Intelligent fault diagnosis method for variable operating conditions based on distillation learning and feature fusion

By combining distillation learning and feature fusion in intelligent fault diagnosis, using amplitude phase enhancement and characterization self-challenge learning, the problem of insufficient fault feature extraction under variable operating conditions is solved, and the generalization performance and diagnostic effect of the model are improved.

CN118760985BActive Publication Date: 2024-11-26JIANGSU HE MICROELECTRONICS TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411255717.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-09
Publication Date
2024-11-26
Estimated Expiration
2044-09-09

AI Technical Summary

Technical Problem

Under variable operating conditions, the generalization performance of the existing intelligent fault diagnosis model is poor, making it difficult to learn the characteristics of mechanical faults, especially the pulse fault information in the time domain signal and the side-band fault information in the frequency domain signal are difficult to be effectively utilized.

Method used

Using a method based on distillation learning and feature fusion, the signal frequency domain information is obtained through fast Fourier transform, and amplitude phase enhancement is performed. Combined time domain features and frequency domain features, characterization self-challenge learning is used to extract mixed features, and the distillation learning mechanism is used. Guide students to learn the domain unchanging characteristics online and improve the generalization performance of the model.

Benefits of technology

By introducing time domain features, frequency domain features and enhancement features, the model can more accurately understand vibration signals, alleviate the problems of insufficient fault feature extraction and poor generalization performance of the model, and improve the diagnostic performance and generalization capabilities of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118760985B_ABST
    Figure CN118760985B_ABST
Patent Text Reader

Abstract

The present invention discloses a variable working condition intelligent fault diagnosis method based on distillation learning and feature fusion, including: the first step, performing amplitude and phase enhancement on the vibration signal to obtain an enhanced signal. The second step, inputting the enhanced signal into the distillation learning module to obtain enhanced features. The third step, extracting frequency domain features from the vibration signal after fast Fourier transform. The fourth step, inputting the original vibration signal into the time domain feature extraction module, and extracting domain invariant features through the knowledge distillation method according to the enhanced features. The fifth step, after splicing the enhanced features, time domain features and frequency domain features, the cross-domain generalization ability of the features is enhanced through characterization self-challenge learning, and the features are sent to the classifier for classification. The sixth step, the fault signal is input into the trained network for diagnosis. The present invention extracts and fuses richer features, strengthens the model's learning of domain invariant representation, and significantly enhances the generalization performance compared to the existing model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of rotating machinery fault diagnosis and data enhancement technology, and in particular to a variable operating condition intelligent fault diagnosis method based on distillation learning and feature fusion. Background Art

[0002] Mechanical failure is an inevitable problem in industrial production. If it is not discovered and handled in time, it will have a serious impact on industrial safety and production efficiency. Therefore, mechanical fault diagnosis plays a very important role in ensuring industrial safety and promoting industrial development. With the continuous development of artificial intelligence technology, intelligent fault diagnosis methods combined with deep learning and other related technologies have attracted widespread attention. It can extract fault features from mechanical fault data through analysis and learning, and classify and diagnose faults. Compared with traditional fault diagnosis methods, intelligent fault diagnosis methods have a higher degree of automation and accuracy, and can effectively improve the safety and efficiency of industrial production.

[0003] At present, thanks to the emergence of a large number of innovative technologies in artificial intelligence technology, intelligent fault diagnosis methods can achieve a high diagnostic effect under fault diagnosis conditions without unknown working conditions. Under conditions without unknown working conditions, deep learning algorithms have learned the characteristics and laws of mechanical faults through training data, so they can effectively classify and diagnose new fault data. However, under variable working conditions, due to changes in working conditions and environment, the data of different working conditions vary greatly in distribution, and there is a domain offset problem that causes performance degradation.

[0004] At present, in order to solve the problem of poor generalization performance of models under variable working conditions, some research on variable working condition fault diagnosis methods based on transfer learning has been carried out, and distillation learning is one of the methods. Distillation learning is a model compression technology that aims to transfer the knowledge of a large and complex neural network model (teacher model) to a small and simple neural network model (student model) to reduce the computational and storage overhead of the model while maintaining the model performance. In terms of signal data, the current mainstream methods generally only use a single time domain signal as the input of the diagnostic model. The learning of time domain vibration signals focuses on fault pulses, while the learning of signal frequency domain focuses on fault sidebands. For intelligent diagnosis models, the former is rich in information but difficult to learn, while the latter has obvious sideband fault information, but discards some features of fault pulses in time domain signals.

[0005] Therefore, when faced with variable machine conditions, the existing trained models are difficult to learn the characteristics of mechanical failures. Researching models with better generalization performance is of far-reaching significance and will have a profound impact on future industrial applications. Summary of the invention

[0006] Purpose of the invention: In view of the above problems, based on the experimental scenarios of actual production and further considering the application scenarios of equipment with variable operating conditions, the present invention proposes a variable operating condition intelligent fault diagnosis method based on distillation learning and feature fusion, and constructs a variable operating condition fault diagnosis model based on information fusion and knowledge distillation to alleviate the problem of poor generalization performance of the model under variable operating conditions.

[0007] Technical solution: To achieve the purpose of the present invention, the technical solution adopted by the present invention is:

[0008] The variable working condition intelligent fault diagnosis method based on distillation learning and feature fusion includes the following steps:

[0009] Step 1: Obtain the original vibration signal, mark the fault type, and build a training data set;

[0010] Step 2: Build a fault diagnosis network model and perform the following operations:

[0011] The collected original vibration signal is subjected to fast Fourier transform to obtain the signal frequency domain information, and the frequency domain features are extracted through the feature extraction module;

[0012] The collected original vibration signal is enhanced in amplitude and phase to obtain the signal sample after amplitude and phase enhancement; the sample after amplitude and phase enhancement is subjected to distillation learning to obtain the corresponding amplitude and phase enhancement features;

[0013] The amplitude and phase enhancement features, time domain features, and frequency domain features are spliced ​​to obtain mixed features, and feature extraction is performed through representation self-challenge learning; the extracted features are input into the classifier for fault type identification;

[0014] Step 3: Use the training data set to train the fault diagnosis network model, and input the collected vibration signal into the trained network model for fault diagnosis.

[0015] Furthermore, during the training process, the features and labels of several samples are linearly interpolated, and the amplitude and phase enhancement method is used to synthesize new domain samples;

[0016] The amplitude and phase information in the frequency domain are regarded as two features and processed separately;

[0017] For the amplitude information, linear interpolation is used;

[0018] ,

[0019] in, To enhance the sample, is the input signal data sample, For Samples belonging to different domains; Represents different weights when performing linear interpolation, obeys Beta distribution, and has a value range of , Represents the amplitude in the frequency domain;

[0020] For phase information, first calculate and The phase difference between , so that the enhanced sample The phase of is interpolated based on the input samples;

[0021] ,

[0022] in Represents the phase of the signal data in the frequency domain; then according to the phase difference Calculate the shortest phase difference ,

[0023] ;

[0024] Compute enhanced samples Phase information ,

[0025] ;

[0026] in is a control parameter used to adjust the degree of phase change, and its value range is ;

[0027] According to the obtained amplitude information and phase information, the signal is reconstructed by using inverse Fourier transform to obtain the signal samples after amplitude and phase enhancement;

[0028] .

[0029] Furthermore, the samples after amplitude and phase enhancement are distilled and the domain invariant features are learned through a distillation learning framework. The method is as follows:

[0030] The samples after amplitude and phase enhancement are used as the input of the teacher network, so that the model can learn the corresponding enhanced features; the teacher network training process is expressed as follows:

[0031] ,

[0032] In the formula, Represents all inputs in the sample space The expected value of Indicates that the data distribution The training samples are obtained by sampling from is the input data, is the corresponding label, Represents the expected value of all training samples; is the classification loss, that is, the cross entropy loss function, and are the classifier and feature extractor of the teacher network, respectively. and are samples and labels after amplitude and phase enhancement, and are the learnable parameters of the teacher network feature extractor and classifier, respectively;

[0033] After training the teacher network, we get the feature extractor of the teacher network , and guide the student network to learn the sample feature information after data enhancement through feature knowledge distillation, and learn the domain invariant representation, which is expressed as follows:

[0034] ,

[0035] In the formula, and They are the feature extractors in the student network. and classifier The learnable parameters of is the mean square error loss, is the mean square error loss weight.

[0036] Furthermore, the structure of each feature extractor in the network is consistent, and a fully connected layer is added after the feature extractor of the teacher network to reduce the size of the features extracted by the teacher network to half of the original size, and the mean square error loss is constructed with the first half of the extracted time domain features.

[0037] Furthermore, the amplitude and phase enhancement features, time domain features, and frequency domain features are spliced ​​together to obtain mixed features. The ability of the self-challenge learning enhancement model to extract and utilize different features is characterized. The specific process is as follows:

[0038] Calculating mixed features Gradient of loss :

[0039] ,

[0040] In the formula Represents the loss function, where the loss function is the cross entropy loss function, that is , is the true label, Predict labels for the model;

[0041] Characterize the self-challenge learning, the neurons that need to be suppressed are determined by the gradient size; After sorting, the proportion of neurons that need to be inhibited is , for the gradient Each neuron in is sorted from large to small to obtain the characteristic The corresponding order of the elements in , determine the position of the neuron to be inhibited according to formula (9);

[0042] ,

[0043] in, Indicates The position of the neuron, 0 means it is inhibited, and 1 means it is retained; Indicates the length of the gradient sequence;

[0044] Finally, the position With features Multiply to get new features.

[0045] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0046] The present invention introduces time domain features, frequency domain features and enhanced features, combines the learning of the three features, and simultaneously learns pulse fault information and variable frequency band fault information, providing the model with richer information and helping the model to understand the vibration signal more accurately. Combined with the distillation learning mechanism, the features of the fault signal under variable working conditions are extracted and learned, which alleviates the problem of insufficient fault feature extraction and the problem of poor generalization performance of the model under variable working conditions, enhances the generalization performance of the model, and improves the diagnostic performance of the model. Experiments show that in the method of the present invention, the boundaries between various fault features are relatively clear, and features of the same category are better aggregated, which can effectively improve the diagnostic performance of the diagnostic model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the variable operating condition fault diagnosis network based on information fusion and knowledge distillation of the present invention. DETAILED DESCRIPTION

[0048] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0049] The variable working condition intelligent fault diagnosis method based on distillation learning and feature fusion described in the present invention takes into account the problem that the fault information features provided by a single data are insufficient, and designs a combination of multiple features to be input into the model. First, the first signal is the original signal that remains unchanged without any processing, that is, the original time domain vibration signal. The second signal is to perform a fast Fourier transform on the time domain signal to obtain its frequency domain information. The third signal is to enhance the sample quality by performing amplitude and phase enhancement on the time domain signal to obtain a cross-domain enhanced signal. In addition, the time domain signal is used as the input of the time domain feature extraction module to extract time domain related features; the signal frequency domain information is used as the input of the frequency domain feature extraction module to extract frequency domain related features; the amplitude and phase enhanced signal is input into the distillation learning module, and the amplitude and phase enhanced features are learned through a distillation learning framework. Secondly, in the distillation learning framework, the student network is guided to learn the sample feature information after data enhancement by means of knowledge distillation, so that the features extracted by the student network are close to the features of the teacher network, that is, the extracted time domain features learn the amplitude and phase enhanced features and move closer to them. Finally, the amplitude and phase enhancement features, time domain features, and frequency domain features are concatenated to obtain mixed features. More generalized features are extracted through representation self-challenge learning and sent to the classifier for fault diagnosis.

[0050] Figure 1 It is a flow chart of the method of the present invention. The specific implementation process of the method of the present invention is as follows:

[0051] Step 1: Get the original vibration signal.

[0052] Time domain vibration signal As input, It is a signal of a type of fault under one working condition, containing data points; Indicates different fault category index numbers. is the total number of fault categories, is the index number of different working condition categories, is the total number of working condition categories; For the Under each working condition Fault type label for each category of fault.

[0053] Step 2: Get signal frequency domain information.

[0054] The collected original vibration signal is subjected to fast Fourier transform to obtain its frequency domain information. The fast Fourier transform formula is as follows:

[0055] ,

[0056] in, represents the frequency domain signal after fast Fourier transform, is the discrete value of frequency; represents a discrete time domain signal, is a discrete value of time; It represents the length of the discrete time domain signal and is also the number of sampling points of the frequency domain signal.

[0057] Step 3: Enhance the amplitude and phase of the original signal.

[0058] According to the main idea of ​​Mixup data enhancement: linearly interpolate the features and labels of several samples during the training process, considering that the relevant features of vibration signals used for fault classification in the frequency domain are often more obvious, the present invention uses the amplitude and phase enhancement method to synthesize new domain samples. This method regards the amplitude and phase information in the frequency domain as two features and processes them separately.

[0059] For amplitude information, linear interpolation is still used.

[0060] ,

[0061] in, To enhance the sample, is the input signal data sample, For Samples belonging to different domains (same fault under different working conditions); Represents different weights when performing linear interpolation, obeys Beta distribution, and has a value range of , Represents the amplitude in the frequency domain.

[0062] For phase information, first calculate and The phase difference between , so that the enhanced sample The phase of is interpolated from the input samples.

[0063] ,

[0064] in Represents the phase of the signal data in the frequency domain. Then according to the phase difference Calculate the shortest phase difference .

[0065] ,

[0066] Compute enhanced samples Phase information ,

[0067] ,

[0068] in is a control parameter used to adjust the degree of phase change, and its value range is .

[0069] According to the obtained amplitude information and phase information, the signal is reconstructed by using inverse Fourier transform to obtain signal samples with amplitude and phase enhancement.

[0070] .

[0071] Step 4: Perform distillation learning on the samples after amplitude and phase enhancement.

[0072] Domain invariant features are learned through a distillation learning framework. Specifically, the samples after amplitude and phase enhancement are used as the input of the teacher network so that the model can learn the corresponding enhanced features. The teacher network training process is shown as follows:

[0073] ,

[0074] In the formula, Represents all inputs in the sample space The expected value of Indicates that the data distribution The training samples are obtained by sampling from is the input data, is the corresponding label, represents the expected value of all training samples. is the classification loss, that is, the cross entropy loss function, and are the classifier and feature extractor of the teacher network, respectively. and are samples and labels after amplitude and phase enhancement, and are the learnable parameters of the teacher network feature extractor and classifier, respectively.

[0075] After training the teacher network, we get the feature extractor of the teacher network , and guide the student network to learn the sample feature information after data enhancement through feature knowledge distillation, and learn the domain invariant representation, which is expressed as follows:

[0076] ,

[0077] In the formula, and They are the feature extractors in the student network. and classifier The learnable parameters of is the mean squared error loss (MSE), which makes the features extracted by the student network close to the features of the teacher network. is the mean square error loss weight.

[0078] In the specific implementation part, the structure of each feature extractor in the network is kept consistent, and a fully connected layer is added after the feature extractor of the teacher network to reduce the size of the features extracted by the teacher network to half of the original size, and the mean square error loss is constructed with the first half of the extracted time domain features to extract relevant features.

[0079] Step 5: Frequency domain feature extraction.

[0080] The signal data after fast Fourier transform will be used as the input of the frequency domain feature extraction module to extract frequency domain related features.

[0081] Step 6: Feature representation self-challenge learning.

[0082] The mixed features are obtained by splicing the amplitude and phase enhancement features, time domain features, and frequency domain features. The self-challenge learning enhancement model’s ability to extract and utilize different features is enhanced by characterizing the self-challenge learning, thus avoiding the model’s overfitting of a few significant features and enhancing the generalization of the model. The specific process is as follows:

[0083] Calculating mixed features Gradient of loss :

[0084] ,

[0085] In the formula Represents the loss function, where the loss function is the cross entropy loss function, that is , is the true label, Predict labels for the model.

[0086] The selection of neurons to be inhibited is determined by the size of the gradient, which represents the selection of features with large gradients for inhibition in self-challenge learning. The gradient is calculated based on the features, and when inhibiting, the neurons are set to 1 or 0.

[0087] In the gradient After sorting, the proportion of neurons that need to be inhibited is , for the gradient Each neuron in is sorted from large to small to obtain the characteristic The corresponding order of the elements in , determine the location of the neuron to be inhibited according to the following formula;

[0088] ,

[0089] in, Indicates The position of the neuron, 0 means it is inhibited, and 1 means it is retained; Indicates the length of the gradient sequence.

[0090] Finally, the position With features Multiply to get new features.

[0091] Step 7: Troubleshooting.

[0092] The new features obtained after characterization self-challenge learning are input into the classifier composed of a fully connected layer and SoftMax for fault diagnosis. By inputting the vibration data collected by the equipment into the trained network, its health status type can be output.

Claims

1. Intelligent fault diagnosis method for variable working conditions based on distillation learning and feature fusion, It is characterized in that The method comprises the following steps: Step 1: Obtain the original vibration signal, mark the fault type, and build a training data set; Step 2: Build a fault diagnosis network model and perform the following operations: The time domain signal is used as the input of the time domain feature extraction module to extract time domain related features; The collected original vibration signal is subjected to fast Fourier transform to obtain the signal frequency domain information, and the frequency domain features are extracted through the feature extraction module; The collected original vibration signal is enhanced in amplitude and phase to obtain the signal sample after amplitude and phase enhancement; the sample after amplitude and phase enhancement is subjected to distillation learning to obtain the corresponding amplitude and phase enhancement features; During the training process, the features and labels of several samples are linearly interpolated, and the amplitude and phase enhancement method is used to synthesize new domain samples; The amplitude and phase information in the frequency domain are regarded as two features and processed separately; For the amplitude information, linear interpolation is used; Among them, x aug is the enhanced sample, x is the input signal data sample, is a sample from a different domain than x; A Represents different weights when performing linear interpolation, obeys Beta distribution, and has a value range of [0,1]. A(·) represents the amplitude in the frequency domain; For phase information, first calculate x and The phase difference θ between them makes the enhanced sample x aug The phase of is interpolated based on the input samples; Where P(·) represents the phase of the signal data in the frequency domain; then the shortest phase difference Δθ is calculated based on the phase difference θ; Calculate the augmented sample x aug The phase information P(x aug ); where λ p It is a control parameter used to adjust the degree of phase change, and its value range is 0≤λ p ≤1; According to the obtained amplitude information and phase information, the signal is reconstructed by using inverse Fourier transform to obtain the signal samples after amplitude and phase enhancement; Distillation learning is performed on the samples after amplitude and phase enhancement, and domain invariant features are learned through a distillation learning framework. The method is as follows: The samples after amplitude and phase enhancement are used as the input of the teacher network, so that the model can learn the corresponding enhanced features; the teacher network training process is expressed as follows: In the formula, E represents the expected value of all input x in the sample space, (x, y) ~ P data (x,y) represents the distribution from data P data The training samples are obtained by sampling from (x, y), where x is the input data and y is the corresponding label. represents the expected value of all training samples; L c is the classification loss, that is, the cross entropy loss function, C T and F T are the classifier and feature extractor of the teacher network, respectively. and are samples and labels after amplitude and phase enhancement, and are the learnable parameters of the teacher network feature extractor and classifier, respectively; After the teacher network is trained, the feature extractor F of the teacher network is obtained. T , and guide the student network to learn the sample feature information after data enhancement through feature knowledge distillation, and learn the domain invariant representation, which is expressed as follows: In the formula, and They are the feature extractor F in the student network respectively. S and classifier C S The learnable parameters, L mse is the mean square error loss, λ 1 is the mean square error loss weight; The structure of each feature extractor in the network is consistent. A fully connected layer is added after the feature extractor of the teacher network to reduce the size of the features extracted by the teacher network to half of the original size, and the mean square error loss is constructed with the first half of the extracted time domain features; The amplitude and phase enhancement features, time domain features and frequency domain features are spliced ​​to obtain mixed features, and feature extraction is performed through representation self-challenge learning; The extracted features are input into the classifier to identify the fault type; Step 3: Use the training data set to train the fault diagnosis network model, and input the collected vibration signal into the trained network model for fault diagnosis.

2. According to claim 1, the variable operating condition intelligent fault diagnosis method based on distillation learning and feature fusion, It is characterized in that The mixed features are obtained by splicing the amplitude and phase enhancement features, time domain features, and frequency domain features. The ability of the self-challenge learning enhancement model to extract and utilize different features is characterized. The specific process is as follows: Calculate the gradient g of the mixed feature z with respect to the loss z : Where loss represents the loss function, where the loss function is the cross entropy loss function, i.e. CrossEntropy, y true is the true label, y pred Predict labels for the model; Characterize the self-challenge learning, the neurons that need to be suppressed are determined by the gradient size; z After sorting, the gradient g is adjusted according to the proportion p of neurons that need to be suppressed. z Each neuron in is sorted from large to small to obtain the ranking corresponding to the elements in feature z. g , determine the position of the neuron to be inhibited according to formula (9); Among them, location(i) represents the location of the i-th neuron, 0 means it is suppressed, and 1 means it is retained; len(g z ) represents the length of the gradient sequence; Finally, multiply the location (i) by the feature z to obtain a new feature.