Rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation

Abstract

The invention requests to protect a rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation. The method comprises the following steps: obtaining a rolling bearing vibration signal, and carrying out sample division and normalization processing; the spectral kurtosis of the source working condition sample is calculated, the center frequency and bandwidth of envelope demodulation are determined according to the spectral kurtosis, envelope demodulation is carried out, and the demodulation sample is sent to a teacher network phi for pre-training; respectively inputting an original vibration sample and an envelope spectrum sample under a source working condition into a student network psi and a pre-trained teacher network phi, and extracting bearing health state representations learned by the two networks to establish envelope spectrum knowledge distillation loss; calculating the prediction loss of the student network, the cross-working-condition sample probability distribution alignment loss and the diversity loss represented by the health state; carrying out multi-loss back propagation, and optimizing the weight and bias of the fault diagnosis network by adopting an Adam optimizer; and finally, inputting the target working condition sample into the trained fault diagnosis model for performance testing.

Description

Technical Field

[0001] The present invention relates to the field of rolling bearing fault diagnosis and classification under variable working conditions, and designs a domain invariant representation learning model based on envelope spectrum knowledge distillation, which is suitable for rolling bearing fault diagnosis under variable working conditions. Background Art

[0002] Rolling bearing fault diagnosis has made significant progress with the help of deep learning technology, improving the accuracy and reliability of diagnosis. These technologies usually assume that the training and test samples are independent and identically distributed. However, this condition is usually difficult to meet in practical applications, especially when different speeds and loads are involved in rolling bearing health status monitoring. This change in working conditions puts forward unique requirements for rolling bearing fault mode diagnosis.

[0003] After searching, the closest prior art is application publication number CN110160791B, an induction motor bearing fault diagnosis system and method based on wavelet-spectral kurtosis, including a wavelet-spectral kurtosis analysis module, a time domain parameter statistical analysis module, an FFT analysis module, a Hilbert envelope demodulation analysis module, a spectral kurtosis analysis module, etc., wherein the wavelet-spectral kurtosis analysis module includes two data preprocessing methods, namely wavelet and wavelet packet, which can realize the identification and feature extraction of multiple faults in the vibration signal of motor bearing coupled fault; the time domain parameter statistical analysis module monitors whether the motor bearing is abnormal by calculating the kurtosis value of the time domain sensitive parameter; the spectrum analysis module and the Hilbert envelope demodulation analysis module mainly determine the feature extraction of a single bearing fault of the motor. From the perspective of problem solving, patent CN110160791B does not consider the challenges brought to the diagnostic method by changes in bearing operating conditions, while this application mainly solves problems such as internal covariance drift in bearing vibration signals across operating conditions, and has stronger universality; from the perspective of technical implementation path, patent CN110160791B adopts traditional signal analysis and processing methods, which are highly dependent on manual experience, while this application adopts an end-to-end deep learning model, which is more convenient to deploy and use. Summary of the invention

[0004] The present invention aims to solve the above problems of the prior art. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation is proposed. The technical solution of the present invention is as follows:

[0005] A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation, characterized by comprising the following steps:

[0006] S1: Obtain the vibration signal of the rolling bearing, divide the vibration signal into samples, and perform normalization processing;

[0007] S2: Calculate the spectral kurtosis of the source working condition sample, and determine the center frequency and bandwidth of the envelope demodulation according to the spectral kurtosis, perform envelope demodulation on the source working condition sample according to the demodulation parameters, and send the demodulated sample to the teacher network Φ for pre-training;

[0008] S3: Extracting envelope spectrum knowledge representation from source condition samples using a pre-trained teacher network;

[0009] S4: Input the original signal under the source working condition into the student network Ψ, and use the student network to learn the knowledge representation of the health state of the source working condition signal;

[0010] S5: Calculate the spectral knowledge distillation loss based on the envelope spectrum knowledge representation and the original signal knowledge representation, calculate the high-order statistics between the health status representations of multi-source domain samples to obtain the alignment loss of cross-condition representations, use the L2 norm to determine the representation diversity loss, and obtain the prediction loss based on the cross entropy between the predicted health status and the actual rolling bearing health status;

[0011] S6: Perform loss back propagation and use the Adam optimizer to optimize the weights and biases of the fault diagnosis network;

[0012] S7: Input the target operating condition samples into the trained fault diagnosis model for performance testing.

[0013] Furthermore, the step S2 calculates the spectral kurtosis of the source working condition sample, and determines the center frequency and bandwidth of the envelope demodulation according to the spectral kurtosis, performs envelope demodulation on the source working condition sample according to the demodulation parameters, and sends the demodulated sample to the teacher network Φ for pre-training, specifically including:

[0014] S21: Perform short-time Fourier transform (STFT) on the signal x to obtain its time-frequency representation. STFT divides the signal into multiple overlapping small segments, and Fourier transform is performed on each segment.

[0015] S22: Square each time window of the STFT result to obtain the power spectral density, and calculate the mean and standard deviation of the power spectral density;

[0016] S23: Calculate the spectral kurtosis SK(f) of each signal segment according to the following formula, and determine the center frequency f of the corresponding signal segment based on the maximum value of the spectral kurtosis: s And the bandwidth parameter b n ;

[0017]

[0018] Where X(f,t) is the STFT result of the signal at frequency f and time t, Indicates expected value.

[0019] Center frequency f sis the position where the spectral kurtosis value reaches its maximum value, that is: f s =argmin f SK(f). Bandwidth b n It reflects the spectrum expansion range of the rolling bearing vibration signal, which is calculated as follows: First, a threshold SK is set threshold =η×SK(f s ), usually η is taken as 0.5, that is, 50% kurtosis peak; find the left and right boundary frequency points equal to the threshold in the spectrum kurtosis, that is, SK(f low )=SK(f high )=SK threshold , then the bandwidth is calculated as: b n =f high -f low .

[0020] Furthermore, step S3 is based on the center frequency f s And the bandwidth parameter b n The source condition samples are envelope demodulated and the demodulated samples are sent to the teacher network Φ for pre-training, which includes:

[0021] S31: Design a center frequency f s And the bandwidth parameter is b n Bandpass filter to filter out other frequency components in the signal and retain the components within the target frequency range;

[0022] S32: for the filtered signal x bp (t) Perform Hilbert transform on the bandpass filtered signal to obtain the analytical signal of the signal. The Hilbert transform converts the signal from the real domain to the complex domain, so that the real part of the signal is the original signal and the imaginary part is its Hilbert transform, which is expressed as follows:

[0023] x a (t) = x bp (t)+j*H{x bp (t)} (2)

[0024] S33: Obtain the signal envelope and obtain the corresponding phase spectrum. The envelope and phase spectrum are calculated as shown in formulas (3) and (4);

[0025]

[0026] Phase(E(f))=arg{FFT[e(t)]} (4)

[0027] Where e(t) represents the modulus of the Hilbert analytic signal, Phase(E(f)) represents the envelope demodulation phase spectrum, and FFT is the fast Fourier transform;

[0028] S34: Feed the envelope demodulated phase spectrum into the teacher model Φ for training.

[0029] Furthermore, the step S4: using the pre-trained teacher network Φ to extract the envelope spectrum knowledge representation rep from each source working condition sample phase (E(f)) Φ , rep Φ The calculation is shown in formula (5);

[0030] rep Φ =Featurizer Φ {Phase(E(f))} (5)

[0031] Featurizer is the network health status representation extraction module, Φ represents the teacher network; Φ The bearing health status representation extraction module representing the teacher network; the following Featurizer Ψ Then the health status of the student network is represented by the extraction module.

[0032] Furthermore, the step S5 inputs the original signal under the source working condition into the student network Ψ, and uses the student network to learn the health state knowledge representation rep of the source working condition signal. Ψ , and its calculation formula is as follows:

[0033] rep Ψ =Featurizer Ψ {x} (6)

[0034] Furthermore, the step S6 respectively calculates the spectral knowledge distillation loss, alignment loss, diversity loss and prediction loss of the health status prediction model training process, specifically including:

[0035] S61: Calculate the spectral knowledge distillation loss based on the envelope spectrum knowledge representation and the original signal knowledge representation, the distillation loss L spe The calculation formula is as follows:

[0036]

[0037] Where n represents the dimension of the deep representation of the health status obtained by the fault diagnosis model; They correspond to the rolling bearing health status representation learned by the teacher network and the student network respectively;

[0038] S62: Calculate the high-order statistics between the health status representations of multi-source domain samples to obtain the alignment loss of cross-condition representations, the alignment loss L ali The calculation formula is as follows:

[0039]

[0040] Where F represents the Frobenius norm, They represent the learned health status representations under the i-th and j-th source conditions, respectively, and C is the covariance matrix;

[0041] S63: Using L2 norm to determine the representation diversity loss L div , and its calculation formula is as follows:

[0042]

[0043] S64: The prediction loss L is obtained based on the cross entropy between the predicted health state and the actual rolling bearing health state. cla ,

[0044]

[0045] Where N is the number of health status categories, y i is the distribution of the true health status labels, where the position corresponding to the health status is 1 and the others are 0. is the probability distribution predicted by the diagnostic model.

[0046] Furthermore, the teacher network and the student network use the same network architecture parameters, and each network includes a feature extraction module Featurizer, a bottleneck module Bottleneck and a classification module Classification. The feature extraction module Featurizer is a front-end module for designing a deep network, which is used to convert the original input data (including the original vibration signal or the envelope demodulation spectrum) into a health status representation in the feature space; the bottleneck module Bottleneck is an intermediate layer between Featurizer and Classification in the deep network, which is mainly used for dimensionality reduction and fusion of the health status representation of the rolling bearing, and filtering out redundant or unimportant feature representations; the classification module Classification is the terminal part of the deep learning network, which is used to perform the final classification of the feature representation compressed by Bottleneck.

[0047] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation as described in any one of the items is implemented.

[0048] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation as described in any one of the items.

[0049] The advantages and beneficial effects of the present invention are as follows:

[0050] The present invention aims at the fact that the traditional fault diagnosis model relies on the stability of the device signal generation mechanism. However, in the process of rolling bearing fault identification, it is difficult to meet the above conditions due to the changes in speed and load. Therefore, the present invention proposes a domain-invariant representation learning strategy based on envelope spectrum knowledge distillation for diagnosing rolling bearing faults under different working conditions. It aims to be able to effectively identify the rolling bearing fault state and adapt to the changes in working conditions, thereby promoting the migration and application of marked fault knowledge under different working conditions.

[0051] The present invention utilizes envelope spectrum knowledge distillation to capture Fourier features as invariant features within the domain, and ensures robust health status characterization by aligning high-order statistics of samples under different working conditions. In addition, an innovative loss function is designed to enrich the diversity of characterization. Unlike traditional models that require training and test samples to follow independent and identical distribution conditions, the present invention provides a more flexible method. The average AUC on the Paderborn rolling bearing dataset was increased from 0.68 to 0.99, and the accuracy of rolling bearing fault diagnosis under variable working conditions was 98.5%, an increase of 3.4 percentage points compared to the traditional machine learning model decision tree.

[0052] The innovation of the bearing diagnosis model proposed in the present invention lies in that the loss function included is composed of four parts. The first is the classification loss function, which is responsible for completing the preliminary identification of the health status of the rolling bearing; the second is the envelope spectrum knowledge distillation loss, which enables the trained diagnosis network to still learn the fault characteristics in the signal when the input is the original vibration signal; the third is the diversity loss of the health status representation, which reduces the redundant information between the health status representations and enhances the diversity of the representation; the fourth is the sample probability distribution adaptation loss, which adapts the probability distribution state of the cross-working condition samples. On the basis of the classification loss, the envelope spectrum knowledge, the representation diversity metric and the probability distribution adaptation are introduced to improve the generalization performance of the diagnosis model. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The present invention provides a preferred embodiment of a domain-invariant representation learning framework based on envelope spectrum knowledge distillation;

[0054] Figure 2 This is the principle diagram of envelope spectrum knowledge distillation;

[0055] Figure 3 is the spectral kurtosis diagram of the source condition sample;

[0056] Figure 4 3D graph for model parameter mesh optimization;

[0057] Figure 5 The loss change and verification accuracy curve of the model training process;

[0058] Figure 6 ROC curve of the fault diagnosis model (a) the model proposed by the present invention (b) the spectral knowledge distillation model;

[0059] Figure 7 Confusion matrix of the fault diagnosis model on the working condition generalization task (a) Decision tree model (b) One-dimensional convolution model based on MMD (c) Fourier spectrum knowledge distillation model (d) Model proposed by the present invention; Figure 8 A flow chart of the rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation of the present invention is shown. DETAILED DESCRIPTION

[0060] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0061] The technical solution of the present invention to solve the above technical problems is:

[0062] Figure 1 The model framework of the fault diagnosis network is given. From the perspective of model operation, the framework is mainly divided into model training and model testing. Structurally, the framework consists of a teacher network and a student network. Both the teacher network and the student network use the same network architecture parameters. Each network contains a feature extraction module Featurizer, a bottleneck module Bottleneck and a classification module Classification.

[0063] The present invention is based on a deep learning model constructed by a feature extraction module, a bottleneck module and a classification module. First, knowledge distillation is used to extract a robust representation of the health status of rolling bearings that reflects frequency domain features; then, high-order statistics of rolling bearing samples across working conditions are calculated to improve the consistency of sample probability distributions under different working conditions, so as to promote the model's learning of universal representations of the health status of rolling bearings; finally, an additional loss function is designed to enhance the diversity of the learned representations of the health status of rolling bearings, further improving the generalization ability of the rolling bearing fault diagnosis model. Specifically, the following steps are included:

[0064] S1: Obtain rolling bearing vibration samples and normalize them.

[0065] S2: Calculate the spectral kurtosis of the source working condition samples, and determine the center frequency and bandwidth of the envelope demodulation according to the spectral kurtosis, perform envelope demodulation on the source working condition samples according to the demodulation parameters, and send the demodulated samples to the teacher network Φ for pre-training.

[0066] S21: Perform a short-time Fourier transform (STFT) on the signal x to obtain its time-frequency representation. STFT divides the signal into multiple overlapping small segments, and performs a Fourier transform on each segment.

[0067] S22: Square each time window of the STFT result to obtain the power spectral density, and calculate the mean and standard deviation of the power spectral density.

[0068] S23: Calculate the spectral kurtosis of each signal segment according to the following formula, and determine the center frequency f of the corresponding signal segment based on the maximum value of the spectral kurtosis: s And the bandwidth parameter b n .

[0069]

[0070] Where X(f,t) is the STFT result of the signal at frequency f and time t, Indicates expected value.

[0071] S3: Based on the center frequency f s And the bandwidth parameter b n The source working condition samples are envelope demodulated and the demodulated samples are sent to the teacher network Φ for pre-training.

[0072] S31: Design a center frequency f s And the bandwidth parameter is b n Bandpass filter to filter out other frequency components in the signal and retain the components within the target frequency range.

[0073] S32: for the filtered signal x bp (t) Perform Hilbert transform on the bandpass filtered signal to obtain the analytical signal of the signal. The Hilbert transform converts the signal from the real domain to the complex domain, so that the real part of the signal is the original signal and the imaginary part is its Hilbert transform. It is expressed as follows:

[0074] x a (t) = x bp (t)+j*H{x bp (t)} (12)

[0075] S33: Obtain the signal envelope and obtain the corresponding phase spectrum. The envelope and phase spectrum are calculated as shown in formulas (3) and (4).

[0076]

[0077] Phase(E(f))=arg{FFT[e(t)]} (14)

[0078] S34: Feed the envelope demodulated phase spectrum into the teacher model Φ for training.

[0079] S4: Use the pre-trained teacher network Φ to extract the envelope spectrum knowledge representation rep from each source condition sample Phase(E(f)) Φ , the teacher network structure parameters are shown in the following table. Φ The calculation is shown in formula (5).

[0080] rep Φ =Featureizer Φ {Phase(E(f))} (15)

[0081] Table 1 Teacher network structure parameters

[0082]

[0083]

[0084] S5: Input the original signal under the source working condition into the student network Ψ, and use the student network to learn the health status knowledge representation rep of the source working condition signal Ψ , and its calculation formula is as follows:

[0085] rep Ψ =Featureizer Ψ {x} (16)

[0086] S6: Calculate the spectral knowledge distillation loss, alignment loss, diversity loss, and prediction loss of the health status prediction model training process respectively.

[0087] S61: Calculate the spectral knowledge distillation loss based on the envelope spectrum knowledge representation and the original signal knowledge representation. The basic principle of spectral knowledge distillation is as follows: Figure 2 As shown. Distillation loss L spe The calculation formula is as follows:

[0088]

[0089] Where n represents the dimension of the deep representation of the health status obtained by the fault diagnosis model. They correspond to the health status representations learned by the teacher network and the student network respectively.

[0090] S62: Calculate the high-order statistics between the health status representations of multi-source domain samples to obtain the alignment loss of cross-condition representations, the alignment loss L ali The calculation formula is as follows:

[0091]

[0092] Where F represents the Frobenius norm. denote the learned health state representations under the i-th and j-th source conditions, respectively, and C(·) denotes the covariance matrix.

[0093] S63: Using L2 norm to determine the representation diversity loss L div , and its calculation formula is as follows:

[0094]

[0095] S64: The prediction loss L is obtained based on the cross entropy between the predicted health state and the actual rolling bearing health state. cla .

[0096]

[0097] Where N is the number of health status categories, y i is the distribution of the true health status labels (usually a one-hot encoded vector where the position corresponding to the health status is 1 and the others are 0), is the probability distribution predicted by the diagnostic model.

[0098] S7: Perform loss backpropagation and use the Adam optimizer to optimize the weights and bias of the fault diagnosis network.

[0099] S8: Input the target operating condition samples into the trained fault diagnosis model for performance testing.

[0100] Embodiment 1

[0101] This embodiment provides a fault diagnosis case based on the Paderborn rolling bearing dataset, which uses the six health states in the dataset. The fault locations include inner ring, outer ring, and mixed inner and outer ring faults. The fault types include pitting, deformation, etc. The data details are shown in the table below.

[0102] Table 2 Details of Paderborn rolling bearing dataset

[0103]

[0104] S1: Dataset division: A 4-second vibration signal is selected for model building, the source working condition samples are used as model training samples, and the target working condition samples are used as model testing samples.

[0105] S2: Demodulation parameter determination: Spectral kurtosis method is used to analyze the source condition samples, such as Figure 3 As shown, based on the extreme value in the spectral kurtosis diagram, the center frequency of the new segment rich in fault information is determined to be 4.8kHz and the bandwidth is 1.0667kHz

[0106] S3: Teacher model pre-training: The teacher model is trained using the resonant demodulation phase spectrum.

[0107] S4: Model training parameter optimization: The grid search method is used to optimize the learning rate and batch size of the rolling bearing fault diagnosis model, such as Figure 4 As shown. It is worth noting that different combinations of learning rate and batch size lead to different accuracy levels. The x-axis parameter (learning rate) ranges between 0.001 and 0.01, while the y-axis parameter (batch size) varies from 32 to 1024. Analysis shows that a learning rate of 0.001 and a batch size of 256 can achieve higher diagnostic accuracy. On the contrary, a larger learning rate will lead to reduced accuracy due to its amplifying effect on the parameter update step size, potentially reducing the parameter search space of the fault diagnosis model and causing the search parameters of the diagnosis model to swing on both sides of the optimal value, making it difficult to achieve the best results.

[0108] S5: Feed the source condition into the student network and calculate the spectral knowledge distillation loss, alignment loss, diversity loss, and prediction loss of the health status prediction model training process. The diagnostic network model is built based on Python 3.9, the core framework includes PyTorch 1.9.1, and the integrated development environment used is PyCharm 2021. The model runs on hardware equipped with an i9 processor and an RTX 4060 GPU. Figure 5 The overall loss value of the training process and the change in validation accuracy are given. Obviously, as more iterative training is carried out, the loss value steadily decreases and eventually reaches stability. This means that the deep diagnosis model gradually converges to the equilibrium point, indicating that its convergence efficiency is high. In addition, the validation accuracy gradually increases, which indicates that the predicted health state is gradually approaching its actual health state.

[0109] S6: Analyzing the role of envelope spectrum knowledge distillation in domain-invariant representation learning, Figure 6 The ROC curves of the method of the present invention and the knowledge distillation model without envelope spectrum are compared. The calculation formulas of the true positive rate TPR and the false positive rate FPR are as follows:

[0110]

[0111] Where TP is the number of positive examples correctly predicted as positive examples; FN is the number of positive examples incorrectly predicted as negative examples; FP is the number of negative examples incorrectly predicted as positive examples; and TN is the number of negative examples correctly predicted as negative examples.

[0112] For the multi-class rolling bearing fault diagnosis task, the average area under ROC of A1 is 0.68. In contrast, the proposed method achieves an average area under ROC of 0.99, showing near-perfect accuracy on the Paderborn dataset. This means that there is a suitable threshold that enables this model to achieve better fault diagnosis results.

[0113] S7: The ability of the confusion matrix analysis model to identify the health status of rolling bearings, such as Figure 7 This embodiment carefully selects three methods, namely, decision tree in traditional machine learning methods Figure 7 a) One-dimensional convolution model based on MMD Figure 7 (b) and Fourier spectrum knowledge distillation model Figure 7 (c) to illustrate the advantages of the method of the present invention. The comparison methods represent benchmark models in the fields of deep learning and transfer learning. The confusion matrix shows that the fault diagnosis performance of the Fourier spectrum knowledge distillation model is the worst. The collected rolling bearing vibration signals usually show signal modulation, in which the rolling bearing fault characteristic frequency is modulated by the natural characteristic frequency of the system. Simply applying fast Fourier transform to the rolling bearing vibration signal cannot effectively highlight the rolling bearing fault characteristics. Therefore, spectrum knowledge distillation will impair the performance of the fault diagnosis model. Compared with the Fourier spectrum knowledge distillation model, the fault diagnosis performance of the decision tree is improved by 3.5%. However, traditional machine learning methods mainly rely on hand-crafted features, and rules need to be defined to filter these features to achieve competitive performance. The one-dimensional convolution model based on MMD enhances the cross-condition generalization ability of the model by combining MMD in the deep learning architecture. MMD aligns the probability distribution of samples under different working conditions, so that the diagnostic accuracy of the one-dimensional convolution model based on MMD reaches 97.9%; while the present invention aligns the high-order statistics of samples under different working conditions to achieve the highest diagnostic accuracy of the model.

[0114] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0115] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0116] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0117] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation, characterized in that: The following steps are involved: S1: Obtain the vibration signal of the rolling bearing, divide the vibration signal into samples, and perform normalization processing; S2: Calculate the spectral kurtosis of the source working condition sample, and determine the center frequency and bandwidth of the envelope demodulation according to the spectral kurtosis, perform envelope demodulation on the source working condition sample according to the demodulation parameters, and send the demodulated sample to the teacher network Φ for pre-training; S3: Using the pre-trained teacher network to extract the rolling bearing health status representation based on envelope spectrum knowledge from the source working condition samples; S4: Input the original signal under the source working condition into the student network Ψ, and use the student network to learn the health status knowledge representation of the source working condition signal; S5: Calculate the spectral knowledge distillation loss based on the envelope spectrum knowledge representation and the original signal knowledge representation, calculate the high-order statistics between the health status representation covariance matrices of multi-source domain samples to obtain the probability distribution alignment loss of cross-condition representation, use the L2 norm to determine the representation diversity loss, and obtain the prediction loss based on the cross entropy between the predicted health status and the actual rolling bearing health status; S6: Perform loss back propagation and use the Adam optimizer to optimize the weights and biases of the rolling bearing fault diagnosis network; S7: Input the target working condition samples into the trained fault diagnosis model for performance testing.

2. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation according to claim 1, characterized in that: The step S2 calculates the spectral kurtosis of the source working condition sample, and determines the center frequency and bandwidth of the envelope demodulation according to the spectral kurtosis, performs envelope demodulation on the source working condition sample according to the demodulation parameters, and sends the demodulated sample to the teacher network Φ for pre-training, specifically including: S21: Perform short-time Fourier transform (STFT) on the signal x to obtain its time-frequency representation. STFT divides the signal into multiple overlapping small segments, and Fourier transform is performed on each segment. S22: Square each time window of the STFT result to obtain the power spectral density, and calculate the mean and standard deviation of the power spectral density; S23: Calculate the spectral kurtosis SK(f) of each signal segment according to the following formula, and determine the center frequency f of the corresponding signal segment based on the maximum value of the spectral kurtosis: s And the bandwidth parameter b n ; Where X(f, t) is the STFT result of the signal at frequency f and time t, Indicates expected value. Center frequency f s is the position where the spectral kurtosis value reaches the maximum value, that is: f s =argmin f SK(f). Bandwidth b n It reflects the spectrum expansion range of the rolling bearing vibration signal, which is calculated as follows: First, a threshold SK is set threshold =η×SK(f s ), usually η is taken as 0.5, that is, 50% kurtosis peak; find the left and right boundary frequency points equal to the threshold in the spectrum kurtosis, that is, SK(f low )=SK(f high )=SK threshold , then the bandwidth is calculated as: b n =f high -f low .

3. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation according to claim 2, characterized in that: The step S3 is based on the center frequency f s And the bandwidth parameter b n The source condition samples are envelope demodulated and the demodulated samples are sent to the teacher network Φ for pre-training, which includes: S31: Design a center frequency f s And the bandwidth parameter is b n Bandpass filter to filter out other frequency components in the signal and retain the components within the target frequency range; S32: for the filtered signal x bp (t) Perform Hilbert transform on the bandpass filtered signal to obtain the analytical signal of the signal. The Hilbert transform converts the signal from the real domain to the complex domain, so that the real part of the signal is the original signal and the imaginary part is its Hilbert transform, which is expressed as follows: x a (t)=x bp (t)+j*H{x bp (t)} (2) S33: Obtain the signal envelope and obtain the corresponding phase spectrum. The envelope and phase spectrum are calculated as shown in formulas (3) and (4); Phase(E(f))=arg{FFT[e(t)]} (4) Where e(t) represents the modulus of the Hilbert analytic signal, Phase(E(f)) represents the envelope demodulation phase spectrum, and FFT is the fast Fourier transform; S34: Feed the envelope demodulated phase spectrum into the teacher model Φ for training.

4. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation according to claim 3, characterized in that: Step S4: extracting the envelope spectrum knowledge representation rep from each source working condition sample Phase(E(f)) using the pre-trained teacher network Φ Φ , rep Φ The calculation is shown in formula (5); rep Φ =Featurizer Φ {Phase(E(f))} (5) Featurizer is the network health status representation extraction module, Φ represents the teacher network; Φ The bearing health status representation extraction module representing the teacher network; the following Featurizer Ψ Then the health status of the student network is represented by the extraction module.

5. The rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation according to claim 4 is characterized in that: The step S5 inputs the original vibration signal under the source working condition into the student network Ψ, and uses the student network to learn the health state knowledge representation rep of the source working condition signal. Ψ , and its calculation formula is as follows: rep Ψ =Featurizerq Ψ {x} (6) 6. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation according to claim 5, characterized in that: The step S6 respectively calculates the spectral knowledge distillation loss, probability distribution alignment loss, diversity loss and prediction loss of the health status prediction model training process, specifically including: S61: Calculate the spectral knowledge distillation loss based on the envelope spectrum knowledge representation and the original signal knowledge representation, the distillation loss L spe The calculation formula is as follows: Where n represents the dimension of the health state representation obtained by the fault diagnosis model; Corresponding to the health status representation learned by the teacher network and the student network respectively; S62: Calculate the high-order statistics between the covariance matrices of the health status representation of samples in multiple source domains to obtain the alignment loss of cross-condition representation, the alignment loss L ali The calculation formula is as follows: Where F represents the Frobenius norm, They represent the rolling bearing health status representation learned under the i-th and i-th source conditions respectively, and C is the covariance matrix; S63: Using L2 norm to determine the representation diversity loss L div , and its calculation formula is as follows: S64: The prediction loss L is obtained based on the cross entropy between the predicted health state and the actual rolling bearing health state. cla as follows: Where N is the number of rolling bearing health status categories, y i is the distribution of the true health status labels, where the position corresponding to the health status is 1 and the others are 0. is the probability distribution predicted by the diagnostic model.

7. A rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation according to any one of claims 1 to 6, characterized in that: The teacher network and the student network both use the same network architecture parameters. Each network includes a feature extraction module Featurizer, a bottleneck module Bottleneck and a classification module Classification. The feature extraction module Featurizer is a front-end module for designing a deep network, which is used to convert the original input data (including the original vibration signal or the envelope demodulation spectrum) into a health status representation in the feature space; the bottleneck module Bottleneck is an intermediate layer between Featurizer and Classification in the deep network, which is mainly used for dimensionality reduction and fusion of the health status representation of the rolling bearing, and filtering out redundant or unimportant feature representations; the classification module Classification is the terminal part of the deep learning network, which is used to perform the final classification of the feature representation compressed by Bottleneck.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation as described in any one of claims 1 to 7.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the rolling bearing fault diagnosis method based on envelope spectrum knowledge distillation as described in any one of claims 1 to 7 is implemented.

Images