Rolling bearing fault diagnosis method based on wavelet change and improved TL-ResNet

Abstract

The invention discloses a bearing fault diagnosis method based on wavelet variation and an improved TL-ResNet. Comprising the following steps of data preprocessing, model establishment and model training. In the aspect of data preprocessing, a time domain vibration signal is converted into an image data set through a wavelet transform processing mode, and the advantages of a neural network in image processing are fully utilized. In the model building process, a pre-trained ResNet18 (residual network) model is utilized, and a network structure and characteristic parameters are migrated to a target domain for training, so that the training time is shortened, and the problems of insufficient data and poor model training effect are solved. In addition, by introducing an SENet module, generating a channel weight by adopting compression and excitation operations, and re-weighting a feature response, the model can pay more attention to important features, and the accuracy of fault diagnosis is improved.

Description

Technical Field

[0001] The present invention relates to the field of bearing fault diagnosis, and in particular to a vehicle bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet. Background Art

[0002] Rolling bearings are key components of vehicle transmission systems and electromechanical systems and are widely used in vehicle transportation. However, since rolling bearings often operate in complex and harsh environments, the probability of failure increases significantly. According to statistics, bearing failures account for between 40% and 70% of all vehicle transmission system failures. Therefore, it is particularly important to effectively detect rolling bearing failures and accurately determine the type of failure. By regularly inspecting and maintaining rolling bearings, mechanical accidents caused by bearing failures can be prevented and equipment safety can be ensured.

[0003] Traditional fault diagnosis methods usually include three main steps: signal acquisition, feature extraction, and fault classification. Traditional methods mainly rely on Fourier transform, empirical mode decomposition, Hilbert-Huang transform and other technologies to extract features by converting the original time domain vibration signal into the time-frequency domain, and then use these features for fault diagnosis. However, these methods have the following shortcomings: they rely on artificial prior knowledge, and researchers need to choose appropriate diagnostic techniques according to the specific environment, which lacks universality; it is difficult to design features that are applicable to all working conditions, resulting in limited diagnostic accuracy under different working conditions.

[0004] With the rapid development of computing power and technology, deep learning technology has overcome many limitations of traditional methods with its end-to-end learning capabilities, significantly improving the accuracy and efficiency of fault diagnosis. However, deep learning still faces some challenges in the field of fault diagnosis, such as the need for a large amount of sample data for training, and the possibility of gradient vanishing or gradient exploding during the training process, which affects the accuracy of the diagnosis results. In this context, the application of transfer learning technology is particularly important, especially in bearing fault diagnosis under different working conditions and environments.

[0005] In this process, the signal conversion into images plays a crucial role. According to the characteristics of the bearing vibration signal, the signal is first preprocessed to remove noise interference. Then, the one-dimensional time domain signal is converted into a two-dimensional time-frequency matrix using the time-frequency analysis method wavelet transform, thereby simultaneously showing the time characteristics and frequency characteristics of the signal. This representation method more intuitively reflects the fault characteristics of the signal. In order to further improve the visualization effect of the features, the time-frequency matrix is ​​normalized and a pseudo-color processing method is applied to generate a color image in RGB format. This graphical signal representation enhances the expression ability of the features. The generated color image is directly used as the input of the deep learning model, thereby achieving end-to-end fault classification and diagnosis.

[0006] The processing method of converting vibration signals into images can not only improve the feature expression ability, but also greatly reduce the dependence on artificial prior knowledge and simplify the complex feature design process in traditional methods. At the same time, this method can provide a unified feature representation under different working conditions, making transfer learning technology more efficient in cross-domain fault diagnosis. By fine-tuning the pre-trained model of the source domain data and migrating the high-level features extracted after signal imaging to the target domain, the fault diagnosis accuracy and robustness of the model in the target domain can be significantly improved, thereby effectively solving the limitations of traditional fault diagnosis methods under complex working conditions. The processing method of converting signals into images lays the foundation for the application of deep learning and transfer learning in bearing fault diagnosis, significantly improving the diagnostic efficiency and accuracy. Summary of the invention

[0007] In order to solve the above technical problems, an embodiment of the present application provides a bearing fault diagnosis method which converts one-dimensional signal data into a two-dimensional image format, thereby better utilizing the neural network extraction capability to be applicable to different working conditions, improving the computational efficiency of the method, and improving the accuracy of rolling bearing fault diagnosis.

[0008] The present invention solves the above technical problem by providing a technical solution: a rolling bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet, comprising the following steps:

[0009] (1) A vehicle bearing fault diagnosis method based on wavelet transform and improved TL-ResNet, characterized by comprising the following steps:

[0010] Step S1, data preprocessing: preprocessing the collected bearing vibration data and the bearing vibration data of the public data set to prepare for subsequent model training;

[0011] Step S2, model establishment: design a bearing fault diagnosis method based on wavelet transform and improved TL-ResNet. By using the pre-trained ResNet18 (residual network) model, its network structure and feature parameters are transferred to the target domain for training, which can reduce the training time of the model; then the SEResNet module is introduced to enhance the expressiveness of the model. Through the two operations of compression and excitation, the weights of each channel are generated, and then these weights are used to re-weigh the feature response of each channel, so that the model pays more attention to important feature maps. The specific process of the method is as follows:

[0012] Step S21: Use the ResNet18 model pre-trained on the ImageNet dataset, freeze some layers, perform parameter migration, and obtain features

[0013] Step S22: Construct a residual layer SENet, and learn the interdependence between different feature maps (channels) through squeezing and excitation to enhance the expression of useful features and suppress unimportant features.

[0014] Step S23: Migrate the pre-trained ResNet18 model after fine-tuning to the source domain for high-level feature extraction.

[0015] Step S24: Migrate the high-level features extracted from the source domain to the target domain to complete cross-domain fault diagnosis.

[0016] Step S3, model training: training the method model constructed in step S2 until the entire model converges;

[0017] (2) According to claim 1, the bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that the specific process of step S1 is:

[0018] Step S11: data set division;

[0019] This dataset contains the operating data of bearings under different fault states and working conditions. Specifically, the dataset covers the operation of bearings under four different fault types: normal state, inner ring fault, outer ring fault, and rolling element fault. Each fault type introduces three defects of different diameters to represent different degrees of fault. In addition, the dataset also includes the operating data of bearings under four different working conditions, one of which is a normal working condition, and the other three simulate different environments and load changes. Finally, the dataset contains a total of ten sets of data, covering a combination of various fault types and working conditions, which is suitable for bearing fault diagnosis and model training. Through this process, a bearing vibration dataset covering a variety of fault types and working conditions is generated. The composition of this dataset is as follows:

[0020] X 1 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0021] X 2 ={classI 1 ,classI 2 ,classI 3 ,classO1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0022] X 2 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0023] X 2 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0024] Among them, X i represents the bearing data set obtained under different working conditions, and class represents different types of fault data. Fault categories are distinguished by subscripts:

[0025] I indicates inner ring fault, I 1 ,I 2 ,I 3 Represents three different defect sizes of inner race faults.

[0026] O indicates outer ring fault, O 1 , O 2 , O 3 Represents three different defect sizes of outer race faults.

[0027] B indicates rolling element failure, B 1 , B 2 , B 3 Represent three different defect sizes of rolling element failure.

[0028] N indicates normal status.

[0029] Each set of data X i It includes all fault types (inner ring fault, outer ring fault, rolling element fault) collected under corresponding working conditions and normal operating data, totaling 10 categories.

[0030] Step S12: Data preprocessing

[0031] For the original data, wavelet transform is used to process it. The specific process is as follows:

[0032] The original data is the time series of the bearing vibration signal, recorded as x(t). According to the experimental settings, the time series signal is divided into multiple samples according to the sampling length of 1024 to form signal blocks of equal length. Each signal block is represented as

[0033] x i (t) = {x(t 1 ),x(t 2 ),...,x(t N )}i=1,2,...,M

[0034] Where M is the number of samples after segmentation.

[0035] Then the segmented signal is normalized to enhance the feature contrast and reduce the impact of amplitude on wavelet transform. The normalization formula is:

[0036]

[0037] Where μ is the signal mean and σ is the signal standard deviation.

[0038] Then the normalized data is subjected to wavelet transformation. The wavelet transformation converts the time domain signal into the time-frequency domain and analyzes the local characteristics of the signal by adjusting the scale (frequency) and time position. The formula of wavelet transformation is:

[0039]

[0040] Where x(t) is the time series signal, ψ(t) is the wavelet basis function. In this example, the complex Gaussian wavelet cgau8 is used. a is the scale parameter that controls the frequency range of the wavelet, and b is the translation parameter that determines the position of the wavelet on the time axis.

[0041] The amplitude (modulus) of the wavelet coefficient is taken as the energy distribution of the signal, and the calculation formula is:

[0042] Amplitude(a,b)=|W(a,b)|

[0043] Then use the amplitude matrix of the wavelet coefficients |W(a,b)| to draw the time-frequency diagram. The specific steps are as follows:

[0044] Time axis division: construct a time axis t, whose length is equal to the number of sampling points of the signal block: 1024;

[0045] Frequency axis division: According to the selected scale range a, calculate the corresponding frequency range f;

[0046] Plot: Display the magnitude matrix using a 2D pseudo-color plot (e.g., heat map) where:

[0047] The horizontal axis represents time t, the vertical axis represents frequency f, and the color represents the energy amplitude |W(a,b)|.

[0048] The image size is fixed to 224×224, and the time-frequency graph is normalized to the specified size through interpolation. The generated time-frequency graph is saved as an image file for subsequent model training. The time-frequency graph of each sample corresponds to a fault category, and the path naming contains the fault type and sample number.

[0049] (3). According to claim 1, the bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that the specific process of step S21 is: loading the pre-trained ResNet18 model: using the ResNet18 model that has been trained on the ImageNet dataset as the pre-trained model. The model has learned many common features of images during the training process. The ResNet network alleviates the gradient vanishing problem in deep network training by introducing residual connections. The structure of ResNet18 is as follows:

[0050] Y=F(X,{W i})+X

[0051] Among them, X is the input feature map, Y is the output feature map, F represents the residual function, and W i is the weight of each layer. The residual connection allows the input of each layer to be directly passed to the output to avoid information loss. The pre-trained weights of this structure are used to fine-tune the target task (bearing fault diagnosis), and the first three convolutional layers of ResNet18 (Conv1 and the first two residual layers) are frozen, and their parameters are not updated and remain in the initial state. The freezing operation means that the parameters of these layers will not be updated during training, and only the subsequent layers will be adjusted.

[0052] Freezing can be done by setting the gradient updates to zero:

[0053]

[0054] Among them, L is the loss function, W iRepresents the weights of the frozen layer. In this case, the parameters of the frozen layer will not be updated according to the gradient descent, which speeds up the training process and reduces the model complexity.

[0055] The weights of Conv1 and the first 3 residual layers are fixed so that they act as shallow feature extractors, thereby reducing training time and avoiding overfitting of these layers. The Conv1 layer and the first 3 residual layers in ResNet18 learn low-level features (such as edges, corners, etc.), which are usually universal and contribute greatly to all images (including the bearing fault dataset), so we fix the weights of these layers. At this stage, by keeping these shallow feature extractors unchanged, the model can focus on learning deep features and improve training efficiency. Mathematically, it can be expressed as:

[0056] X input →Conv1→Residual Block1→Residual Block2→Frozen Parameters

[0057] In this way, by leveraging the pre-trained features of the original ResNet18 model, the model can share this shallow knowledge between different tasks and avoid repeated learning of this part of knowledge.

[0058] Step S22: Construct a residual layer SENet, and learn the interdependence between different feature maps (channels) through squeezing and excitation, so as to enhance the expression of useful features and suppress unimportant features. The fourth residual layer is replaced by SENet. The SENet module mainly includes two operations: Squeeze and Excitation, which are used to model the global information between channels and adjust the response weight of the channel respectively. Specifically, for the input feature map Where H and W are the height and width of the feature map, respectively, and C is the number of channels. Global Average Pooling (GAP) is used to obtain the global information of each channel. The specific steps are:

[0059] 1) Squeeze operation, on the input feature map Global Average Pooling (GAP) is used to obtain the global information of each channel. The calculation formula of GAP is:

[0060]

[0061] Among them, z represents the global average of the cth channel, H and W are the height and width of the feature map respectively, and X C (i,j) represents the pixel value of the cth channel at position (i,j).

[0062] 2) Excitation operation: excitation of the global information vector after squeezing Two layers of full connection operations are performed to model the interdependence between channels and generate channel weights. The first layer of full connection: maps z to a lower-dimensional feature space to reduce computational complexity:

[0063] s=σ(W 2 ·δ(W 1 ·z))

[0064] in, and The weight matrix of the fully connected layer, r is the scaling factor, set to 16; δ represents the ReLU activation function, σ represents the Sigmoid activation function, is the generated channel weight.

[0065] 3) Reweighting: The generated channel weight s is applied to the input feature map X through channel-by-channel multiplication to adjust the response strength of each channel:

[0066]

[0067] in, Represents the adjusted c-th channel feature map.

[0068] 4) Replace the fourth residual layer and apply the above-constructed SENet module to the fourth residual layer of ResNet18, so that the model can adaptively strengthen key features and suppress redundant information. The output of the replaced residual block is:

[0069]

[0070] in, is the input feature map after SENet weighting, is the mapping result of the convolutional layer

[0071] Step S23: Migrate the fine-tuned pre-trained ResNet18 model to the source domain for high-level feature extraction. The specific process is:

[0072] 1) Load the fine-tuned model weights and load the ResNet18 model fine-tuned in step S22 into the target environment. At this point, the first four layers of the model (Conv1, Block1, Block2, Block3) and the SENet-enhanced Block4 have been optimized to adapt to the fault diagnosis task.

[0073] 2) The goal of source domain feature extraction, the data in the source domain Contains sample xi and the corresponding label y i , extract its high-level semantic features f(x i ), which is used for subsequent fault mode identification. The mathematical representation of the extracted features is:

[0074] f(xi)=ResNet(xi;Θ)

[0075] Among them, Θ represents the parameter set of the model, including the frozen part and the fine-tuned part

[0076] 3) High-level feature extraction process: first input the data, and then convert the sample data x in the source domain into i Through the fine-tuned ResNet18 model, when the data flows through the lower layers of the model (Conv1 to Block3), the basic features are extracted; when passing through the fourth residual layer (Block4), the SENet module is used to enhance the high-level semantic features. Then the fully connected layer of the model is used to output the feature vector z i , used to represent the sample x i The fully connected layer of the model outputs the feature vector z i It is expressed as:

[0077] z i =FC(GlobalAvgPool(Block4(...Block1(Conv1(x i )))))

[0078] Among them, FC represents the fully connected layer. The extracted high-level features Stored as a feature matrix for subsequent classification or transfer learning tasks.

[0079] Step S24: Migrate the high-level features extracted from the source domain to the target domain to complete cross-domain fault diagnosis. The specific process is:

[0080] 1) Preparation of target domain data: Unlabeled or small number of labeled data samples are input into the model. There are certain differences in the distribution of target domain and source domain data, and the distribution of the target domain is adapted through transfer learning methods.

[0081] 2) Feature mapping and alignment: Input the target domain sample xj into the fine-tuned ResNet18 model to extract the corresponding high-level features zj:

[0082] z j =FC(GlobalAvgPool(Block4(...Block1(Conv1(x j )))))

[0083] 3) Training of transfer learning tasks, using features extracted from the source domain and target domain unlabeled data The target domain data is adapted through transfer learning, so that predictions can be made on the test set of the target domain and the classification accuracy and recall rate can be calculated.

[0084] (4). According to claim 1, the bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that the specific process of step S3 is as follows: for the fault diagnosis problem of rolling bearings under different working conditions, the improved TL-ResNet model is adjusted using the test data under different working conditions, and the migration effect of the method between different working conditions is studied through transfer learning, wherein the migration task "0-1" means that the TL-ResNet model generated under working condition 0 (working condition 0 is the source domain) is migrated and applied to the rolling bearing fault diagnosis problem under working condition 1 (working condition 1 is the target domain). Through this migration training method, the model can effectively obtain features from the source domain and adapt to the target domain, further improving the diagnostic performance and robustness under changing working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0086] Figure 1 A schematic diagram of a transfer learning principle provided in an embodiment of the present application;

[0087] Figure 2 A wavelet change flow chart provided for an embodiment of the present application;

[0088] Figure 3 A flow chart of the method provided in the embodiment of the present application;

[0089] Figure 4 SENet flow chart provided for the embodiment of the present application; DETAILED DESCRIPTION

[0090] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0091] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0092] In order to make the description of the technical solution in the embodiment of the present invention more clear and complete, the specific implementation method of the embodiment of the present invention will be described in detail with reference to the accompanying drawings. Figure 3 As shown:

[0093] (2) A vehicle bearing fault diagnosis method based on wavelet transform and improved TL-ResNet, characterized in that it includes the following steps:

[0094] Step S1, data preprocessing: preprocessing the collected bearing vibration data and the bearing vibration data of the public data set to prepare for subsequent model training;

[0095] Step S2, model establishment: design a bearing fault diagnosis method based on wavelet transform and improved TL-ResNet. By using the pre-trained ResNet18 (residual network) model, its network structure and feature parameters are transferred to the target domain for training, which can reduce the training time of the model; then the SEResNet module is introduced to enhance the expressiveness of the model. Through the two operations of compression and excitation, the weights of each channel are generated, and then these weights are used to re-weigh the feature response of each channel, so that the model pays more attention to important feature maps. The specific process of the method is as follows:

[0096] Step S21: Use the ResNet18 model pre-trained on the ImageNet dataset, freeze some layers, perform parameter migration, and obtain features

[0097] Step S22: Construct a residual layer SENet, and learn the interdependence between different feature maps (channels) through squeezing and excitation to enhance the expression of useful features and suppress unimportant features.

[0098] Step S23: Migrate the pre-trained ResNet18 model after fine-tuning to the source domain for high-level feature extraction.

[0099] Step S24: Migrate the high-level features extracted from the source domain to the target domain to complete cross-domain fault diagnosis.

[0100] Step S3, model training: training the method model constructed in step S2 until the entire model converges;

[0101] (2) According to claim 1, the bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that the specific process of step S1 is:

[0102] Step S11: data set division;

[0103] This dataset contains the operating data of bearings under different fault states and working conditions. Specifically, the dataset covers the operation of bearings under four different fault types: normal state, inner ring fault, outer ring fault, and rolling element fault. Each fault type introduces three defects of different diameters to represent different degrees of fault. In addition, the dataset also includes the operating data of bearings under four different working conditions, one of which is a normal working condition, and the other three simulate different environments and load changes. Finally, the dataset contains a total of ten sets of data, covering a combination of various fault types and working conditions, which is suitable for bearing fault diagnosis and model training. Through this process, a bearing vibration dataset covering a variety of fault types and working conditions is generated. The composition of this dataset is as follows:

[0104] X 1 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0105] X 2 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0106] X2 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0107] X 2 ={classI 1 ,classI 2 ,classI 3 ,classO 1 ,classO 2 ,classO 3 ,classB 1 ,classB 2 ,classB 3 ,classN,}

[0108] Among them, X i represents the bearing data set obtained under different working conditions, and class represents different types of fault data. Fault categories are distinguished by subscripts:

[0109] I indicates inner ring fault, I 1 ,I 2 ,I 3 Represents three different defect sizes of inner race faults.

[0110] O indicates outer ring fault, O 1 , O 2 , O 3 Represents three different defect sizes of outer race faults.

[0111] B indicates rolling element failure, B 1 , B 2 , B 3 Represent three different defect sizes of rolling element failure.

[0112] N indicates normal status.

[0113] Each set of data X i It includes all fault types (inner ring fault, outer ring fault, rolling element fault) collected under corresponding working conditions and normal operating data, totaling 10 categories.

[0114] Step S12: Data preprocessing

[0115] For the original data, wavelet transform is used to process it. The specific process is as follows:

[0116] The original data is the time series of the bearing vibration signal, recorded as x(t). According to the experimental settings, the time series signal is divided into multiple samples according to the sampling length of 1024 to form signal blocks of equal length. Each signal block is represented as

[0117] x i (t) = {x(t 1 ),x(t 2 ),...,x(t N )}i=1,2,...,M

[0118] Where M is the number of samples after segmentation.

[0119] Then the segmented signal is normalized to enhance the feature contrast and reduce the impact of amplitude on wavelet transform. The normalization formula is:

[0120]

[0121] Where μ is the signal mean and σ is the signal standard deviation.

[0122] Then the normalized data is subjected to wavelet transformation. The wavelet transformation converts the time domain signal into the time-frequency domain and analyzes the local characteristics of the signal by adjusting the scale (frequency) and time position. The formula of wavelet transformation is:

[0123]

[0124] Where x(t) is the time series signal, ψ(t) is the wavelet basis function. In this example, the complex Gaussian wavelet cgau8 is used. a is the scale parameter that controls the frequency range of the wavelet, and b is the translation parameter that determines the position of the wavelet on the time axis.

[0125] The amplitude (modulus) of the wavelet coefficient is taken as the energy distribution of the signal, and the calculation formula is:

[0126] Amplitude(a,b)=|W(a,b)|

[0127] Then use the amplitude matrix of the wavelet coefficients |W(a,b)| to draw the time-frequency diagram. The specific steps are as follows:

[0128] Time axis division: construct a time axis t, whose length is equal to the number of sampling points of the signal block: 1024;

[0129] Frequency axis division: According to the selected scale range a, calculate the corresponding frequency range f;

[0130] Plot: Display the magnitude matrix using a 2D pseudo-color plot (e.g., heat map) where:

[0131] The horizontal axis represents time t, the vertical axis represents frequency f, and the color represents the energy amplitude |W(a,b)|.

[0132] The image size is fixed to 224×224, and the time-frequency graph is normalized to the specified size through interpolation. The generated time-frequency graph is saved as an image file for subsequent model training. The time-frequency graph of each sample corresponds to a fault category, and the path naming contains the fault type and sample number.

[0133] (3). According to claim 1, the bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that the specific process of step S21 is: loading the pre-trained ResNet18 model: using the ResNet18 model that has been trained on the ImageNet dataset as the pre-trained model. The model has learned many common features of images during the training process. The ResNet network alleviates the gradient vanishing problem in deep network training by introducing residual connections. The structure of ResNet18 is as follows:

[0134] Y=F(X,{W i})+X

[0135] Among them, X is the input feature map, Y is the output feature map, F represents the residual function, and W i is the weight of each layer. The residual connection allows the input of each layer to be directly passed to the output to avoid information loss. The pre-trained weights of this structure are used to fine-tune the target task (bearing fault diagnosis), and the first three convolutional layers of ResNet18 (Conv1 and the first two residual layers) are frozen, and their parameters are not updated and remain in the initial state. The freezing operation means that the parameters of these layers will not be updated during training, and only the subsequent layers will be adjusted.

[0136] Freezing can be done by setting the gradient updates to zero:

[0137]

[0138] Among them, L is the loss function, W i Represents the weights of the frozen layer. In this case, the parameters of the frozen layer will not be updated according to the gradient descent, which speeds up the training process and reduces the model complexity.

[0139] The weights of Conv1 and the first 3 residual layers are fixed so that they act as shallow feature extractors, thereby reducing training time and avoiding overfitting of these layers. The Conv1 layer and the first 3 residual layers in ResNet18 learn low-level features (such as edges, corners, etc.), which are usually universal and contribute greatly to all images (including the bearing fault dataset), so we fix the weights of these layers. At this stage, by keeping these shallow feature extractors unchanged, the model can focus on learning deep features and improve training efficiency. Mathematically, it can be expressed as:

[0140] X input →Conv1→Residual Block1→Residual Block2→Frozen Parameters

[0141] In this way, by leveraging the pre-trained features of the original ResNet18 model, the model can share this shallow knowledge between different tasks and avoid repeated learning of this part of knowledge.

[0142] Step S22: Construct a residual layer SENet, and learn the interdependence between different feature maps (channels) through squeezing and excitation, so as to enhance the expression of useful features and suppress unimportant features. The fourth residual layer is replaced by SENet. The SENet module mainly includes two operations: Squeeze and Excitation, which are used to model the global information between channels and adjust the response weight of the channel respectively. Specifically, for the input feature map Where H and W are the height and width of the feature map, respectively, and C is the number of channels. Global Average Pooling (GAP) is used to obtain the global information of each channel. The specific steps are:

[0143] 1) Squeeze operation, on the input feature map Global Average Pooling (GAP) is used to obtain the global information of each channel. The calculation formula of GAP is:

[0144]

[0145] Among them, z represents the global average of the cth channel, H and W are the height and width of the feature map respectively, and X C (i,j) represents the pixel value of the cth channel at position (i,j).

[0146] 2) Excitation operation: the global information vector after squeezing Two layers of full connection operations are performed to model the interdependence between channels and generate channel weights. The first layer of full connection: maps z to a lower-dimensional feature space to reduce computational complexity:

[0147] s=σ(W 2 ·δ(W 1 ·z))

[0148] in, and The weight matrix of the fully connected layer, r is the scaling factor, set to 16; δ represents the ReLU activation function, σ represents the Sigmoid activation function, is the generated channel weight.

[0149] 3) Reweighting: The generated channel weight s is applied to the input feature map X through channel-by-channel multiplication to adjust the response strength of each channel:

[0150]

[0151] in, Represents the adjusted c-th channel feature map.

[0152] 4) Replace the fourth residual layer and apply the above-constructed SENet module to the fourth residual layer of ResNet18, so that the model can adaptively strengthen key features and suppress redundant information. The output of the replaced residual block is:

[0153]

[0154] in, is the input feature map after SENet weighting, The mapping result of the convolution layer

[0155] Step S23: Migrate the fine-tuned pre-trained ResNet18 model to the source domain for high-level feature extraction. The specific process is:

[0156] 1) Load the fine-tuned model weights and load the ResNet18 model fine-tuned in step S22 into the target environment. At this point, the first four layers of the model (Conv1, Block1, Block2, Block3) and the SENet-enhanced Block4 have been optimized to adapt to the fault diagnosis task.

[0157] 2) The goal of source domain feature extraction, the data in the source domain Contains sample x i and the corresponding label y i , extract its high-level semantic features f(x i), which is used for subsequent fault mode identification. The mathematical representation of the extracted features is:

[0158] f(xi)=ResNet(xi;Θ)

[0159] Among them, Θ represents the parameter set of the model, including the frozen part and the fine-tuned part

[0160] 3) High-level feature extraction process: first input the data, and then convert the sample data x in the source domain into i Through the fine-tuned ResNet18 model, when the data flows through the lower layers of the model (Conv1 to Block3), the basic features are extracted; when passing through the fourth residual layer (Block4), the SENet module is used to enhance the high-level semantic features. Then the fully connected layer of the model is used to output the feature vector z i , used to represent the sample x i The high-level features of the model’s fully connected layer output feature vector zi is expressed as:

[0161] z i =FC(GlobalAvgPool(Block4(...Block1(Conv1(x i )))))

[0162] Among them, FC represents the fully connected layer. The extracted high-level features Stored as a feature matrix for subsequent classification or transfer learning tasks.

[0163] Step S24: Migrate the high-level features extracted from the source domain to the target domain to complete cross-domain fault diagnosis. The specific process is:

[0164] 1) Preparation of target domain data: Unlabeled or small number of labeled data samples are input into the model. There are certain differences in the distribution of target domain and source domain data, and the distribution of the target domain is adapted through transfer learning methods.

[0165] 2) Feature mapping and alignment: Input the target domain sample xj into the fine-tuned ResNet18 model to extract the corresponding high-level features zj:

[0166] z j =FC(GlobalAvgPool(Block4(...Block1(Conv1(x j )))))

[0167] 3) Training of transfer learning tasks, using features extracted from the source domain and target domain unlabeled data The target domain data is adapted through transfer learning, so that predictions can be made on the test set of the target domain and the classification accuracy and recall rate can be calculated.

[0168] (4). According to claim 1, the bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that the specific process of step S3 is as follows: for the fault diagnosis problem of rolling bearings under different working conditions, the improved TL-ResNet model is adjusted using the test data under different working conditions, and the migration effect of the method between different working conditions is studied through transfer learning, wherein the migration task "0-1" means that the TL-ResNet model generated under working condition 0 (working condition 0 is the source domain) is migrated and applied to the rolling bearing fault diagnosis problem under working condition 1 (working condition 1 is the target domain). Through this migration training method, the model can effectively obtain features from the source domain and adapt to the target domain, further improving the diagnostic performance and robustness under changing working conditions.

[0169] Obviously, the above embodiments are merely examples for clearly explaining the technical solutions of the present invention, and are not intended to limit the implementation methods of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the protection of the claims of the present invention.

Claims

1. A vehicle bearing fault diagnosis method based on wavelet transform and improved TL-ResNet, characterized in that: The following steps are involved: S1. Data preprocessing: preprocess the collected bearing vibration data and the bearing vibration data of the public data set to prepare for subsequent model training. The specific steps include: Step S11: Data set division to obtain operating data under different fault states and working conditions Step S12: Data preprocessing, S2. Model establishment: Design a bearing fault diagnosis method based on wavelet transform and improved TL-ResNet. The specific process is as follows: For the original data, use wavelet transform to process it Step S21: Use the ResNet18 model pre-trained on the ImageNet dataset, freeze some layers, perform parameter migration, and obtain features. Step S22: Construct the residual layer SENet, and learn the interdependence between different feature maps (channels) through squeezing and excitation to enhance the expression of useful features and suppress unimportant features. Step S23: Migrate the pre-trained ResNet18 model after fine-tuning to the source domain for high-level feature extraction. Step S24: Migrate the high-level features extracted from the source domain to the target domain to complete cross-domain fault diagnosis. Step S3, model training: training the algorithm model constructed in step S2, and performing migration training between different working conditions.

2. According to claim 1, a vehicle bearing fault diagnosis method based on wavelet transformation and improved TL-ResNet is characterized in that: The step S11 is for the operation of the shaft bearing under four different fault types: normal state, inner ring fault, outer ring fault and rolling element fault. Three defects of different diameters are introduced for each fault type to represent different fault levels. In addition, the data set also includes the operation data of the bearing under four different working conditions, one of which is a normal working condition, and the other three simulate different environments and load changes. Finally, the data set contains a total of ten sets of data, covering a combination of multiple fault types and working conditions, which is suitable for bearing fault diagnosis and model training.

3. The vehicle bearing fault diagnosis method based on wavelet transform and improved TL-ResNet according to claim 1 is characterized in that: The step S12 comprises: For the original data, wavelet transform is used to process it. The specific process is as follows: The original data is the time series of the bearing vibration signal, recorded as x(t). According to the experimental settings, the time series signal is divided into multiple samples according to the sampling length of 1024 to form signal blocks of equal length. Each signal block is represented as x i (t)={x(t1),x(t2),...,x(t N )}i=1,2,...,M Where M is the number of samples after segmentation. Then the segmented signal is normalized to enhance the feature contrast and reduce the impact of amplitude on wavelet transform. The normalization formula is: Where μ is the signal mean and σ is the signal standard deviation. Then the normalized data is subjected to wavelet transformation. The wavelet transformation converts the time domain signal into the time-frequency domain and analyzes the local characteristics of the signal by adjusting the scale (frequency) and time position. The formula of wavelet transformation is: Where x(t) is the time series signal, ψ(t) is the wavelet basis function. In this example, the complex Gaussian wavelet cgau8 is used. a is the scale parameter that controls the frequency range of the wavelet, and b is the translation parameter that determines the position of the wavelet on the time axis. The amplitude (modulus) of the wavelet coefficient is taken as the energy distribution of the signal, and the calculation formula is: Amplitude(a,b)=|W(a,b)| Then use the amplitude matrix of the wavelet coefficients |W(a,b)| to draw the time-frequency diagram. The specific steps are as follows: Time axis division: construct a time axis t, whose length is equal to the number of sampling points of the signal block: 1024; Frequency axis division: According to the selected scale range a, calculate the corresponding frequency range f; Plot: Display the magnitude matrix using a 2D pseudo-color plot (e.g., heat map) where: The horizontal axis represents time t, the vertical axis represents frequency f, and the color represents the energy amplitude |W(a,b)|. The image size is fixed to 224×224, and the time-frequency graph is normalized to the specified size through interpolation. The generated time-frequency graph is saved as an image file for subsequent model training. The time-frequency graph of each sample corresponds to a fault category, and the path naming contains the fault type and sample number.

4. The bearing fault diagnosis method based on wavelet transform and improved TL-ResNet according to claim 1 is characterized in that The specific process of step S21 is as follows: Load the pre-trained ResNet18 model: Use the ResNet18 model that has been trained on the ImageNet dataset as the pre-trained model. The model has learned many common features of images during the training process. The ResNet network alleviates the gradient vanishing problem in deep network training by introducing residual connections. The structure of ResNet18 is as follows: Y=F(X,{W i })+X Among them, X is the input feature map, Y is the output feature map, F represents the residual function, and W i is the weight of each layer. The residual connection allows the input of each layer to be directly passed to the output to avoid information loss. The pre-trained weights of this structure are used to fine-tune the target task (bearing fault diagnosis), and the first three convolutional layers of ResNet18 (Conv1 and the first two residual layers) are frozen, and their parameters are not updated and remain in the initial state. The freezing operation means that the parameters of these layers will not be updated during training, and only the subsequent layers will be adjusted. Freezing can be done by setting the gradient updates to zero: Among them, L is the loss function, W i Represents the weights of the frozen layer. In this case, the parameters of the frozen layer will not be updated according to the gradient descent, which speeds up the training process and reduces the model complexity. The weights of Conv1 and the first 3 residual layers are fixed so that they act as shallow feature extractors, thereby reducing training time and avoiding overfitting of these layers. The Conv1 layer and the first 3 residual layers in ResNet18 learn low-level features (such as edges, corners, etc.), which are usually universal and contribute greatly to all images (including the bearing fault dataset), so we fix the weights of these layers. At this stage, by keeping these shallow feature extractors unchanged, the model can focus on learning deep features and improve training efficiency. Mathematically, it can be expressed as: X input →Conv1→ResidualBlock1→ResidualBlock2→FrozenParameters In this way, by leveraging the pre-trained features of the original ResNet18 model, the model can share this shallow knowledge between different tasks and avoid repeated learning of this part of knowledge.

5. According to the vehicle bearing fault diagnosis method based on wavelet transform and improved TL-ResNet of claim 1, the specific process of step S22 is: learning the interdependence between different feature maps (channels) through squeezing and excitation, so as to enhance the expression of useful features and suppress unimportant features. The fourth residual layer is replaced by SENet, and the SENet module mainly includes: The two operations of squeeze and excitation are used to model the global information between channels and adjust the response weight of the channels respectively. Specifically, for the input feature map Where H and W are the height and width of the feature map, respectively, and C is the number of channels. Global average pooling (GAP) is used to obtain the global information of each channel. The specific steps are: a) Introducing the Squeeze-and-Excitation (SENet) module based on the residual layer of the ResNet18 model; b) Perform a "squeeze" operation on each feature map, that is, obtain the global features of each channel through global average pooling; c) Perform an "excitation" operation on the features of each channel and learn the dependencies between channels through a fully connected layer and a ReLU activation function; d) According to the channel weights obtained by the excitation, the feature map is weighted and the feature response is readjusted to enhance the expression of important features and suppress the influence of unimportant features.

6. The bearing fault diagnosis method based on wavelet transform and improved TL-ResNet according to claim 1 is characterized in that The step S23 migrates the pre-trained ResNet18 model after fine-tuning to the source domain for high-level feature extraction. The specific process is: migrate the fine-tuned ResNet18 model to the source domain data of the target bearing fault diagnosis task for high-level feature extraction. Using the strategy of transfer learning, the weights of the model are applied to the source domain data to extract high-level feature representations of the input bearing fault data. The core of this process is to provide accurate and meaningful high-level features for the fault diagnosis task in the target domain through feature extraction of the source domain data. These features can be used for further classification or regression tasks, thereby enhancing the performance and adaptability of the model in the target task. Through transfer learning, the knowledge of the source domain data can be effectively transferred to the target domain, thereby shortening the training time of the target task and improving the learning efficiency and accuracy of the model.

7. The bearing fault diagnosis method based on wavelet transform and improved TL-ResNet according to claim 1 is characterized in that The step S24 includes: the target domain data is unlabeled or a small number of labeled samples, and the data distribution difference between the target domain and the source domain is adapted through transfer learning. The target domain samples are input into the fine-tuned ResNet18 model to extract the corresponding high-level features and realize feature mapping and alignment. Transfer learning training is performed using source domain features and target domain unlabeled data, prediction is performed on the target domain test set, and the classification accuracy and recall rate are calculated to complete the diagnosis task.

8. The bearing fault diagnosis method based on wavelet transform and improved TL-ResNet according to claim 1 is characterized in that The step S3 includes: for the fault diagnosis problem of rolling bearings under different working conditions, the improved TL-ResNet model is adjusted using the test data under different working conditions, and the migration effect of the algorithm between different working conditions is studied through transfer learning, wherein the migration task "0-1" means migrating the TL-ResNet model generated under working condition 0 (working condition 0 is the source domain) and applying it to the rolling bearing fault diagnosis problem under working condition 1 (working condition 1 is the target domain). Through this migration training method, the model can effectively acquire features from the source domain and adapt to the target domain, further improving the diagnostic performance and robustness under changing working conditions.

Images