Knowledge distillation-based passive domain data domain adaptive fault diagnosis method and system

Abstract

The invention discloses a knowledge distillation-based passive domain data domain adaptive fault diagnosis method and system, and the method comprises the following steps: collecting fault acceleration data of a generator bearing of a wind turbine generator in different working conditions, dividing the data into source domain data and target domain data, and randomly dividing the data into training samples and test samples; establishing a source domain model, training the model by using source domain data, and storing trained model parameters; establishing a target domain model of knowledge distillation by taking the source domain model as a reference, including a teacher model and a student model, and initializing the target domain model by using the stored source domain model parameters; and adopting the target domain data training sample to adapt to the student model, and inputting the test sample into the trained student model to realize diagnosis of the generator bearing fault. According to the method, the dependence on the source domain tagged data in cross-domain fault diagnosis is reduced, the problem of cross-domain fault diagnosis when only part of source domain fault data can be obtained or the source domain fault data cannot be obtained is solved, and meanwhile, the data privacy is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of mechanical fault diagnosis, and in particular to a passive domain data domain adaptive fault diagnosis method and system based on knowledge distillation. Background Art

[0002] With the improvement of precision and complexity of rotating machinery and equipment, the reliability requirements for the operation of rotating machinery and equipment have become more stringent. As a large-scale rotating machinery and equipment, wind turbines are under complex conditions of variable working conditions for a long time. Any minor fault may cause the failure of parts or the whole machine, thereby affecting the normal operation of the wind turbine and may even cause devastating damage and catastrophic accidents. Intelligent fault diagnosis can effectively identify and evaluate the health status of rotating machinery and equipment, which is crucial to improving the safety and reliability of rotating machinery and equipment. In order to ensure the normal operation of wind turbines, the research on health monitoring and fault diagnosis technology of wind turbines is becoming more and more important. Common unsupervised transfer learning fault diagnosis methods include mapping-based unsupervised transfer learning (MK-MMD), adversarial-based unsupervised transfer learning (DANN), and instance-based unsupervised transfer learning (AdaBN). Although these methods have achieved certain results in fault diagnosis, the above methods not only require a large amount of labeled source domain data and target domain data to close the distance at the distribution level to achieve cross-domain fault diagnosis, but also require large computing resources to train the model with the data of the source domain and the target domain, which limits its application in actual engineering practice. In order to reduce the dependence on labeled source domain data, the present invention discloses a domain-adaptive fault diagnosis method for passive domain data based on knowledge distillation. This method can enable the target domain model to handle cross-domain fault diagnosis tasks without using source domain data for training, thereby achieving effective identification of wind turbine faults under different working conditions and having good generalization performance. Summary of the invention

[0003] In response to the problems and needs raised above, this solution proposes a passive domain data domain adaptive fault diagnosis method and system based on knowledge distillation. Due to the adoption of the following technical features, it can achieve the above technical objectives and bring about many other technical effects.

[0004] An object of the present invention is to propose a passive domain data domain adaptive fault diagnosis method based on knowledge distillation, comprising:

[0005] S10: installing an acceleration sensor on the rotating mechanical equipment, collecting vibration signal data corresponding to different faults of the rotating mechanical equipment under different working conditions, dividing the vibration signal data into source domain data and target domain data, and then preprocessing the source domain data and the target domain data respectively and randomly dividing them into training samples and test samples respectively;

[0006] S20: Establish a source domain network model in the source domain according to the fault data, initialize and train the source domain network model, use the training samples of the source domain data to train the source domain network model, and input the source domain test samples into the trained source domain network model for testing, so as to realize the classification of the source domain data by the source domain network model and save the trained source domain network model parameters;

[0007] S30: Establish a target domain knowledge distillation model in the target domain. The target domain knowledge distillation model is constructed based on the source domain network model and combined with the knowledge distillation technology. The target domain knowledge distillation model is divided into a teacher model and a student model. The target domain knowledge distillation model is initialized using the source domain network model parameters saved in the above steps.

[0008] S40: adapt the target domain knowledge distillation model in the target domain, initialize the training parameters of the target domain knowledge distillation model, fix the classifier parameters of the target domain knowledge distillation model, train the target domain knowledge distillation model using the target domain training samples, input the test samples into the trained student model, and realize fault diagnosis of the target domain data.

[0009] In addition, the passive domain data domain adaptive fault diagnosis method based on knowledge distillation according to the present invention may also have the following technical features:

[0010] In an example of the present invention, in step S20, the source domain network model includes: a feature extraction module and a classification module, wherein the feature extraction module includes a residual network module, a batch processing layer and a fully connected layer connected in series in sequence;

[0011] The residual network module includes a convolution module, two residual modules, two residual modules with downsampling, an average pooling layer and a batch processing layer connected in series, wherein the convolution module is configured to learn signal feature information through the convolution layer, and suppress mean deviation through the maximum pooling layer to better retain the signal feature information; the residual module is configured to learn signal residual feature information, and prevent gradient vanishing and gradient explosion problems as the network depth increases and prevent network performance degradation through an identity mapping method; the residual module with downsampling is configured to reduce the dimension of feature information by adding a downsampling module to reduce the amount of model calculation; the average pooling layer is configured to reduce the variance of the estimated value to retain the features of the input information to the greatest extent, reduce information loss, and at the same time reduce the feature dimension of the residual network to speed up the training speed; the batch processing layer is configured to use the mean and variance of the feature information to adjust the data distribution to conform to the Gaussian distribution, so that the data distribution is more stable and the convergence speed of the network is accelerated.

[0012] The classification module includes a first fully connected layer, which is configured to map the extracted feature information into a sample label space for fault classification.

[0013] In one example of the present invention, the residual module includes two first convolutional layers, two first processing layers and two first activation function layers, wherein the input features are connected to the two first convolutional layers, the two first processing layers and one first activation function layer by a shortcut connection, and are configured to perform superposition processing on the input features after being processed by the two first convolutional layers, the two first processing layers and one first activation function layer; another first activation function layer is connected to the above-mentioned superposition processing results, and is configured to process the above-mentioned superposition results to obtain the final output result.

[0014] In an example of the present invention, the residual module with downsampling includes two second convolutional layers, two second batch processing layers, two second activation function layers and a downsampling module, wherein the input features are connected to the two second convolutional layers, two second batch processing layers, one second activation function layer and a downsampling module by a shortcut connection, and are configured to perform superposition processing on the input features after being processed by the two second convolutional layers, two second batch processing layers, one second activation function layer and a downsampling module, and another second activation function layer is connected to the above-mentioned superposition processing results, and is configured to process the above-mentioned superposition results to obtain the final output result.

[0015] In one example of the present invention, in step S20, when training the source domain network model using the training samples of the source domain data, the cross entropy loss is used as the loss function, and the difference between the predicted label and the true label is used to measure the quality of the model prediction, wherein the cross entropy loss function formula is as follows:

[0016]

[0017] In the formula, is the model prediction label, y n is the true label, and C represents the fault type.

[0018] In an example of the present invention, in step S20, when training the source domain network model using the training samples of the source domain data, it also includes updating the parameters of the source domain network model, wherein the source domain model network parameters are automatically updated by the SGD algorithm, and the specific updating process is as follows:

[0019] t←t+1

[0020]

[0021] θ t ←θ t-1 -rg t

[0022] Where t is the time step, and the initial value is t = 1; is the gradient operation; θ is a random parameter, the initial value θ 0 is the model parameter; f t (θ t-1 ) is the objective function at time t; r is the learning rate.

[0023] In an example of the present invention, in step S40, training the target knowledge distillation model using the target domain training samples specifically includes the following steps:

[0024] The teacher model fixes the network parameters to perform preliminary classification of training samples and generate pseudo labels, while the student model fixes the classifier parameters and uses pseudo labels to perform adaptive learning on training samples.

[0025] In one example of the present invention, in step S40, when the target domain training sample is used to train the target domain knowledge distillation model, the loss function of the student model adaptation process includes a self-label loss function l between the pseudo label generated by the teacher model and the label predicted by the student model. sl and the information maximization loss function l of the student model prediction label im ;

[0026] The loss function formula of the student model during adaptation is:

[0027] l=γl sl +ηl im

[0028] In the formula, γ represents the coefficient of the self-label loss function, and η represents the coefficient of the information maximization loss function;

[0029] The self-label loss function formula is:

[0030]

[0031] In the formula, is the student model prediction label, y n is the teacher model pseudo label, and C represents the fault type.

[0032] The information maximization loss function formula is:

[0033]

[0034] In the formula, is the student model prediction label, σ(·) represents the softmax function, p represents the output mean of the softmax function of the current batch samples, and K represents the dimension of the student model output prediction label.

[0035] In an example of the present invention, in step S40, the target domain knowledge distillation model training process also includes updating the teacher model parameters by the EMA method. The EMA method formula is as follows:

[0036] v t =βv t-1 +(1-β)θ t

[0037] Where t is the number of updates, v t represents the moving average of all parameters updated by the teacher model for the tth time, θ t represents the parameter weight obtained by updating the student model for the tth time, and β represents the weight parameter.

[0038] Another object of the present invention is to propose a system for the passive domain data domain adaptive fault diagnosis method based on knowledge distillation as described above, comprising:

[0039] The signal acquisition and division unit is configured to install an acceleration sensor on the rotating mechanical equipment, collect vibration signal data corresponding to different faults of the rotating mechanical equipment under different working conditions, divide the vibration signal data into source domain data and target domain data, and then pre-process the source domain data and the target domain data respectively and randomly divide them into training samples and test samples respectively;

[0040] A source domain network model building unit is configured to build a source domain network model in the source domain according to the fault data, initialize and train the source domain network model, train the source domain network model using training samples of the source domain data, and input the source domain test samples into the trained source domain network model for testing, so as to realize the classification of the source domain data by the source domain network model and save the trained source domain network model parameters;

[0041] A target domain knowledge distillation model building unit is configured to build a target domain knowledge distillation model in the target domain, and the target domain knowledge distillation model is built based on the source domain network model and combined with the knowledge distillation technology. The target domain knowledge distillation model is divided into a teacher model and a student model, and the target domain knowledge distillation model is initialized using the source domain network model parameters saved in the above steps;

[0042] The target domain fault diagnosis unit is configured to adapt the target domain knowledge distillation model in the target domain, initialize the training parameters of the target domain knowledge distillation model, fix the classifier parameters of the target domain knowledge distillation model, train the target domain knowledge distillation model using the target domain training samples, input the test samples into the trained student model, and realize the fault diagnosis of the target domain data.

[0043] The best embodiment for carrying out the present invention will be described in more detail below with reference to the accompanying drawings so that the features and advantages of the present invention can be easily understood. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings of the embodiments of the present invention, wherein the drawings are only used to illustrate some embodiments of the present invention, but not to limit all embodiments of the present invention thereto.

[0045] Figure 1 A flowchart of a passive domain data domain adaptation fault diagnosis method based on knowledge distillation according to an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of the structure of a source domain network model according to an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of the structure of a residual module according to an embodiment of the present invention;

[0048] Figure 4 2 is a schematic diagram of the structure of a residual module with downsampling according to an embodiment of the present invention;

[0049] Figure 5 A flowchart of the process of adapting the knowledge distillation model in the target domain according to an embodiment of the present invention;

[0050] Figure 6 is the diagnostic accuracy of different deep learning models according to the embodiments of the present invention. DETAILED DESCRIPTION

[0051] In order to make the purpose, technical solution and advantages of the technical solution of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described in conjunction with the drawings of specific embodiments of the present invention. The same figure marks in the drawings represent the same parts. It should be noted that the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] Unless otherwise defined, the technical terms or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present invention belongs. The words "first", "second" and similar words used in the patent application specification and claims of the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not necessarily indicate a quantity limitation. Words such as "include" or "comprise" mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Words such as "connect" or "connected" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0053] According to the first aspect of the present invention, a passive domain data domain adaptation fault diagnosis method based on knowledge distillation is provided. Figure 1 As shown, including:

[0054] S10: Install an acceleration sensor on the rotating mechanical equipment, collect vibration signal data corresponding to different faults of the rotating mechanical equipment under different working conditions, divide the vibration signal data into source domain data and target domain data, and then pre-process the source domain data and the target domain data respectively and randomly divide them into training samples and test samples respectively; for example, collect acceleration data corresponding to different faults (inner ring fault, outer ring fault, rolling element fault) of a wind turbine generator under different working conditions;

[0055] S20: Establish a source domain network model in the source domain according to the fault data, initialize and train the source domain network model, use the training samples of the source domain data to train the source domain network model, and input the source domain test samples into the trained source domain network model for testing, so as to realize the classification of the source domain data by the source domain network model and save the trained source domain network model parameters;

[0056] S30: Establish a target domain knowledge distillation model in the target domain. The target domain knowledge distillation model is constructed based on the source domain network model and combined with the knowledge distillation technology. The target domain knowledge distillation model is divided into a teacher model and a student model. The target domain knowledge distillation model is initialized using the source domain network model parameters saved in the above steps.

[0057] S40: adapt the target domain knowledge distillation model in the target domain, initialize the training parameters of the target domain knowledge distillation model, fix the classifier parameters of the target domain knowledge distillation model, train the target domain knowledge distillation model using the target domain training samples, input the test samples into the trained student model, and realize fault diagnosis of the target domain data.

[0058] This fault diagnosis method combines the domain adaptation method of passive domain data with the knowledge distillation model to cope with cross-domain fault diagnosis tasks where source domain data is unavailable. Compared with the traditional unsupervised transfer learning fault diagnosis method, the domain adaptation method of passive domain data reduces the dependence on source domain data and solves the problem that source domain fault data is difficult to obtain in engineering practice. At the same time, this fault diagnosis method improves the adaptability of the model by constructing a network model based on knowledge distillation in the target domain, thereby improving the accuracy of fault diagnosis. Combining the domain adaptation method of passive domain data with the knowledge distillation model ensures data privacy and security, improves the ability to diagnose cross-domain faults, and facilitates engineering applications.

[0059] In an example of the present invention, the data preprocessing and random division into training samples and test samples in step S10 refers to using a sliding window function to cut the collected fault data into multiple samples x(L) of a window length, where L is the sliding window length. After normalization, the sample x(L) is randomly divided into training samples x(L) according to 8:2. 1 (L) and test sample x 2 (L).

[0060] In one example of the present invention, Figure 2 As shown, in the step S20, the source domain network model includes: a feature extraction module and a classification module, wherein the feature extraction module includes a residual network module, a batch processing layer and a fully connected layer connected in series in sequence;

[0061] The residual network module includes a convolution module, two residual modules, two residual modules with downsampling, an average pooling layer and a batch processing layer connected in series, wherein the convolution module is configured to learn signal feature information through the convolution layer, and suppress mean deviation through the maximum pooling layer to better retain the signal feature information; the residual module is configured to learn signal residual feature information, and prevent gradient vanishing and gradient explosion problems as the network depth increases and prevent network performance degradation through an identity mapping method; the residual module with downsampling is configured to reduce the dimension of feature information by adding a downsampling module to reduce the amount of model calculation; the average pooling layer is configured to reduce the variance of the estimated value to retain the features of the input information to the greatest extent, reduce information loss, and at the same time reduce the feature dimension of the residual network to speed up the training speed; the batch processing layer is configured to use the mean and variance of the feature information to adjust the data distribution to conform to the Gaussian distribution, so that the data distribution is more stable and the convergence speed of the network is accelerated.

[0062] The classification module includes a first fully connected layer, which is configured to map the extracted feature information into a sample label space for fault classification.

[0063] In one example of the present invention, Figure 3 As shown, the residual module includes two first convolutional layers, two first batch processing layers and two first activation function layers, wherein the input feature is connected to the two first convolutional layers, the two first batch processing layers and one first activation function layer by a shortcut connection, and is configured to perform superposition processing on the input feature after the input feature is processed by the two first convolutional layers, the two first batch processing layers and one first activation function layer; another first activation function layer is connected to the above-mentioned superposition processing result, and is configured to process the above-mentioned superposition result to obtain the final output result.

[0064] In one example of the present invention, Figure 4 As shown, the residual module with downsampling includes two second convolutional layers, two second batch processing layers, two second activation function layers and a downsampling module, wherein the input features are connected to the two second convolutional layers, the two second batch processing layers, the second activation function layer and the downsampling module by a shortcut connection, and are configured to perform superposition processing on the input features after being processed by the two second convolutional layers, the two second batch processing layers, the second activation function layer and the downsampling module, and another second activation function layer is connected to the above-mentioned superposition processing result, and is configured to process the above-mentioned superposition result to obtain the final output result.

[0065] In an example of the present invention, in step S20, the model parameters are initialized including the number of initialization iterations E=30, the initialization learning rate r=0.01, and the initialization batch size s=64.

[0066] In one example of the present invention, in step S20, when training the source domain network model using the training samples of the source domain data, the cross entropy loss is used as the loss function, and the difference between the predicted label and the true label is used to measure the quality of the model prediction, wherein the cross entropy loss function formula is as follows:

[0067]

[0068] In the formula, is the model prediction label, y n is the true label, and C represents the fault type.

[0069] In an example of the present invention, in step S20, when training the source domain network model using the training samples of the source domain data, it also includes updating the parameters of the source domain network model, wherein the source domain model network parameters are automatically updated by the SGD algorithm, and the specific updating process is as follows:

[0070] t←t+1

[0071]

[0072] θ t ←θ t-1 -rg t

[0073] Where t is the time step, and the initial value is t = 1; is the gradient operation; θ is a random parameter, the initial value θ 0 is the model parameter; f t (θ t-1 ) is the objective function at time t; r is the learning rate.

[0074] In one example of the present invention, in step S40, the fixed knowledge distillation model classifier parameters refer to the fixed teacher model and student model classifier parameters, and the classifier parameters of the two models are the trained source domain model classifier parameters. The purpose of fixing the classifier parameters is to preserve the distribution characteristics of the source domain data, and use the knowledge learned in the source domain to classify the target domain unlabeled samples, and it is no longer necessary to close the distance between the source domain data and the target domain data at the distribution level.

[0075] In one example of the present invention, Figure 5 As shown, in step S40, using the target domain training samples to train the target knowledge distillation model specifically includes the following steps:

[0076] The teacher model fixes the network parameters to perform preliminary classification of the training samples and generate pseudo labels, while the student model fixes the classifier parameters and uses pseudo labels to adapt the training samples. Finally, the test samples are input into the trained student model to achieve cross-domain fault diagnosis.

[0077] The use of target domain training samples to train the knowledge distillation model refers to the teacher model fixing the network parameters to perform preliminary classification of the training samples to generate pseudo labels, and the student model fixing the classifier parameters and using pseudo labels to perform adaptive learning on the training samples to improve the diagnostic ability. Finally, the test samples are input into the trained student model to achieve cross-domain fault diagnosis.

[0078] In one example of the present invention, Figure 5 As shown, in step S40, when the target domain training sample is used to train the target domain knowledge distillation model, the loss function of the student model adaptation process includes the self-label loss function l between the pseudo labels generated by the teacher model and the labels predicted by the student model. sl and the information maximization loss function l of the student model prediction label im ;

[0079] The loss function formula of the student model during adaptation is:

[0080] l=γl sl +ηl im

[0081] In the formula, γ represents the coefficient of the self-label loss function, and η represents the coefficient of the information maximization loss function;

[0082] The self-label loss function formula is:

[0083]

[0084] In the formula, is the student model prediction label, y n is the teacher model pseudo label, and C represents the fault type.

[0085] The information maximization loss function formula is:

[0086]

[0087] In the formula, is the student model prediction label, σ(·) represents the softmax function, p represents the output mean of the softmax function of the current batch samples, and K represents the dimension of the student model output prediction label.

[0088] In an example of the present invention, in step S40, the target domain knowledge distillation model training process also includes updating the teacher model parameters by the EMA method. The EMA method formula is as follows:

[0089] v t =βv t-1 +(1-β)θ t

[0090] Where t is the number of updates, v t represents the moving average of all parameters updated by the teacher model for the tth time, θ t represents the parameter weight obtained by updating the student model for the tth time, and β represents the weight parameter.

[0091] According to the second aspect of the present invention, a system for the passive domain data domain adaptive fault diagnosis method based on knowledge distillation as described above comprises:

[0092] The signal acquisition and division unit is configured to install an acceleration sensor on the rotating mechanical equipment, collect vibration signal data corresponding to different faults of the rotating mechanical equipment under different working conditions, divide the vibration signal data into source domain data and target domain data, and then pre-process the source domain data and the target domain data respectively and randomly divide them into training samples and test samples respectively;

[0093] A source domain network model building unit is configured to build a source domain network model in the source domain according to the fault data, initialize and train the source domain network model, train the source domain network model using training samples of the source domain data, and input the source domain test samples into the trained source domain network model for testing, so as to realize the classification of the source domain data by the source domain network model and save the trained source domain network model parameters;

[0094] A target domain knowledge distillation model building unit is configured to build a target domain knowledge distillation model in the target domain, and the target domain knowledge distillation model is built based on the source domain network model and combined with the knowledge distillation technology. The target domain knowledge distillation model is divided into a teacher model and a student model, and the target domain knowledge distillation model is initialized using the source domain network model parameters saved in the above steps;

[0095] The target domain fault diagnosis unit is configured to adapt the target domain knowledge distillation model in the target domain, initialize the training parameters of the target domain knowledge distillation model, fix the classifier parameters of the target domain knowledge distillation model, train the target domain knowledge distillation model using the target domain training samples, input the test samples into the trained student model, and realize the fault diagnosis of the target domain data.

[0096] The fault diagnosis system combines the domain adaptation method of passive domain data with the knowledge distillation model to cope with cross-domain fault diagnosis tasks where source domain data is unavailable. Compared with the traditional unsupervised transfer learning fault diagnosis method, the domain adaptation method of passive domain data reduces the dependence on source domain data and solves the problem of difficulty in obtaining source domain fault data in engineering practice. At the same time, the fault diagnosis system improves the adaptability of the model by building a network model based on knowledge distillation in the target domain and using knowledge distillation technology to improve the accuracy of fault diagnosis. Combining the domain adaptation method of passive domain data with the knowledge distillation model ensures data privacy and security, improves the ability to diagnose cross-domain faults, and facilitates engineering applications.

[0097] Specific examples

[0098] This case takes the doubly-fed wind power generation fault simulation test bench as the research object. The test bench is mainly composed of a drive motor, a parallel shaft gearbox, a parallel wheel gearbox and a doubly-fed generator. During the operation of the test bench, the piezoelectric accelerometers deployed on the right side of the generator shaft, the left side of the generator shaft and the generator base were used to collect vibration acceleration data of three operating states (inner ring fault, outer ring fault, rolling element fault) of the doubly-fed generator at different constant speeds (1000rpm, 1500rpm, 1800rpm), respectively, with a sampling frequency of 20KHz. The sliding window size is set to L=4096, the moving step is s=4096, 150 samples are collected for each type of fault, and the length of each sample is 4096 points. The collected samples are randomly divided into 360 training samples and 90 test samples at a ratio of 8:2. One of the three different speed fault data is selected as the source domain data, and the other is selected as the target domain data. The fault diagnosis task is described in Table 1 below, and training is performed according to the experimental steps.

[0099] Table 1 shows the fault data of three different speeds

[0100]

[0101] In order to prove the effectiveness and superiority of the proposed method in wind turbine generator fault diagnosis, the proposed model and three common unsupervised transfer learning models (MK-MMD, DANN, AdaBN) are trained using the same training samples, and the trained transfer learning models are tested with test data to obtain the fault diagnosis results of wind turbine generators, such as Figure 6 As shown. Figure 6 It can be clearly seen that the proposed model has achieved a diagnostic accuracy of more than 99%. Compared with other transfer learning models, the proposed method can achieve the highest diagnostic accuracy, which shows that the model proposed in this invention has obvious diagnostic advantages over traditional unsupervised transfer learning models, and can realize intelligent identification and diagnosis of the operating status of wind turbine generators under different working conditions, and has good generalization.

[0102] The exemplary implementation scheme of the passive domain data domain adaptive fault diagnosis method and system based on knowledge distillation proposed in the present invention is described in detail above with reference to the preferred embodiments. However, those skilled in the art will appreciate that, without departing from the concept of the present invention, various modifications and variations can be made to the above-mentioned specific embodiments, and various technical features and structures proposed in the present invention can be combined in various ways without exceeding the protection scope of the present invention, which is determined by the attached claims.

Claims

1. A passive domain data domain adaptation fault diagnosis method based on knowledge distillation, characterized in that: include: S10: installing an acceleration sensor on the rotating mechanical equipment, collecting vibration signal data corresponding to different faults of the rotating mechanical equipment under different working conditions, dividing the vibration signal data into source domain data and target domain data, and then preprocessing the source domain data and the target domain data respectively and randomly dividing them into training samples and test samples respectively; S20: Establish a source domain network model in the source domain according to the fault data, initialize and train the source domain network model, use the training samples of the source domain data to train the source domain network model, and input the source domain test samples into the trained source domain network model for testing, so as to realize the classification of the source domain data by the source domain network model and save the trained source domain network model parameters; S30: Establish a target domain knowledge distillation model in the target domain. The target domain knowledge distillation model is constructed based on the source domain network model and combined with the knowledge distillation technology. The target domain knowledge distillation model is divided into a teacher model and a student model. The target domain knowledge distillation model is initialized using the source domain network model parameters saved in the above steps. S40: adapt the target domain knowledge distillation model in the target domain, initialize the training parameters of the target domain knowledge distillation model, fix the classifier parameters of the target domain knowledge distillation model, train the target domain knowledge distillation model using the target domain training samples, input the test samples into the trained student model, and realize fault diagnosis of the target domain data.

2. The passive domain data domain adaptation fault diagnosis method based on knowledge distillation according to claim 1 is characterized in that: In the step S20, the source domain network model includes: a feature extraction module and a classification module, wherein the feature extraction module includes a residual network module, a batch processing layer and a fully connected layer connected in series in sequence; The residual network module includes a convolution module, two residual modules, two residual modules with downsampling, an average pooling layer and a batch processing layer connected in series, wherein the convolution module is configured to learn signal feature information through the convolution layer, and suppress mean deviation through the maximum pooling layer to better retain the signal feature information; the residual module is configured to learn signal residual feature information, and prevent gradient vanishing and gradient explosion problems as the network depth increases and prevent network performance degradation through an identity mapping method; the residual module with downsampling is configured to reduce the dimension of feature information by adding a downsampling module to reduce the amount of model calculation; the average pooling layer is configured to reduce the variance of the estimated value to retain the features of the input information to the greatest extent, reduce information loss, and at the same time reduce the feature dimension of the residual network to speed up the training speed; the batch processing layer is configured to use the mean and variance of the feature information to adjust the data distribution to conform to the Gaussian distribution, so that the data distribution is more stable and the convergence speed of the network is accelerated. The classification module includes a first fully connected layer, which is configured to map the extracted feature information into a sample label space for fault classification.

3. The passive domain data domain adaptation fault diagnosis method based on knowledge distillation according to claim 2 is characterized in that: The residual module includes two first convolutional layers, two first batch processing layers and two first activation function layers, wherein the input features are connected to the two first convolutional layers, the two first batch processing layers and one first activation function layer by a shortcut connection method, and are configured to perform superposition processing on the input features after being processed by the two first convolutional layers, the two first batch processing layers and one first activation function layer; another first activation function layer is connected to the above-mentioned superposition processing result, and is configured to process the above-mentioned superposition result to obtain the final output result.

4. The passive domain data domain adaptation fault diagnosis method based on knowledge distillation according to claim 2 is characterized in that: The residual module with downsampling includes two second convolutional layers, two second batch processing layers, two second activation function layers and a downsampling module, wherein the input features are connected to the two second convolutional layers, the two second batch processing layers, the second activation function layer and the downsampling module by a shortcut connection method, and are configured to perform superposition processing on the input features after being processed by the two second convolutional layers, the two second batch processing layers, the second activation function layer and the downsampling module, and another second activation function layer is connected to the above-mentioned superposition processing results, and is configured to process the above-mentioned superposition results to obtain the final output result.

5. The passive domain data domain adaptive fault diagnosis method based on knowledge distillation according to claim 1 is characterized in that: In step S20, when training the source domain network model using the training samples of the source domain data, the cross entropy loss is used as the loss function, and the difference between the predicted label and the true label is used to measure the quality of the model prediction, wherein the cross entropy loss function formula is as follows: In the formula, is the model prediction label, y n is the true label, and C represents the fault type.

6. The passive domain data domain adaptation fault diagnosis method based on knowledge distillation according to claim 1 is characterized in that: In the step S20, when the source domain network model is trained using the training samples of the source domain data, the parameters of the source domain network model are also updated, wherein the network parameters of the source domain model are automatically updated by the SGD algorithm, and the specific updating process is as follows: t←t—1 i t ←θ t-1 -rg t Where t is the time step, and the initial value is t = 1; is the gradient operation; θ is a random parameter, and the initial value θ0 is the model parameter; f t (θ t-1 ) is the objective function at time t; r is the learning rate.

7. The passive domain data domain adaptation fault diagnosis method based on knowledge distillation according to claim 1 is characterized in that: In step S40, training the target knowledge distillation model using the target domain training samples specifically includes the following steps: The teacher model fixes the network parameters to perform preliminary classification of training samples and generate pseudo labels, while the student model fixes the classifier parameters and uses pseudo labels to perform adaptive learning on training samples.

8. The passive domain data domain adaptive fault diagnosis method based on knowledge distillation according to claim 1 is characterized in that: In step S40, when the target domain training sample is used to train the target domain knowledge distillation model, the loss function of the student model adaptation process includes the self-label loss function l between the pseudo labels generated by the teacher model and the labels predicted by the student model. sl and the information maximization loss function l of the student model prediction label im ; The loss function formula of the student model during adaptation is: l=γl sl +ηl im In the formula, γ represents the coefficient of the self-label loss function, and η represents the coefficient of the information maximization loss function; The self-label loss function formula is: In the formula, is the student model prediction label, y n is the teacher model pseudo label, and C represents the fault type. The information maximization loss function formula is: In the formula, is the student model prediction label, σ(·) represents the softmax function, p represents the output mean of the softmax function of the current batch samples, and K represents the dimension of the student model output prediction label.

9. The passive domain data domain adaptive fault diagnosis method based on knowledge distillation according to claim 1 is characterized in that: In the step S40, during the target domain knowledge distillation model training process, the parameters are also updated by the EMA method. The formula of the EMA method is as follows: v t =βv t-1 +(1-β)θ t Where t is the number of updates, v t represents the moving average of all parameters updated by the teacher model for the tth time, θ t represents the parameter weight obtained by updating the student model for the tth time, and β represents the weight parameter.

10. A system for the passive domain data domain adaptive fault diagnosis method based on knowledge distillation as claimed in any one of claims 1 to 9, characterized in that: include: The signal acquisition and division unit is configured to install an acceleration sensor on the rotating mechanical equipment, collect vibration signal data corresponding to different faults of the rotating mechanical equipment under different working conditions, divide the vibration signal data into source domain data and target domain data, and then pre-process the source domain data and the target domain data respectively and randomly divide them into training samples and test samples respectively; A source domain network model building unit is configured to build a source domain network model in the source domain according to the fault data, initialize and train the source domain network model, train the source domain network model using training samples of the source domain data, and input the source domain test samples into the trained source domain network model for testing, so as to realize the classification of the source domain data by the source domain network model and save the trained source domain network model parameters; A target domain knowledge distillation model building unit is configured to build a target domain knowledge distillation model in the target domain, and the target domain knowledge distillation model is built based on the source domain network model and combined with the knowledge distillation technology. The target domain knowledge distillation model is divided into a teacher model and a student model, and the target domain knowledge distillation model is initialized using the source domain network model parameters saved in the above steps; The target domain fault diagnosis unit is configured to adapt the target domain knowledge distillation model in the target domain, initialize the training parameters of the target domain knowledge distillation model, fix the classifier parameters of the target domain knowledge distillation model, train the target domain knowledge distillation model using the target domain training samples, input the test samples into the trained student model, and realize the fault diagnosis of the target domain data.

Images