DNN bearing fault diagnosis method based on joint optimization of multi-modal feature fusion and alpha evolutionary algorithm

Abstract

The invention belongs to the technical field of mechanical equipment fault diagnosis, and particularly relates to a DNN bearing fault diagnosis method based on multi-modal feature fusion and alpha evolutionary algorithm joint optimization, which comprises the following steps: firstly, processing vibration signals in different states into three different images of a two-dimensional time domain grey-scale map, a two-dimensional frequency domain spectrogram and a two-dimensional time frequency map; respectively extracting a time domain feature, a frequency domain feature and a time frequency feature by using the strong feature extraction capability of the CNN; then, the extracted three features are used for generating feature representations with better distinction by utilizing the powerful feature fusion capability of KPCA, and parameters and structures of DNN are optimized by adopting an alpha evolutionary algorithm; and finally, training the DNN model by using the fault feature set, and carrying out fault diagnosis on the bearing by using the trained model. On the basis, the influence of noise in the vibration signal on the fault diagnosis accuracy is effectively reduced, and the bearing fault diagnosis accuracy can be remarkably improved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of mechanical equipment fault diagnosis, in particular to a bearing fault diagnosis method based on multi-modal feature fusion and alpha evolutionary algorithm optimization of DNN. BACKGROUND

[0002] Bearing is a core mechanical component of industrial equipment. Under the current advocacy of promoting intelligent operation and maintenance, bearing fault diagnosis is still a key link to ensure stable operation of equipment and a strategic support for the development of industrial intelligence. According to statistics, bearing faults account for 40% in rotating machinery. Therefore, bearing fault diagnosis technology has become the key to ensuring the reliable operation of rotating machinery, and its importance is increasingly prominent.

[0003] Bearing fault diagnosis is crucial to ensure the safety of equipment operation and reduce maintenance costs. In traditional bearing fault diagnosis, excessive reliance on prior knowledge often leads to inaccurate diagnosis, which in turn causes equipment safety accidents and unnecessary economic losses, seriously restricting the widespread application of fault diagnosis technology in complex industrial environments. SUMMARY

[0004] To solve the technical problem of inaccurate diagnosis due to excessive reliance on prior knowledge in traditional bearing fault diagnosis, the application provides the following technical solutions:

[0005] A bearing fault diagnosis method based on multi-modal feature fusion and alpha evolutionary algorithm optimization of DNN includes the following specific steps:

[0006] Step S1, data preprocessing: process the vibration signals under different states into two-dimensional time-domain gray images, two-dimensional frequency spectrum images and two-dimensional time-frequency images Figure Three Different images;

[0007] Step S2, CNN feature extraction: based on step S1, the powerful feature extraction capability of CNN is used to extract time-domain features, frequency domain features and time-frequency features respectively;

[0008] Step S3, feature fusion and parameter optimization: the three features extracted in step S2 are used to generate more distinctive feature representations using the powerful feature fusion capability of KPCA, and the alpha evolutionary algorithm is used to optimize the parameters and structure of DNN;

[0009] Step S4: fault diagnosis: first, use the fault feature set to train the DNN model, and then use the trained model to diagnose the bearing fault.

[0010] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein the specific steps of step S1 are as follows:

[0011] Step S11, draw the time domain gray scale map;

[0012] Step S12, draw the frequency domain log spectrum;

[0013] Step S13, draw the time-frequency map.

[0014] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein the specific steps of step S11 are as follows:

[0015] Step S111, process the time domain signal:

[0016] The time domain amplitude mapping method is used to construct the two-dimensional visual representation of the vibration signal;

[0017]

[0018] The preprocessed multi-channel vibration data matrix X is , wherein N is the number of sampling points, C is the number of sensor channels, and first, global normalization processing is performed;

[0019] Step S112, draw the gray scale map:

[0020] Through the gray scale image mapping technology, the time series data is converted into a two-dimensional space representation: the horizontal axis represents the time dimension (t=n / Fs), the vertical axis corresponds to the sensor channel number, and the gray value intensity is determined by the normalized amplitude; and the viridis color mapping scheme is used in the image generation process, and after gray scale conversion, an 8-bit depth image is formed.

[0021] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein the specific steps of step S12 are as follows:

[0022] Step S121, process the frequency domain signal:

[0023] The two-dimensional frequency domain log spectrum map is constructed by using the signal analysis method based on real fast Fourier transform;

[0024]

[0025] Wherein F represents the Fourier transform operator, and f is the frequency variable;

[0026] Step S122, draw the log spectrum map:

[0027] By taking the modulus |X i (f) | extracting the spectral amplitude information to eliminate the influence of phase difference on visualization; secondly, dynamic range compression and numerical stability processing are carried out;

[0028]

[0029] In the formula, ε= It is a smoothing factor to prevent overflow of logarithmic operation caused by zero value.

[0030] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein: the specific steps of step S13 are as follows:

[0031] Step S131, the two-dimensional time-frequency graph is decomposed in time-frequency by continuous wavelet transform on the first single-channel signal, and the wavelet coefficient amplitude is calculated in the scale-time plane by using Morlet wavelet;

[0032] Step S132, map it to the pseudo-color graph of the frequency-time grid, and the color intensity reflects the local energy distribution of the signal.

[0033] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein: in step S2, the specific steps of feature extraction are as follows:

[0034] Step S21, the original feature map of 128x128 is operated by the first convolutional layer to become the feature map of 128x128 Figure 1 ;

[0035] Step S22, operated by the first pooling layer to become the feature map of 64x64 Figure 2 ;

[0036] Step S23, operated by the second convolutional layer to become the feature map of 32x32 Figure 3 ;

[0037] Step S24, operated by the second pooling layer to become the feature map of 16x16 Figure 4 ;

[0038] Step S25, the multiple feature maps output by the pooling layer are unfolded into a one-dimensional vector through the full connection layer.

[0039] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein: in step S3, the specific processing process of multi-modal feature fusion is as follows:

[0040] Step SA31, normalize the extracted texture and edge features;

[0041] Step SA32, decompose the principal components of the equalized image using KPCA technology;

[0042] Step SA33, extract the principal components of the three features, which can realize the histogram matching between the feature image and the corresponding principal component;

[0043] Step S34, after matching, replace the corresponding principal component with the extracted feature;

[0044] Step S35, reconstruct the multi-modal feature through KPCA inverse transformation;

[0045] In the step S3, the specific processing process of the alpha evolution algorithm is as follows:

[0046] Step SB31, alpha operation;

[0047] Step SB32, boundary constraint;

[0048] Step SB33, selection strategy.

[0049] As a preferred scheme of the multi-modal feature fusion and alpha evolution algorithm based on the DNN bearing fault diagnosis method, the specific steps of the step SB31 are as follows:

[0050] Step SB311, obtain the evolution matrix:

[0051] The candidate matrix is sampled with replacement to obtain the evolution matrix;

[0052]

[0053] The symbol "→" represents a sampling operation with a replacement mechanism, which will be repeated N times; wherein N represents the population size; the candidate matrix M is used to store the current candidate solution, and the evolution matrix H is used to generate a new solution, and When the jth solution H j in the evolution matrix is successfully optimized, the solution in the corresponding index position in the candidate matrix is replaced and updated;

[0054] Step SB312, efficient search with alpha operator:

[0055]

[0056] Wherein H j represents the jth evolution solution; Z j and U j represent the jth sampled solution from M, and satisfy f(Zj )≤f(H j )≤f(U j ); t represents the number of iterations; Q represents the basis vector; alpha represents the attenuation coefficient; theta represents the control parameter; represents the jth random step;

[0057] Step SB313: obtaining an adaptive basis vector:

[0058] The double computer mechanism defined by the basis vector determines the initial point of the evolution process;

[0059]

[0060] wherein diagonal represents the diagonal element extraction operation, matrix A is a G-order square matrix; matrix B is a K*G dimensional structure; G represents the problem size; omega represents a non-negative weight; the weight parameter of the jth solution in B is ;

[0061] Step SB314, calculating the weight parameter :

[0062]

[0063] wherein j:K represents the jth solution obtained from the K solutions sampled, f(M j∶K ) represents the objective function value of returning M j∶K ;

[0064] Step SB315, constructing an evolution path:

[0065] Two different evolution paths are constructed for Q;

[0066]

[0067] wherein Q b represents the evolution path b; Q c represents the evolution path c; d b represents the learning rate of Q b .

[0068] As a preferred scheme of the DNN bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization, the specific steps of the step SB32 are as follows:

[0069] The alpha evolution algorithm framework adopts the distance semi-mapping strategy as the boundary processing core mechanism;

[0070]

[0071] wherein ub represents the upper limit; lb represents the lower limit.

[0072] As a preferred scheme of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN of the application, wherein the specific steps of step SB33 are as follows:

[0073] The selection operator is used as the evolution rule for determining the replacement of the parent and child populations in the evolution algorithm.

[0074] .

[0075] Compared with the prior art:

[0076] The original vibration signal is processed into time domain gray scale image, frequency domain log spectrum image and time-frequency image by using time domain amplitude mapping method, signal analysis method based on real number fast Fourier transform and continuous wavelet transform respectively; the ability of CNN to extract features and the ability of KPCA to fuse features are combined to generate more distinctive feature representation; and the alpha evolution algorithm is used to optimize the imperfect DNN parameters, which effectively reduces the influence of noise in the vibration signal on the fault diagnosis accuracy and can significantly improve the accuracy of bearing fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0077] Figure 1 is the implementation process schematic diagram of the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN in the embodiment of the application;

[0078] Figure 2 is the time domain gray scale image, frequency domain log spectrum image and time-frequency image generated in the embodiment of the application;

[0079] Figure 3 is the implementation process schematic diagram of CNN feature extraction in the embodiment of the application;

[0080] Figure 4 is the implementation process schematic diagram of the alpha evolution algorithm in the embodiment of the application;

[0081] Figure 5 is the diagnosis result image after diagnosis by the fault diagnosis model in the embodiment of the application. DETAILED DESCRIPTION

[0082] In order to make the purpose, technical scheme and advantages of the application clearer, the embodiments of the application will be further described in detail below with reference to the drawings.

[0083] Embodiment one

[0084] As Figure 1As shown, the application provides a bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm optimization DNN, which comprises the following specific steps:

[0085] Step S1, data preprocessing: the vibration signals in different states are processed into two-dimensional time domain gray maps, two-dimensional frequency domain spectrum maps and two-dimensional time-frequency Figure Three Different images Figure 2 As shown);

[0086] The specific steps of the step S1 are as follows:

[0087] Step S11, draw the time domain gray map;

[0088] The specific steps of the step S11 are as follows:

[0089] Step S111, process the time domain signal:

[0090] The two-dimensional visual representation of the vibration signal is constructed by using the time domain amplitude mapping method;

[0091]

[0092] The preprocessed multi-channel vibration data matrix X , wherein N is the number of sampling points, and C is the number of sensor channels, and first, global normalization processing is performed;

[0093] Step S112, draw the gray map:

[0094] Through the gray image mapping technology, the time series data is converted into a two-dimensional space representation: the horizontal axis represents the time dimension (t=n / Fs), the vertical axis corresponds to the sensor channel number, and the gray value intensity is determined by the normalized amplitude (black corresponds to the minimum value, and white corresponds to the maximum value); and the viridis color mapping scheme is used in the image generation process, and an 8-bit depth image is formed after gray conversion.

[0095] Step S12, draw the frequency domain log spectrum;

[0096] The specific steps of the step S12 are as follows:

[0097] Step S121, process the frequency domain signal:

[0098] The construction process of the two-dimensional frequency domain log spectrum map adopts a signal analysis method based on real fast Fourier transform;

[0099]

[0100] Wherein F represents the Fourier transform operator, and f is the frequency variable;

[0101] Step S122, draw the log spectrum map:

[0102] By taking the modulus of the transform result |X i (f) Extracting the spectral amplitude information to eliminate the influence of phase difference on visualization; secondly, dynamic range compression and numerical stability processing are carried out;

[0103]

[0104] In the formula, ε= is a smoothing factor to prevent overflow of logarithmic operation caused by zero value.

[0105] Step S13, drawing a time-frequency graph;

[0106] The specific steps of the step S13 are as follows:

[0107] Step S131, a two-dimensional time-frequency graph is obtained by time-frequency decomposition of the first channel single-channel signal through continuous wavelet transform (CWT), and the wavelet coefficient amplitude is calculated on the scale-time plane by using Morlet wavelet.

[0108] Step S132, mapping it into a pseudo-color graph of a frequency-time grid, and the color intensity reflects the local energy distribution of the signal.

[0109] Step S2, CNN feature extraction: on the basis of step S1, the time domain features, frequency domain features and time-frequency features are extracted by using the powerful feature extraction capability of CNN.

[0110] In the step S2, the specific steps of feature extraction are as follows: Figure 3

[0111] Step S21, the original feature map of 128x128 is operated through the first convolutional layer to become the feature map of 128x128 Figure 1 .

[0112] Step S22, operated through the first pooling layer to become the feature map of 64x64 Figure 2 .

[0113] Step S23, operated through the second convolutional layer to become the feature map of 32x32 Figure 3 .

[0114] Step S24, operated through the second pooling layer to become the feature map of 16x16 Figure 4 .

[0115] Step S25, the multiple feature maps output by the pooling layer are unfolded into a one-dimensional vector through the full connection layer.

[0116] ​Step S3, feature fusion and parameter optimization: the three features extracted in step S2 are fused into a more discriminative feature representation using the powerful feature fusion capability of KPCA, and the parameters and structure of the DNN are optimized using the alpha evolutionary algorithm;

[0117] In step S3, the specific processing procedure of the multi-modal feature fusion is as follows:

[0118] Step SA31, normalization processing is performed on the extracted texture and edge features;

[0119] Step SA32, the principal components of the equalized image are decomposed using the KPCA technique;

[0120] Step SA33, the principal components of the three features are extracted, and histogram matching between the feature images and the corresponding principal components can be achieved;

[0121] Step S34, after the matching is completed, the corresponding principal components are replaced with the extracted features;

[0122] Step S35, the multi-modal features are reconstructed through KPCA inverse transformation;

[0123] In step S3, the specific processing procedure of the alpha evolutionary algorithm is as follows:

[0124] Step SB31, alpha operation;

[0125] The specific steps of step SB31 are as follows:

[0126] Step SB311, obtain the evolution matrix:

[0127] The candidate matrix is sampled with replacement to obtain the evolution matrix;

[0128]

[0129] The symbol "→" represents a sampling operation with a replacement mechanism, which will be repeated N times; where N represents the population size; the candidate matrix M is used to store the current candidate solution, and the evolution matrix H is used to generate a new solution, and When the jth solution H j is successfully optimized, the solution at the corresponding index position in the candidate matrix is updated and replaced;

[0130] Step SB312, efficient search is performed using the alpha operator:

[0131]

[0132] Where H j represents the jth evolution solution; Z j and U jdenotes the jth sampled solution from M and satisfies f(Z j ) ≤ f(H j ) ≤ f(U j ); t denotes the iteration number; Q denotes the basis vector; a denotes the attenuation coefficient; and q denotes the control parameter; denotes the jth random step;

[0133] Step SB313: obtaining an adaptive basis vector:

[0134] The double-computing mechanism defined by the basis vector determines the initial point of the evolution process;

[0135]

[0136] where diagonal denotes the diagonal element extraction operation, matrix A is a G-order matrix (generated by sampling G times with replacement from the candidate solution set M); matrix B is a KxG structure (generated by sampling K times without replacement from M, where K = ⌈N x rand(0, 1)⌉ and satisfies 1 ≤ K ≤ N); G denotes the problem size; and w denotes the non-negative weight; the weight parameter of the jth solution in B is ;

[0137] Step SB314: calculating the weight parameter :

[0138]

[0139] where j:K denotes the jth solution obtained from the sampled K solutions, f(M j∶K ) denotes the objective function value returned by M j∶K ;

[0140] Step SB315: constructing an evolution path:

[0141] Two different evolution paths are constructed for Q;

[0142]

[0143] where Q b denotes the evolution path b; Q c denotes the evolution path c; d b denotes the learning rate of Q b .

[0144] Step SB32: boundary constraint;

[0145] The specific steps of the step SB32 are as follows:

[0146] The alpha evolution algorithm framework adopts a distance semi-mapping strategy as the core mechanism for boundary processing;

[0147]

[0148] wherein ub denotes the upper limit; lb denotes the lower limit.

[0149] Step SB33, selecting a strategy;

[0150] The specific steps of the step SB33 are as follows:

[0151] The selection operator is selected as an evolution rule for determining the replacement of the parent and child populations in the evolution algorithm.

[0152] .

[0153] Step S4: fault diagnosis: first, the DNN model is trained using the fault feature set, and then the trained model is used for fault diagnosis of the bearing.

[0154] Example two

[0155] The data set of this embodiment adopts the Fuyang Normal University data set. The categories are divided into normal state, rolling body fault, inner ring fault, outer ring fault and composite fault. The operating conditions are set to 1200 rpm (load 0%) and 1800 rpm (load 50%). Therefore, this task contains 10 files, each file corresponding to one working condition. Based on 1024000 data points, 205 samples are constructed, each sample consisting of 5000 data points. Each file contains 4 columns of signals, representing: 1- vibration signal of one-way acceleration sensor; 2, 3, 4- vibration signals of three-way acceleration sensor: x, y, z. The sampling frequency is 10240 Hz. And the training set, the validation set and the test set are divided according to the ratio of 8:1:1. See Table 1 for details.

[0156] In the alpha evolution algorithm, the population size is 30, the problem size is 30, the upper limit is 1.5, the lower limit is -1.5, and the decay coefficient is 0.5. The flow chart of the alpha evolution algorithm is shown in Figure 4 .

[0157] Table 1 Fault type category information of Fuyang Normal University data set

[0158] The diagnosis results are shown in Figure 5 , wherein the horizontal axis is the original label of the fault type, and the vertical axis is the label after model diagnosis. It can be seen that the classification accuracy of all labels is very high, and the accuracy rate is more than 90%.

[0159] In summary, the bearing fault diagnosis method based on multi-modal feature fusion and alpha evolution algorithm of the above-mentioned embodiments respectively utilizes the time-domain amplitude mapping method, the signal analysis method based on real fast Fourier transform, the continuous wavelet transform to process the original vibration signal into time-domain gray image, frequency-domain log spectrum image and time-frequency image and other advanced images; and combines the feature extraction capability of CNN and the feature fusion capability of KPCA to generate more distinctive feature representation; and aiming at the problem of imperfect DNN parameters, the alpha evolution algorithm is used for optimization; effectively reduces the influence of noise in the vibration signal on the fault diagnosis accuracy, and can significantly improve the accuracy of bearing fault diagnosis.

[0160] Although the present application has been described with reference to the embodiments above, various modifications can be made to it without departing from the scope of the present application, and equivalent components thereof can be substituted. In particular, the features of the embodiments disclosed in the present application can be combined with each other in any manner as long as there is no structural conflict, and the combinations thereof are not exhaustively described in the present specification only for the purpose of omitting the length and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, characterized in that, The specific steps are as follows: Step S1, Data preprocessing: The vibration signals under different states are processed into three different images: two-dimensional time-domain grayscale image, two-dimensional frequency-domain spectrum image, and two-dimensional time-frequency image. Step S2, CNN feature extraction: Based on step S1, the powerful feature extraction capability of CNN is used to extract time-domain features, frequency-domain features, and time-frequency features respectively; Step S3, Feature Fusion and Parameter Optimization: The three features extracted in step S2 are used to generate more discriminative feature representations using the powerful feature fusion capability of KPCA, and the parameters and structure of the DNN are optimized using the Alpha Evolution algorithm. Step S4: Fault diagnosis: First, the DNN model is trained using the fault feature set, and then the trained model is used to diagnose bearing faults.

2. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 1, is characterized in that... The specific steps of step S1 are as follows: Step S11: Draw a time-domain grayscale image; Step S12: Plot the logarithmic spectrum in the frequency domain; Step S13: Draw a time-frequency diagram.

3. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 2, is characterized in that... The specific steps of step S11 are as follows: Step S111, process the time-domain signal: A two-dimensional visualization representation of the vibration signal is constructed using the time-domain amplitude mapping method; ； For the preprocessed multi-channel vibration data matrix X∈ Where N is the number of sampling points and C is the number of sensor channels, global normalization is performed first; Step S112, draw a grayscale image: By using grayscale image mapping technology, time series data is transformed into a two-dimensional spatiotemporal representation: the horizontal axis represents the time dimension (t=n / Fs), the vertical axis corresponds to the sensor channel number, and the grayscale intensity is determined by the normalized amplitude; and the Viridis color mapping scheme is used in the image generation process, and an 8-bit depth image is formed after grayscale conversion.

4. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 2, is characterized in that... The specific steps of step S12 are as follows: Step S121, process the frequency domain signal: The construction of the two-dimensional frequency domain logarithmic spectrum adopts a signal analysis method based on real number fast Fourier transform; ； Where F represents the Fourier transform operator, and f is the frequency variable; Step S122, plot the logarithmic spectrum: By taking the modulus |X of the transformation result i (f) Extract spectral amplitude information to eliminate the impact of phase differences on visualization; then perform dynamic range compression and numerical stability processing. ； In the formula ε= It acts as a smoothing factor to prevent zero values ​​from causing overflow in logarithmic operations.

5. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 2, is characterized in that... The specific steps of step S13 are as follows: Step S131: The two-dimensional time-frequency graph is decomposed into time-frequency values ​​of the first single-channel signal by continuous wavelet transform, and the wavelet coefficient amplitude is calculated in the scale-time plane using Morlet wavelet. Step S132: Map it as a pseudo-color map of a frequency-time grid, where the color intensity reflects the local energy distribution of the signal.

6. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 1, is characterized in that... In step S2 The specific steps for feature extraction are as follows: Step S21: The original 128×128 feature map is processed through the first convolutional layer to become a 128×128 feature map 1; Step S22: After the first pooling layer operation, a 64×64 feature map 2 is obtained; Step S23: After the second convolutional layer operation, the feature map becomes 32×32. Step S24: After the second pooling layer operation, the feature map becomes 16×16; Step S25: Expand the multiple feature maps output by the pooling layer into a one-dimensional vector through a fully connected layer.

7. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 1, is characterized in that... In step S3, the specific processing procedure for multimodal feature fusion is as follows: Step SA31: Normalize the extracted texture and edge features; Step SA32: Use KPCA to decompose the principal components of the equalized image; Step SA33 extracts the principal components of the three features, which enables histogram matching between the feature image and the corresponding principal components; Step S34: After matching is completed, the corresponding principal components are replaced with the extracted features; Step S35: Reconstruct multimodal features using inverse KPCA transform; In step S3, the specific processing procedure of the Alpha Evolutionary Algorithm is as follows: Step SB31, Alpha operation; Step SB32, boundary constraints; Step SB33, select a strategy.

8. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 7, is characterized in that... The specific steps of step SB31 are as follows: Step SB311, obtain the evolution matrix: The candidate matrix is ​​sampled with replacement to obtain the evolution matrix; ； The symbol "→" indicates a sampling operation with a replacement mechanism, which will be repeated N times; where N represents the population size; the candidate matrix M is used to store the current candidate solutions, and the evolution matrix H is used to generate new solutions. When the j-th solution H in the evolution matrix j Upon successful optimization, the solution at the corresponding index position in the candidate matrix will be replaced and updated; Step SB312, perform an efficient search using the α operator: ； Among them, H j Z represents the solution of the j-th evolution; j and U j Let f(Z) be the solution sampled from M for the j-th time, and satisfy f(Z) j )≤f(H j )≤f(U j ); t represents the number of iterations; Q represents the basis vector; α represents the decay coefficient; θ represents the control parameter; Indicates the j-th random step size; Step SB313: Obtain the adaptive basis vectors: The dual computation mechanism defined by the basis vectors determines the initial point of the evolutionary process; ； Where, diagonal represents the diagonal element extraction operation, matrix A is a G-order square matrix; matrix B is a K×G dimensional structure; G represents the problem size; ω represents the non-negative weight; and the weight parameter of the j-th solution in B is... ; Step SB314: Calculate the weight parameters : ； Where j∶K represents the j-th solution obtained from the K sampled solutions, f(M j∶K ) indicates that M is returned. j∶K The objective function value; Step SB315, construct the evolutionary path: Two different evolutionary paths were constructed for Q; ； Among them, Q b Indicates evolutionary path b; Q c Indicates evolutionary path c;d b Q represents b The learning rate.

9. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 7, is characterized in that... The specific steps of step SB32 are as follows: The Alpha Evolutionary algorithm framework employs a distance semi-mapping strategy as the core mechanism for boundary handling. ； Where ub represents the upper limit and lb represents the lower limit.

10. The bearing fault diagnosis method based on multimodal feature fusion and alpha evolution algorithm jointly optimized by DNN, as described in claim 7, is characterized in that... The specific steps of step SB33 are as follows: Selection operators are used as evolutionary rules in evolutionary algorithms to determine the succession of parent and offspring populations; 。

Images