Bearing fault diagnosis method and system across operating condition unbalanced data

Abstract

The bearing fault diagnosis method and system across working condition unbalance data provided by the application abandon the existing structure based on a convolutional neural network, select a Swin Transformer model based on an attention mechanism and improve the same, introduce CBAM and IECA attention modules in the improved deep convolutional generative adversarial network, propose a bearing diagnosis method based on the attention mechanism and used for small samples for deep parameter transfer learning of the same. Based on transfer learning and the improved deep convolutional generative adversarial network, the small sample parameters of the target domain are expanded by improving the deep convolutional generative adversarial network, the generated simulation signals retain the complete high-frequency and low-frequency characteristics of the real signals; the features of the multiple source domains and the target domain are mapped to the same feature space, the multi-domain feature extraction is realized and the feature alignment operation is performed; the unknown label rolling bearing health state under the cross working condition is identified by improving the Swin Transformer network.

Description

Technical Field

[0001] This application proposes a novel bearing fault diagnosis method and system. The method and system are based on transfer learning and improved deep convolutional generative adversarial networks for imbalanced data across operating conditions, and belong to the field of deep learning and fault diagnosis technology. Background Technology

[0002] Bearings are an essential and core structure of transmission devices. Common failures caused by bearing malfunctions under load include fatigue failure, wear, and fracture, with these failures often concentrated between the inner and outer rings. Bearings are more prone to failure under conditions of excessive load over extended periods, poor lubrication, and improper installation. Rolling bearings have poor impact resistance and are more susceptible to failure due to factors such as excessive speed and corrosion. As a crucial and easily damaged component of rotating parts, effective monitoring and diagnosis of bearing conditions are of paramount importance.

[0003] Currently, deep learning-based bearing diagnostic methods face small sample problems such as insufficient training data and missing sample labels. Furthermore, the distribution of bearing condition data varies under different operating conditions and models, making it impossible to use the same network for classification. To address these issues, research into transfer learning is needed. Specifically, the theoretical basis of transfer learning is based on discriminative and adversarial methods. Discriminative learning uses a mapping function to map source and target domain data to a shared space, reducing the distributional differences between the two domains after mapping. Adversarial learning borrows from the idea of ​​Generative Adversarial Networks (GANs), where a generator produces features, and a domain discriminator determines whether the proposed features belong to the source or target domain. The generator and discriminator jointly enhance each other, enabling the features generated by the generator to deceive the discriminator, and the discriminator to accurately determine whether the features belong to the source or target domain.

[0004] Existing bearing diagnostic methods typically employ data augmentation and improved network structures. For example, auxiliary classification generative adversarial networks are used to generate a large number of samples with the same probability distribution as the small amount of data; furthermore, order tracking and resampling are used to process bearing data at different speeds, and adaptive batch normalization networks are used for fault classification across operating conditions; still others are designed to use multi-scale compact convolutional neural networks to extract multi-scale signal features with fewer network layers.

[0005] The existing technologies mentioned above are limited to feature extraction using invariant scale convolution kernels, and still cannot solve the problems of small sample size and differences in data distribution under different working conditions and models. The performance and generalization ability of bearing diagnosis methods are still low.

[0006] In view of the above, this patent application is hereby filed. Summary of the Invention

[0007] The bearing fault diagnosis method and system proposed in this application for unbalanced data under different operating conditions aims to solve the problems existing in the prior art by abandoning the existing structure based on convolutional neural networks. Instead, it selects and improves the SwinTransformer model based on the attention mechanism, introduces CBAM and IECA attention modules into the improved deep convolutional generative adversarial network, and proposes a bearing diagnosis method based on the attention mechanism for small samples through deep parameter transfer learning. This aims to cope with small sample problems and the differences in data distribution under different operating conditions and models, thereby effectively improving the performance and generalization ability of the bearing diagnosis method through data expansion, network structure improvement and parameter transfer.

[0008] To achieve the above design objectives, the bearing fault diagnosis method proposed in this application for unbalanced data under different operating conditions is based on transfer learning and an improved deep convolutional generative adversarial network. By improving the deep convolutional generative adversarial network, the small sample parameters of the target domain are expanded, and the generated simulated signal retains the complete high-frequency and low-frequency features of the real signal. The features of the multi-source domain and the target domain are mapped to the same feature space to achieve multi-domain feature extraction and feature alignment. The health status of unknown labeled rolling bearings under different operating conditions is identified by improving the SwingTransformer network.

[0009] This includes the following implementation steps:

[0010] Step 1: Obtain the original multi-source vibration signals under multiple fault conditions by sampling from multiple vibration sensors under various working conditions, save them by channel and name them t0;

[0011] Step 2: Perform data processing on the multi-source vibration signal, including filtering, denoising, data alignment, and signal correction, and name the result t1.

[0012] Step 3: For the time series data t1, extract time segments sequentially, apply Fourier transform to them, save the transformed complex number result, and name it F1;

[0013] Step 4: For time series data t1, perform wavelet transform and reconstruction at different scales using complex wavelet functions; select the optimal scale by comparing errors, and calculate the time-frequency features under the current preferred and unique scale, and save them in the form of a coefficient matrix in a binary file, naming it C1;

[0014] Step 5: Read the above F1 and C1 files and perform feature concatenation to generate a data matrix. Map the matrix values ​​to generate a grayscale image dataset I1 of wavelet transform. Set the grayscale image dataset I1 as the source domain dataset O1 and divide the source domain dataset O1 into a training set and a test set.

[0015] Step 6: Change the initial parameter settings of the vertical motor and proceed to step 2. Then, construct the time series dataset mu-t2. Perform steps 2 to 5 on the above dataset in sequence and finally obtain the grayscale dataset I2.

[0016] The grayscale dataset I2 is used as the target domain D1 to divide into training and test sets, which are named D1_train and D1_test, respectively.

[0017] Step 7: Construct an improved DCGAN network architecture;

[0018] Step 8: Build an improved Swing Transformer network architecture;

[0019] Step 9: The IDCGAN network is pre-trained and new synthetic samples are generated;

[0020] Step 10: Train and build an improved Swing Transformer network model for fault diagnosis process.

[0021] Furthermore, in step 1, different fault types are classified as data samples belonging to different faults, and the original data of different working conditions under the same fault type are grouped to construct the original time-series information acquisition sample t0.

[0022] Furthermore, in step 2, a filter is used to remove unwanted frequency components or noise, and the filtered vibration signal is subjected to mean denoising processing. Correlation analysis and other methods are used to calculate the time offset of the signal. If there is a time deviation or phase difference between the multi-source vibration signals, data alignment is required to ensure that the signals are consistent in time. The preprocessed vibration signal data is stored in the form of time-series data, and the processed data is saved as a .mat file named t1.

[0023] Furthermore, in step 3, the Fourier transform formula for the non-periodic continuous-time signal x(t) is as follows:

[0024]

[0025] We obtain discrete sampled values ​​x(nT) of a continuous signal x(t), and use the discrete signal x(nT) to calculate the following spectrum of the signal x(t).

[0026]

[0027] The complex results of the amplitude spectrum and phase spectrum after Fourier transform are saved as a .csv text file and named F1.

[0028] Furthermore, in step 4, the formula for the Complex Morlet wavelet function is as follows:

[0029]

[0030] In the above formula, f(t) is the time-domain signal, ψ(x) is the wavelet basis function, a is the positioning frequency, and b is the positioning time;

[0031] Choose the Complex Morlet wavelet basis function with the following expression.

[0032]

[0033] In the above formula, a is the scaling factor and b is the time shift factor;

[0034] The wavelet transform of the wavelet basis function is applied to the heterogeneous signal at different scales of 128, 64...8, and the reconstruction error is calculated.

[0035] A lower reconstruction error indicates that the wavelet basis function at the current scale better characterizes the original signal.

[0036] The time-frequency characteristics, including the frequency band coefficient, energy spectrum, and amplitude spectrum, are calculated under the selected unique preferred scale and stored in a binary file in the form of a coefficient matrix, which is named C1.

[0037] Further, in step 5, features belonging to the same time domain information are concatenated row by row, and average segment interpolation sampling is used for longer array parameters to align the feature parameters; the matrix values ​​are mapped to the gray value range, converted to three channels and mapped to generate a wavelet transform grayscale image dataset I1, and the grayscale image dataset I1 is set as the source domain dataset O1; it is divided into a training set and a test set in a 4:1 ratio, and named O1_train and O1_test respectively.

[0038] Furthermore, step 6 includes the following steps:

[0039] S601. Cross-condition operation maintains constant ambient temperature and humidity in the working environment, sets bearing load; sets three different bearing speed setting methods; simulates the actual operating conditions of a vertical motor with irregular speed changes; presets a constant speed state; and sets multiple bearing states under the same speed operating conditions.

[0040] S602. Mix the multi-source vibration signals after changing the rotational speed conditions with random data of equal length average, and construct a dataset of different rotational speeds for the same damage under multiple channels. Perform data processing such as filtering, denoising, data alignment, and signal correction on the above signals. Store the preprocessed data in the form of time-series data (.mat) and name it "mu-t2".

[0041] S603. Perform feature concatenation on the dataset "mu-t2" by concatenating the spectral features and time-frequency features under the same time domain information row by row; if the feature parameter lengths are inconsistent, use average segment interpolation sampling to interpolate the longer array parameters to achieve data alignment;

[0042] S604. Adjust the concatenated feature vectors into a two-dimensional matrix, specifying the number of rows and columns as needed. Map the matrix values ​​to the grayscale range and convert it into a three-channel grayscale image dataset "I2";

[0043] S605. Divide the dataset "I2" into a training set and a test set in a 4:1 ratio, and name them the target domain datasets "D1_train" and "D1_test" respectively.

[0044] Furthermore, in step 7, the generator network structure includes an input layer, a multi-transposed convolutional layer, a multi-head self-attention module, an SN spectrum normalization layer, a ReLU activation function layer, an IECA module, and an output layer.

[0045] The discriminator network structure includes an input layer, SN spectrum normalization, convolutional layers, multi-head self-attention module, BN normalization layer, LeakyReLU activation function layer, MK-MMD loss function, IECA module, and output layer;

[0046] The built-in parameter setting process for the aforementioned IDCGAN network architecture includes the following steps:

[0047] The discriminator network structure is shown in the following expression.

[0048] L1_Discriminator=Sigmoid(ConvBlock(MHA(ConvBlockn(IECA(leakyRELU(Conv2d(x))))))

[0049] Where x is the real fusion dataset N1 (n≤1) and the synthesized new feature map (n>1); Conv2d(*) is a two-dimensional convolution operation; leakyRELU activation layer; IECA channel attention mechanism layer, which introduces global max pooling weighting on the basis of global mean pooling; ConvBlock block includes (in order, Conv2d(), SN(*) spectrum normalization, leakyRELU(*) activation function); n is n stacked ConvBlock blocks; MHA is a multi-head self-attention module, which is the intrinsic attention weight allocation before calculating the features of multi-channel vibration information; Sigmoid(*) discriminator judges whether the input parameter conforms to the distribution metric of MK-MMD. If it does, Sigmoid(*) maps true(1), otherwise false(0). When true(1), the auxiliary dataset A1 is saved.

[0050] The generator network structure is shown in the following expression.

[0051] L1_Generator=Concat(T-ConvBlock(MHA(T-ConvBlockn(IECA(RELU(T-Conv2d(x))))))

[0052] Where x is the spatial random variable N1 (n≤1) and the newly generated feature parameters after network parameter tuning (n>1); T-Conv2d(*) is a two-dimensional transposed convolution operation; ReLU activation layer; IECA channel attention mechanism layer; T-ConvBlock block includes (in order: T-Conv2d(), SN(*) spectral normalization, ReLU(*) activation function); n is n stacked T-ConvBlock blocks; MHA is a multi-head self-attention module, which is the intrinsic attention weight allocation before calculating the features of multi-channel vibration information; Concat(*) synthesizes a new feature map after updating its self-parameter settings after feedback from the discriminator.

[0053] Furthermore, step 8, the improved Swing Transformer network architecture includes an input layer, an IECA channel attention layer, a Patch Embedding image block embedding layer, a position encoding layer, a Transformer encoder layer, a downsampling layer between Transformer encoder modules, a cross-layer window-based local attention mechanism, a global weighted pooling layer, a fully connected layer, and a Softmax output layer.

[0054] The improved Swing Transformer network model expression is as follows:

[0055] L2_Swin_T=softmax(MLP(SwinBlock2(Patch Merging

[0056] (SwinBlock1(IECA(PE((PatchEmbed(x))Ni)))))Nj)

[0057] Wherein, during pre-training, x is the fusion dataset N1+A1 (n=1); during fine-tuning after the first transfer learning, x is a small amount of auxiliary dataset A2 (n=2); after the second transfer learning and parameter tuning, x is the remaining auxiliary dataset A2 (n>2); PatchEmbed(·) is a function that segments the input image (x) into image patches and converts them into embedding vectors through linear projection; PE(·) is a position encoding function that adds position encoding to the embedded data to preserve the position information of the sequence; Ni is the number of channels for multi-channel data processing; IECA channel attention mechanism layer; SwinBlock1 block (in order: LN(*) layer normalization, W-MSA(*) window-type local multi-head self-attention mechanism, CBAM(*) channel spatial attention mechanism, weighted node, LN(*) layer normalization, FeedForward(*) feedforward network layer, weighted node); Patch Merging(*) consists of a downsampling layer, a cross-layer windowed local attention mechanism, and a global weighted pooling layer (weighted global max and average pooling); SwinBlock2 blocks (in sequence: LN(*) layer normalization, SW-MSA(*) windowed multi-head self-attention mechanism, CBAM(*) channel spatial attention mechanism, weighted nodes, LN(*) layer normalization, FeedForward(*) feedforward network layer, weighted nodes); Nj is a stack of N pairs of SwinBlock blocks and the Patch Merging(*) layer between them; MLP(·) is a multilayer perceptron function consisting of a fully connected layer and an activation function to map the transformed representation to the output layer of the classification task; the softmax(*) function is used to activate the output layer and obtain the predicted result of the classification label.

[0058] Based on the design of the bearing fault diagnosis method based on the above cross-operating condition unbalanced data, this application also proposes the following bearing fault diagnosis system based on cross-operating condition unbalanced data, including multiple lower-level machines connected in parallel and corresponding to the upper-level machine, or a single lower-level machine connected to the upper-level machine, as well as a processor and a memory.

[0059] The lower-level machine uses a multi-channel data acquisition instrument, which is connected to a dual-head BNC terminal signal input line and various adapters to acquire signals of various data types. Various sensors at the end are fixedly installed in the vertical bearing part.

[0060] The host computer consists of a server-side program and a monitoring interface that integrates the bearing fault diagnosis method for cross-operating condition imbalance data, including a speed test module, a core processor module, a data acquisition module, and an interactive communication module;

[0061] The memory includes a computer-readable storage medium for storing a computer-executable program for a bearing diagnostic method based on cross-condition multimodal data fusion;

[0062] The processor is used to read the computer executable program from memory and execute it.

[0063] In summary, the bearing fault diagnosis method and system based on cross-condition imbalance data has the following advantages:

[0064] 1. The bearing fault diagnosis method proposed in this application, which uses unbalanced data across operating conditions, is a multi-channel signal fusion method. It can effectively reduce the impact of unbalanced data and improve diagnostic performance, thereby improving the accuracy and efficiency of fault diagnosis and thus improving the reliability and maintenance effect of the equipment.

[0065] 2. This application performs FFT and time-frequency conversion on synchronous vibration signals measured by multiple vibration sensors, and then performs pooling and splicing operations to fuse features and map them into a time-frequency image to characterize richer fault features.

[0066] 3. This application proposes a data augmentation method based on IDCGAN to solve the imbalanced data problem. It introduces IECA and CBAM attention mechanisms into DCGAN to improve the weighting of generated samples, thereby improving the quality of generated samples. It also applies spectral normalization (SN), MK-MMD metric and gradient penalty to stabilize the training process of DCGAN.

[0067] 4. This application proposes an improved Swin Transformer fault diagnosis classification network based on an attention mechanism, which significantly improves diagnostic accuracy. IECA and CBAM attention modules are introduced into the pre-trained residual network, enabling better capture of important spatial and channel features. Attached Figure Description

[0068] Figure 1 This is a flowchart of the bearing fault diagnosis method based on cross-operating condition unbalance data proposed in this application;

[0069] Figure 2 This is a schematic diagram illustrating the construction of a bearing unbalanced dataset across operating conditions;

[0070] Figure 3 This is a schematic diagram of data augmentation processing for bearing unbalanced datasets across operating conditions;

[0071] Figure 4 This is a schematic diagram of the built-in data feature parameter settings of the lDCGAN network;

[0072] Figure 5 This is a schematic diagram of the built-in network layer of the lDCGAN network;

[0073] Figure 6 This is a schematic diagram of the improved Swing Transformer network architecture;

[0074] Figure 7 This is a schematic diagram illustrating the improved settings for built-in data feature parameters in the Swing Transformer network.

[0075] Figure 8 This is a schematic diagram of the modules of the bearing local defect information acquisition system. Detailed Implementation

[0076] The technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described solutions are some embodiments of this application, but not all embodiments. Based on the embodiments proposed in this application, those skilled in the art can make improvements or modifications without creative effort, and all such improvements and modifications should fall within the protection scope of the appended claims.

[0077] Example 1, such as Figure 1 As shown, this application proposes a bearing fault diagnosis method based on unbalanced data across operating conditions. This method is based on transfer learning and improved deep convolutional generative adversarial networks, thus enabling it to be applied to small sample problems such as insufficient training data or missing sample labels for vertical motor bearing defects.

[0078] The bearing fault diagnosis method based on unbalanced data across operating conditions improves the small-sample parameters of the target domain by using a deep convolutional generative adversarial network to simulate and expand the target domain. The generated simulated signal retains the complete high-frequency and low-frequency features of the real signal. It maps features from multiple source domains and the target domain to the same feature space, enabling multi-domain feature extraction and feature alignment. Finally, it uses an improved Swing Transformer network to identify the health status of rolling bearings with unknown labels under different operating conditions. The method includes the following implementation steps:

[0079] Step 1: Obtain the original multi-source vibration signals under multiple fault conditions by sampling from multiple vibration sensors under various working conditions, save them by channel and name them t0;

[0080] Specifically, different fault types are categorized into different fault types for data samples, and raw data from different operating conditions under the same fault type are grouped to construct raw time-series information acquisition samples t0; refer to Figure 2As shown, the original multi-fault state multi-source vibration signal includes three different bearing speed setting methods: stable constant speed, uniformly variable speed or periodically changing speed, and irregular speed change to simulate the actual operating conditions of a vertical motor, including but not limited to sudden acceleration, deceleration, and emergency turning to control the output direction and torque of the motor. This application takes constant speed state data as an example, reasonably adjusting the speed settings of three gradient speed differences within the speed range, such as 1000 r / min, 1500 r / min, and 2000 r / min. Under the same speed operating conditions, five bearing states are set: inner ring damage, outer ring damage, cage damage, roller damage, and normal. In a fully connected form, a total of 15 basic datasets constitute three different speed states of the same fault type, with a total of 5 dataset categories.

[0081] Step 2: Perform data processing on the multi-source vibration signal, including filtering, denoising, data alignment, and signal correction, and name the result t1.

[0082] Specifically, filters are used to remove unwanted frequency components or noise. The filtered vibration signal undergoes mean denoising. Correlation analysis and other methods are used to calculate the signal's time offset. If time deviations or phase differences exist between multiple vibration signals, data alignment is required to ensure that all signals are consistent in time. The preprocessed vibration signal data is stored as time-series data, and the processed data is saved as a .mat file named t1.

[0083] Step 3: For the time series data t1, extract time segments sequentially, apply Fourier transform to them, save the transformed complex number result, and name it F1;

[0084] Specifically, the formula for the Fourier transform of a non-periodic continuous-time signal x(t) is as follows:

[0085]

[0086] We obtain discrete sampled values ​​x(nT) of a continuous signal x(t), and use the discrete signal x(nT) to calculate the following spectrum of the signal x(t).

[0087]

[0088] The complex results of its Fourier transform amplitude spectrum and phase spectrum are saved as a .csv text file and named F1;

[0089] Step 4: For time series data t1, perform wavelet transform and reconstruction at different scales using complex wavelet functions; select the optimal scale by comparing errors, and calculate the time-frequency features under the current preferred and unique scale, and save them in the form of a coefficient matrix in a binary file, naming it C1;

[0090] Specifically, the formula for the Complex Morlet wavelet function is as follows:

[0091]

[0092] In the above formula, f(t) is the time-domain signal, ψ(x) is the wavelet basis function, a is the positioning frequency, and b is the positioning time;

[0093] Choose the Complex Morlet wavelet basis function with the following expression.

[0094]

[0095] In the above formula, a is the scaling factor and b is the time shift factor;

[0096] The wavelet transform of the wavelet basis function is applied to the heterogeneous signal at different scales of 128, 64...8, and the reconstruction error is calculated.

[0097] A lower reconstruction error indicates that the wavelet basis function at the current scale better characterizes the original signal.

[0098] Calculate the time-frequency characteristics, including frequency band coefficients, energy spectrum, amplitude spectrum, etc., under the selected unique preferred scale, and save them in the form of a coefficient matrix in a binary file, named C1.

[0099] Step 5, refer to Figure 3 As shown, read the above F1 and C1 files and perform feature concatenation to generate a data matrix. Map the matrix values ​​to generate a grayscale image dataset I1 of wavelet transform. Set the grayscale image dataset I1 as the source domain dataset O1 and divide the source domain dataset O1 into a training set and a test set.

[0100] Specifically, features belonging to the same time domain information are concatenated row by row, and average segment interpolation sampling is used for longer array parameters to align the feature parameters; matrix values ​​are mapped to grayscale range, converted to three channels and mapped to generate wavelet transform grayscale image dataset I1, and grayscale image dataset I1 is set as source domain dataset O1; it is divided into training set and test set in a 4:1 ratio, and named O1_train and O1_test respectively.

[0101] The above process includes the following steps:

[0102] S501. Read the spectral features and time-frequency features "F1" and "C1" contained in the files of steps 3 and 4, and perform feature splicing operation to splice the spectral features and time-frequency features under the same time domain information line by line.

[0103] S502. Since the parameter lengths of the two arrays may be inconsistent, a segmented interpolation sampling method is used to interpolate the longer array parameters to achieve data alignment. This results in a new, wider feature vector.

[0104] S503. If the dimensionality of the feature vector is high, the number of rows and columns of the matrix can be specified as needed to readjust the feature vector into a two-dimensional matrix. For ease of visualization and processing, the matrix values ​​are mapped to the grayscale range and converted into a three-channel grayscale image dataset "I1", which serves as the source domain dataset "O1".

[0105] S504. Further, the dataset "O1" is divided into a training set and a test set in a 4:1 ratio, and named the source domain datasets "O1_train" and "O1_test" respectively;

[0106] Step 6, refer to Figure 3 As shown, the initial parameter settings of the vertical motor are changed and step 2 is performed. Then, the time series dataset mu-t2 is constructed. The above dataset is then processed through steps 2 to 5 in sequence to finally obtain the grayscale dataset I2.

[0107] Divide the grayscale dataset I2 into a training set and a test set (e.g., in a 4:1 ratio) as the target domain D1, and name them D1_train and D1_test respectively.

[0108] Specifically, the above process includes the following steps:

[0109] S601. Cross-condition operation maintains constant ambient temperature and humidity in the working environment and sets bearing loads: the bearing bears different initial conditions in terms of axial and radial load magnitude and direction; sets three different bearing speed setting methods, mainly stable constant speed, uniform speed change or periodic speed change, and irregular speed change to simulate the actual operating conditions of a vertical motor, including but not limited to sudden acceleration, deceleration, and emergency turning to control the output direction and torque of the motor; preset constant speed state, and reasonably adjust the speed settings of three gradient speed differences within the speed range, such as 1300r / min, 1600r / min, and 1900r / min, under the same speed operating conditions, and set five bearing states: inner ring damage, outer ring damage, cage damage, roller damage, and normal;

[0110] S602. Mix the multi-source vibration signals after changing the rotational speed conditions, such as vibration datasets of three different rotational speed conditions for inner ring damage, with random data of equal length average, and construct different rotational speed datasets of the same damage under multiple channels. Perform data processing such as filtering, denoising, data alignment, and signal correction on the above signals. Store the preprocessed data in the form of time-series data (.mat) and name it "mu-t2".

[0111] S603. Perform feature concatenation on the dataset "mu-t2", concatenating the spectral features and time-frequency features under the same time domain information row by row. If the feature parameter lengths are inconsistent, use average segment interpolation sampling to interpolate the longer array parameters to achieve data alignment;

[0112] S604. Adjust the concatenated feature vectors into a two-dimensional matrix, specifying the number of rows and columns as needed. Map the matrix values ​​to the grayscale range and convert them into a three-channel grayscale image dataset "I2".

[0113] S605. Divide the dataset "I2" into a training set and a test set in a 4:1 ratio, and name them the target domain datasets "D1_train" and "D1_test" respectively;

[0114] Step 7, refer to Figure 5 As shown, an improved DCGAN network architecture is constructed;

[0115] Specifically, the network layer is configured as follows:

[0116] The generator network structure includes an input layer, a multi-transposed convolutional layer, a multi-head self-attention module, an SN spectrum normalization layer, a ReLU activation function layer, an IECA module, and an output layer.

[0117] The discriminator network structure is the same, including an input layer, SN spectrum normalization, convolutional layer, multi-head self-attention module, BN normalization layer, LeakyReLU activation function layer, MK-MMD loss function, IECA module, and output layer;

[0118] Reference Figure 4 The built-in parameter setting process for the aforementioned IDCGAN network architecture includes the following steps:

[0119] The discriminator network structure is shown in the following expression.

[0120] L1_Discriminator=Sigmoid(ConvBlock(MHA(ConvBlockn(IECA(leakyRELU(Conv2d(x))))))

[0121] Where x is the real fusion dataset N1 (n≤1) and the synthesized new feature map (n>1); Conv2d(*) is a two-dimensional convolution operation; leakyRELU activation layer; IECA channel attention mechanism layer, which introduces global max pooling weighting on the basis of global mean pooling; ConvBlock block includes (in order, Conv2d(), SN(*) spectrum normalization, leakyRELU(*) activation function); n is n stacked ConvBlock blocks; MHA is a multi-head self-attention module, which is the intrinsic attention weight allocation before calculating the features of multi-channel vibration information; Sigmoid(*) discriminator judges whether the input parameter conforms to the distribution metric of MK-MMD. If it does, Sigmoid(*) maps true(1), otherwise false(0). When true(1), the auxiliary dataset A1 is saved.

[0122] The generator network structure is shown in the following expression.

[0123] L1_Generator=Concat(T-ConvBlock(MHA(T-ConvBlockn(IECA(RELU(T-Conv2d(x))))))

[0124] Where x is the spatial random variable N1 (n≤1) and the newly generated feature parameters after network parameter tuning (n>1); T-Conv2d(*) is a two-dimensional transposed convolution operation; ReLU activation layer; IECA channel attention mechanism layer; T-ConvBlock block includes (in order: T-Conv2d(), SN(*) spectral normalization, ReLU(*) activation function); n is n stacked T-ConvBlock blocks; MHA is a multi-head self-attention module, which is the intrinsic attention weight allocation before calculating the features of multi-channel vibration information; Concat(*) synthesizes a new feature map after updating its self-parameter settings after feedback from the discriminator;

[0125] Step 8, refer to Figure 6 As shown, an improved Swing Transformer network architecture is constructed;

[0126] Specifically, such as Figure 7As shown, the improved Swin Transformer network architecture includes an input layer, an IECA channel attention layer, a Patch Embedding image block embedding layer, a positional encoding layer, a Transformer encoder layer (a. multi-head self-attention mechanism, b. CBAM module, c. position-based feedforward network, d. residual connection and layer normalization), a downsampling layer between Transformer encoder modules, a cross-layer windowed local attention mechanism, a global weighted pooling layer (weighted global max and average pooling), a fully connected layer, and a Softmax output layer.

[0127] The improved Swing Transformer network model expression is as follows:

[0128] L2_Swin_T=softmax(MLP(SwinBlock2(Patch Merging

[0129] (SwinBlock1(IECA(PE((PatchEmbed(x))Ni)))))Nj)

[0130] Wherein, during pre-training, x is the fusion dataset N1+A1 (n=1); during fine-tuning after the first transfer learning, x is a small amount of auxiliary dataset A2 (n=2); after the second transfer learning and parameter tuning, x is the remaining auxiliary dataset A2 (n>2); PatchEmbed(·) is a function that segments the input image (x) into image patches and converts them into embedding vectors through linear projection; PE(·) is a position encoding function that adds position encoding to the embedded data to preserve the position information of the sequence; Ni is the number of channels for multi-channel data processing; IECA channel attention mechanism layer; SwinBlock1 block (in order: LN(*) layer normalization, W-MSA(*) window-type local multi-head self-attention mechanism, CBAM(*) channel spatial attention mechanism, weighted node, LN(*) layer normalization, FeedForward(*) feedforward network layer, weighted node); Patch Merging(*) consists of a downsampling layer, a cross-layer windowed local attention mechanism, and a global weighted pooling layer (weighted global max and average pooling); SwinBlock2 blocks (in sequence: LN(*) layer normalization, SW-MSA(*) windowed multi-head self-attention mechanism, CBAM(*) channel spatial attention mechanism, weighted nodes, LN(*) layer normalization, FeedForward(*) feedforward network layer, weighted nodes); Nj is a stack of N pairs of SwinBlock blocks and the Patch Merging(*) layer between them; MLP(·) is a multilayer perceptron function composed of fully connected layers and activation functions to map the transformed representation to the output layer of the classification task; the softmax(*) function is used to activate the output layer and obtain the predicted result of the classification label;

[0131] Step 9, refer to Figure 5 As shown, the IDCGAN network is pre-trained and generates new synthetic samples;

[0132] Specifically, the IDCGAN generator constructed in step 7 is used for pre-training. Data alignment is performed using the source domain O1_train dataset and the target domain dataset D1_train, and feature fusion is performed. The fused dataset N1 is used to train the IDCGAN model to capture the distribution of real samples of different categories. After training, the model can generate and distinguish images in the source domain dataset. The source domain knowledge is transferred to the target domain using methods such as domain adaptation to generate new synthetic samples, which are named auxiliary dataset A1. Feature parameters that meet the MK-MMD metric are extracted and interpolated and aligned with the target domain D1_test, and feature fusion is performed. This is named auxiliary dataset A2 to enhance the imperfect dataset.

[0133] The above-mentioned pre-training and new synthetic sample generation process includes the following steps:

[0134] S901. Discriminator internal data flow structure:

[0135] The real fusion dataset N1 is used as input and enters the Conv2D convolutional layer. After passing through the LeakyRELU activation layer and the IECA channel attention mechanism, the channel weights are assigned. After passing through multiple ConvBlock blocks (including Conv2D convolutional layers, SN2d spectral normalization, and LeakyRELU activation layers), after passing through the second to last ConvBlock block, the channel attention weights are calculated by the multi-head self-attention module. Then, the final ConvBlock block is passed. After the MK-MMD loss function is calculated, the sigmoid mapping is used to determine whether the output is true (1) or false (0). If the output is false (0), the parameters of the generator and the discriminator are updated. The gradient penalty supervision mechanism is introduced. Then, the backpropagation and optimization algorithm (such as Adam) are used to minimize the adversarial loss. The training is carried out until the discriminator considers the generator to generate data that conforms to the feature distribution of the real fusion dataset N1 as true (1). Then, the specified batch that meets the output requirements is saved and named the auxiliary dataset A1.

[0136] S902. Generator Internal Data Flow Structure:

[0137] The random variable enters the Conv2D transposed convolutional layer, passes through the ReLU activation layer, and undergoes channel weighting through the IECA channel attention mechanism. After passing through multiple Transpose ConvBlock blocks (including Conv2D transposed convolutional layers, SN2d spectral normalization, and ReLU activation layers), and after the penultimate Transpose ConvBlock block, it undergoes channel attention weighting calculation through a multi-head self-attention module, and then enters the final Transpose ConvBlock block. After concat feature concatenation, a composite grayscale image is generated through mapping. If the preset conditions are not met, the parameters of the generator and discriminator are further adjusted and updated. The generated data that does not meet the requirements continues to be adjusted and updated through the discriminator network until the generator can generate feature distributions that conform to the real fusion dataset N1. The specified batch that meets the output requirements is saved and named auxiliary dataset A1.

[0138] S903.IDCGAN overall training process:

[0139] S903.1 pre-training:

[0140] - Iterate through the following steps:

[0141] - Randomly sample a batch of real grayscale mapping sample data from the fusion dataset N1.

[0142] - The generator generates a batch of synthetic samples. The real mapping samples are interpolated and sampled and then input into the generator network. After a series of network layer transformations, synthetic samples are generated.

[0143] The discriminator classifies real and synthetic samples and calculates adversarial loss. For real samples, the discriminator aims to classify them as real samples (label 1), and for synthetic samples, the discriminator aims to classify them as synthetic samples (label 0).

[0144] - Update the parameters of the generator and discriminator, minimizing the adversarial loss through backpropagation and optimization algorithms (such as Adam). The generator aims to deceive the discriminator, making it unable to accurately distinguish between generated synthetic samples and real samples, while the discriminator aims to accurately distinguish between generated synthetic samples and real samples.

[0145] - Following the steps above, if the distribution difference of the preset MK-MMD loss function is not met, the hyperparameters are adjusted to reduce the neighborhood difference between the source and target domains until the predetermined number of training iterations or loss convergence is reached.

[0146] - After all training steps have been iterated, the IDCGAN generator will learn to generate synthetic samples similar to the fused dataset N1. S903.2. Generating New Samples:

[0147] - Obtain a batch of datasets that conform to the N1 parameter of the fused dataset from the generator as input.

[0148] - The above is converted into synthetic samples through a generator network.

[0149] - The features of the generated samples can be controlled by adjusting the generator's parameters.

[0150] - Repeat the above steps to generate the required number of new sample datasets as auxiliary dataset A1.

[0151] S903.3. Enhancement of source and target domains:

[0152] During pre-training, either the generator or discriminator network of IDCGAN can be used as a feature extractor. The feature representations of the intermediate layers are extracted by inputting the fused dataset N1 into the generator or discriminator network. The extracted feature representations are then aligned and interpolated with the mapping feature map of the target domain D1_test, and finally fused to form the auxiliary dataset A2.

[0153] Step 10, refer to Figure 6As shown, an improved Swing Transformer network model is trained and constructed for fault diagnosis process;

[0154] Specifically, the improved Swing Transformer network model constructed in step 8 is used for pre-training. The fusion dataset N1 and the auxiliary dataset A1 are fused, and then the improved Swing Transformer network is pre-trained. After fifty iterations or meeting the preset classification accuracy, the pre-trained weights, model parameters, and other data are transferred once, and a small amount of data from the auxiliary dataset A2 is used to fine-tune the model. The auxiliary dataset A2 is then used to perform a fault diagnosis process on the fine-tuned improved Swing Transformer model. The hyperparameters are adjusted until the preset accuracy, recall, and other indicators are met, resulting in the trained fault diagnosis model. Finally, the model accuracy is tested using data from the target domain D1_test (unlabeled), and the model generalization ability is tested using data from the source domain O1_test (unlabeled).

[0155] The above process includes the following steps:

[0156] S1001. First, merge the fusion dataset "N1" and the auxiliary dataset "A1". Then, pre-train the improved Swin Transformer network using the merged dataset. The pre-training process can be iterated fifty times or until a preset classification accuracy condition is met.

[0157] S1002. After completing pre-training, the weights, model parameters, and other data obtained from pre-training are used for a data transfer exercise. Calculate metrics such as classification accuracy, precision, recall, and F1 score to evaluate the model's performance. This allows knowledge from the source domain to be transferred to the target domain, enabling better adaptation to the characteristics of the target domain.

[0158] S1003. Next, fine-tune the model using a small amount of data from the auxiliary dataset "A2". Fine-tuning is to further optimize the model's performance in the target domain. Steps such as adjusting hyperparameters can be performed as needed until the model meets preset metrics such as accuracy and recall.

[0159] S1004. The finely tuned improved Swin Transformer model can be used in a fault diagnosis process. The auxiliary dataset "A2" is used for fault diagnosis to detect and diagnose faults. During this process, hyperparameters can be adjusted as needed to further improve the model's performance and accuracy.

[0160] S1005. Finally, the accuracy of the trained fault diagnosis model is tested using the target domain dataset "D1_test" (unlabeled). This evaluates the model's performance in the target domain. Simultaneously, the model's generalization ability is tested using the source domain dataset "O1_test" (unlabeled) to evaluate its transferability across different domains.

[0161] Based on the above-mentioned bearing fault diagnosis method for cross-condition unbalanced data, this application also proposes the following bearing fault diagnosis system for cross-condition unbalanced data based on transfer learning and improved deep convolutional generative adversarial networks:

[0162] like Figure 8 As shown, the bearing fault diagnosis system includes multiple lower-level machines connected in parallel to a corresponding upper-level machine, or a single lower-level machine connected to a corresponding upper-level machine, as well as a processor and a memory;

[0163] The lower-level machine uses a multi-channel data acquisition instrument, which is connected to a dual-head BNC terminal signal input line and various adapters to acquire signals of various data types. Various sensors at the end are fixedly installed in the vertical bearing part.

[0164] The host computer consists of a server-side program and a monitoring interface that integrates the bearing fault diagnosis method for cross-operating condition imbalance data, including a speed test module, a core processor module, a data acquisition module, and an interactive communication module;

[0165] The server-side program is responsible for data processing, analysis, and decision-making, while the monitoring interface is responsible for providing a user-friendly display and operation interface.

[0166] The data acquisition module is used to acquire raw signals such as multi-source vibrations during bearing rotation, including various vibration acceleration sensors. It employs DDS high-precision frequency synthesis technology to ensure that all channels achieve parallel synchronous acquisition. Each input channel is connected to a sensor responsible for acquiring vibration signals and converting them into digital signals. Through the input terminal of the processor module, the host computer achieves real-time data transmission via an interactive communication module. The monitored sound, vibration, and temperature data are input to the host computer through the interactive module. The host computer receives the raw acquired data and performs signal data analysis using modern signal analysis techniques. The above overview can be considered as part of the information processing section for multi-source heterogeneous data fusion and stored as a database file. The data acquisition module is powered and grounded by the ARM core module, and the output signal is an analog voltage signal.

[0167] The interactive communication module is used to enable communication between the data acquisition instrument and external devices. It transmits data and performs control operations with external devices through different communication interfaces. Wired communication can be used to avoid signal interference in the working environment. It typically provides one or more serial ports for data exchange with serial devices (such as computers, sensors, and actuators). Communication can be asynchronous or synchronous. Each communication module is configured with ID information; the chip ID information is added to the data frame transmitting monitoring data for the host computer to verify the information source, thus supporting confidentiality and security during data transmission, remote monitoring, and remote control functions.

[0168] The core processor module is used by the host computer to perform signal analysis in the field of modern signal analysis technology after receiving the raw acquired data. Specifically, this includes data storage, analysis, time / frequency domain visualization, wavelet transform, and other data analysis methods. The core processor module includes a minimum system module, a real-time clock module, a backup storage module, a debugging module, and a signal conditioning module. The power supply module, reset module, clock module, program download module, and startup setting module in the minimum system module are all connected to the microprocessor. The real-time clock module is connected to the microprocessor in the minimum system module and is used to read time information during signal acquisition. It provides accurate clock information for timestamps and recording the time information of the acquired data. The backup storage module is connected to the microprocessor in the minimum system module and stores all monitoring data locally. In this embodiment, the backup storage module has a capacity of 16GB and stores data in binary format, with filenames named according to date information. The debugging module is connected to the microprocessor in the minimum system module and is used for debugging the lower-level machine program; the input terminal of the signal conditioning module is connected to the output terminal of the signal acquisition module, and the output terminal of the signal conditioning module is connected to the microprocessor in the minimum system module; the input of the signal conditioning module is the analog voltage signal output by the signal acquisition module, and it converts the analog voltage signal into a digital signal for output.

[0169] The memory includes a computer-readable storage medium for storing a computer-executable program for a bearing diagnostic method based on cross-condition multimodal data fusion;

[0170] The processor is used to read the computer executable program from memory and execute it.

[0171] As described above, the bearing fault diagnosis method proposed in this application employs simultaneous acquisition of multi-source vibration information. To improve signal quality and avoid interference, vibration sensors are integrated and installed together inside the rotating machinery to acquire various modal information during the operation of the rolling bearing under test. Based on transfer learning and improved deep convolutional generative adversarial networks for cross-condition unbalanced data, this application can accurately and efficiently determine the lubrication state of a small sample of rolling bearings under test. By comprehensively analyzing and processing information data from different modalities, precise monitoring and diagnosis of the lubrication state of bearings with small sample data can be achieved, improving the reliability of bearing operation.

[0172] Although the present invention has been shown and described in the above embodiments, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bearing fault diagnosis method based on cross-condition imbalance data, characterized in that: Based on transfer learning and improved deep convolutional generative adversarial networks, the small sample parameters of the target domain are expanded by improving the deep convolutional generative adversarial network, and the generated simulated signal retains the complete high-frequency and low-frequency features of the real signal. Multi-source domain and target domain features are mapped to the same feature space to achieve multi-domain feature extraction and feature alignment; the health status of unknown labeled rolling bearings under cross-working conditions is identified by improving the Swing Transformer network. The implementation steps include the following: Step 1: Obtain the original multi-source vibration signals under multiple fault conditions by sampling from multiple vibration sensors under various working conditions, save them by channel and name them t0; Step 2: Filter, denoise, align, and correct the multi-source vibration signal, and name the result t1. Step 3: For the time series data t1, extract time segments sequentially, apply Fourier transform to them, save the transformed complex number result, and name it F1; Step 4: For time series data t1, perform wavelet transform and reconstruction at different scales using complex wavelet functions; select the optimal scale by comparing errors, and calculate the time-frequency features under the current preferred and unique scale, and save them in the form of a coefficient matrix in a binary file, naming it C1; Step 5: Read the above F1 and C1 files and perform feature concatenation to generate a data matrix. Map the matrix values ​​to generate a grayscale image dataset I1 of wavelet transform. Set the grayscale image dataset I1 as the source domain dataset O1 and divide the source domain dataset O1 into a training set and a test set. Step 6: Change the initial parameter settings of the vertical motor and proceed to step 2. Then construct the time series dataset mu-t2. Perform steps 2 to 5 on the above dataset in sequence and finally obtain the grayscale dataset I2. Divide the grayscale dataset I2 into training set and test set as the target domain D1, and name them D1_train and D1_test respectively. Step 7: Introduce the IECA channel attention mechanism layer into the DCGAN network architecture to construct an improved IDCGAN network architecture. Step 8: Build an improved Swing Transformer network architecture; Step 9: The IDCGAN network is pre-trained and new synthetic samples are generated; Step 10: Train and build an improved Swing Transformer network model for fault diagnosis process.

2. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: Step 1 involves classifying different fault types into different fault categories, and grouping the raw data of different operating conditions under the same fault type to construct the raw time-series information acquisition sample t0.

3. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: Step 2 involves using a filter to remove unwanted frequency components or noise, performing mean denoising on the filtered vibration signal, and using correlation analysis to calculate the time offset of the signal. If there is a time deviation or phase difference between the multi-source vibration signals, data alignment is required to ensure that the signals are consistent in time. The preprocessed vibration signal data is stored in time-series data format, and the processed data is saved as a .mat file named t1.

4. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: Step 3, the Fourier transform formula for the non-periodic continuous-time signal x(t) is as follows: We obtain discrete sampled values ​​x(nT) of a continuous signal x(t), and use the discrete signal x(nT) to calculate the following spectrum of the signal x(t). The complex results of the amplitude spectrum and phase spectrum after Fourier transform are saved as a .csv text file and named F1.

5. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: In step 4, the formula for the Complex Morlet wavelet function is as follows: In the above formula, f(t) is the time-domain signal, ψ(x) is the wavelet basis function, a is the scaling factor used to locate the frequency, and b is the translation factor used to locate the time. Choose the Complex Morlet wavelet basis function with the following expression. In the above formula, a is the scale factor used for positioning frequency, and b is the translation factor used for positioning time; The wavelet transform of the wavelet basis function is applied to the heterogeneous signal at different scales of 128, 64...8, and the reconstruction error is calculated. A lower reconstruction error indicates that the wavelet basis function at the current scale better characterizes the original signal. The time-frequency characteristics, including the frequency band coefficient, energy spectrum, and amplitude spectrum, are calculated at the selected single preferred scale and stored in a binary file in the form of a coefficient matrix, which is named C1.

6. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: In step 5, features belonging to the same time domain information are concatenated row by row, and average segment interpolation sampling is used for longer array parameters to align the feature parameters; the matrix values ​​are mapped to the gray value range, converted to three channels and mapped to generate a wavelet transform grayscale image dataset I1, and the grayscale image dataset I1 is set as the source domain dataset O1; it is divided into training set and test set in a 4:1 ratio, and named O1_train and O1_test respectively.

7. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: The aforementioned Step 6 includes the following steps: S601. Cross-condition operation involves maintaining constant ambient temperature and humidity in the working environment and setting the bearing load; setting three different bearings. Speed ​​setting method; irregular speed variation to simulate actual vertical motor operating conditions; preset constant speed state; in speed condition bar Multiple bearing states can be set for the same components; S602. Mix the multi-source vibration signals after changing the rotational speed conditions with random data of equal length average, and construct a dataset of different rotational speeds for the same damage under multiple channels. Perform data processing such as filtering, denoising, data alignment, and signal correction on the above signals. Store the preprocessed data in the form of time-series data .mat and name it "mu-t2". S603. Perform feature concatenation on the dataset "mu-t2" by concatenating the spectral features and time-frequency features under the same time domain information row by row; if the feature parameter lengths are inconsistent, use average segment interpolation sampling to interpolate the longer array parameters to achieve data alignment; S604. Adjust the concatenated feature vector into a two-dimensional matrix, specify the number of rows and columns of the matrix as needed, map the matrix values ​​to the grayscale range, and convert it into a three-channel grayscale image dataset "I2"; S605. Divide the dataset "I2" into a training set and a test set in a 4:1 ratio, and name them the target domain datasets "D1_train" and "D1_test" respectively.

8. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: Step 7, the generator network structure includes an input layer, a multi-transposed convolutional layer, a multi-head self-attention module, an SN spectrum normalization layer, a ReLU activation function layer, an IECA module, and an output layer; The discriminator network structure includes an input layer, SN spectrum normalization, convolutional layers, multi-head self-attention module, BN normalization layer, LeakyReLU activation function layer, MK-MMD loss function, IECA module, and output layer; The built-in parameter setting process for the aforementioned IDCGAN network architecture includes the following steps: The discriminator network structure is shown in the following expression. L1_Discriminator=Sigmoid(ConvBlock(MHA(ConvBlockn(IECA(leakyRELU(Conv2d(x_d))))))) Where x_d is the real fusion dataset N1 (n≤1) and the synthesized new feature map (n>1); Conv2d(*) is a two-dimensional convolution operation; leakyRELU activation layer; IECA channel attention mechanism layer, which introduces global max pooling weighting on the basis of global mean pooling; ConvBlock block includes Conv2d(), SN(*) spectrum normalization, and leakyRELU(*) activation function in sequence; n is n stacked ConvBlock blocks; MHA is a multi-head self-attention module, which is the intrinsic attention weight allocation before calculating the features of multi-channel vibration information; Sigmoid(*) discriminator judges whether the input parameter conforms to the distribution metric of MK-MMD. If it does, Sigmoid(*) maps true(1), otherwise false(0). When true(1), the auxiliary dataset A1 is saved; The generator network structure is shown in the following expression. L1_Generator=Concat(T-ConvBlock(MHA(T-ConvBlockn(IECA(RELU(T-Conv2d(x_g))))))) Where x_g is the spatial random variable N2 (n≤1) and the newly generated feature parameters after network parameter tuning (n>1); T-Conv2d(*) is a two-dimensional transposed convolution operation; ReLU activation layer; IECA channel attention mechanism layer; T-ConvBlock block includes T-Conv2d(), SN(*) spectrum normalization, and ReLU(*) activation function in sequence; n is n stacked T-ConvBlock blocks; MHA is a multi-head self-attention module, which is the intrinsic attention weight allocation before calculating the features of multi-channel vibration information; Concat(*) synthesizes a new feature map after updating its self-parameter settings after feedback from the discriminator.

9. The bearing fault diagnosis method based on cross-condition imbalance data according to claim 1, characterized in that: Step 8, the improved Swing Transformer network architecture includes an input layer, an IECA channel attention layer, a Patch Embedding image block embedding layer, a positional encoding layer, a Transformer encoder layer, a downsampling layer between Transformer encoder modules, a cross-layer window-based local attention mechanism, a global weighted pooling layer, a fully connected layer, and a Softmax output layer. The improved Swing Transformer network model expression is as follows: L2_Swin_T=softmax(MLP(SwinBlock2(Patch Merging (SwinBlock1 (IECA( PE((PatchEmbed(x))Ni) ))))Nj)) Wherein, during pre-training, x is the fused dataset N1+A1; during fine-tuning after the first transfer learning, x is a small amount of auxiliary dataset A2; after the second transfer learning and parameter tuning, x is the remaining auxiliary dataset A2; PatchEmbed(•) is a function that segments the input image (x) into image patches and converts them into embedding vectors through linear projection; PE(•) is a position encoding function that adds position encoding to the embedded data to preserve the positional information of the sequence; Ni is the number of channels for multi-channel data processing; IECA channel attention mechanism layer; SwinBlock1 block includes LN(*) layer normalization, W-MSA(*) window-type local multi-head self-attention mechanism, CBAM(*) channel spatial attention mechanism, weighted nodes, LN(*) layer normalization, FeedForward(*) feedforward network layer, weighted nodes; Patch Merging(*) consists of a downsampling layer, a cross-layer windowed local attention mechanism, and a global weighted pooling layer; the SwinBlock2 block includes an LN(*) layer for normalization, a SW-MSA(*) windowed multi-head self-attention mechanism, a CBAM(*) channel spatial attention mechanism, weighted nodes, an LN(*) layer for normalization, a FeedForward(*) feedforward network layer, and weighted nodes; Nj is a stack of N pairs of SwinBlock blocks and the Patch Merging(*) layer between them; MLP(•) is a multilayer perceptron function consisting of a fully connected layer and an activation function to map the transformed representation to the output layer of the classification task; the softmax(*) function is used to activate the output layer and obtain the predicted result of the classification label.

10. A bearing diagnostic system applying the bearing fault diagnosis method based on cross-condition imbalance data as described in any one of claims 1 to 9, characterized in that: This includes multiple lower-level machines connected in parallel to a corresponding upper-level machine, or a single lower-level machine connected to a corresponding upper-level machine, as well as a processor and memory; The lower-level machine uses a multi-channel data acquisition instrument, which is connected to a dual-head BNC terminal signal input line and various adapters to acquire signals of various data types. Various sensors at the end are fixedly installed in the vertical bearing part. The host computer consists of a server-side program and a monitoring interface that integrates the bearing fault diagnosis method for cross-operating condition imbalance data, including a speed test module, a core processor module, a data acquisition module, and an interactive communication module; The memory includes a computer-readable storage medium for storing a computer-executable program for a bearing diagnostic method based on cross-condition multimodal data fusion; The processor is used to read the computer executable program from memory and execute it.

Images