Complex working condition-oriented high-noise-resistance bearing fault diagnosis method

Abstract

The invention provides a high-noise-resistance bearing fault diagnosis method for complex working conditions, relates to the technical field of bearing fault detection, and constructs a progressive technical system of'dynamic noise reduction-multi-modal fusion-cross-domain association '. According to the method, firstly, working condition self-adaptive noise reduction is achieved through sliding window dynamic principal component analysis, and redundant noise caused by rotating speed fluctuation is stripped while periodic impact characteristics are reserved; meanwhile, on this basis, a time domain-frequency domain multi-mode complementary feature system is constructed, and the denoised signals are respectively input into a bidirectional gating circulation unit and an improved residual network added with FFT transform to extract bidirectional time sequence dependence and frequency domain texture features. And finally, performing feature enhancement on a time sequence signal and a time-frequency image extracted by the multi-modal feature network by adopting eight self-attention, constructing a dynamic interaction channel by taking a time sequence feature as Query and an image feature as Key-Value, and adaptively screening a cross-modal association feature most related to the context of the current rotating speed.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault detection, and in particular to a high-noise-resistance bearing fault diagnosis method for complex working conditions. Background Art

[0002] As industrial equipment continues to advance toward higher speeds and greater integration, the health of rotating machinery, a critical transmission component, directly determines the reliability and safety of production systems. However, under complex operating conditions, equipment is constantly exposed to challenges such as high loads, temperature fluctuations, and lubrication degradation. This leads to frequent faults such as gear pitting, bearing wear, and motor winding short circuits. Failure to diagnose these faults in a timely manner can lead to serious accidents.

[0003] The bearing assembly is structurally composed of an outer ring, inner ring, rolling elements, and a cage, and the joints between these components are susceptible to mechanical damage. When a bearing component experiences local damage, its rotational motion excites a periodic impact vibration signal with a specific frequency characteristic. The mathematical representation of this characteristic frequency forms a nonlinear mapping relationship with the fault location, damage geometry, shafting speed, and load conditions. Existing bearing fault detection methods often rely on time-frequency analysis of single-source signals. While these methods perform well under stable operating conditions, they face the dual limitations of incomplete information and insufficient noise immunity. When the equipment load or speed changes, the model suffers from weak noise immunity and poor adaptability across operating conditions.

[0004] Therefore, it is necessary to propose an enhanced fault diagnosis algorithm that uses dynamic principal component analysis to reduce data dimension and fuse multimodal features for identification and classification to solve the above problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention provides a highly noise-resistant bearing fault diagnosis method for complex operating conditions. To address signal fluctuations caused by equipment load or speed variations, a progressive technical system combining dynamic noise reduction, multimodal fusion, and cross-domain correlation is constructed. The method first implements condition-adaptive noise reduction through sliding window sampling and dynamic principal component analysis (DPCA). This method removes redundant noise caused by speed fluctuations while preserving periodic impact features. Furthermore, a time-domain and frequency-domain multimodal complementary feature system is constructed. The denoised signal is fed into a bidirectional gated recurrent unit and an improved residual network with a fast Fourier transform (FFT) to extract bidirectional temporal dependencies and frequency-domain texture features. Finally, an eight-head self-attention algorithm is used to enhance the time series signal and time-frequency image extracted by the multimodal feature network. A dynamic interaction channel is constructed using time series features as queries and image features as key-value pairs to adaptively select cross-modal correlation features most relevant to the current speed context.

[0006] The present invention is achieved through the following technical solution: a high noise-resistant bearing fault diagnosis method for complex working conditions, specifically comprising the following steps: Step S1: divide the vibration signal of the sample into practical windows: Use sliding windows to split the continuous vibration signal into subsequences of fixed length; assign a label to each window according to the fault type to which it belongs. ,The minimum-maximum normalization is used to normalize the signal so that the signal amplitude is mapped in the interval [0,1]; Step S2: Perform DPCA operation on the fixed-length subsequence to reduce the data dimension: Dynamically select the number of principal components for normalized time series signals partitioned by sliding windows k , so that the proportion of the retained cumulative variance is not less than 95%, based on the normalized data , calculate the covariance matrix ; Then, perform eigendecomposition on the covariance matrix to obtain the eigenvalues λ i and the corresponding eigenvectors v i , arrange the eigenvalues in descending order, and finally obtain the data after DPCA dimension reduction.

[0007] Step S3: Extract time-frequency domain features from the data after dimensionality reduction: On the one hand, nine-dimensional statistical features including mean, variance, skewness, kurtosis, peak-to-peak value, shape factor, peak factor, impulse factor and margin factor are extracted from each sample after DPCA dimension reduction. On the other hand, the data after DPCA dimension reduction is subjected to fast Fourier transform, and after the steps of signal time-frequency domain conversion, amplitude spectrum calculation, mean frequency domain feature extraction, and standard deviation domain feature extraction, the frequency domain distribution features are extracted; Step S4: Multimodal dataset construction: The training set and test set are randomly divided into 8:2 ratios. Among them, the time series data generates a 25-bit random ID to uniquely identify the data sample and save it as a Pickle file. The time series signal is converted into a 32×32 grayscale image, and the pixel value is linearly mapped from the signal amplitude to [0,255] to generate the image data. For the multimodal input constructed by this dataset, the feature matrix is: (3) in, X final 、X pca 、F time 、F freq Represent the multimodal dataset feature matrix, data dimension reduction feature map, time series features, and frequency domain features respectively. M, k Represents the feature dimension of the original data and the target dimension of dimensionality reduction; Step S5: Extracting time series features and image features: A bidirectional gated recurrent unit (BiGRU) is used as a temporal feature extractor. The input sequence is processed separately through two independent forward and backward GRU layers. The forward layer captures the positive temporal dependency from the start to the end, while the backward layer captures the reverse dependency from the end to the start. The hidden state at each moment is concatenated and bidirectionally output to form a temporal feature vector containing global context information. The specific parameters of BiGRU are: enter , where B is the batch size, sequence length T=16, feature dimension D=64; the number of hidden units in each direction is 128, and the bidirectional splicing is 256-dimensional; An improved residual network is used as the image feature extractor. The residual block contains two convolutional layers and batch normalization, and the input and the convolution result are added through identity mapping. Residual block level configuration: (4) in, convN Represents different residual layers; for the input image, adjust the step size of the initial convolution layer and the pooling layer, and finally output the feature vector ; Step S6: Single-modal internal feature enhancement: The features are split into 8 subspaces, and attention weights are calculated independently for each subspace to capture dependencies from different perspectives. Finally, the results of all subspaces are concatenated and fused through a linear layer to generate enhanced unimodal features. The specific steps are as follows: For the input sequence Q (Query), K (key) 、V (value), each header i Generate subspace representations through independent linear transformations: (5) Single-head attention calculation: (6) After the outputs of all heads are concatenated, the final result is obtained through linear transformation; in, To merge the weight matrix: (7) Where: To learn the parameters, head i For different attention heads, d k is the dimension of each attention head, d model is the model dimension, d k = d model / h,h is the number of heads, scaling factor Used to alleviate the problem of excessively large dot product values; Step S7: Cross-modal dynamic interaction fault diagnosis: Using temporal features as queries and image features as keys, a cross-attention module is constructed. By calculating the similarity matrix between temporal features and image features, attention weights are dynamically assigned to select image region features that are most relevant to the current temporal context. The cross-modal features are then further fused with the projected DPCA features through weighted summation or concatenation to form a unified joint representation. The specific parameters of the cross attention module are: 8 parallel computation attention heads, each with a dimension of d k =128 / 8=16, the fused features ; Finally, the fused multimodal features are input into the fully connected classifier, and the model parameters are optimized with the cross-entropy loss function to output the probability distribution of the fault category, completing the end-to-end diagnostic process from data preprocessing to classification decision-making.

[0008] As a preferred solution, in step S1 , the window length L is set to 1024 and the step length S is set to 100.

[0009] As a preferred solution, step S2 specifically includes the following steps: Step S2.1, normalize the original data X; Step S2.2, calculating the covariance matrix; Step S2.3, perform eigenvalue decomposition; Step S2.4: Calculate the cumulative variance explained, so that its threshold is not less than 0.95. The calculation formula is as follows: (1) in, CV is the cumulative variance explained, k To dynamically select the number of principal components, λ is the eigenvalue of the covariance matrix; Step 2.5: Project the processed data into the principal component space using the following formula: (2) in, X pca For the front k The feature maps on the principal components are k <n,X norm To normalize the data, V are the eigenvalues and eigenvectors of the covariance matrix.

[0010] As a preferred solution, the specific parameters of the improved residual network in step S5 are: Convolutional layer: grayscale image input channel 1, output channel 64, 7×7 convolution kernel, stride 2, maximum pooling layer (3×3 kernel, stride 2), output size 8×8; Residual block: Two 3×3 convolutional layers, including batch normalization layer and ReLU activation function. When the input and output dimensions do not match, the number of channels is adjusted through 1×1 convolution.

[0011] Due to the adoption of the above technical solution, the present invention has the following beneficial effects compared with the prior art: The high-noise-resistance bearing fault diagnosis method proposed in the present invention can maintain a high level of diagnostic accuracy under noise interference, and is improved by 14%-17% compared with the single-mode model when the noise level is 1; the average diagnostic accuracy under variable speed conditions reaches 98.2%, which is 4%-6.1% higher than the single-mode model.

[0012] Compared with existing technologies, this method offers advantages in combining dynamic data dimensionality reduction with multimodal data feature fusion. It also constructs a hierarchical attention design framework that uses a multi-head attention mechanism to enhance local features of single-modal features and cross-attention to achieve feature fusion. This approach addresses the model's poor noise immunity and cross-condition adaptability under complex working conditions.

[0013] At the same time, an enhanced diagnostic model is developed to specifically improve noise resistance and adaptability across working conditions: DPCA is used as the preprocessing method, bidirectional gated recurrent units and residual networks are used as single-channel feature extraction models, and a cross-attention mechanism is used for feature fusion and positioning of key modes and feature dimensions. This significantly improves the accuracy and stability of fault diagnosis under complex working conditions, meeting the requirements of industrial scenarios for the credibility of diagnostic results.

[0014] Additional aspects and advantages of the invention will become apparent from the description which follows, or may be learned by practice of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments with reference to the accompanying drawings, in which: Figure 1 This is the overall architecture diagram of the system of the present invention. DETAILED DESCRIPTION

[0016] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0017] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0018] The following combination Figure 1 The high noise resistance bearing fault diagnosis method for complex working conditions according to an embodiment of the present invention is described in detail.

[0019] like Figure 1 As shown, the present invention proposes a highly noise-resistant bearing fault diagnosis method for complex working conditions. It introduces dynamic principal component analysis and multimodal feature fusion to achieve the purpose of comprehensive extraction of bearing fault information, improve noise resistance and adaptability across working conditions. Specifically, it includes the following steps: Step S1: divide the vibration signal of the sample into practical windows: Use sliding windows to split the continuous vibration signal into fixed-length subsequences. Set the window length L = 1024 and the step size S = 100. Assign a label to each window according to the fault type to which it belongs. ,At the same time, in order to adjust the different ,characteristics of the fault signal to the same dimension, the ,minimum-maximum normalization is used to normalize the signal so that ,the signal amplitude is mapped in the [0,1] interval; Step S2: Perform dynamic principal component analysis on the fixed-length subsequence: Dynamically select the number of principal components for normalized time series signals partitioned by sliding windows k , so that the proportion of the retained cumulative variance is not less than 95%, based on the normalized data , calculate the covariance matrix ; Then, perform eigendecomposition on the covariance matrix to obtain the eigenvalues λ i and the corresponding eigenvectors v i , sort the eigenvalues in descending order, and finally obtain the data after DPCA dimension reduction; The specific steps include: Step S2.1, normalize the original data X; Step S2.2, calculating the covariance matrix; Step S2.3, perform eigenvalue decomposition; Step S2.4: Calculate the cumulative variance explained, so that its threshold is not less than 0.95. The calculation formula is as follows: (1) in, CV is the cumulative variance explained, k To dynamically select the number of principal components, λ i is the eigenvalue of the covariance matrix; Step 2.5: Project the processed data into the principal component space using the following formula: (2) in, X pca For the front k The feature maps on the principal components are k <n,X norm To normalize the data, V are the eigenvalues and eigenvectors of the covariance matrix.

[0020] Step S3: Extract time-frequency domain features from the data after dimensionality reduction: On the one hand, nine-dimensional statistical features including mean, variance, skewness, kurtosis, peak-to-peak value, shape factor, peak factor, impulse factor and margin factor are extracted from each sample after DPCA dimension reduction. On the other hand, the data after DPCA dimension reduction is subjected to fast Fourier transform, and after the steps of signal time-frequency domain conversion, amplitude spectrum calculation, mean frequency domain feature extraction, and standard deviation domain feature extraction, the frequency domain distribution features are extracted; Step S4: Multimodal dataset construction: The training set and test set are randomly divided into 8:2 ratios to ensure balanced distribution of samples in each category. A 25-bit random ID is generated for the time series data to uniquely identify the data sample and save it as a pickle file. The time series signal is converted into a 32×32 grayscale image, and the pixel value is linearly mapped from the signal amplitude to [0,255] to generate the image data. For the multimodal input constructed by this dataset, the feature matrix is: (3) in, X final 、X pca 、F time 、F freq Represent the multimodal dataset feature matrix, data dimension reduction feature map, time series features, and frequency domain features respectively. M, k Represents the feature dimension of the original data and the target dimension of dimensionality reduction; Step S5: Extracting time series features and image features: A bidirectional gated recurrent unit (BiGRU) is used as a temporal feature extractor. The input sequence is processed separately through two independent forward and backward GRU layers. The forward layer captures the positive temporal dependency from the start to the end, while the backward layer captures the reverse dependency from the end to the start. The hidden state at each moment is concatenated and bidirectionally output to form a temporal feature vector containing global context information. The specific parameters of BiGRU are: enter , where B is the batch size, sequence length T=16, feature dimension D=64; the number of hidden units in each direction is 128, and the bidirectional splicing is 256-dimensional; An improved residual network is used as the image feature extractor. The residual block contains two convolutional layers and batch normalization. The input and the convolution result are added through identity mapping to ensure stable training of deep networks. Residual block level configuration: (4) in, convNRepresents different residual layers; for the input image, adjust the step size of the initial convolution layer and the pooling layer, and finally output the feature vector ; The specific parameters of the improved residual network are: Convolutional layer: grayscale image input channel 1, output channel 64, 7×7 convolution kernel, stride 2, maximum pooling layer (3×3 kernel, stride 2), output size 8×8; Residual block: Two 3×3 convolutional layers, including batch normalization layer and ReLU activation function. When the input and output dimensions do not match, the number of channels is adjusted through 1×1 convolution.

[0021] Step S6: Single-modal internal feature enhancement: After feature extraction to form unimodal features, a multi-head self-attention mechanism is applied to both time series and image features to reduce the possibility of limited expressiveness due to local noise or information sparsity. The features are split into eight subspaces, and attention weights are calculated independently for each subspace to capture dependencies from different perspectives. Finally, the results of all subspaces are concatenated and fused through a linear layer to generate enhanced unimodal features. The specific steps are as follows: For the input sequence Q (Query), K (key) 、V (value), each header i Generate subspace representations through independent linear transformations: (5) Single-head attention calculation: (6) After the outputs of all heads are concatenated, the final result is obtained through linear transformation; in, To merge the weight matrix: (7) Where: To learn the parameters, head i For different attention heads, d k is the dimension of each attention head, d model is the model dimension, d k = d model / h,h is the number of heads, scaling factor Used to alleviate the problem of excessively large dot product values; Step S7: Cross-modal dynamic interaction fault diagnosis: Using temporal features as queries and image features as keys, a cross-attention module is constructed. By calculating the similarity matrix between temporal features and image features, attention weights are dynamically assigned to select image region features that are most relevant to the current temporal context. The cross-modal features are then further fused with the projected DPCA features through weighted summation or concatenation to form a unified joint representation. The specific parameters of the cross attention module are: 8 parallel computation attention heads, each with a dimension of d k =128 / 8=16, the fused features ; Finally, the fused multimodal features are input into the fully connected classifier, and the model parameters are optimized with the cross-entropy loss function to output the probability distribution of the fault category, completing the end-to-end diagnostic process from data preprocessing to classification decision-making. Example

[0022] The effects of the present invention can be further illustrated by the following examples.

[0023] The present invention uses the variable speed bearing fault data set in Ottawa, Canada as the initial data set. The data contains vibration signals collected from bearings with different health conditions under time-varying speed conditions, with a total of 36 data sets. For each data set, the health status of the bearing includes healthy, inner ring defects, and outer ring defects, and the operating speed conditions are increasing speed, decreasing speed, increasing and then decreasing speed, and decreasing and then increasing speed. Each data set contains two channels: vibration data measured by an accelerometer and speed data measured by an encoder. These data are sampled at a frequency of 200KHz and a sampling time of 10 seconds. A total of 10 data sets at different speeds under three fault types were selected, and the specific descriptions are shown in Table 1: Table 1 Health status serial number Speed normal HA-1 Increase normal HB-2 reduce normal HC-3 Increase first and then decrease normal HD-1 Decrease first and then increase Inner race fault IA-1 Increase Inner race fault IB-1 reduce Inner race fault IC-1 Increase first and then decrease Outer ring failure OA-1 Increase Outer ring failure OB-1 reduce Outer ring failure OC-1 Increase first and then decrease Table 2 compares the single-channel BiGru, ResNet, and F-PCA feature training models with the fusion models CNN-GRU and PCA-MFF. The results are shown in Table 2. The single-channel BiGru, ResNet, and F-PCA models achieved average test accuracy of 92.3%, 94.1%, and 93.6%, respectively. The DPCA-MFF model achieved an average test accuracy of 98.1%, representing improvements of 6.1%, 4.0%, and 4.5%, respectively, compared to the single-modality models. Table 2 category BiGRU ResNet F-PCA DPCA-MFF HA1 93.5% 95.2% 93.8% 98.8% HB2 90.1% 91.6% 89.6% 97.3% HC3 93.3% 95.0% 95.0% 98.6% HD1 92.0% 93.0% 92.8% 97.0% IA1 92.6% 94.2% 93.7% 98.4% IB1 92.4% 93.5% 92.7% 97.9% IC1 92.8% 94.5% 93.6% 98.7% OA1 91.6% 94.2% 93.5% 97.8% OB1 93.7% 94.8% 94.0% 98.5% OC1 91.0% 95.0% 93.5% 98.0% AVG 92.3% 94.1% 93.4% 98.1% Noise at different levels was added to the original fault signal. The results of the noise reduction experiment under the same load are shown in Table 3. Compared with the single-channel model, the recognition accuracy of the proposed method decreases more steadily as the noise level increases, and the recognition accuracy remains at 65.8% when the noise level is 1.

[0024] Table 3 λ BiGRU ResNet F-PCA DPCA-MFF 0.2 80.4% 82.5% 79.8% 96.2% 0.4 70.3% 71.6% 69.6% 88.9% 0.6 66.4% 67.3% 66.1% 77.4% 0.8 64.5% 65.1% 60.2% 71.2% 1.0 60.2% 64.3% 57.2% 65.8% It shows that this aspect can remove noise redundancy through dynamic dimensionality reduction, suppress local interference through attention mechanism, and comprehensively extract fault information through complementary fusion of multimodal features, significantly improving the model's noise resistance and cross-working condition adaptability under complex working conditions, and providing a high noise resistance solution for intelligent operation and maintenance of industrial equipment.

[0025] Throughout this specification, terms such as "one embodiment," "some embodiments," and "specific embodiments" mean that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0026] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A high noise-resistant bearing fault diagnosis method for complex working conditions, characterized by , specifically including the following steps: Step S1: divide the vibration signal of the sample into practical windows: Use sliding windows to split the continuous vibration signal into subsequences of fixed length; assign a label to each window according to the fault type to which it belongs. ,The minimum-maximum normalization is used to normalize the signal so that the signal amplitude is mapped in the interval [0,1]; Dynamically select the number of principal components for normalized time series signals partitioned by sliding windows k , so that the proportion of the retained cumulative variance is not less than 95%, based on the normalized data , calculate the covariance matrix ; Then, perform eigendecomposition on the covariance matrix to obtain the eigenvalues λ i and the corresponding eigenvectors v i , sort the eigenvalues in descending order, and finally obtain the data after DPCA dimension reduction; Step S3: Extract time-frequency domain features from the data after dimensionality reduction: On the one hand, nine-dimensional statistical features including mean, variance, skewness, kurtosis, peak value, shape factor, peak factor, impulse factor and margin factor are extracted from each sample after DPCA dimension reduction. On the other hand, the data after DPCA dimension reduction is subjected to fast Fourier transform, and after the steps of signal time-frequency domain conversion, amplitude spectrum calculation, mean frequency domain feature extraction, and standard deviation domain feature extraction, the frequency domain distribution features are extracted; Step S4: Multimodal dataset construction: The training set and test set are randomly divided into 8:2 ratios. Among them, the time series data generates a 25-bit random ID to uniquely identify the data sample and save it as a Pickle file. The time series signal is converted into a 32×32 grayscale image, and the pixel value is linearly mapped from the signal amplitude to [0,255] to generate the image data. For the multimodal input constructed by this dataset, the feature matrix is: (3) in, X final 、X pca 、F time 、F freq Represent the multimodal dataset feature matrix, data dimension reduction feature map, time series features, and frequency domain features respectively. M, k Represents the feature dimension of the original data and the target dimension of dimensionality reduction; Step S5: Extracting time series features and image features: A bidirectional gated recurrent unit (BiGRU) is used as a temporal feature extractor. The input sequence is processed separately through two independent forward and backward GRU layers. The forward layer captures the positive temporal dependency from the start to the end, while the backward layer captures the reverse dependency from the end to the start. The hidden state at each moment is concatenated and bidirectionally output to form a temporal feature vector containing global context information. The specific parameters of BiGRU are: enter , where B is the batch size, sequence length T=16, feature dimension D=64; the number of hidden units in each direction is 128, and the bidirectional splicing is 256-dimensional; An improved residual network is used as the image feature extractor. The residual block contains two convolutional layers and batch normalization, and the input and the convolution result are added through identity mapping. Residual block level configuration: (4) in, convN Represents different residual layers; for the input image, adjust the step size of the initial convolution layer and the pooling layer, and finally output the feature vector ; Step S6: Single-modal internal feature enhancement: The features are split into 8 subspaces, and attention weights are calculated independently for each subspace to capture dependencies from different perspectives. Finally, the results of all subspaces are concatenated and fused through a linear layer to generate enhanced unimodal features. Step S7: Cross-modal dynamic interaction fault diagnosis: Using temporal features as queries and image features as keys, a cross-attention module is constructed. By calculating the similarity matrix between temporal features and image features, attention weights are dynamically assigned to select image region features that are most relevant to the current temporal context. The cross-modal features are then further fused with the projected DPCA features through weighted summation or concatenation to form a unified joint representation. The specific parameters of the cross attention module are: 8 parallel computation attention heads, each with a dimension of d k =128 / 8=16, the fused features ; Finally, the fused multimodal features are input into the fully connected classifier, and the model parameters are optimized with the cross-entropy loss function to output the probability distribution of the fault category, completing the end-to-end diagnostic process from data preprocessing to classification decision-making.

2. A high noise resistance bearing fault diagnosis method for complex working conditions according to claim 1, characterized in that In step S1, the window length L is set to 1024 and the step length S is set to 100.

3. A high noise-resistant bearing fault diagnosis method for complex working conditions according to claim 1, characterized in that , the step S2 specifically includes the following steps: Step S2.1, normalize the original data X; Step S2.2, calculating the covariance matrix; Step S2.3, perform eigenvalue decomposition; Step S2.4: Calculate the cumulative variance explained, so that its threshold is not less than 0.

95. The calculation formula is as follows: (1) in, CV is the cumulative variance explained, k To dynamically select the number of principal components, λ is the eigenvalue of the covariance matrix; Step 2.5: Project the processed data into the principal component space using the following formula: (2) in, X pca For the front k The feature maps on the principal components are k <n,X norm To normalize the data, V are the eigenvalues and eigenvectors of the covariance matrix.

4. A high noise-resistant bearing fault diagnosis method for complex working conditions according to claim 1, characterized in that ,The specific parameters of the improved residual network in step S5 are: Convolutional layer: grayscale image input channel 1, output channel 64, 7×7 convolution kernel, stride 2, maximum pooling layer (3×3 kernel, stride 2), output size 8×8; Residual block: Two 3×3 convolutional layers, including batch normalization layer and ReLU activation function. When the input and output dimensions do not match, the number of channels is adjusted through 1×1 convolution.

5. A high noise-resistant bearing fault diagnosis method for complex working conditions according to claim 1, characterized in that , the step S6 specifically includes the following steps: For input sequence query Q ,key K、 value V , each head i Generate subspace representations through independent linear transformations: (5) Single-head attention calculation: (6) After the outputs of all heads are concatenated, the final result is obtained through linear transformation; in, To merge the weight matrix: (7) Where: To learn the parameters, head i For different attention heads, d k is the dimension of each attention head, d model is the model dimension, d k = d model / h,h is the number of heads, scaling factor Used to alleviate the problem of excessively large dot product values.

Images