supervised learning, and learning can be performed to determine the fault states of a device with high accuracy.

[0016] Further advantages and effects of one embodiment of this disclosure will be clearly presented by the specification and accompanying drawings. These advantages and / or effects are provided by various embodiments and the features described in the specification and drawings, but not all of them need to be provided to obtain one or more of the same features. Attached Figure Description

[0017] Figure 1 This is a structural diagram illustrating an example of a fault prediction system in an embodiment of the present disclosure.

[0018] Figure 2 This is a flowchart illustrating an example of a learning method in an embodiment of this disclosure.

[0019] Figure 3 This is a diagram illustrating an example of device harmonics in an embodiment of this disclosure.

[0020] Figure 4 This is a diagram illustrating an example of a device model in an embodiment of this disclosure.

[0021] Figure 5 This is a graph illustrating an example of the time fluctuation of the amplitude of device harmonics in an embodiment of this disclosure.

[0022] Figure 6 This is a diagram illustrating an example of the time-varying pattern of the device harmonic amplitude in an embodiment of this disclosure.

[0023] Figure 7 This is a diagram illustrating an example of training data generation in an embodiment of this disclosure.

[0024] Figure 8 This is a flowchart illustrating an example of a fault prediction method in an embodiment of the present disclosure.

[0025] Figure 9 This is a diagram illustrating an example of fault prediction in an embodiment of this disclosure. Detailed Implementation

[0026] Hereinafter, embodiments of the present disclosure will be described in detail with appropriate reference to the accompanying drawings. However, sometimes overly detailed descriptions will be omitted. For example, detailed descriptions of well-known matters or repetitive descriptions of substantially the same structures may be omitted. This is to avoid making the following description unnecessarily lengthy and to facilitate understanding by those skilled in the art.

[0027] Furthermore, the purpose of providing the accompanying drawings and the following description is to enable those skilled in the art to fully understand this disclosure, and there is no intention to limit the subject matter set forth in the claims.

[0028] (Implementation Method)

[0029] <Fault Prediction System>

[0030] First, refer to Figure 1 The fault prediction system 100 in the embodiments of this disclosure is described. The fault prediction system 100 monitors equipment in a factory or similar facility and performs machine learning based on data representing past equipment states and results obtained through simulation to generate (build) a classification model for determining future equipment fault states. Then, the fault prediction system 100 uses data representing current equipment states and the generated classification model to determine future equipment fault states and displays the determination result to the user.

[0031] The fault prediction system 100 includes a device 101 as the monitoring object, a sensor 110, a learning device 150, a status determination unit 112, and a display unit 113. At least some of these functional units can communicate with each other, for example, via a communication network (wired network, wireless network, or a combination of wired and wireless networks).

[0032] Device 101 includes a motor 102 and a load 103. Load 103 represents a gearbox or mechanism driven by motor 102. Furthermore, multiple devices 101 may exist. In this case, devices 101-1 to 101-M (where M is an integer of 2 or more) each include motors 102-1 to 102-M and loads 103-1 to 103-M.

[0033] Sensor 110 is connected to or installed on motors and loads of various devices, namely devices 101-1 to 101-M. Sensor 110 senses (measures) the state of the devices over time (e.g., every minute, every 5 minutes, every 30 minutes, or every hour), and generates a sensing signal representing the sensed state of the devices. Then, sensor 110 outputs the generated sensing signal to learning device 150. Furthermore, although in Figure 1 The diagram shows one sensor 110, but multiple sensors may also be present. For example, various sensors such as a current sensor or harmonic sensor that measures the power supply current of a motor, a torque sensor that measures the torque of a motor, and an acceleration sensor or vibration sensor that measures the vibration of a device can be used as sensor 110. Alternatively, these sensors can be used in combination.

[0034] The learning device 150 accepts the sensing signal output from the sensor 110 as input. Based on the received sensing signal and results obtained through simulation, the learning device 150 performs machine learning and generates a classification model for determining the state of the device (more specifically, the future fault state of the device). Additionally, the learning device 150 outputs state observation signal data, generated by preprocessing the sensing signal input from the sensor 110, to the state determination unit 112. Details of the learning device 150 will be described later.

[0035] The state determination unit 112 uses the classification model, state observation signal data, and training data generated by the learning device 150 to determine the state of the device. Specifically, the state determination unit 112 detects the features shown in the state observation signal data (the amplitude of the characteristic frequencies described later), and uses the classification model, training data, and detected features to determine the state of the device. Such device states include, for example, normal, deteriorated, abnormal, signs of malfunction, malfunction, and related internal parts of the device. In this embodiment, in particular, the state determination unit 112 uses the classification model, training data, and detected features to determine the future malfunction state of the device. Then, the state determination unit 112 outputs the determined result to the display unit 113.

[0036] The display unit 113 accepts the determination result output from the status determination unit 112 as input and displays the accepted determination result to the user. For example, a user interface such as a display with a touch screen can be used as the display unit 113. The user can determine the status of the device through the display unit 113 and determine the parts of the device that need maintenance or repair.

[0037] <Learning Device>

[0038] Next, continue to refer to Figure 1 This describes the learning device 150 included in the fault prediction system 100.

[0039] The learning device 150 includes a state observation unit 111, a simulation unit 114, a feature frequency storage unit 115, a pattern extraction unit 116, a training data generation unit 117, a learning unit 118, a model storage unit 119, and a device model 121.

[0040] The state observation unit 111 receives the sensing signal output from the sensor 110 as input and preprocesses the received sensing signal to generate (obtain) state observation signal data representing the operating state of the device in operation. Preprocessing of the sensing signal may include, for example, filtering, noise reduction, moving average, Fourier transform, and wavelet transform to the frequency spectrum. For example, the state observation unit 111 outputs frequency domain data as state observation signal data, which is obtained by performing a Fourier transform on time-series sensing data from a current sensor that detects the power supply current of the motor within a defined time window. Alternatively, the state observation unit 111 may also output data as state observation signal data, which is obtained by performing a moving average on sensing data from a harmonic sensor that directly detects the harmonic components of the power supply current of the motor within a defined time window.

[0041] Next, the state observation unit 111 outputs the generated state observation signal data to the state determination unit 112 and the pattern extraction unit 116, and stores it in the model storage unit 119. In addition, the state observation unit 111 outputs the state observation signal data to the pattern extraction unit 116 for pre-learning until the classification model described later is generated.

[0042] The simulation unit 114 simulates the operation of the device 101 corresponding to the device model 121 by causing the device model 121 to perform simulation operations. The simulation unit 114 generates (obtains) simulated state observation signal data, which simulates the operation state of the device model during the simulation operation. The simulated state observation signal data includes the characteristic frequencies of each component of the device, which will be described later. Here, as will be described later, it has been found that each part of the device is closely related to the characteristic frequencies.

[0043] Next, the simulation unit 114 stores the characteristic frequencies of each structure of the device in the generated simulated state observation signal data in association with each part of the device in the characteristic frequency storage unit 115. Furthermore, the simulated state observation signal data generated by the simulation unit 114 can also be used as input data for the training data generated by the training data generation unit 117. Therefore, the simulation unit 114 can also store the generated simulated state observation signal data in the model storage unit 119 for use by the training data generation unit 117.

[0044] The characteristic frequency storage unit 115 stores the characteristic frequencies of each component of the device generated by the simulation unit 114 in association with each part of each device. Furthermore, although in Figure 1The example shown is that the characteristic frequency storage unit 115 is located inside the learning device 150. However, this embodiment is not limited to this. The characteristic frequency storage unit 115 may also be located outside the learning device 150 and inside the fault prediction system 100.

[0045] The pattern extraction unit 116 receives state observation signal data output from the state observation unit 111 as input. The pattern extraction unit 116 extracts time fluctuation patterns from the received state observation signal data. For example, such time fluctuation patterns include time fluctuation patterns of signal amplitude, time fluctuation patterns of signal amplitude in a specific frequency band, etc. For example, various patterns such as linear fitting, nth-order curve fitting, and exponential function fitting can be used as time fluctuation patterns.

[0046] The pattern extraction unit 116 obtains feature frequencies from the feature frequency storage unit 115. Using the obtained feature frequencies, the pattern extraction unit 116 extracts a time fluctuation pattern of amplitude for each device's feature frequency, thereby creating a set of time fluctuation patterns of amplitude extracted from each device. Then, the pattern extraction unit 116 outputs the created set of time fluctuation patterns of amplitude to the training data generation unit 117.

[0047] The training data generation unit 117 receives a set of amplitude time fluctuation patterns output from the pattern extraction unit 116 as input. The training data generation unit 117 obtains feature frequencies from the feature frequency storage unit 115. Using the received set of time fluctuation patterns and the obtained feature frequencies, the training data generation unit 117 generates training data for machine learning in the learning unit 118. Specifically, the training data is data that groups input data, frequency labels (or device part labels), and time point labels. The input data is simulated state observation signal data generated through simulation, which simulates the operating state of the device after the current time point (or, a specific time point, or a first time point) (during learning). The frequency labels (or device part labels) are labels representing feature frequencies associated with virtual fault parts of the device, and the time point labels are labels representing the elapsed time from the current time point (or, a specific time point, or a first time point).

[0048] Here, "virtual fault location" refers to a device part that may fail in the future. The training data generation unit 117 generates such grouped training data in a variety and large amount through simulation. Then, the training data generation unit 117 stores the generated training data in the model storage unit 119 and notifies the learning unit 118 that the training data has been generated and stored in the model storage unit 119.

[0049] After receiving the aforementioned notification from the training data generation unit 117, the learning unit 118 obtains the feature frequencies and training data from the feature frequency storage unit 115 and the model storage unit 119, respectively. Using the obtained feature frequencies and training data, the learning unit 118 performs machine learning on the features (time fluctuation patterns of the amplitude of the feature frequencies) shown in the simulated state observation signal data that are associated with the state of the equipment to be determined by the state determination unit 112 (especially the future fault state of the equipment).

[0050] In this type of machine learning, various well-known algorithms can be used, such as deep learning based on neural networks, support vector machines, random forests, and ensemble learning that combines these. Furthermore, the learning unit 118 generates a classification model for determining the future fault state of the equipment by performing machine learning in the manner described above. Next, the learning unit 118 stores the generated classification model in the model storage unit 119. In this way, by performing machine learning using training data generated by the training data generation unit 117, the learning unit 118 can characterize the relationship between simulated state observation signal data and the state of the equipment (especially fault states), and classify the state of the equipment based on the simulated state observation signal data.

[0051] The model storage unit 119 stores the state observation signal data generated by the state observation unit 111, the simulated state observation signal data generated by the simulation unit 114, the training data generated by the training data generation unit 117, and the classification model generated by the learning unit 118. Furthermore, although in Figure 1 The example shown is that the model storage unit 119 is located inside the learning device 150. However, this embodiment is not limited to this. The model storage unit 119 may also be located outside the learning device 150 and inside the fault prediction system 100 (including inside the state determination unit 112).

[0052] Furthermore, the state observation signal data generated by the state observation unit 111 can also be stored in a state observation signal data storage unit (not shown) different from the model storage unit 119. Additionally, the simulated state observation signal data generated by the simulation unit 114 can also be stored in a simulated state observation signal data storage unit (not shown) different from the model storage unit 119. Furthermore, the training data generated by the training data generation unit 117 can also be stored in a training data storage unit (not shown) different from the model storage unit 119. The state determination unit 112 refers to the model storage unit 119 when performing the above determination. The display unit 113 refers to the model storage unit 119 when performing the above display.

[0053] Equipment model 121 is a model obtained by modeling equipment 101. Equipment model 121 includes a motor model 122 obtained by modeling motor 102 and a load model 123 obtained by modeling load 103. In the case of multiple devices, each of the equipment models 121-1 to 121-M corresponding to the multiple devices, namely equipment 101-1 to equipment 101-M, includes a motor model 122 obtained by modeling motor and a load model 123 obtained by modeling load. This modeling can be performed at various levels of detail, and can reproduce, for example, shape, mass, material, and electromagnetic circuits from a 1-D model that approximates the action, or can use a 3-D model that simulates electromagnetic or mechanical actions.

[0054] <Learning Methods>

[0055] Next, refer to Figures 2-7 This describes an example of the learning method 200 in the embodiments of this disclosure. Figure 2 In this process, the learning method 200 is executed by the learning device 150.

[0056] In step S201, the training data generation unit 117 obtains at least one characteristic frequency of each device from each device model via the simulation unit 114 and the characteristic frequency storage unit 115.

[0057] In step S202, the state observation unit 111 receives a sensing signal from the sensor 110 that senses the state of the device, and performs preprocessing on the sensing signal as described above to generate state observation signal data representing the operating state of the device in operation. Alternatively, in step S202, the state observation unit 111 may also obtain state observation signal data representing the operating state of the device in operation from the device via the sensor 110. Then, the state observation unit 111 outputs the obtained state observation signal data to the pattern extraction unit 116.

[0058] In step S203, the pattern extraction unit 116 obtains characteristic frequencies from the characteristic frequency storage unit 115, and uses the obtained characteristic frequencies to extract, according to the past history of the state observation signal data output by the state observation unit 111, the time fluctuation patterns of the amplitude of each characteristic frequency or its surrounding area (hereinafter collectively referred to as "characteristic frequencies"). Furthermore, in step S203, the pattern extraction unit 116 creates a set of the time fluctuation patterns of the amplitude extracted according to the device. Next, the pattern extraction unit 116 outputs the created set of time fluctuation patterns of amplitude to the training data generation unit 117.

[0059] In step S204, the training data generation unit 117 generates simulated state observation signal data that simulates the operating state of the device after the current time point (or, the specific time point, the first time point) based on the time fluctuation pattern of the amplitude of each characteristic frequency associated with each part of each device up to the current time point (or, the specific time point, the first time point). Specifically, the training data generation unit 117 applies an extrapolation of the time fluctuation pattern of any amplitude selected from the set of amplitude time fluctuation patterns for any characteristic frequency per device, thereby generating simulated state observation signal data in multiple combinations. Therefore, the simulated state observation signal data represents the time fluctuation pattern of the amplitude generated by simulation after the specific time point (the first time point) of each characteristic frequency associated with each part of each device.

[0060] In step S205, the training data generation unit 117 generates training data that groups the input data, frequency labels, and time point labels. The input data is the generated simulated state observation signal data, the frequency labels are labels that represent the characteristic frequencies of the time fluctuation pattern of the amplitude applied to the part of the device, and the time point labels are labels that represent the elapsed time from the current time point (or, a specific time point, a first time point).

[0061] In step S206, the learning unit 118 uses training data that groups the simulated state observation signal data, frequency labels, and time point labels to learn fault determination conditions and generate a classification model for determining the future fault state of the equipment.

[0062] Next, learning method 200 returns from step S206 to step 202, and learning device 150 repeatedly executes steps S202 to S206, thereby updating the time fluctuation pattern of the amplitude from the current time point and generating training data to continue learning.

[0063] Next, refer to Figure 3 ,and Figure 2 Step S201 illustrates, in connection with, an example of device harmonics (“characteristic frequencies” of each device) in an embodiment of this disclosure.

[0064] exist Figure 3In this context, state observation signal data 301 represents signal data indicating the normal state when the motor 102 of device 101 is operating normally. On the other hand, state observation signal data 302 to 304 represent signal data indicating the fault state when different parts of the same motor of the same device experience failures. Furthermore, state observation signal data 302 to 304 can also represent the possibility of fluctuations in state observation signal data when a fault occurs. Moreover, these state observation signal data, namely state observation signal data 301 to 304, are signal data generated by the state observation unit 111 through Fourier transform and frequency analysis of the sensed signals from sensor 110.

[0065] Here, we find the following insight: if these parts malfunction, the characteristic frequencies associated with these parts will fluctuate.

[0066] exist Figure 3 In the normal state state observation signal data 301, at least the characteristic frequencies f0, f1, and f2 of each device generate harmonic components. These frequencies depend on the construction of the motor or load. For example, the motor's speed, power supply frequency, rotor structure, number of stator slots, and number of stator poles are related to these frequencies. Additionally, the load's gear number and gear ratio are related to these frequencies. How the frequencies fluctuate depends on the depth of the anomaly or fault. For example, the fault state state observation signal data 302 indicates the possibility that the amplitude of characteristic frequency f1 becomes greater than the amplitude of characteristic frequency f1 in the normal state state observation signal data 301. Furthermore, for example, the fault state state observation signal data 303 indicates the possibility that the amplitude of characteristic frequency f2 becomes greater than the amplitude of characteristic frequency f2 in the normal state state observation signal data 301. Furthermore, for example, the fault state state observation signal data 304 indicates the possibility that characteristic frequency f1 is modulated.

[0067] If a device has actually failed, state observation signal data for this failure state can be derived. However, from a failure prediction perspective, it is desirable to estimate signs of these possibilities before a failure occurs. Based on this reasoning and the above insights, in step S201, the characteristic frequencies of each part of the device associated with the device are obtained from the device model.

[0068] Next, refer to Figure 4 ,and Figure 2 Step S201 describes an example of a device model in an embodiment of this disclosure.

[0069] exist Figure 4In the diagram, device 401 simulates device 101 including motor 102 and load 103, while device model 402 simulates device model 121 including motor model 122 and load model 123. Additionally, in Figure 4 In the example, simulated state observation signal data 403 represents a characteristic frequency derived from the device model.

[0070] In this embodiment, in order to generate the training data as described above, the simulation unit 114 performs simulation using a device model obtained by modeling the device, thereby estimating the characteristic frequencies.

[0071] First, the simulation unit 114 obtains structural information about the motor or gear from the design data and measurement data of the device 401 via user input from the learning device 150. The simulation unit 114 obtains mechanical information such as shape, size, mass, material, number of gears, and gear ratio, as well as electromagnetic circuit information such as power supply frequency, number of slots in the motor, and number of poles in the motor as this structural information.

[0072] Next, the simulation unit 114 performs simulation using the device model 402 obtained by modeling this construction information through model parameters, generating (obtaining) simulated state observation signal data. For example, load torque, electromagnetic torque, rotor current / electromotive force / magnetic flux density, stator current / electromotive force / magnetic flux density, inertial torque, friction coefficient, rotational speed, and shaft frequency are used as model parameters.

[0073] Simulation unit 114 uses device model 402 to simulate various operating states, thereby obtaining simulated state observation signal data. The accuracy of the simulated state observation signal data varies greatly depending on the level of detail of the simulation and model (e.g., the number of model parameters). However, regardless of this level of detail, characteristic frequencies dependent on the structure can be estimated. On the other hand, the amplitude of each characteristic frequency is highly dependent on the simulation conditions or level of detail because it depends on the state of each part.

[0074] High-detail simulations can obtain simulated state observation signal data with the accuracy to reproduce amplitude. On the other hand, low-detail simulations can obtain simulated state observation signal data in a short time. Figure 4 The simulated state observation signal data 403 illustrates an example of signal data generated from a low-detail simulation. In this example, the characteristic frequency was estimated.

[0075] exist Figure 4In the example shown by the simulated state observation signal data 403, six characteristic frequencies, namely characteristic frequencies f0 to f5, were estimated. As described above, each of the six characteristic frequencies f0 to f5 is associated with a different part of the device 401. For example, f0 is the frequency associated with the power supply, f1 is the frequency associated with the stator, and f2 is the frequency associated with the rotor. Additionally, for example, f3 is the frequency associated with the first gear, f4 is the frequency associated with the second gear, and f5 is the frequency associated with the third gear.

[0076] Next, the simulation unit 114 stores the characteristic frequencies of each configuration of the device in the generated simulated state observation signal data in the characteristic frequency storage unit 115. As a result, the training data generation unit 117 can obtain the characteristic frequencies of each configuration of the device from the characteristic frequency storage unit 115.

[0077] Next, refer to Figure 5 ,and Figure 2 Step S203 is described in connection with an example of the time fluctuation of the amplitude of device harmonics in the embodiments of this disclosure.

[0078] exist Figure 5 In the system, the normal state observation signal data 301 represents the state at time t0 (e.g., the current state). The fault state observation signal data 302, representing the state at time t2 (e.g., the future state), indicates a fault state where the amplitude of the characteristic frequency f1 increases. The fault prediction system 100 determines which time point in the sequence t0 < t1 < t2 the state observation signal data representing the operating state of the device in operation is closest to. Therefore, the learning device 150 uses training data to learn how to extract and extrapolate the temporal fluctuation pattern 501 of the amplitude of the characteristic frequency f1 from the past history of the state observation signal data to predict and classify the state observation signal data at time t1. Furthermore, this temporal fluctuation pattern of amplitude may vary by device or by location (characteristic frequency), therefore, the learning device 150 learns multiple temporal fluctuation patterns of amplitude.

[0079] Next, refer to Figure 6 ,and Figure 2 Step S203 describes, in connection with, an example of the time fluctuation pattern of the device harmonic amplitude in an embodiment of the present disclosure.

[0080] exist Figure 6In the diagram, amplitude time fluctuation patterns 601 to 603 represent examples of changes in time fluctuation patterns related to the amplitude of any characteristic frequency. For example, amplitude time fluctuation pattern 601 represents a pattern that increases linearly over time. Additionally, amplitude time fluctuation pattern 602 represents a pattern that increases exponentially. Furthermore, amplitude time fluctuation pattern 603 represents a pattern that decreases quadraticly.

[0081] exist Figure 4 In the simulation of device model 402, it is sometimes difficult to fully obtain these time fluctuation patterns. Therefore, in embodiments of this disclosure, time fluctuation patterns are extracted from actual state observation signal data. The learning device 150 extracts time fluctuation patterns of amplitude separated by device or by characteristic frequency and applies them to the simulation, thereby enabling the simulation of multiple fluctuation possibilities. Thus, the learning device 150 can learn time fluctuation patterns of multiple amplitudes. The learning of such time fluctuation patterns of multiple amplitudes will be described below.

[0082] Next, refer to Figure 7 ,and Figure 2 Steps S204 and S205 illustrate, in connection with, an example of the generation of training data in an embodiment of this disclosure.

[0083] First, set the status observation signal data 710, 720, 730, etc., of the M devices up to the present normal state as...

[0084] [Mathematical Expression 1]

[0085] S m (t), m=1…M.

[0086] If these state observation signal data are segmented according to the characteristic frequencies obtained from the corresponding device models, they can be represented as follows:

[0087] [Mathematical Expression 2]

[0088]

[0089] Here, n = 1…N m a represents the characteristic frequency index of device m. mn (t) represents the time fluctuation of the amplitude of the characteristic frequency n of device m, f mn (t) represents the characteristic frequency n of device m. For example, as... Figure 7 As shown, frequencies f10 (7101), f11 (7102), and f12 (7103) exist as characteristic frequencies of device 1. Similarly, as Figure 7As shown, frequencies t20 (7201), f21 (7202), f22 (7203), f30 (7301), f31 (7302), and f32 (7303) exist as characteristic frequencies of other devices.

[0090] If we take equation (1) as the vector representation of the composite function, then it can be expressed as:

[0091] [Mathematical Expression 3]

[0092]

[0093]

[0094] S m (t)=a m f m (t).

[0095] The set of time fluctuation patterns of amplitude extracted from M devices, 700, can be represented as:

[0096] [Mathematical Expression 4]

[0097] A={a mn}, m=1…M, n=1…N m .

[0098] The training data generation unit 117 generates training data related to the virtual fault location (associated with the characteristic frequency n) of the device m based on the time fluctuation pattern set 700 of the amplitude. Specifically, the training data is generated in the following manner.

[0099] The training data generation unit 117 extracts an arbitrary amplitude time fluctuation pattern from the amplitude time fluctuation pattern set 700.

[0100] [Mathematical Expression 5]

[0101] α′ i ∈A.

[0102] Next, the training data generation unit 117 obtains the time fluctuation pattern vector of the aforementioned amplitude from the device m.

[0103] [Mathematical Expression 6]

[0104]

[0105] Choose any characteristic frequency n as the virtual fault location, and generate a time fluctuation pattern a of its amplitude. mn It was multiplied by an arbitrary gain.

[0106] [Mathematical Expression 7]

[0107] g

[0108] of

[0109] [Mathematical Expression 8]

[0110] a′ i

[0111] Replaced vector

[0112] [Mathematical Expression 9]

[0113]

[0114] Here, the aforementioned gain is a coefficient set to accelerate the attainment of a fault state.

[0115] Next, the training data generation unit 117 extrapolates the time fluctuations of the amplitude to generate an arbitrary future time point t. k Simulated state observation signal data

[0116] [Mathematical Expression 10]

[0117] S′ mn (t k )=a′ mn f m (t k ).

[0118] For example, Figure 7 The simulated state observation signal data 713 is obtained by extrapolating the time fluctuation of the amplitude of the characteristic frequency f10 of device 1, the simulated state observation signal data 723 is obtained by extrapolating the time fluctuation of the amplitude of the characteristic frequency f21 of device 2, and the simulated state observation signal data 733 is obtained by extrapolating the time fluctuation of the amplitude of the characteristic frequency f31 of device 3.

[0119] The training data generation unit 117 generates simulated state observation signal data that serves as input data by repeatedly performing the above processing.

[0120] [Mathematical Expression 11]

[0121] S′ mn (t k )

[0122] Frequency label n and time point label t k A set D of training data for device m as shown below. m The frequency label n is a label indicating which characteristic frequency changed and associated with the virtual fault location; the time point label t k It is a label that indicates the elapsed time from the current point in time (when learning).

[0123] [Mathematical Expression 12]

[0124] D m ={(S′ mn (t k ), n, t k )}

[0125] The training data generation unit 117 sets the amplitude of the characteristic frequency and the time point t2 of the virtual fault location under the fault state in the following manner.

[0126] When the simulation unit 114 performs a highly detailed simulation using the equipment model, the training data generation unit 117 sets the amplitude and time point of the characteristic frequency under the fault state obtained after the simulation to the amplitude and time point t2 of the characteristic frequency under the fault state of the virtual fault location. Furthermore, the training data generation unit 117 can also extrapolate the time fluctuation of the amplitude in a way that converges to the simulated state observation signal data obtained through the simulation of the fault state when extrapolating the time fluctuation of the amplitude. Such highly detailed simulations include, for example, using CAD (Computer Aided Design) to simulate the actual faulty location by placing it into the equipment model as a shape or parameter change; or simulating the location where a fault is expected to occur by adding detailed parameter changes. In this case, training data can be generated in a way that converges to the fault state simulated by the simulation unit 114. In addition, the training data generation unit 117 not only sets the amplitude and time point of the characteristic frequency under the fault state obtained by the simulation unit 114, but can also use the simulated state observation signal data itself generated during the highly detailed simulation performed by the simulation unit 114 as input data for the training data.

[0127] When the simulation unit 114 performs a low-detail simulation using the device model, or when the simulation unit 114 can determine the characteristic frequency based on the device's structural parameters without performing a low-detail simulation, the training data generation unit 117, for example, sets the amplitude of the characteristic frequency pre-input by the user of the learning device 150 to the amplitude of the characteristic frequency under the fault state of the virtual fault location, and sets the time point corresponding to the set characteristic frequency amplitude to the time point t2 under the fault state of the virtual fault location.

[0128] Furthermore, when replacing the elements of the amplitude time fluctuation pattern vector (amplitude time fluctuation pattern), the training data generation unit 117 can also select multiple feature frequencies associated with multiple virtual fault locations respectively to replace multiple amplitude time fluctuation patterns. This allows the generation of training data that considers the possibility of multiple virtual fault locations failing simultaneously.

[0129] Moreover, instead of replacing the time fluctuation mode a with an amplitude multiplied by an arbitrary gain, the time fluctuation mode a with an amplitude is... mn The training data generation unit 117 can also perform the following operations. Furthermore, the training data generation unit 117 can also replace the amplitude time fluctuation pattern a with the following amplitude time fluctuation pattern. mn The aforementioned amplitude time fluctuation pattern is an amplitude time fluctuation pattern a obtained based on the actual state observation signal data generated by the state observation unit 111. mn The amplitude time fluctuation pattern obtained by weighted synthesis is obtained from the simulation state observation signal data obtained by simulation using the device model through simulation unit 114.

[0130] Here, the weighting can also be set in such a way that the closer to the current time point, the higher the weight of the actual state observation signal data, and the further away from the current future time point, the higher the weight of the simulated state observation signal data. For example, it can also be set such that in the first half of the time interval between the current time point t0 and the time point t2 under the fault state, the weight of the actual state observation signal data is higher than the weight of the simulated state observation signal data, while in the second half of the time interval between the current time point t0 and the time point t2 under the fault state, the weight of the simulated state observation signal data is higher than the weight of the actual state observation signal data.

[0131] Therefore, training data can be generated in a way that smoothly converges to the fault state obtained through simulation. Alternatively, if the state observation signal data of the actual fault state of other devices has already been obtained by the state observation unit 111, the weighting can be set to make the weight of the state observation signal data higher. Therefore, training data can be generated in a way that converges to the actual fault state.

[0132] In addition, if the status observation signal data of other devices up to the actual fault state has been obtained by the status observation unit 111, the training data generation unit 117 can set the amplitude of the characteristic frequency and the time point t2 of the virtual fault location under the fault state based on the status observation signal data up to the actual fault state, or use the status observation signal data up to the actual fault state itself as the simulated status observation signal data.

[0133] Although the time point label t was explained above k This indicates the elapsed time since the current time point t0 (during learning), but this implementation is not limited to this; the time point label t... k It can also be set to the value obtained by subtracting the elapsed time from the current time point t0 from the time point t2 associated with the fault state.

[0134] <Variation Example>

[0135] In addition, although Figure 1 The example shown includes a simulation unit 114 and a device model 121. However, in this embodiment, the desired effect can still be achieved even without the simulation unit 114 and the device model 121. This variation corresponds to the case described above where the characteristic frequency can be determined based on the device's structural parameters. In this variation, the characteristic frequency is stored in the characteristic frequency storage unit 115, for example, through user input processing of the learning device 150. According to this variation, instead of the simulation unit 114 performing simulation using the device model 121, the pattern extraction unit 116, the training data generation unit 117, and the learning unit 118 perform the aforementioned processing using the characteristic frequencies stored in the characteristic frequency storage unit 115. Therefore, the processing load on the learning device 150 can be reduced.

[0136] (Effects of the learning device and learning method in the implementation)

[0137] The learning device 150 and its learning method 200 in the embodiments of this disclosure have the above-described structure. Therefore, by predicting the time fluctuation pattern of the amplitude of the characteristic frequency, it is possible to generate simulated state observation signal data that simulates the possibility of multiple fluctuations after the current time point (or, a specific time point, a first time point), and generate training data that includes simulated state observation signal data that matches the structure of each device.

[0138] Here, the simulated state observation signal data includes amplitude data of characteristic frequencies that clearly represent the fault state of the equipment. Therefore, the learning device 150 can easily obtain the data clearly representing the fault state used in the training data and can perform learning for high-precision determination of the equipment's fault state. Furthermore, given the aforementioned structure, the learning device 150 and its learning method 200 can generate training data that converges to the simulated fault state when simulating the fault state with high detail. Thus, learning for determining the equipment's fault state with even higher precision is possible.

[0139] <Fault Prediction Methods>

[0140] Next, refer to Figure 8 An example of a fault prediction method 800 in an embodiment of this disclosure is described. The fault prediction method 800 is executed by the fault prediction system 100.

[0141] In step S801, the state observation unit 111 receives a sensing signal from the sensor 110 that senses the device state, and performs preprocessing on the sensing signal as described above to generate state observation signal data representing the operating state of the device in operation. Alternatively, in step S801, the state observation unit 111 may also obtain state observation signal data representing the operating state of the device in operation from the device via the sensor 110. Then, the state observation unit 111 outputs the obtained state observation signal data to the state determination unit 112.

[0142] In step S802, the state determination unit 112 classifies the state observation signal data output from the state observation unit 111 according to the classification model generated by the learning unit 118 and stored in the model storage unit 119 in step S205 of the learning method 200. Specifically, the state determination unit 112 calculates (obtains) the similarity or probability between the state observation signal data and the simulated state observation signal data according to the classification model, and classifies the state observation signal as a state observation signal belonging to the amplitude time fluctuation pattern associated with the simulated state observation signal data that is closest in similarity or has the highest probability.

[0143] In step S803, the state determination unit 112 determines the fault state (fault probability) of each part based on the tag associated with the simulated state observation signal data that has the closest similarity or the highest probability. The specific determination is as follows: When the tag is a characteristic frequency, the state determination unit 112 refers to the characteristic frequency storage unit 115 to determine the fault part corresponding to the characteristic frequency (frequency tag). When the tag is a time point, the state determination unit 112 determines the fault part based on the corresponding amplitude's time fluctuation pattern; specifically, it subtracts the tag t from the time point t2 associated with the fault state. k Determine the predicted time until the faulty part fails.

[0144] In addition, at the time point label t k When the time point label t is set to be the value obtained by subtracting the elapsed time from the current time point (or, a specific time point, the first time point) t0 during learning from the time point t2 associated with the fault state, the time point label t is... k This is essentially the prediction time until the faulty part fails. In this way, the status determination unit 112 can use the current status observation signal data representing the current operating state of the equipment and the classification model to determine the fault state of each part.

[0145] In step S804, the display unit 113 displays the determination result of the fault state determined by the state determination unit 112 in step S803. Specifically, the display unit 113 displays to the user the determined fault location and the predicted time until the fault location fails.

[0146] The fault prediction system 100 performs fault prediction and judgment by repeatedly executing the above steps and displays the results to the user.

[0147] Furthermore, the fault state determined by the state determination unit 112 and displayed by the display unit 113 can be not only the fault state associated with the analog state observation signal data that is closest in similarity or has the highest probability, but also the fault state associated with a predetermined number of analog state observation signal data in order of similarity from high to low or in order of probability from high to low.

[0148] <Fault Prediction Display>

[0149] Next, refer to Figure 9 This describes an example of fault prediction in the embodiments of this disclosure. Figure 9 This is a diagram illustrating an example of fault prediction in an embodiment of this disclosure.

[0150] In step S804 of the fault prediction method 800, such as Figure 9 As shown, in addition to displaying the current state observation signal data 901, the display unit 113 also displays the simulated state observation signal data 902 and 903, which serve as the basis for its determination. Furthermore, the display unit 113 displays the simulated state observation signal data 902 and 903 side by side in order of similarity or probability from high to low.

[0151] In addition, such as Figure 9 As shown by the vertical arrow 904, the display area for the simulated condition observation signal data can also be a scrollable user interface that displays the judgment results of the fault status of multiple parts. This allows for the comparison of the probabilities of multiple simulated condition observation signal data.

[0152] Moreover, such as Figure 9 As shown by the horizontal arrow 905, in response to the user's horizontal scrolling instruction, the display unit 113 can also change the time point of the simulated state observation signal data to display the temporal fluctuations of the characteristic frequency amplitude. For example, the display unit 113 can also display the simulated state observation signal data at the indicated time point by rewinding the simulation time point for a left scrolling instruction and advancing the simulation time point for a right scrolling instruction. This allows the user to be informed how the simulated state observation signal data fluctuates to reach the fault state. Furthermore, for the left scrolling instruction, the display unit 113 can display not only the simulated state observation signal data but also the current state observation signal data, and even past state observation signal data.

[0153] (Effects of the fault prediction system and fault prediction method in the implementation plan)

[0154] The fault prediction system 100 and fault prediction method 800 in the embodiments of this disclosure, in addition to having the effects of the learning device 150 and the learning method 200, also have the above-described structure. Therefore, the fault state of the equipment can be determined with high accuracy using the classification model generated by the learning device 150. Furthermore, because the fault prediction system 100 and fault prediction method 800 have the above-described structure, they can display simulated state observation signal data with a similarity close to or high probability to the current state observation signal, allowing the user to judge the basis of the judgment and the reliability of the fault prediction.

[0155] In the above embodiments, the table using terms such as "...part" for each constituent element can also be replaced with other expressions such as "...circuitry", "...assembly", "...device", "...unit", or "...module".

[0156] The embodiments have been described above with reference to the accompanying drawings; however, this disclosure is not limited to these examples. Those skilled in the art will obviously conceive of various modifications or alterations within the scope of the claims. It should be understood that these modifications or alterations also fall within the technical scope of this disclosure. Furthermore, the constituent elements of the above embodiments can be combined arbitrarily without departing from the spirit of this disclosure.

[0157] This disclosure can be implemented in software, hardware, or software in cooperation with hardware. The functional blocks used in the above embodiments are implemented partially or wholly as LSIs (Large Scale Integration) of integrated circuits, and the processes described in the above embodiments can also be controlled partially or wholly by a single LSI or a combination of LSIs. An LSI can be composed of a single chip, and can include some or all of the functional blocks on a single chip. An LSI can also include data input and output. Depending on the degree of integration, an LSI can also be called an IC (Integrated Circuit), a system LSI, a super LSI, or an ultra-large LSI.

[0158] The method of integrating LSIs is not limited to LSIs; it can also be implemented using dedicated circuits, general-purpose processors, or special-purpose processors. Alternatively, it can utilize FPGAs (Field Programmable Gate Arrays) that can be programmed after LSI fabrication, or reconfigurable processors that can reconfigure the connections or settings of the circuit blocks within the LSI. This disclosure can also be implemented for digital or analog processing.

[0159] Furthermore, if advancements in semiconductor technology or the emergence of other derivative technologies lead to integrated circuit technologies that can replace LSIs, these technologies could also be used to integrate functional blocks. There are also possibilities for applications such as biotechnology.

[0160] <Summary of this disclosure>

[0161] One embodiment of the learning apparatus of this disclosure includes: a pattern extraction unit that extracts a time fluctuation pattern of the amplitude of a characteristic frequency associated with a part of the device from state observation signal data representing the operating state of the device up to a first time point; a training data generation unit that generates simulated state observation signal data representing a time fluctuation pattern of the amplitude of the characteristic frequency after the first time point based on the time fluctuation pattern of the amplitude of the characteristic frequency, and generates training data including the simulated state observation signal data; and a learning unit that uses the training data to generate a classification model for determining the fault state of a part of the device.

[0162] In the above-mentioned learning device, the training data generation unit generates training data that groups the simulated state observation signal data, frequency labels, and time point labels. The frequency label is a label representing the characteristic frequency, and the time point label is a label representing the elapsed time from the first time point.

[0163] In the above-described learning device, the training data generation unit generates the simulated state observation signal data by extrapolating the time fluctuation pattern of the amplitude extracted from the state observation signal data up to the first time point.

[0164] The learning device described above also includes a simulation unit that performs a simulation using a device model obtained by modeling the device, thereby estimating the characteristic frequency.

[0165] In the learning device described above, when the simulation unit uses a device model obtained by modeling the device to simulate the fault state of a part of the device, the training data generation unit generates the training data in a manner that converges to the simulated fault state of the part of the device.

[0166] One embodiment of the learning method disclosed herein is executed by a learning device and includes the following steps performed by the learning device: extracting a time fluctuation pattern of the amplitude of a characteristic frequency associated with a part of the device from state observation signal data representing the operating state of the device up to a first time point; generating simulated state observation signal data representing a time fluctuation pattern of the amplitude of the characteristic frequency after the first time point based on the time fluctuation pattern of the amplitude of the characteristic frequency; generating training data including the simulated state observation signal data; and using the training data to generate a classification model for determining the fault state of a part of the device.

[0167] One embodiment of the fault prediction system disclosed herein includes: the learning device described above; and a state determination unit, which uses current state observation signal data representing the current operating state of the device and the classification model to determine the fault state of a part of the device.

[0168] The aforementioned fault prediction system also includes a display unit that displays the determination result of the fault status of a part of the device.

[0169] In the aforementioned fault prediction system, the display unit shows the temporal fluctuation of the amplitude of the characteristic frequency.

[0170] In the above-mentioned fault prediction system, there are multiple parts of the device and multiple characteristic frequencies, and the display unit displays the fault status determination results of multiple parts of the device.

[0171] The entire contents of the specification, drawings and abstract of the specification contained in Japanese Patent Application No. 2020-198545, filed on November 30, 2020, are incorporated herein by reference.

[0172] Industrial applicability

[0173] One embodiment of this disclosure is useful for a fault prediction system.

[0174] Explanation of reference numerals in the attached figures

[0175] 100 Fault Prediction System

[0176] 101 equipment

[0177] 110 sensor

[0178] 111 Status Observation Department

[0179] 112 Status Determination Unit

[0180] 113 Display Unit

[0181] 114 Simulation Department

[0182] 115 Feature Frequency Storage Unit

[0183] 116 Pattern Extraction Department

[0184] 117 Training Data Generation Department

[0185] 118 Study Department

[0186] Model 119 Storage Department

[0187] 121 Equipment Model

[0188] 150 learning devices

Claims

1. A learning device characterized by comprising: including: a pattern extraction section that extracts a time fluctuation pattern of an amplitude of a characteristic frequency associated with each of parts of a device from state observation signal data up to a first time point indicating an operation state of the device; a training data generation section that generates simulated state observation signal data indicating a time fluctuation pattern of the amplitude of the characteristic frequency after the first time point based on the time fluctuation pattern of the amplitude of the characteristic frequency, and generates training data including the simulated state observation signal data; and a learning section that generates a classification model for predicting a future failure state of the parts of the device using the training data.

2. The learning apparatus according to claim 1, wherein the training data generation section generates the training data grouping the simulated state observation signal data, a frequency label indicating the characteristic frequency, and a time point label indicating an elapsed time from the first time point.

3. The learning apparatus according to claim 1, wherein the training data generation section generates the simulated state observation signal data by extrapolating the time fluctuation pattern of the amplitude extracted from the state observation signal data up to the first time point.

4. The learning apparatus according to claim 1, further comprising: a simulation section that simulates using a device model modeling the device, thereby estimating the characteristic frequency.

5. The learning apparatus according to claim 4, wherein in a case where the simulation section simulates a failure state of the parts of the device using the device model modeling the device, the training data generation section generates the training data in a manner converging to the simulated failure state of the parts of the device.

6. A learning method executed by a learning device, the learning method characterized by comprising: including the following steps performed by the learning apparatus: extracting a time fluctuation pattern of an amplitude of a characteristic frequency associated with each of parts of a device from state observation signal data up to a first time point indicating an operation state of the device; generating simulated state observation signal data indicating a time fluctuation pattern of the amplitude of the characteristic frequency after the first time point based on the time fluctuation pattern of the amplitude of the characteristic frequency; generating training data including the simulated state observation signal data; and generating a classification model for predicting a future failure state of the parts of the device using the training data.

7. A failure prediction system characterized by, including: the learning apparatus according to claim 1; and a state determination section that predicts a future failure state of the parts of the device using current state observation signal data indicating a current operation state of the device and the classification model.

8. The failure prediction system according to claim 7, further comprising: a display section that displays a prediction result of the failure state of the parts of the device.

9. The failure prediction system according to claim 8, wherein the display section displays a time fluctuation of the amplitude of the characteristic frequency.

10. The failure prediction system according to claim 8, wherein there are a plurality of parts of the device and a plurality of characteristic frequencies, ​ The display section displays a prediction result of a failure state of a part of each of a plurality of the devices.

Images