Method and device for constructing motor fault diagnosis model and motor fault diagnosis method

Abstract

The invention provides a method and device for constructing a motor fault diagnosis model and a motor fault diagnosis method. The method provided by the invention comprises the following steps: constructing an initial model; acquiring multiple groups of health state data of the motor in a health state, and taking the multiple groups of health state data as a training sample set; based on an adversarial training technology, training the initial model by using the training sample set to obtain a motor fault diagnosis model; and the motor fault diagnosis model is used for outputting the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data so as to carry out fault diagnosis based on the coding loss and the classification result. According to the method and device for constructing the motor fault diagnosis model and the motor fault diagnosis method provided by the invention, training is carried out only by relying on the health state data, and the generator optimizes the multiple loss functions, so that the modeling of the model on the health state data is more accurate, and potential faults can be accurately detected and diagnosed.

Description

Technical Field

[0001] The present application relates to the technical field of motor fault diagnosis, and in particular to a method and device for constructing a motor fault diagnosis model and a motor fault diagnosis method. Background Art

[0002] With the continuous development of industrial automation, servo motors are widely used in military, aerospace, mining equipment, industrial manufacturing and other fields. However, servo motors may have various faults during operation, such as winding short circuit, power component failure, etc., which will seriously affect the normal operation of the equipment. Therefore, it is very important to implement effective health management and fault diagnosis of servo motors. In the aviation field, the motor is the power source of the hydraulic pump, and its working state directly affects the performance of the entire airborne hydraulic system. When the motor has an early fault, fault isolation measures must be taken immediately to avoid serious flight accidents.

[0003] At present, the main research methods for motor fault diagnosis tasks are model-based fault diagnosis methods and data-driven fault diagnosis methods. This method establishes a mathematical model of the motor, simulates the current, voltage, and vibration response laws of the motor in healthy and various fault states, and then extracts and screens the collected current, voltage, and vibration signals based on signal analysis methods, and then classifies and identifies the extracted features based on certain rules and thresholds to achieve the fault diagnosis task. This method has extremely high requirements on the modeling accuracy of the motor and the quality of the monitoring signal, and is easily interfered by noise. The data-driven method directly relies on labeled sample data to train a fault diagnosis neural network model, which can adaptively extract and identify fault-sensitive feature vectors hidden in the original monitoring signal, and can achieve end-to-end diagnosis from original monitoring data to fault mode. However, the data-driven intelligent diagnosis method for motor faults requires a labeled fault data set under actual working conditions, but it is difficult to obtain a full fault type data set with sufficient sample size in actual industrial scenarios, which seriously restricts the engineering application scenarios of intelligent fault diagnosis algorithms based on neural networks.

[0004] In actual industrial scenarios, there is sufficient monitoring sample data of the objects being tested in healthy working conditions. How to monitor the abnormal working conditions of the motor based only on healthy data under zero-fault sample conditions is a difficult problem that needs to be solved urgently in actual engineering scenarios. Summary of the invention

[0005] In view of this, the present application provides a method, device and motor fault diagnosis method for constructing a motor fault diagnosis model, so as to realize the monitoring of motor faults under the condition of zero fault samples.

[0006] Specifically, the present application is implemented through the following technical solutions: A first aspect of the present application provides a method for constructing a motor fault diagnosis model, the method being used to construct a motor fault diagnosis model under a zero fault sample condition, the method comprising: Constructing an initial model; wherein the initial model includes a generator and a discriminator; the generator includes a first encoder, a decoder and a second encoder; the first encoder is used to extract features from the input original state data to obtain a first feature vector, the decoder is used to reconstruct data based on the first feature vector to obtain reconstructed state data, and the second encoder is used to extract features from the reconstructed state data to obtain a second feature vector; the discriminator is used to output a classification result based on the original state data and the reconstructed state data; wherein the discriminator outputs a classification loss based on the authenticity of the input data to distinguish between real data and reconstructed data; the generator uses a weighted loss of data reconstruction loss, coding loss and the classification loss as a loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the coding loss is characterized by the difference between the second feature vector and the first feature vector; Acquire multiple sets of health status data of the motor in a healthy state, and use the multiple sets of health status data as training sample sets; Based on the adversarial training technology, the initial model is trained using the training sample set to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data, so as to jointly perform fault diagnosis based on the coding loss and the classification result.

[0007] A second aspect of the present application provides a motor fault diagnosis method, the motor fault diagnosis method is implemented based on the motor fault diagnosis model, and the motor fault diagnosis model is constructed based on the method for constructing a motor fault diagnosis model described in the first aspect of the present application; the method comprises: Get the real-time status data of the motor; Inputting the real-time status data into the motor fault diagnosis model, so that the motor fault diagnosis model outputs the real-time encoding loss and the real-time classification result corresponding to the real-time status data; When the real-time classification result is false and the real-time coding loss is greater than a preset threshold, the real-time status data is determined to be abnormal data and the motor is faulty; wherein the preset threshold is the sum of a baseline coding loss and a preset margin of the motor fault model; and the baseline coding loss is the average value of a preset number of coding losses obtained when the motor fault model processes a preset number of groups of health status data.

[0008] The third aspect of the present application provides a device for constructing a motor fault diagnosis model, the device comprising a construction module, an acquisition module and a training module; wherein: The construction module is used to construct an initial model; wherein the initial model includes a generator and a discriminator; the generator includes a first encoder, a decoder and a second encoder; the first encoder is used to extract features from the input original state data to obtain a first feature vector, the decoder is used to reconstruct data based on the first feature vector to obtain reconstructed state data, and the second encoder is used to extract features from the reconstructed state data to obtain a second feature vector; the discriminator is used to output a classification result according to the original state data and the reconstructed state data; wherein the discriminator outputs a classification loss according to the authenticity of the input data to distinguish between real data and reconstructed data; the generator uses a weighted loss of data reconstruction loss, coding loss and the classification loss as a loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the coding loss is characterized by the difference between the second feature vector and the first feature vector; The acquisition module is used to acquire multiple groups of health status data of the motor in a healthy state, and use the multiple groups of health status data as training sample sets; The training module is used to train the initial model using the training sample set based on adversarial training technology to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data, so as to perform fault diagnosis based on the coding loss and the classification result.

[0009] The method, device and motor fault diagnosis method for constructing a motor fault diagnosis model provided by the present application construct an initial model and obtain multiple groups of health status data of the motor in a healthy state, and then use these multiple groups of health status data as training sample sets, based on adversarial training technology, use the training sample sets to train the initial model to obtain a motor fault diagnosis model. In this way, on the first hand, through the adversarial training mechanism of the generator and the discriminator, the model can identify potential abnormal states by learning the health status data, and provides a method for constructing a motor fault diagnosis model under zero fault sample conditions, which can train a motor with abnormal motor conditions by relying only on the rich and easily accessible health status data in actual industrial scenarios as training samples. A motor fault diagnosis model with state recognition ability is proposed to realize fault monitoring tasks. In addition, the motor fault diagnosis model only relies on healthy status data for training, which overcomes the requirement of intelligent fault diagnosis algorithm based on deep learning technology for the completeness of prior data set, greatly reduces the development cost of the algorithm, and improves its applicability and practicality. Secondly, the motor fault diagnosis model only relies on healthy status data for training, and the generator optimizes multiple loss functions to make the modeling of healthy status data more accurate, so as to accurately detect and diagnose potential faults. Thirdly, the model outputs the coding loss and classification results corresponding to the status data, and can better identify faults by combining the coding loss and classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A flowchart of a method for constructing a motor fault diagnosis model according to an embodiment of the present invention; Figure 2 A schematic diagram of the structure of an initial model shown as an exemplary embodiment of the present application; Figure 3 A flowchart of a motor fault diagnosis method embodiment 1 provided in the present application; Figure 4 This is a structural schematic diagram of embodiment 1 of the device for constructing a motor fault diagnosis model provided in the present application. DETAILED DESCRIPTION

[0011] Here, exemplary embodiments are described in detail, and examples thereof are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application.

[0012] The terms used in this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The singular forms of "a", "said" and "the" used in this application are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in this article refers to and includes any or all possible combinations of one or more associated listed items.

[0013] It should be understood that although the terms first, second, third, etc. may be used in the present application to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0014] Specific embodiments are given below to introduce the technical solution of the present application in detail.

[0015] Figure 1 This is a flowchart of the first embodiment of the method for constructing a motor fault diagnosis model provided by this application. Please refer to Figure 1 The method provided in this embodiment is used to construct a motor fault diagnosis model under zero sample conditions, and the method includes: S101. Construct an initial model; wherein the initial model includes a generator and a discriminator; the generator includes a first encoder, a decoder and a second encoder; the first encoder is used to extract features from the input original state data to obtain a first feature vector, the decoder is used to reconstruct data based on the first feature vector to obtain reconstructed state data, and the second encoder is used to extract features from the reconstructed state data to obtain a second feature vector; the discriminator is used to output a classification result based on the original state data and the reconstructed state data; wherein the discriminator outputs a classification loss based on the authenticity of the input data to distinguish between real data and reconstructed data; the generator uses a data reconstruction loss, a coding loss and a weighted loss of the classification loss as a loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the coding loss is characterized by the difference between the second feature vector and the first feature vector.

[0016] Specifically, the initial model is a neural network based on adversarial learning, including a generator and a discriminator. The generator is used to extract features and reconstruct signals of the input original state data, and the discriminator is used to identify whether the input data is real data or generated data.

[0017] Specifically, Figure 2This is a schematic diagram of the structure of the initial model shown in an exemplary embodiment of the present application. Figure 2 The generator includes two encoders and a decoder, and adopts the form of encoder-decoder-encoder series connection, that is, the generator includes a cascaded first encoder, a decoder and a second encoder; wherein the first encoder is used to extract features from the input original state data, convert the high-dimensional original state data into a low-dimensional potential representation, remove noise and redundant information in the original state data, highlight the sensitive feature vector in the original state data that reflects the health status of the motor, and obtain the first feature vector of the original state data; further, the decoder is used to reconstruct data based on the first feature vector, reconstruct the input data, and obtain reconstructed state data; further, the second encoder is used to extract features from the reconstructed state data and obtain the second feature vector of the reconstructed state data.

[0018] Optionally, in a possible implementation, the first encoder, the decoder, and the second encoder use a fully connected neural network as a basic network structure.

[0019] Furthermore, when the state data is in the form of picture data, the first encoder and the second encoder use convolutional neural networks; and the decoder uses a deconvolutional neural network.

[0020] Specifically, the discriminator is used to output a classification result according to the state data and the reconstructed state data. In specific implementation, the original state data at the current moment and the reconstructed state data corresponding to the data are input into the discriminator, and the discriminator outputs a classification loss according to the authenticity of the input data, which is used to distinguish between real data and reconstructed data.

[0021] The classification loss label output by the discriminator is 0 or 1, where 0 represents reconstructed data and 1 represents real data. It can be understood that there are differences in some features between the original state data and the reconstructed state data, and the discriminator determines whether the input data is real state or reconstructed data based on these differences.

[0022] For further information, please refer to Figure 2 , the generator uses the weighted loss of data reconstruction loss, encoding loss and the classification loss as the loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the encoding loss is characterized by the difference between the second eigenvector and the first eigenvector.

[0023] Specifically, the loss of the generator is a weighted loss of data reconstruction loss, coding loss, and classification loss. The weighted coefficients of data reconstruction loss, coding loss, and classification loss are set according to actual needs and are not limited in this embodiment. For example, in a possible implementation, the loss of the generator is the sum of data reconstruction loss, coding loss, and classification loss.

[0024] Optionally, in a possible implementation, the difference between the reconstructed state data and the original state data and / or the difference between the second eigenvector and the first eigenvector is Manhattan distance or Euclidean distance. For example, in a possible implementation, the data reconstruction loss is the Manhattan distance between the original state data and the reconstructed state data; the encoding loss is the Euclidean distance between the first eigenvector and the second eigenvector.

[0025] S102: Acquire multiple groups of health status data of the motor in a healthy state, and use the multiple groups of health status data as training sample sets.

[0026] Optionally, the set of state data of the motor includes specific state parameters among all state parameters of the motor, and the specific state parameters are state parameters that are strongly correlated with the health state of the motor.

[0027] It should be noted that the specific state parameter may be determined based on experience. Optionally, in a possible implementation, the method for determining the specific state parameter includes: (1) Establishing a simulation model of the motor according to the physical characteristics and working principle of the motor, and obtaining the physical response characteristics of the motor under various operating conditions based on the simulation model; wherein the various operating conditions include healthy state conditions and fault state conditions, and the fault state conditions include various mechanical fault conditions and electrical fault conditions.

[0028] (2) Screening a monitoring signal reflecting a fault from the physical response characteristics, and performing time domain, frequency domain, and time-frequency domain analysis on the monitoring signal to extract a plurality of characteristic quantities of the monitoring signal.

[0029] (3) For each characteristic quantity, calculate the degree of change of the characteristic quantity under the fault state condition and the healthy state condition, and based on the degree of change, evaluate the sensitivity of the characteristic quantity to the fault.

[0030] (4) According to the sensitivity of each characteristic quantity to the fault, a target characteristic quantity sensitive to the fault is determined by using a statistical method, and a monitoring signal corresponding to the target characteristic quantity is determined as the specific state parameter.

[0031] Specifically, a simulation model that can simulate the operation of the motor can be constructed based on the physical characteristics and working principle of the motor itself, using relevant modeling software or mathematical methods. The simulation model can simulate the actual operating conditions of the motor. It should be noted that the actual operating conditions of the motor include healthy state conditions (normal working conditions of the motor when there is no fault) and fault state conditions (working conditions when the motor fails). By running the established model, the operating state of the motor under different conditions can be simulated. The motor will show different physical response characteristics under different conditions. For example, when the motor is running under electrical fault conditions, the current, voltage and other characteristics of the motor will change. By simulating different working conditions of the motor, the physical response characteristics of the motor under different conditions can be obtained.

[0032] Furthermore, the physical response characteristics under different working conditions can be compared, and the physical response characteristics that can reflect the motor fault condition can be selected as monitoring signals. By performing time domain, frequency domain and time-frequency domain analysis on the monitoring signals, multiple feature quantities that can characterize the signal characteristics can be extracted from the monitoring signals.

[0033] Specifically, time domain analysis analyzes the changes of monitoring signals on the time axis and calculates the time domain characteristic parameters of the monitoring signals, such as the mean, variance, standard deviation and peak value; frequency domain analysis converts the monitoring signals from the time domain to the frequency domain by using methods such as Fourier transform to analyze the frequency components of the monitoring signals, such as the amplitude and harmonics of the monitoring signal frequency; time-frequency domain analysis obtains the changes of monitoring signals in both time and frequency dimensions through methods such as wavelet transform, which helps to better understand the characteristics of monitoring signals.

[0034] Furthermore, the degree of change of each characteristic quantity under faulty state conditions and healthy state conditions can be calculated, and the sensitivity of the characteristic quantity to faults can be judged according to the degree of change corresponding to each characteristic quantity. For example, the greater the degree of change of the characteristic quantity, the higher the sensitivity of the characteristic quantity to faults. Finally, a statistical method is used to comprehensively analyze the fault sensitivity of each characteristic quantity, and a target characteristic quantity is selected from them. The monitoring signal corresponding to the target characteristic quantity is the specific state parameter.

[0035] Referring to the previous description, after determining the composition of a set of state data, the corresponding acquisition system can be built, and the acquisition system can include various sensors to collect initial state data based on various sensors. It should be noted that in the actual process, according to the specific type and signal form of each state parameter, a suitable high-precision and high-reliability sensor can be selected to collect the state parameter. For example, in one possible implementation, a temperature sensor, a current sensor, a voltage sensor, a vibration sensor, etc. can be used to ensure that the selected sensor can accurately obtain the state data in the motor operation state.

[0036] It should be noted that the acquisition system may also include a signal conditioning circuit for amplifying, filtering and converting the weak signals collected by the sensor so that they can be effectively identified and processed by subsequent equipment.

[0037] In specific implementation, you can also develop a dedicated data acquisition program, set the corresponding data acquisition frequency, data acquisition channel, data storage format and other functions to meet the acquisition needs under various working conditions. Furthermore, you can also install and configure the data transmission module to transmit the collected data to the central processing unit or remote server in real time to ensure the real-time and integrity of the data. In addition, after the acquisition system is built, the system is debugged and calibrated. By comparing the measurement results of the standard test equipment, the acquisition system is calibrated for accuracy and error correction, and finally the acquisition system can be used to collect data.

[0038] In this step, multiple groups of health status data of the motor in a healthy state can be obtained from the acquisition system, and the multiple groups of health status data are used as training sample sets.

[0039] S103. Based on adversarial training technology, the initial model is trained using the training sample set to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data, so as to perform fault diagnosis based on the coding loss and the classification result.

[0040] Specifically, based on the adversarial training technology, the training sample set is used to train the initial model, and the trained model is the motor fault diagnosis model.

[0041] In this application, the core of using adversarial training technology to train the initial model is to improve the diagnostic ability of the model through the game process between the generator and the discriminator. Especially in the absence of fault samples (zero fault sample conditions), adversarial training technology can effectively model the health status data and generate similar samples for learning potential abnormal patterns.

[0042] Specifically, the generator is responsible for extracting features from the input health status data and generating reconstructed data that is close to the original health data through data reconstruction. In this way, the generator can learn to capture the characteristic patterns of the motor health status data and reconstruct similar health status data. In addition, as the other party in adversarial training, the discriminator's main task is to distinguish between the real health status data and the health data reconstructed by the generator. The discriminator outputs the classification result based on the authenticity of the input data, judging whether the data is real (real health data) or reconstructed (health data generated by the generator).

[0043] In specific implementation, the core idea of ​​adversarial training is that the generator and the discriminator compete with each other and continuously improve each other's capabilities. The specific process is as follows: The goal of the discriminator is to accurately distinguish whether the input data is real healthy data or reconstructed data generated by the generator. By minimizing the classification loss of the discriminator (i.e., the accuracy of judging whether it is real data), the discriminator becomes more and more adept at identifying subtle differences between reconstructed data and real data.

[0044] The goal of the generator is to make the discriminator unable to distinguish the generated health data from the real health data. The task of the generator is to "fool" the discriminator into thinking that the generated health data is as real as the original health data.

[0045] Through this game, the generator gradually improves its ability to generate real data, and the discriminator gradually improves its accuracy in identifying and reconstructing data. Eventually, the generator can generate data that is very close to the real healthy state. In this way, even if the model has never seen a fault sample, it can detect anomalies or potential faults by learning from healthy samples.

[0046] It should be noted that in a possible implementation, after the initial model is trained using the training sample set to obtain the motor fault diagnosis model, the motor fault diagnosis model can also be tested using the test set to adjust and optimize the motor fault diagnosis model.

[0047] In specific implementation, an experimental method or a simulation method can be used to obtain the fault state data of the motor under various typical fault states, and the fault state data and the healthy state data obtained in step S102 are used as test set data, and then the test set data is input into the motor fault diagnosis model, and the motor fault diagnosis model outputs the classification loss and the coding loss, and then the fault diagnosis is performed based on the output classification loss and the coding loss, and the input sample is judged to be a fault signal or a healthy signal.

[0048] Further, referring to the previous description, the input of the motor fault diagnosis model is the real-time status data of the motor, and the output is the coding loss and classification result corresponding to the input status data. In this way, the coding loss and classification result can be combined for fault diagnosis.

[0049] Optionally, in a possible implementation, when fault diagnosis is performed jointly based on coding loss and classification results, the motor fault diagnosis model is based on the overfitting characteristic. When the input state data is fault state data that has not been learned, the output coding loss is much larger than the baseline coding loss of the motor fault diagnosis model; the baseline coding loss is the average of multiple coding losses when the trained motor fault diagnosis model processes a preset number of groups of healthy status data.

[0050] The following is a brief description of the specific principles of fault diagnosis: Specifically, referring to the previous description, it can be seen that the motor fault diagnosis model only uses the motor's healthy state data in the training stage. The generator generates the reconstructed data by learning the features of the healthy state data, and the discriminator is trained by distinguishing between the real data and the reconstructed data. Since the model has fully learned the healthy state data, it can well reconstruct the healthy data and extract the corresponding feature vector.

[0051] In this process, the generator can accurately capture the characteristics of the health status data and generate reconstructed data that is very close to the real health data. After multiple rounds of adversarial training, the encoding loss of the health status data (i.e., the feature difference between the input data and the reconstructed data) will be relatively small.

[0052] Furthermore, after the model is trained, by inputting multiple sets of health status data into the motor fault diagnosis model, the motor fault diagnosis model will output multiple coding losses. Furthermore, the average value of these multiple coding losses is used as the benchmark coding loss, which represents the "normal" reconstruction error of the model when processing health status data.

[0053] Furthermore, when a motor fails, the state data deviates from the patterns learned by the model in a healthy state. Faulty data may have significantly different characteristics from healthy data, especially when the model has not seen these faulty data (i.e., it has not been trained using faulty data).

[0054] Since the model is trained well on healthy data but does not adapt well enough to unseen faulty data, it performs poorly when reconstructing faulty data.

[0055] When fault state data is input into the model, the generator attempts to reconstruct the data according to the features it has learned from the healthy data. However, since the features of the fault state data are different from those of the healthy data, the reconstruction effect of the generator will be poor, resulting in an increase in the difference between the reconstructed data and the original fault data, which in turn causes a large coding loss. This coding loss is significantly greater than the baseline coding loss of the healthy data. The substantial increase in coding loss is a key signal for the model to detect potential faults.

[0056] In this way, if the coding loss and classification results are obviously abnormal (very different from the characteristics of the healthy state data), it may mean that the input state data is abnormal or faulty. The model can detect potential faults without fault samples by jointly analyzing the coding loss and classification results.

[0057] In summary, it is understandable that when the input state data is fault data, the generator cannot effectively reconstruct it, resulting in a coding loss that is much greater than the baseline coding loss. This significantly exceeded coding loss indicates that there is a significant difference between the input data and the healthy data features learned by the model, indicating a potential fault.

[0058] Furthermore, the discriminator determines whether it is in a healthy state by classifying the input data. If the classification result indicates that the input data is abnormal, combined with a significant increase in coding loss, the model can be more certain that this is a fault state. The model provides more reliable fault diagnosis by jointly analyzing coding loss and classification results. The larger the coding loss and the more abnormal the classification result, the higher the confidence of the fault diagnosis.

[0059] That is, when the model is trained in a healthy state, the reconstruction error of healthy data is small (baseline coding loss), but when faced with unseen fault data, the generator's reconstruction ability decreases due to the overfitting characteristics, resulting in a significant increase in coding loss. By comparing the coding loss of the input data with the baseline coding loss and combining it with the classification results of the discriminator, the model can effectively detect potential faults in the motor.

[0060] Optionally, in a possible implementation, when the real-time classification result is false and the real-time encoding loss is greater than a preset threshold, the real-time status data is determined to be abnormal data and the motor is faulty; wherein the preset threshold is the sum of the baseline encoding loss and a preset margin of the motor fault model.

[0061] Referring to the previous description, it can be understood that the method provided in this embodiment improves the network structure of the generative adversarial neural network, adds an encoder to the generator, extracts and encodes features of the generated data, and obtains the distribution difference between the feature encoding of the generated data and the feature encoding of the original input data, and introduces the feature distribution difference into the loss function of the generative network, which can effectively enhance the feature extraction and encoding capabilities of the generative network; at the same time, during the neural network training process, the generative network can effectively learn the feature distribution of healthy data, and the difference between the feature encoding of healthy data and that of generated data is small. When the output is abnormal data, the encoding loss value will increase sharply. Based on the size of the encoding loss value, it can be determined whether the input data is abnormal data, thereby achieving effective monitoring and diagnosis of abnormal motor conditions.

[0062] The method for constructing a motor fault diagnosis model provided in this embodiment constructs an initial model and obtains multiple groups of health status data of the motor in a healthy state, and then uses these multiple groups of health status data as training sample sets. Based on the adversarial training technology, the initial model is trained using the training sample sets to obtain a motor fault diagnosis model. In this way, on the first hand, through the adversarial training mechanism of the generator and the discriminator, the model can identify potential abnormal states by learning from the health status data, and a method for constructing a motor fault diagnosis model under zero fault sample conditions is provided, which can only rely on the rich and easily obtained health status data in actual industrial scenarios as training samples, and can train a model with the ability to recognize abnormal motor states. The motor fault diagnosis model realizes the fault monitoring task. In addition, the motor fault diagnosis model only relies on healthy status data for training, which overcomes the requirement of intelligent fault diagnosis algorithm based on deep learning technology for the completeness of prior data set, greatly reduces the development cost of the algorithm, and improves its applicability and practicality; secondly, the motor fault diagnosis model only relies on healthy status data for training, and the generator optimizes multiple loss functions to make the modeling of healthy status data more accurate, so that potential faults can be accurately detected and diagnosed; thirdly, the model outputs the coding loss and classification results corresponding to the status data, and by combining the coding loss and classification results, it can better identify faults.

[0063] Corresponding to the aforementioned embodiment of a method for constructing a motor fault diagnosis model, the present application also provides an embodiment of a motor fault diagnosis method, wherein the motor fault diagnosis method is implemented based on the motor fault diagnosis model, and the motor fault diagnosis model is constructed based on the method for constructing a motor fault diagnosis model described in the first aspect of the present application.

[0064] Figure 3 This is a flow chart of the first embodiment of the motor fault diagnosis method provided by this application. Please refer to Figure 3 , the method provided in this embodiment includes: S301, obtaining real-time status data of the motor.

[0065] S302 , inputting the real-time status data into the motor fault diagnosis model, so that the motor fault diagnosis model outputs a real-time encoding loss and a real-time classification result corresponding to the real-time status data.

[0066] S303. When the real-time classification result is false and the real-time coding loss is greater than a preset threshold, it is determined that the real-time status data is abnormal data and the motor is faulty; wherein the preset threshold is the sum of a baseline coding loss and a preset margin of the motor fault model; and the baseline coding loss is an average value of a preset number of coding losses obtained when the motor fault model processes a preset number of groups of health status data.

[0067] In specific implementation, the real-time status data of the motor can be obtained from the sensor in the motor. Furthermore, after the real-time status data is input into the motor fault diagnosis model, the motor fault diagnosis model outputs the real-time encoding loss and real-time classification result corresponding to the real-time status data.

[0068] Furthermore, fault diagnosis can be performed based on real-time coding loss and real-time classification structure. When the real-time classification result is false and the real-time coding loss is greater than a preset threshold, it is determined that the real-time status data is abnormal data and the motor is faulty.

[0069] It should be noted that, referring to the previous introduction, after the motor fault diagnosis model is trained, the baseline coding loss of the motor fault diagnosis model can be determined, and an appropriate preset margin can be set, and the sum of the baseline coding loss and the preset margin can be determined as the preset threshold for fault diagnosis.

[0070] In this way, during fault diagnosis, the real-time coding loss and classification results can be combined for joint fault diagnosis. When the real-time classification result of the motor is false (the output is 0) and the real-time coding loss value is greater than the preset threshold, the monitoring data is judged to be abnormal data and the motor is currently in an abnormal working state.

[0071] It should be noted that the preset margin is set according to actual needs and is not limited in this embodiment.

[0072] Optionally, referring to the above description, it can be understood that, in a possible implementation, the process of determining the reference coding loss may include: (1) for each set of health status data in a preset number of sets of health status data, inputting the set of health status data into the motor fault diagnosis model, so that the motor fault diagnosis model outputs a coding loss corresponding to the set of health status data; (2) Calculating an average value of a preset number of encoding losses corresponding to the preset number of groups of health status data, and determining the average value as a reference encoding loss of the motor fault diagnosis model.

[0073] Specifically, the number of groups of health status data of the preset number of groups is set according to actual needs and is not limited in this embodiment.

[0074] The method provided in this embodiment is based on real-time status data and can perform fault diagnosis in real time through a pre-trained motor fault diagnosis model, which can quickly discover potential motor faults and reduce reaction time when faults occur; in addition, fault diagnosis is performed by combining coding loss and classification structure to improve accuracy and reduce misjudgment; further, the preset threshold can be adjusted according to the health status data to adapt to the operating characteristics of different motors and enhance flexibility.

[0075] Corresponding to the above-mentioned embodiment, the present application also provides a device for constructing a motor fault diagnosis model. Figure 4 This is a schematic diagram of the structure of the first embodiment of the device for constructing a motor fault diagnosis model provided by the present application. Figure 4 The device provided in this embodiment includes a construction module 410, an acquisition module 420 and a training module 430; wherein: The construction module 410 is used to construct an initial model; wherein the initial model includes a generator and a discriminator; the generator includes a first encoder, a decoder, and a second encoder; the first encoder is used to extract features from the input original state data to obtain a first feature vector, the decoder is used to reconstruct data based on the first feature vector to obtain reconstructed state data, and the second encoder is used to extract features from the reconstructed state data to obtain a second feature vector; the discriminator is used to output a classification result based on the original state data and the reconstructed state data; wherein the discriminator outputs a classification loss based on the authenticity of the input data to distinguish between real data and reconstructed data; the generator uses a weighted loss of data reconstruction loss, coding loss, and the classification loss as a loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the coding loss is characterized by the difference between the second feature vector and the first feature vector; The acquisition module 420 is used to acquire multiple sets of health status data of the motor in a healthy state, and use the multiple sets of health status data as training sample sets; The training module 430 is used to train the initial model using the training sample set based on adversarial training technology to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data, so as to perform fault diagnosis based on the coding loss and the classification result.

[0076] The device of this embodiment can be used to perform Figure 1 The steps, specific implementation principles and implementation processes of the method embodiment shown are similar and will not be repeated here.

[0077] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0078] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present application scheme. A person of ordinary skill in the art can understand and implement it without paying any creative work.

[0079] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A method for constructing a motor fault diagnosis model, It is characterized in that The method is used to construct a motor fault diagnosis model under a zero-fault sample condition, and the method comprises: Constructing an initial model; wherein the initial model includes a generator and a discriminator; the generator includes a first encoder, a decoder and a second encoder; the first encoder is used to extract features from the input original state data to obtain a first feature vector, the decoder is used to reconstruct data based on the first feature vector to obtain reconstructed state data, and the second encoder is used to extract features from the reconstructed state data to obtain a second feature vector; the discriminator is used to output a classification result based on the original state data and the reconstructed state data; wherein the discriminator outputs a classification loss based on the authenticity of the input data to distinguish between real data and reconstructed data; the generator uses a weighted loss of data reconstruction loss, coding loss and the classification loss as a loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the coding loss is characterized by the difference between the second feature vector and the first feature vector; Acquire multiple sets of health status data of the motor in a healthy state, and use the multiple sets of health status data as training sample sets; Based on the adversarial training technology, the initial model is trained using the training sample set to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data, so as to jointly perform fault diagnosis based on the coding loss and the classification result.

2. The method according to claim 1, It is characterized in that When fault diagnosis is performed jointly based on coding loss and classification results, the motor fault diagnosis model is based on the overfitting characteristic. When the input state data is fault state data that has not been learned, the output coding loss is greater than the baseline coding loss of the motor fault diagnosis model; the baseline coding loss is the average of multiple coding losses when the motor fault diagnosis model processes a preset number of groups of healthy status data.

3. The method according to claim 1, It is characterized in that The first encoder, the decoder and the second encoder adopt a fully connected neural network as a basic network structure.

4. The method according to claim 1 or 3, It is characterized in that When the state data is in the form of picture data, the first encoder and the second encoder adopt convolutional neural networks; and the decoder adopts a deconvolutional neural network.

5. The method according to claim 1, It is characterized in that The set of state data of the motor includes specific state parameters among all state parameters of the motor, and the specific state parameters are state parameters that are strongly correlated with the health state of the motor.

6. The method according to claim 5, It is characterized in that The method for determining the specific state parameter comprises: According to the physical characteristics and working principle of the motor, a simulation model of the motor is established, and based on the simulation model, the physical response characteristics of the motor under various working conditions are obtained; wherein the various working conditions include healthy state conditions and fault state conditions, and the fault state conditions include various mechanical fault conditions and electrical fault conditions; Screening a monitoring signal reflecting a fault from the physical response characteristics, and performing time domain, frequency domain and time-frequency domain analysis on the monitoring signal to extract a plurality of characteristic quantities of the monitoring signal; For each characteristic quantity, calculating the degree of change of the characteristic quantity under the fault state condition and the healthy state condition, and based on the degree of change, evaluating the sensitivity of the characteristic quantity to the fault; According to the sensitivity of each characteristic quantity to the fault, a target characteristic quantity sensitive to the fault is determined by using a statistical method, and a monitoring signal corresponding to the target characteristic quantity is determined as the specific state parameter.

7. The method according to claim 1, It is characterized in that The difference between the reconstructed state data and the original state data and / or the difference between the second eigenvector and the first eigenvector is a Manhattan distance or a Euclidean distance.

8. A motor fault diagnosis method, It is characterized in that The motor fault diagnosis method is implemented based on a motor fault diagnosis model, and the motor fault diagnosis model is constructed based on the method for constructing a motor fault diagnosis model according to any one of claims 1 to 7; the method comprises: Get the real-time status data of the motor; Inputting the real-time status data into the motor fault diagnosis model, so that the motor fault diagnosis model outputs the real-time encoding loss and the real-time classification result corresponding to the real-time status data; When the real-time classification result is false and the real-time coding loss is greater than a preset threshold, the real-time status data is determined to be abnormal data and the motor is faulty; wherein the preset threshold is the sum of a baseline coding loss and a preset margin of the motor fault model; and the baseline coding loss is the average value of a preset number of coding losses obtained when the motor fault model processes a preset number of groups of health status data.

9. The method according to claim 8, It is characterized in that The process of determining the reference coding loss includes: For each set of health status data in the preset number of sets of health status data, the set of health status data is input into the motor fault diagnosis model, so that the motor fault diagnosis model outputs the encoding loss corresponding to the set of health status data; An average value of a preset number of encoding losses corresponding to the preset number of groups of health status data is calculated, and the average value is determined as a reference encoding loss of the motor fault diagnosis model.

10. A device for constructing a motor fault diagnosis model, It is characterized in that The device comprises a construction module, an acquisition module and a training module; wherein: The construction module is used to construct an initial model; wherein the initial model includes a generator and a discriminator; the generator includes a first encoder, a decoder and a second encoder; the first encoder is used to extract features from the input original state data to obtain a first feature vector, the decoder is used to reconstruct data based on the first feature vector to obtain reconstructed state data, and the second encoder is used to extract features from the reconstructed state data to obtain a second feature vector; the discriminator is used to output a classification result according to the original state data and the reconstructed state data; wherein the discriminator outputs a classification loss according to the authenticity of the input data to distinguish between real data and reconstructed data; the generator uses a weighted loss of data reconstruction loss, coding loss and the classification loss as a loss; the data reconstruction loss is characterized by the difference between the reconstructed state data and the original state data; the coding loss is characterized by the difference between the second feature vector and the first feature vector; The acquisition module is used to acquire multiple groups of health status data of the motor in a healthy state, and use the multiple groups of health status data as training sample sets; The training module is used to train the initial model using the training sample set based on adversarial training technology to obtain a motor fault diagnosis model; the motor fault diagnosis model is used to output the coding loss corresponding to the state data and the classification result corresponding to the state data based on the input state data, so as to perform fault diagnosis based on the coding loss and the classification result.

Images