Multi-modal data fusion-based rotor fault diagnosis task-oriented method

Abstract

A multi-modal data fusion-based rotor fault diagnosis task-oriented method comprises the following steps of S1, removing original vibration data noise, and converting de-noised vibration data into a recurrence plot; s2, dividing the de-noised vibration data and the corresponding recurrence plot into a training set and a test set; s3, pre-training a time domain feature extraction module and a recursive feature extraction module by using the training set; s4, extracting time domain features and recursive features of the training set, and training a multi-modal feature fusion and fault diagnosis module; s5, inputting the vibration signals of the test set and the corresponding recursion plots into a time domain feature extraction module and a recursion feature extraction module respectively, and extracting time domain features and recursion features; s6, inputting the features extracted in S5 into a multi-modal feature fusion and fault diagnosis module, and outputting a fault diagnosis result; according to the invention, rotor fault diagnosis is carried out by using the multi-modal data of the fault vibration signal time domain and recursion characteristics, and the method has the advantages of high feasibility, high precision and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of rotor fault diagnosis, and in particular to a method for rotor fault diagnosis tasks based on multi-modal data fusion. Background Art

[0002] At present, rotating machinery is increasingly used in industrial production, and the degradation and failure of rotating machinery are receiving more and more attention. As one of the main types of rotating machinery, the reliability and stability of the rotor are crucial to ensuring production efficiency and safety.

[0003] The patent document with the authorization announcement number CN115081483B discloses a method for diagnosing rotor faults of a hydro-turbine generator based on feature selection and GWO-BP. The method first extracts the mathematical statistical features and frequency domain features of the vibration signal data, then uses the GWO-BP classifier model to select features with higher classification accuracy and lower dimension, and finally inputs them into the GWO-BP model classifier to complete the diagnosis of the rotor fault. This method and other technologies are very dependent on the frequency domain feature extraction effect of the fault vibration signal. When facing signals with strong nonlinearity and complexity, the diagnostic effect will be greatly affected. For complex vibration signals in actual working conditions, it is impossible to fully characterize the characteristics of different fault states by relying solely on statistical features and frequency domain features. Therefore, the present invention proposes a method for diagnosing rotor faults by using a multimodal data fusion model to analyze the multidimensional features of rotor faults. Through the time-frequency analysis of the rotor vibration signal and the transformation of two-dimensional data, the time domain and recursive characteristics of the vibration signal are integrated and analyzed, so that the results of rotor fault diagnosis have better stability and applicability, providing a feasible new solution for intelligent fault diagnosis. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for rotor fault diagnosis tasks based on multimodal data fusion, so as to realize the use of a multimodal data fusion model to analyze the multidimensional characteristics of rotor faults and diagnose rotor faults.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A method for rotor fault diagnosis based on multimodal data fusion, comprising the following steps:

[0007] S1, removing noise from the original fault vibration data, and converting the denoised vibration data into a recursive graph;

[0008] S2, dividing the denoised vibration data and the corresponding recursive graph into a training set and a test set;

[0009] S3, using the training set to pre-train the time domain feature extraction module and the recursive feature extraction module;

[0010] S4, inputting the training set into the pre-trained time domain feature extraction module and recursive feature extraction module in S3, extracting time domain features and recursive features, and training the multimodal feature fusion and fault diagnosis module;

[0011] S5, inputting the denoised vibration data of the test set and the corresponding recursive graph into the time domain feature extraction module and the recursive feature extraction module respectively, extracting the time domain features and the recursive features;

[0012] S6. Input the time domain features and recursive features extracted in S5 into the multimodal feature fusion and fault diagnosis module, and output the fault diagnosis results.

[0013] The above step S1 includes the following steps:

[0014] S1.1. Use wavelet threshold filtering with different wavelet bases and decomposition levels to denoise the original fault vibration data, calculate the signal-to-noise ratio of the denoised vibration data, and select the wavelet threshold filtering with the highest signal-to-noise ratio to complete data denoising;

[0015] S1.2, select a suitable threshold based on prior experience and convert the denoised vibration data into a recursive graph;

[0016] S1.3. Save each denoised vibration data and the corresponding recursive graph, and add the corresponding fault label.

[0017] In the above step S2, the denoised vibration data and the corresponding recursive graph of each fault vibration data are taken as a fault sample, and the fault sample is divided into a training set and a test set according to a set ratio.

[0018] The above step S3 includes the following steps:

[0019] S3.1, divide the training set in S2 into a pre-training set and a validation set according to a certain ratio;

[0020] S3.2. The step of obtaining and screening IMF components after EEMD of the input signal is used as the time domain feature extraction module, the denoised vibration data in the pre-training set is input, the sample entropy and the coefficient of variation of the sample entropy of each IMF component after EEMD are calculated, the first few IMF components with the smallest coefficient of variation are selected and input into SVM for training, the first few IMF components with the highest fault diagnosis accuracy in the verification set are obtained, and the pre-training of the time domain feature extraction module is completed; the calculation expression of the coefficient of variation is as follows:

[0021]

[0022] Among them, CV i represents the coefficient of variation of the ith IMF component, μ i and σ iRespectively represent the mean and standard deviation of the coefficient of variation of the i-th IMF component, j∈(1,2,…,N) represents the number of samples, represents the sample entropy value of the i-th IMF component in the j-th sample;

[0023] S3.3, construct a residual space attention network as a recursive feature extraction module, train the recursive feature extraction module with the recursive graph in the pre-training set, save the network weight file with the highest diagnostic accuracy of the validation set during the training process of the recursive feature extraction module, and complete the pre-training of the recursive feature extraction module;

[0024] S3.4. The established residual space attention network is:

[0025] Input the recursive graph into the network → Perform a 10*10 convolution operation on the input image → Perform two 3*3 convolution operations on the extracted feature map → Perform ReLU activation after each convolution → Concatenate the feature maps obtained by the first and third convolutions in the channel dimension → Perform ReLu activation to obtain the residual feature map → Perform maximum pooling and global average pooling operations along the channel axis → Concatenate the two extracted feature maps → Perform a 7*7 convolution operation → Perform Sigmoid activation → Generate a spatial attention map → Apply the spatial attention map to the residual feature map → Perform a 2*2 maximum pooling operation → Flatten the feature map into one-dimensional data → Perform a full connection operation → Perform a batch normalization operation → Perform a Dropout operation → The classification layer outputs the fault classification.

[0026] The above step S4 includes the following steps:

[0027] S4.1, input the denoised vibration data of the training set in S2 to the time domain feature extraction module, and select the first few IMF components obtained by pre-training in S3.2 as the time domain features of the fault samples;

[0028] S4.2, load the pre-trained weights in S3.3 into the recursive feature extraction module, input the recursive graph of the training set in S2, and use the output of the batch normalization operation in the recursive feature extraction module as the recursive feature of the fault sample;

[0029] S4.3. A multi-core support vector machine constructed by a custom kernel function including a Gaussian kernel and a Sigmoid kernel is used as a multimodal feature fusion and fault diagnosis module. The time domain features and recursive features of the fault samples in the training set in S4.1 and S4.2 are spliced ​​and the multimodal feature fusion and fault diagnosis module is trained. Cross-validation is used to optimize parameters, and the optimal parameters with the best accuracy are saved to complete the training of the multimodal feature fusion and fault diagnosis module.

[0030] S4.4. In the kernel function, when the input feature vector is x i and x jWhen , the Gaussian kernel expression is:

[0031]

[0032] in, You can scale the vector x i and x j The distance between ||x i -x j || determines the degree of fit of the Gaussian kernel and can be defined as the bandwidth parameter G r ;

[0033] The Sigmoid kernel expression is:

[0034]

[0035] Among them, β and θ are scaling parameters and constant offsets respectively. The scaling parameters control the input vector The degree of scaling determines the degree of fitting of the kernel function, which can be defined as the bandwidth parameter G s ;

[0036] The combination of kernel functions in multi-core support vector machines is:

[0037] K(x i ,x j )=(1-α)K r (x i ,x j )+αK s (x i ,x j ), 0≤γ≤1;

[0038] Among them, K r and K s Represent Gaussian kernel and Sigmoid kernel respectively, α represents the weight coefficient;

[0039] The cross-validation parameters in S4.5 and S4.3 include selecting the fusion weight coefficient α of the two kernel functions, the bandwidth parameter G of the Gaussian kernel and the Sigmoid kernel r and G s , the penalty coefficient C of multi-core support vector machine.

[0040] In the above step S5, the denoised vibration data of the test set and the corresponding recursive graph are respectively input into the time domain feature extraction module and the recursive feature extraction module, and steps S4.1 and S4.2 are repeated to obtain the time domain features and recursive features of the fault samples in the test set.

[0041] The above step S6 includes the following steps:

[0042] S6.1, load the optimal parameters of the multimodal feature fusion and fault diagnosis module trained in S4.3;

[0043] S6.2, concatenate the time domain features and recursive features of the fault samples in the test set in S5, input them into the multimodal feature fusion and fault diagnosis module, and obtain the diagnosed fault results.

[0044] The present invention provides a method for rotor fault diagnosis based on multimodal data fusion, which has the following beneficial effects:

[0045] 1. The present invention constructs a residual space attention network to fully extract the recursive characteristics of the fault vibration signals in key areas of different scales of the recursive graph, such as periodicity, complexity, and state variability.

[0046] 2. The present invention calculates the coefficient of variation of the entropy of different IMF samples and selects more and more stable fault information as the time domain characteristics of the fault signal, thereby effectively reducing the interference of irrelevant characteristic signals on the time domain characteristic fault diagnosis.

[0047] 3. The present invention utilizes Gaussian kernel function and Sigmoid kernel function to perform high-dimensional mapping on time domain and recursive features respectively, thereby improving the ability of kernel function to capture nonlinear features in two modal data, and finding the optimal feature fusion and fault diagnosis parameters through cross-validation, thereby optimizing single-modal fault diagnosis and making multi-modal feature fusion fault classification more stable and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0049] Figure 1 It is a flowchart of a method for rotor fault diagnosis task based on multimodal data fusion;

[0050] Figure 2 A time domain diagram of denoised vibration data and a corresponding recursive diagram for each fault category of the present invention;

[0051] Figure 3 This is a flow chart of pre-training of the time domain feature extraction module of the present invention;

[0052] Figure 4 This is a structural diagram of the residual space attention network of the present invention;

[0053] Figure 5 It is a training flow chart of the multimodal feature fusion and fault diagnosis module of the present invention;

[0054] Figure 6 It is a confusion matrix diagram of the fault diagnosis results of the test set of the present invention. DETAILED DESCRIPTION

[0055] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and embodiments.

[0056] like Figure 1 As shown in , a method for rotor fault diagnosis task based on multimodal data fusion is presented. The specific steps are as follows:

[0057] S1. Collecting experimental data of the rotor in the mechanical equipment under normal conditions and fault conditions as original fault vibration data, removing noise from the original fault vibration data, and converting the denoised vibration data into a recursive graph;

[0058] S2, dividing the denoised vibration data and the corresponding recursive graph into a training set and a test set;

[0059] S3, using the training set to pre-train the time domain feature extraction module and the recursive feature extraction module;

[0060] S4, inputting the training set into the pre-trained time domain feature extraction module and recursive feature extraction module in S3, extracting time domain features and recursive features, and training the multimodal feature fusion and fault diagnosis module;

[0061] S5, inputting the denoised vibration data of the test set and the corresponding recursive graph into the time domain feature extraction module and the recursive feature extraction module respectively, extracting the time domain features and the recursive features;

[0062] S6. Input the time domain features and recursive features extracted in S5 into the multimodal feature fusion and fault diagnosis module, and output the fault diagnosis results.

[0063] The decomposition steps of step S1 are:

[0064] S1.1. In this example, the experimental data is collected from a rotating machinery experimental system, which consists of a rotor test bench, a speed controller, a front-end machine, and a computer. The data set uses rotor vibration data in four states: normal, rubbing, unbalanced, and misaligned. During signal acquisition, the rotor speed is 1200r / min, the sampling frequency is 2048Hz, the sampling time is 1s, and there are 45 sets of data for different rotor states, for a total of 180 sets of samples.

[0065] S1.2. Use wavelet threshold filtering with different wavelet bases and decomposition layers to denoise the original fault vibration data, calculate the signal-to-noise ratio of the denoised vibration data, and select the 'db8' wavelet base with the best denoising effect and 3-layer decomposition to denoise the original rotor signal. The five common wavelet functions and their wavelet bases are shown in the following table.

[0066] Wavelet Function Wavelet basis Orthogonality Continuous Wavelet Transform Daubechies ['db1', 'db4'] have Can Coiflets ['coif1', 'coif5'] have Can Symlets ['sym2', 'sym4'] have Can Discrete Meyer ['dmey'] have Can't Biorthogonal ['bior3.3', 'bior5.5'] none Can

[0067] S1.3, convert the denoised vibration data into a recursive graph, and set the threshold of the rotor to 0.05 based on prior knowledge;

[0068] S1.4. Save each denoised vibration data and the corresponding recursive graph, and add the corresponding fault label. The specific denoised vibration data and its recursive graph are as follows: Figure 2 shown.

[0069] In step S2, each denoised vibration data and corresponding recursive graph obtained in S1 is taken as a fault sample and divided into a training set and a test set. The training set has a total of 100 samples, of which each state contains 25 samples; the other samples are used as a test set, which contains a total of 80 samples.

[0070] The decomposition steps of step S3 are:

[0071] S3.1, divide the training set in S2 into a pre-training set and a validation set in a ratio of 8:2;

[0072] S3.2, the step of obtaining and screening the IMF components after performing EEMD on the input signal is used as the time domain feature extraction module, and input into S3.1 to obtain the denoised vibration data in the training set, and calculate the sample entropy and the coefficient of variation of each IMF component. The calculation expression of the coefficient of variation is as follows:

[0073]

[0074] Among them, CV i represents the coefficient of variation of the ith IMF component, μ i and σ i Respectively represent the mean and standard deviation of the coefficient of variation of the i-th IMF component, j∈(1,2,…,N) represents the number of samples, represents the sample entropy value of the i-th IMF component in the j-th sample;

[0075] S3.3. Select the first few IMF components with the smallest coefficient of variation and input them into SVM for training. When the first 5, 6, 7, 8, and 9 IMF components are selected respectively, the verification set can achieve the highest accuracy and complete the pre-training of the module. The specific flowchart of module pre-training is as follows: Figure 3 As shown;

[0076] S3.4. Construct a residual space attention network as a recursive feature extraction module, and use S3.1 to obtain the recursive graph in the training set to train it. Set the training loss function to the cross entropy function, and use the Adam optimization algorithm to update the network weights. Set the batch size (Batch Size) to 6 and the training round (Epoch) to 15. In order to stabilize the convergence trend of the model, find the global optimal solution faster, and prevent the overfitting of the model. Define the callback function so that it can automatically adjust the learning rate and stop early during the training process, save the network weight file with the highest diagnostic accuracy of the verification set during the network training process, and complete the pre-training of the recursive feature extraction module;

[0077] S3.4. The established residual space attention network is:

[0078] Input recursive graph into the network → Perform 10*10 convolution operation on the input image → Perform two 3*3 convolution operations on the extracted feature map → Perform ReLU activation after each convolution → Concatenate the feature maps obtained by the first and third convolutions in the channel dimension → Perform ReLu activation to obtain the residual feature map → Perform maximum pooling and global average pooling operations along the channel axis → Concatenate the two extracted feature maps → Perform 7*7 convolution operation → Perform Sigmoid activation → Generate spatial attention map → Apply the spatial attention map to the residual feature map → Perform 2*2 maximum pooling operation → Flatten the feature map into one-dimensional data → Perform full connection operation → Perform batch normalization operation → Perform Dropout operation → The classification layer outputs fault classification. The specific structure diagram of the residual spatial attention network is as follows: Figure 4 shown.

[0079] The decomposition steps of step S4 are:

[0080] S4.1. Input the denoised vibration data of the training set in S2 into the time domain feature extraction module. Considering that the model has fewer parameters and faster recognition speed while ensuring the classification accuracy, the first five IMF components are selected as the time domain features of the fault samples.

[0081] S4.2, load the pre-trained weights in S3.3 into the recursive feature extraction module, input the recursive graph of the training set in S2, and use the output of the batch normalization operation in the recursive feature extraction module as the recursive feature of the fault sample;

[0082] S4.3. The multi-core support vector machine constructed by the custom kernel function containing Gaussian kernel and Sigmoid kernel is used as the multimodal feature fusion and fault diagnosis module. The time domain features and recursive features of the training sets in S4.1 and S4.2 are spliced ​​and the multimodal feature fusion and fault diagnosis module is trained. The parameter grid is optimized by five-fold cross validation, and the optimal parameters with the best accuracy are saved to complete the multimodal feature fusion and fault diagnosis module training. The specific flowchart of module training is shown in the figure. Figure 5 As shown;

[0083] S4.4. In the kernel function, when the input feature vector is x i and x j When , the Gaussian kernel expression is:

[0084]

[0085] in, You can scale the vector x i and x j The distance between ||x i -x j || determines the degree of fit of the Gaussian kernel and can be defined as the bandwidth parameter G r .

[0086] The Sigmoid kernel expression is:

[0087]

[0088] Among them, β and θ are scaling parameters and constant offsets respectively. The scaling parameters control the input vector The degree of scaling determines the degree of fitting of the kernel function, which can be defined as the bandwidth parameter G s .

[0089] The combination of kernel functions in multi-core support vector machines is:

[0090] K(x i ,x j )=(1-α)K r (x i ,x j )+αK s (x i ,x j ),0≤γ≤1

[0091] Among them, K r and K s They represent Gaussian kernel and Sigmoid kernel respectively, and α represents the weight coefficient.

[0092] The grid parameters for the five-fold cross validation in S4.5 and S4.3 include the selection of the fusion weight coefficient α of the two kernel functions, the bandwidth parameter G of the Gaussian kernel and the Sigmoid kernel. r and G s , the penalty coefficient C of the multi-core support vector machine. The specific parameter grid and the optimal parameters are shown in the following table:

[0093]

[0094] In step S5, the denoised vibration data and recursive graph of the test set are respectively input into the time domain feature extraction module and the recursive feature extraction module, and steps S4.1 and S4.2 are repeated to obtain the time domain and recursive features of the fault sample.

[0095] The decomposition steps of step S6 are:

[0096] S6.1, load the optimal parameters of the multimodal feature fusion and fault diagnosis module trained in S4.3;

[0097] S6.2, the time domain features and recursive features of the test set in S5 are spliced ​​and input into the multimodal feature fusion and fault diagnosis module, and the accuracy of the fault diagnosis of the test set reaches 100%. The specific diagnosis results are as follows Figure 6 .

[0098] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A method for rotor fault diagnosis based on multimodal data fusion, characterized in that: The following steps are involved: S1, removing noise from the original fault vibration data, and converting the denoised vibration data into a recursive graph; S2, dividing the denoised vibration data and the corresponding recursive graph into a training set and a test set; S3, using the training set to pre-train the time domain feature extraction module and the recursive feature extraction module; S4, inputting the training set into the pre-trained time domain feature extraction module and recursive feature extraction module in S3, extracting time domain features and recursive features, and training the multimodal feature fusion and fault diagnosis module; S5, inputting the denoised vibration data of the test set and the corresponding recursive graph into the time domain feature extraction module and the recursive feature extraction module respectively, extracting the time domain features and the recursive features; S6. Input the time domain features and recursive features extracted in S5 into the multimodal feature fusion and fault diagnosis module, and output the fault diagnosis results.

2. The method for rotor fault diagnosis based on multimodal data fusion according to claim 1, characterized in that: The step S1 comprises the following steps: S1.

1. Use wavelet threshold filtering with different wavelet bases and decomposition levels to denoise the original fault vibration data, calculate the signal-to-noise ratio of the denoised vibration data, and select the wavelet threshold filtering with the highest signal-to-noise ratio to complete data denoising; S1.2, select a suitable threshold based on prior experience and convert the denoised vibration data into a recursive graph; S1.

3. Save each denoised vibration data and the corresponding recursive graph, and add the corresponding fault label.

3. The method for rotor fault diagnosis based on multimodal data fusion according to claim 1, characterized in that: In the step S2, the denoised vibration data and the corresponding recursive graph of each fault vibration data are taken as a fault sample, and the fault sample is divided into a training set and a test set according to a set ratio.

4. The method for rotor fault diagnosis based on multimodal data fusion according to claim 1 is characterized in that: The step S3 comprises the following steps: S3.1, divide the training set in S2 into a pre-training set and a validation set according to a certain ratio; S3.

2. The step of obtaining and screening IMF components after EEMD of the input signal is used as the time domain feature extraction module, the denoised vibration data in the pre-training set is input, the sample entropy and the coefficient of variation of the sample entropy of each IMF component after EEMD are calculated, the first few IMF components with the smallest coefficient of variation are selected and input into SVM for training, the first few IMF components with the highest fault diagnosis accuracy in the verification set are obtained, and the pre-training of the time domain feature extraction module is completed; the calculation expression of the coefficient of variation is as follows: Among them, CV i represents the coefficient of variation of the ith IMF component, μ i and σ i Respectively represent the mean and standard deviation of the coefficient of variation of the i-th IMF component, j∈(1,2,…,N) represents the number of samples, represents the sample entropy value of the i-th IMF component in the j-th sample; S3.3, construct a residual space attention network as a recursive feature extraction module, train the recursive feature extraction module with the recursive graph in the pre-training set, save the network weight file with the highest diagnostic accuracy of the validation set during the training process of the recursive feature extraction module, and complete the pre-training of the recursive feature extraction module; S3.

4. The established residual space attention network is: Input the recursive graph into the network → Perform a 10*10 convolution operation on the input image → Perform two 3*3 convolution operations on the extracted feature map → Perform ReLU activation after each convolution → Concatenate the feature maps obtained by the first and third convolutions in the channel dimension → Perform ReLu activation to obtain the residual feature map → Perform maximum pooling and global average pooling operations along the channel axis → Concatenate the two extracted feature maps → Perform a 7*7 convolution operation → Perform Sigmoid activation → Generate a spatial attention map → Apply the spatial attention map to the residual feature map → Perform a 2*2 maximum pooling operation → Flatten the feature map into one-dimensional data → Perform a full connection operation → Perform a batch normalization operation → Perform a Dropout operation → The classification layer outputs the fault classification.

5. The method for rotor fault diagnosis based on multimodal data fusion according to claim 4 is characterized in that: The step S4 comprises the following steps: S4.1, input the denoised vibration data of the training set in S2 to the time domain feature extraction module, and select the first few IMF components obtained by pre-training in S3.2 as the time domain features of the fault samples; S4.2, load the pre-trained weights in S3.3 into the recursive feature extraction module, input the recursive graph of the training set in S2, and use the output of the batch normalization operation in the recursive feature extraction module as the recursive feature of the fault sample; S4.

3. A multi-core support vector machine constructed by a custom kernel function including a Gaussian kernel and a Sigmoid kernel is used as a multimodal feature fusion and fault diagnosis module. The time domain features and recursive features of the fault samples in the training set in S4.1 and S4.2 are spliced ​​and the multimodal feature fusion and fault diagnosis module is trained. Cross-validation is used to optimize parameters, and the optimal parameters with the best accuracy are saved to complete the training of the multimodal feature fusion and fault diagnosis module. S4.

4. In the kernel function, when the input feature vector is x i and x j When , the Gaussian kernel expression is: in, You can scale the vector x i and x j The distance between ||x i -x j || determines the degree of fitting of the Gaussian kernel and can be defined as the bandwidth parameter G r ; The Sigmoid kernel expression is: Among them, β and θ are scaling parameters and constant offsets respectively. The scaling parameters control the input vector x i T ·x j The degree of scaling determines the degree of fitting of the kernel function, which can be defined as the bandwidth parameter G s ; The combination of kernel functions in multi-core support vector machines is: K(x i ,x j )=(1-α)K r (x i ,x j )+αK s (x i ,x j ),0≤γ≤1; Among them, K r and K s Represent Gaussian kernel and Sigmoid kernel respectively, α represents the weight coefficient; The cross-validation parameters in S4.5 and S4.3 include selecting the fusion weight coefficient α of the two kernel functions, the bandwidth parameter G of the Gaussian kernel and the Sigmoid kernel r and G s , the penalty coefficient C of multi-core support vector machine.

6. The method for rotor fault diagnosis based on multimodal data fusion according to claim 5, characterized in that: In the step S5, the denoised vibration data of the test set and the corresponding recursive graph are respectively input into the time domain feature extraction module and the recursive feature extraction module, and steps S4.1 and S4.2 are repeated to obtain the time domain features and recursive features of the fault samples in the test set.

7. The method for rotor fault diagnosis based on multimodal data fusion according to claim 6, characterized in that: The step S6 comprises the following steps: S6.1, load the optimal parameters of the multimodal feature fusion and fault diagnosis module trained in S4.3; S6.2, concatenate the time domain features and recursive features of the fault samples in the test set in S5, input them into the multimodal feature fusion and fault diagnosis module, and obtain the diagnosed fault results.

Images