Intelligent rotating machine fault diagnosis method

Abstract

The invention provides an intelligent rotating machine fault diagnosis method, and belongs to the technical field of mechanical fault diagnosis, and the method comprises the steps: firstly, obtaining a vibration signal to be subjected to fault diagnosis under a source domain working condition and a target domain working condition, and carrying out the noise reduction processing; secondly, converting the denoised signal into a two-dimensional time-frequency image through short-time Fourier transform, and respectively making a source domain data set and a target domain data set; thirdly, constructing an attention enhancement network (SE-CNN-BiLSTM-MSMHA), and performing pre-training on the model based on the source domain data set to obtain a source domain working condition model; then, freezing part of parameters of the source domain working condition model, and finely adjusting the parameters of the source domain working condition model by using the target domain data to obtain a target domain working condition model; and finally, inputting target domain data to realize fault diagnosis of target domain working conditions. The method can provide a scientific basis for improving the cross-working-condition fault diagnosis accuracy under the strong noise background.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and more particularly to an intelligent fault diagnosis method for rotating machinery. Background Art

[0002] When a rotating machine fails, it will affect the normal production process and cause a decline in product quality at least, and at worst, it will cause huge economic losses to the enterprise, and even lead to major safety and personnel casualties. Moreover, due to its harsh working environment, the characteristics of fault signals are often masked by strong noise, resulting in difficulties in feature extraction and recognition. In actual industrial scenarios, rotating machinery often needs to operate under multi-condition conditions such as variable speed and variable load. The difference in feature distribution between different conditions will lead to feature mismatch problems in traditional diagnostic models. Therefore, it is of great practical significance to effectively diagnose cross-condition faults of rotating machinery in a strong noise environment in a timely and accurate manner.

[0003] In actual industrial scenarios, rotating machinery often needs to operate under multi-condition conditions such as variable speed and variable load. The difference in feature distribution between different conditions will lead to feature mismatch problems in traditional diagnostic models. Research shows that when the working condition changes exceed 15% of the rated parameters, the false alarm rate of existing fixed-threshold diagnostic methods will rise to more than 30%. This problem of insufficient cross-condition adaptability and strong noise interference form a coupling effect, further exacerbating the difficulty of fault feature extraction and pattern recognition. Therefore, it is necessary to study advanced signal denoising techniques to improve the accuracy of fault diagnosis.

[0004] In industrial sites, the vibration signals of rotating machinery are often interfered by complex noises. Especially in extreme working conditions such as wind power gearboxes and mining machinery, the background noise energy may completely mask the fault characteristics. According to the consensus in the field of mechanical vibration diagnosis, when the signal-to-noise ratio (SNR) of the vibration signal in the full frequency band is lower than 0 dB (i.e., the noise power ≥ signal power), traditional time-frequency analysis methods (such as envelope spectrum and fast Fourier transform) fail because they cannot effectively separate noise and fault characteristics. Moreover, rotating machinery often needs to operate under multi-condition conditions such as variable speed and variable load. The difference in feature distribution between different conditions will lead to the problem of insufficient cross-condition adaptability in traditional diagnostic models. This problem of cross-condition adaptability and strong noise interference form a coupling effect, further exacerbating the difficulty of fault feature extraction and pattern recognition. Therefore, it is necessary to study advanced signal denoising techniques, fault feature extraction and bearing intelligent fault diagnosis techniques theory and transfer learning strategies, and continuously enrich the theory and key technologies of health monitoring and fault diagnosis of bearing conditions to improve the accuracy and robustness of cross-condition fault diagnosis methods under strong noise backgrounds. Summary of the Invention

[0005] The main objective of the present invention aims to address the problem that fault signals in the vibration data of rotating machinery measured in a strong noise environment are easily masked, making it difficult to extract and identify fault features. It provides an intelligent fault diagnosis method for rotating machinery, adopting a progressive processing framework of "feature extraction-channel feature enhancement-temporal modeling-multi-scale feature fusion" and a "multi-scale multi-head attention module with parameter self-adaptation". Through modular design, efficient and flexible fault diagnosis is achieved. At the same time, the model architecture proposed in the present invention demonstrates significant advantages in bearing fault diagnosis: the feature extraction-channel feature enhancement mechanism improves the feature representation ability, the multi-scale temporal modeling enhances the adaptability to complex working conditions, and the parameter self-adaptation module ensures strong anti-noise performance, having excellent engineering practical value in the fault diagnosis of rotating machinery in a strong noise environment. Thus, it provides a scientific basis for the fault diagnosis of rotating machinery in a strong noise environment and reliable support for ensuring people's lives and property safety.

[0006] The technical solution to achieve the objective of the present invention is as follows: An intelligent fault diagnosis method for rotating machinery, including the following steps: Step 1: Obtain the vibration signals of the rotating machinery under the source domain working conditions and the target domain working conditions; Step 2: Perform noise reduction processing on the vibration signals of the source domain working conditions and the target domain working conditions through a noise reduction processing module. The noise reduction processing module includes two sub-modules: a variational mode decomposition module optimized by the beluga optimization algorithm, and a screening and reconstruction module for variational mode decomposition by the technique for order preference by similarity to an ideal solution; Step 3: Perform short-time Fourier transform on the noise-reduced vibration signals to obtain two-dimensional time-frequency images. Make a source domain dataset from the two-dimensional time-frequency images obtained under the source domain working conditions, and make a target domain dataset from the two-dimensional time-frequency images obtained under the target domain working conditions; Step 4: Construct an attention-enhanced network model based on SE-CNN-BiLSTM-MSMHA, and use the source domain dataset to complete the pre-training of the network model to obtain a source domain working condition model; Step 5: Through the transfer learning method, freeze all the network model parameters before the first bidirectional long short-term memory layer in the source domain working condition model, and fine-tune the parameters of the source domain working condition model with a small number of data samples in the target domain dataset to obtain a target domain working condition model; Step 6: Adopt the methods of Step 2 and Step 3. After performing noise reduction processing on the vibration signal to be fault-diagnosed and obtaining a two-dimensional time-frequency image, input the time-frequency image into the fault diagnosis model corresponding to the working condition of the vibration signal to achieve fault diagnosis.

[0007] Further, in Step 2, in the noise reduction processing module, for the variational mode decomposition sub-module optimized by the beluga optimization algorithm, its decomposition steps are as follows: Step S21: Set the penalty factor α and the decomposition layer number N The search ranges of these two parameters; Step S22: Initialize the positions of the beluga whale population, set the minimum envelope entropy as the fitness function, and then calculate the fitness values of each whale individual; Step S23: Iteratively update the positions of the whale population based on the beluga whale optimization algorithm until the preset termination condition is met, and then output the optimal N parameter combination of fitness value and α; Step S24: Based on the optimal parameter combination, perform VMD decomposition on the vibration signal to obtain N IMF components; The ideal solution sorting method is used to screen and reconstruct the variational mode decomposition sub-module, and its reconstruction steps are as follows: Step S25: For N intrinsic mode components, when N is even, remove N / 2 high-frequency components through the central frequency; when N is odd, remove ([[]] N +1) / 2 high-frequency components through the central frequency; Step S26: Calculate three indexes of the remaining intrinsic mode components: envelope entropy, kurtosis index, and correlation coefficient with the input signal respectively; Step S27: Calculate the comprehensive score of the three indexes by using the ideal solution sorting method, and remove the component with the minimum score; Step S28: Reconstruct the remaining intrinsic mode components by adding them together as the denoised signal.

[0008] Furthermore, in step four, the attention enhancement network of the SE-CNN-BiLSTM-MSMHA has the following structure: First, the construction of the model starts from the input layer, and passes through three consecutive convolutional layers, a squeeze-and-excitation module, the fourth convolutional layer, a pooling layer, a flattening layer, and the first bidirectional long short-term memory layer in sequence; Second, reshape it into a one-dimensional sequence through the reshaping layer and connect the multi-scale multi-head attention module; Third, connect a flattening layer to flatten it, and after flattening, pass through two fully connected layers with the ReLU activation function and an output layer using softmax activation function for fault classification and recognition; The multi-scale multi-head attention module adopts a three-channel parallel architecture, including: first, constructing a layer group containing three sets of parallel bidirectional long short-term memory networks, where each bidirectional long short-term memory layer has a different number of hidden neurons and all enable full sequence output mode; second, each bidirectional long short-term memory layer is connected to a parameter-adaptive multi-head attention module. In the multi-head attention module, the number of attention heads is fixed to 4, and the key vector dimension is dynamically calculated. The key vector dimension = the number of hidden units / the number of attention heads; finally, a deep splicing strategy is used to fuse the features of the three sets of attention outputs to form three parallel processing streams. The mathematical formula of the multi-scale multi-head attention module is as follows: First are three bidirectional long short-term memory networks: H i =BiLSTM i ( X ) Among them, BiLSTM i For the module i BiLSTM channels, X is the input sequence, H i is the output sequence; Next, we construct the query, key, value matrix and attention score calculation formula for the subsequent attention mechanism; Query Matrix Q i,j : Q i,j = H i W Q (i,j) Bond Matrix K i,j : K i,j = H i W K (i,j) Value Matrix V i,j : V i,j = H i W V (i,j) Attention score Attention i,j : Attentioni,j =V i,j × softmax(Q i,j × K i,j ) / ( d k ( i )) 1 / 2 Multi - head attention score MultiHead(i) : MultiHead(i)=Concat(Attention i,1 ,Attention i,2 ,Attention i,3 ,Attention i,4 ) W o Among them, Concat ( ) is the concatenation transformation function; j represents the j th attention head, j =(1, 2, 3, 4); W Q (i,j) , W K (i ,j) , W V (i,j) is the i th BiLSTM channel, the three learnable weight matrices of the j th attention head; d k ( i ) is an adaptive parameter, representing the output dimension / number of attention heads of the input i th BiLSTM, used to scale the dot product to stabilize training; W o is a learnable weight matrix; the attention scores are calculated by the matrix multiplication of the query and the key, and after being normalized by the softmax function, they are multiplied by the value matrix to obtain the attention output of the j th head; then the attention outputs of 4 heads are concatenated to obtain the multi - head attention output of the i th channel; finally, the multi - head attention outputs of three channels are concatenated to obtain the output sequence of the final multi - scale multi - head attention. The mathematical formula is: F fused =Concat(MultiHead(1), MultiHead(2), MultiHead (3)) 。

[0009] Further, in step five, all the network model parameters before the first bidirectional long short-term memory layer in the frozen source domain condition model include: all the parameters of the first three convolutional layers, SE module, fourth convolutional layer, pooling layer, and flattening layer in the frozen source domain condition model. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention, and thus should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0011] Figure 1 is the flowchart of the rotating machinery fault diagnosis provided by the present invention; Figure 2 is the structure diagram of the attention enhancement network (SE-CNN-BiLSTM-MSMHA); Figure 3 is the classification confusion matrix diagram of the attention enhancement network model using the transfer learning strategy under condition 1; Figure 4 is the classification confusion matrix diagram of the attention enhancement network model using the transfer learning strategy under condition 2; Figure 5 is the classification confusion matrix diagram of the attention enhancement network model using the transfer learning strategy under condition 3. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0012] To make the purpose, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the implementation cases and drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0013] As Figure 1 , an intelligent rotating machinery fault diagnosis method includes the following steps: Step 1: Obtain the vibration signals of the rotating machinery under the source domain conditions and the target domain conditions; Step 2: Denoise the vibration signals of the source domain working condition and the target domain working condition through the denoising processing module. The denoising processing module includes two sub-modules: the beluga optimization algorithm optimized variational mode decomposition module, and the TOPSIS (Technique for Order Preference by Similarity to an Ideal Solution) for variational mode decomposition screening and reconstruction module; Step 3: Perform short-time Fourier transform on the denoised vibration signals to obtain two-dimensional time-frequency images. Make a source domain dataset from the two-dimensional time-frequency images obtained from the source domain working condition, and make a target domain dataset from the two-dimensional time-frequency images obtained from the target domain working condition; Step 4: Construct an attention-enhanced network model based on SE-CNN-BiLSTM-MSMHA, and use the source domain dataset to complete the pre-training of the network model to obtain a source domain working condition model; Step 5: Through the transfer learning method, freeze all the network model parameters before the first bidirectional long short-term memory layer in the source domain working condition model, and fine-tune the parameters of the source domain working condition model with a small number of data samples in the target domain dataset to obtain a target domain working condition model; Step 6: Adopt the methods of Step 2 and Step 3. After denoising the vibration signals to be fault diagnosed and obtaining two-dimensional time-frequency images, input the time-frequency images into the fault diagnosis model of the corresponding working condition of the vibration signals to realize fault diagnosis.

[0014] Furthermore, in Step 2, in the denoising processing module, for the beluga optimization algorithm optimized variational mode decomposition sub-module, its decomposition steps are as follows: Step S21, set the penalty factor α and the decomposition layer number N the search ranges of these two parameters; Step S22, initialize the positions of the beluga group, and set the minimum envelope entropy as the fitness function, and then calculate the fitness value of each whale individual; Step S23, based on the beluga optimization algorithm, iteratively update the positions of the whale group until the preset termination condition is met, and then output the optimal N parameter combination of fitness value and α; Step S24, based on the optimal parameter combination, perform VMD (Variational Mode Decomposition) decomposition on the vibration signals to obtain N IMF components; For the TOPSIS for variational mode decomposition screening and reconstruction sub-module, its reconstruction steps are as follows: Step S25, for N intrinsic mode components, when N is even, remove the high-frequency components of N / 2 through the center frequency; when N is odd, remove the high-frequency components of ( N +1) / 2 through the center frequency; Step S26: Calculate three indicators of the remaining intrinsic mode components, namely, envelope entropy, kurtosis index, and correlation coefficient with the input signal; Step S27: Calculate the comprehensive scores of the three indicators using the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) method, and remove the component with the minimum score; Step S28: Add and reconstruct the remaining intrinsic mode components as the denoised signal.

[0015] Furthermore, in Step 4, the attention enhancement network of SE-CNN-BiLSTM-MSMHA has the following structure: First, the model construction starts from the input layer, and successively passes through three consecutive convolutional layers, a squeeze-and-excitation module, a fourth convolutional layer, a pooling layer, a flattening layer, and the first bidirectional long short-term memory layer; Second, reshape it into a one-dimensional sequence through a reshaping layer and connect the multi-scale multi-head attention module; Third, connect a flattening layer to flatten it, and after flattening, pass through two fully connected layers with the ReLU activation function and an output layer using softmax the activation function for fault classification and recognition; The attention enhancement network of SE-CNN-BiLSTM-MSMHA adopts a progressive processing framework of "feature extraction-channel feature enhancement-temporal modeling-multi-scale feature fusion" and a "parameter-adaptive multi-scale multi-head attention module", which can automatically learn and extract fault information from the original vibration signal, realize efficient and flexible fault diagnosis. At the same time, this model framework can greatly improve the anti-noise interference ability of the model. Even in a high-noise background, the model can maintain accurate fault diagnosis performance; The attention enhancement network of SE-CNN-BiLSTM-MSMHA adopts a collaborative strategy of convolutional layers and SE modules. It realizes spatial downsampling of two-dimensional time-frequency images through three-level convolution while maintaining the integrity of time-frequency local features. After the third convolutional layer, an SE module is connected. Global average pooling is used to capture channel statistics, and combined with a fully connected layer to generate channel attention weights for dynamic recalibration of multi-channel features, strengthening key fault features and suppressing noise interference; After the SE module, a fourth convolutional layer is connected to perform high-order non-linear transformation on the recalibrated channel features, ensuring that the features have been fully screened and enhanced before entering the BiLSTM layer in the temporal modeling stage, avoiding the problem of invalid feature interference in the temporal modeling in the traditional CNN-BiLSTM architecture; Among them, the multi-scale multi-head attention module adopts a three-channel parallel architecture, including: First, construct a layer group containing three groups of parallel bidirectional long short-term memory networks, where each bidirectional long short-term memory layer has different numbers of hidden neurons and all enable the full-sequence output mode; Second, each bidirectional long short-term memory layer is connected to a multi-head attention module with adaptive parameters. In the multi-head attention module, the number of attention heads is fixed at 4, and the dimension of the Key vector is dynamically calculated, where the dimension of the Key vector = the number of hidden units / the number of attention heads; Finally, adopt a depth concatenation strategy to fuse the feature of the attention outputs of the three groups to form three parallel processing streams; The mathematical formula of the multi-scale multi-head attention module is expressed as follows: First are three bidirectional long short-term memory networks: H i =BiLSTM i ( X ) Among them, BiLSTM i is the i th BiLSTM channel in the module, X is the input sequence, H i is the output sequence; Then, construct the query, key, value matrices and the attention score calculation formula for the subsequent attention mechanism; Query matrix Q i,j : Q i,j = H i W Q (i,j) Key matrix K i,j : K i,j = H i W K (i,j) Value matrix V i,j : V i,j = H i W V (i,j) Attention score Attention i,j : Attentioni,j =V i,j × softmax(Q i,j × K i,j ) / ( d k ( i )) 1 / 2 Multi-head attention score MultiHead(i) : MultiHead(i)=Concat(Attention i,1 ,Attention i,2 ,Attention i,3 ,Attention i,4 ) W o Among them, Concat ( ) is the splicing transformation function; j represents the j th attention head, j =(1, 2, 3, 4); W Q (i,j) , W K (i ,j) , W V (i,j) is the i th BiLSTM channel, the three learnable weight matrices of the j th attention head; d k ( i ) is an adaptive parameter, representing the output dimension / number of attention heads of the input i th BiLSTM, used to scale the dot product to stabilize training; W o is a learnable weight matrix; the attention score is calculated by the matrix multiplication of the query and the key, and after being normalized by the softmax function, it is multiplied by the value matrix to obtain the attention output of the j th head; then the attention outputs of 4 heads are spliced to obtain the multi-head attention output of the i th channel; finally, the multi-head attention outputs of three channels are spliced to obtain the output sequence of the final multi-scale multi-head attention. The mathematical formula is: F fused =Concat(MultiHead(1), MultiHead(2), MultiHead (3)) ; The multi-scale multi-head attention module adopts a dynamic adaptive multi-scale architecture strategy. The number of hidden units in the three-channel BiLSTM adopts different neuron configurations, forming a multi-scale observation system with fine-medium-coarse time granularity. This design can capture both local state transition features and long-range dependencies simultaneously; the adaptive Key dimension design, where Key vector dimension = number of hidden units / number of attention heads, enables the parameter scale of the multi-head attention module in each channel to match the feature complexity self-adaptively, capable of capturing feature details at different scales, ensuring that the model can make full use of multi-scale information; the combination of the two greatly improves the robustness and generalization ability of the model under complex working conditions, enabling it to accurately diagnose fault types under complex working conditions.

[0016] Furthermore, in step five, all network model parameters before the first bidirectional long short-term memory layer in the frozen source domain working condition model are frozen, including: all parameters of the first three convolutional layers, SE module, fourth convolutional layer, pooling layer, and flattening layer in the frozen source domain working condition model; the hierarchical freezing strategy is adopted to retain the general time-frequency feature extraction ability learned from the source domain data, such as basic features like edges and textures, to avoid the degradation of low-level features caused by insufficient target domain data volume, allowing the BiLSTM layer and its subsequent MSMHA module to adjust parameters, enabling the network model to adaptively learn the temporal dynamic features under different working conditions, especially to capture the periodic fault features in vibration signals; through the hierarchical freezing strategy, while maintaining the stability of feature extraction, the model can also enhance the adaptability of temporal modeling, and solve the "catastrophic forgetting" problem caused by global parameter adjustment in traditional transfer learning, keeping the diagnostic accuracy stable when the working conditions change.

[0017] Next, the method of the present invention will be described by taking a specific rotating machinery fault case as an example.

[0018] Example 1: Taking the CWRU dataset as an example, the data information of the CWRU dataset is shown in Table 1.

[0019]

[0020] The specific implementation process is as follows: (1) Obtain the vibration signals of the rotating machinery under the source domain working condition (working condition 0) and the target domain working conditions (working conditions 1, 2, 3). To simulate the strong noise environment in actual work, Gaussian noise with an intensity of 15 dB is added to the obtained vibration signals; (2) Perform noise reduction processing on the vibration signals of the source domain working condition and the target domain working conditions through a noise reduction processing module. The noise reduction processing module includes two sub-modules: the beluga optimization algorithm optimized variational mode decomposition module, and the technique for order preference by similarity to ideal solution (TOPSIS) for variational mode decomposition screening and reconstruction module; (3) Perform short-time Fourier transform on the denoised vibration signal to obtain a two-dimensional time-frequency image. Create a source domain dataset using the two-dimensional time-frequency images obtained from the source domain working conditions, and create a target domain dataset using the two-dimensional time-frequency images obtained from the target domain working conditions; (4) Construct an attention-enhanced network model based on SE-CNN-BiLSTM-MSMHA. Use the source domain dataset to complete the pre-training of the network model to obtain a source domain working condition model. The network model structure is as Figure 2 ; (5) Through the transfer learning method, freeze all the network model parameters before the first bidirectional long short-term memory layer in the source domain working condition model, and fine-tune the parameters of the source domain working condition model with a small number of data samples in the target domain dataset to obtain a target domain working condition model; (6) Adopt the methods in steps two and three. After denoising the vibration signal to be fault diagnosed and obtaining a two-dimensional time-frequency image, input the time-frequency image into the fault diagnosis model corresponding to the working condition of the vibration signal to achieve fault diagnosis. The denoising method, diagnosis model, and their corresponding number information are shown in Table 2. The comparison of the final diagnosis results is shown in Tables 3 and 4. The classification confusion matrices of the attention-enhanced network model under various working conditions are as Figure 3 、 Figure 4 、 Figure 5 。

[0021]

[0022]

[0023]

[0024] The above is only the preferred implementation manner of the present invention and is not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for rotating machinery, characterized in that, It includes the following steps: Step 1: Obtain the vibration signals of the rotating machinery under the source domain working conditions and the target domain working conditions; Step 2: Denoise the vibration signals of the source domain working conditions and the target domain working conditions through a denoising processing module. The denoising processing module includes two sub-modules: a beluga optimization algorithm optimized variational mode decomposition sub-module, and an ideal solution sorting method for variational mode decomposition screening and reconstruction sub-module; Step 3: Perform short-time Fourier transform on the denoised vibration signals to obtain two-dimensional time-frequency images. Make a source domain dataset through the two-dimensional time-frequency images obtained from the source domain working conditions, and make a target domain dataset through the two-dimensional time-frequency images obtained from the target domain working conditions; Step 4: Construct an attention-enhanced network model based on SE-CNN-BiLSTM-MSMHA, and use the source domain dataset to complete the pre-training of the network model to obtain a source domain working condition model; Step 5: Through the transfer learning method, freeze all the network model parameters before the first bidirectional long short-term memory layer in the source domain working condition model, and fine-tune the parameters of the source domain working condition model with a small number of data samples in the target domain dataset to obtain a target domain working condition model; Step 6: Adopt the methods of Step 2 and Step 3. After denoising the vibration signals to be fault diagnosed and obtaining two-dimensional time-frequency images, input the time-frequency images into the fault diagnosis model corresponding to the working conditions of the vibration signals to achieve fault diagnosis.

2. The intelligent rotating machinery fault diagnosis method according to claim 1, wherein, In Step 2, in the denoising processing module, for the beluga optimization algorithm optimized variational mode decomposition sub-module, its decomposition steps are: Step S21, set the penalty factor α and the number of decomposition levels N The search ranges of these two parameters; Step S22: Initialize the positions of the beluga group, and set the minimum envelope entropy as the fitness function, and then calculate the fitness values of each whale individual; Step S23, iteratively update the positions of the whales based on the beluga optimization algorithm until the preset termination condition is met, and then output the optimal N parameter combination of and α; Step S24, based on the optimal parameter combination, perform VMD decomposition on the vibration signal to obtain N IMF components; For the ideal solution sorting method for variational mode decomposition screening and reconstruction sub-module, its reconstruction steps are: Step S25, for N intrinsic mode components, when N is even, remove the N / 2 high-frequency components through the center frequency; when N is odd, remove (( N +1) / 2 high-frequency components through the center frequency; Step S26: Calculate three indexes of the remaining intrinsic mode components: namely, envelope entropy, kurtosis index, and correlation coefficient with the input signal; Step S27: Calculate the comprehensive score of the three indexes by using the ideal solution sorting method, and remove the component with the minimum score; Step S28: Reconstruct the remaining intrinsic mode components by adding them together as the denoised signal.

3. An intelligent rotating machinery fault diagnosis method according to claim 1, characterized in that, In Step 4, for the attention-enhanced network of SE-CNN-BiLSTM-MSMHA, its structure is: First, the construction of the model starts from the input layer, and successively passes through three consecutive convolutional layers, a squeeze-and-excitation module, a fourth convolutional layer, a pooling layer, a flattening layer, and a first bidirectional long short-term memory layer; Second, reshape it into a one-dimensional sequence through a reshaping layer and connect a multi-scale multi-head attention module; Next, connect a flattening layer to flatten, and after flattening, perform fault classification and recognition through two fully connected layers with ReLU activation functions and an output layer using softmax an activation function; among them, the multi-scale multi-head attention module adopts a three-channel parallel architecture, including: First, construct a layer group containing three groups of parallel bidirectional long short-term memory networks, where each bidirectional long short-term memory layer has different numbers of hidden neurons and all enable the full-sequence output mode; Second, each bidirectional long short-term memory layer is connected to a parameter-adaptive multi-head attention module. In the multi-head attention module, the number of attention heads is fixed at 4, and the dimension of the Key vector is dynamically calculated. The dimension of the Key vector = the number of hidden units / the number of attention heads; Finally, adopt a deep splicing strategy to fuse the features of the three groups of attention outputs to form three parallel processing streams; The mathematical formula of the multi-scale multi-head attention module is as follows: First, there are three bidirectional long short-term memory networks: H i =BiLSTM i ( X ) Among them, BiLSTM i is the i th BiLSTM channel in the module, X is the input sequence, H i is the output sequence; Next is to construct the query, key, value matrices and the attention score calculation formula for the subsequent attention mechanism; Query matrix Q i,j : Q i,j = H i W Q (i,j) Key matrix K i,j : K i,j = H i W K (i,j) Value matrix V i,j : V i,j = H i W V (i,j) Attention score Attention i,j : Attention i,j =V i,j × softmax(Q i,j × K i,j ) / ( d k ( i )) 1 / 2 Multi-head attention score MultiHead(i) : MultiHead(i)=Concat(Attention i,1 ,Attention i,2 ,Attention i,3 ,Attention i,4 )W o Among them, Concat () is the splicing transformation function; j represents the j th attention head, j =(1,2,3,4); W Q (i,j) , W K (i,j) , W V (i,j) is the i th BiLSTM channel, and the three learnable weight matrices of the j th attention head; d k ( i ) is an adaptive parameter, representing the output dimension / number of attention heads of the input i th BiLSTM, which is used to scale the dot product to stabilize the training; W o is a learnable weight matrix; the attention scores are calculated by the matrix multiplication of the query and the key, and after being normalized by the softmax function, they are multiplied by the value matrix to obtain the attention output of the j th head; then the attention outputs of the 4 heads are spliced to obtain the multi-head attention output of the i th channel; finally, the multi-head attention outputs of the three channels are spliced to obtain the output sequence of the final multi-scale multi-head attention. The mathematical formula is: F fused =Concat(MultiHead(1),MultiHead(2),MultiHead(3)) .

4. The intelligent rotating machinery fault diagnosis method according to claim 1, characterized in that In step five, all network model parameters before the first bidirectional long short-term memory layer in the frozen source domain condition model, including: all parameters of the first three convolutional layers, SE module, fourth convolutional layer, pooling layer, and flattening layer in the frozen source domain condition model.

Images