A Fault Diagnosis Method for Rotating Machinery Based on MP-CNN and SVM

Abstract

The present invention belongs to the field of rotating mechanical equipment fault diagnosis, and specifically relates to a rotating mechanical equipment fault diagnosis method based on MP-CNN and SVM, comprising: obtaining a rotating mechanical equipment image to be diagnosed, preprocessing the rotating mechanical equipment image to be diagnosed, inputting the preprocessed image into a rotating mechanical equipment fault diagnosis model based on MP-CNN and SVM, and obtaining a fault diagnosis result of the mechanical equipment; the present invention combines the strong feature extraction capability of the hierarchical idea applied in the field of mechanical fault diagnosis and the efficient classification function of SVM, and solves the problem of low diagnostic accuracy due to a large number and types of samples when using CNN-SVM.

Description

Technical Field

[0001] The invention belongs to the field of rotating mechanical equipment fault diagnosis, and in particular relates to a rotating mechanical equipment fault diagnosis method based on MP-CNN and SVM. Background Art

[0002] With the rapid development of today's industry and technology, factory equipment, mainly rotating mechanical equipment, has been in high-speed operation for a long time to meet people's production and living needs, and is very prone to failure. Such failures will cause great economic losses and safety threats to the factory. Therefore, accurate fault diagnosis and classification of rotating mechanical equipment is of great significance to maintaining the safety and stable operation of mechanical equipment.

[0003] Fault diagnosis methods can be divided into model-based, vibration signal-based and intelligent methods. Model-based diagnosis methods use mathematical models to achieve diagnosis from physical principles or system identification technology, but they often require a lot of prior knowledge as a prerequisite for the system, which will affect the accuracy of diagnosis. Vibration signal-based diagnosis methods mainly include time domain analysis and frequency domain analysis. The time domain analysis method extracts features from signals in the time domain, and then performs fault diagnosis based on the changes in the extracted features. It has a high diagnostic accuracy for the occurrence and severity of equipment faults, but because the position information in the time domain is not easy to extract, the diagnostic accuracy for the fault location is low. The frequency domain analysis method extracts features from the original signal through Fourier transform to the frequency domain, and then performs fault diagnosis based on the changes in the features. It has a high diagnostic accuracy for the occurrence and location of equipment faults, but cannot determine the severity of the fault. In recent years, intelligent fault diagnosis technology based on machine learning has been increasingly widely used in the diagnosis of rotating machinery. Hu Xiaoyi et al. proposed a deep convolutional neural network based on CNN-SVM to extract fault features from the original signal to achieve a high diagnostic accuracy for bearings of different types and damage degrees. (Hu Xiaoyi, Jing Yunjian, Song Zhikun, Hou Yinqing. Research on bearing fault identification based on deep convolutional neural network of CNN-SVM [J]. Vibration and Shock, 2019, 38 (18): 173-178.). By defining an appropriate kernel function, SVM can realize nonlinear transformation to transform the input space into a linearly separable high-dimensional space and find the optimal linear hyperplane of the space. It has a good effect in multi-classification problems, but the diagnostic accuracy decreases when facing a large number and type of samples. In addition, Kaige Su et al. proposed a fault diagnosis method based on branch convolutional neural network, which introduced the idea of ​​hierarchical diagnosis into the field of fault diagnosis, and can realize the three-level classification of bearing faults from coarse to fine. The classification layer with the highest level will provide prior knowledge for the later layer through feature learning, narrowing the data range that needs to be extracted for feature extraction in the later layer, thereby improving the diagnostic accuracy. (Su Kaige; Liu Jianhua; Xiong Hui Hierarchical diagnosis of bearing faults using branch convolutional neural network considering noise interference and variable working conditions [J] Knowledge-Based Systems, 2021). However, this method uses the Softmax classifier in all three levels of classification instead of the SVM with higher classification accuracy, and the weight distribution of the loss values ​​corresponding to the three levels of classification is not reasonable enough. Summary of the invention

[0004] In order to solve the problems existing in the above prior art, the present invention proposes a rotating mechanical equipment fault diagnosis method based on MP-CNN and SVM, the method comprising: acquiring an image of the rotating mechanical equipment to be diagnosed, preprocessing the image of the rotating mechanical equipment to be diagnosed, inputting the preprocessed image into a rotating mechanical equipment fault diagnosis model based on MP-CNN and SVM, and obtaining a fault diagnosis result of the mechanical equipment;

[0005] Training the rotating machinery fault diagnosis model based on MP-CNN and SVM includes:

[0006] S1: Obtain the original data set, preprocess the data in the original data set, and divide the preprocessed data into a training set and a test set;

[0007] S2: Input the images in the training set into the MP-CNN network for feature extraction to obtain three progressive feature maps;

[0008] S3: Input the three progressive feature maps into SVM respectively to obtain three progressive fault classification results;

[0009] S4: Calculate the loss function of each result according to the three progressive fault classification results, and use the wolf pack algorithm to optimize the attention mechanism to obtain the optimal weight distribution. Weight the three loss functions according to the optimal weight distribution to obtain the final loss function.

[0010] S5: Use back propagation to update the parameters of the model;

[0011] S6: When the final loss function value is the smallest, the model training is completed.

[0012] Preferably, the process of preprocessing the data in the original data set includes: filling the pixels of the image with the time domain signal of the center bearing fault in the original data set in sequence; randomly selecting a pixel with a length of K from the time domain signal; 2 The segmented sample is used as an input signal to obtain a picture with K×K pixels; the K×K picture is converted into a grayscale image; a non-negative matrix decomposition method embedded in an expectation-maximization iterative algorithm is used to perform preliminary feature extraction on the grayscale image, the preliminary features after decomposition are reconstructed to obtain an optimized grayscale image, and the pixel intensity of the optimized grayscale image is calculated.

[0013] Furthermore, the formula for the optimized grayscale image is:

[0014] V≈V'=WH

[0015] V' n-1 =W n-1 H n-1

[0016]

[0017]

[0018] Q H ×H n-1 →H n

[0019] Q W ×W n-1 →W n

[0020] V' n =W n H n

[0021] The iteration stops when the following objective function is satisfied:

[0022] ||VV′ n ||<ε

[0023] Among them, V represents the matrix corresponding to the grayscale image obtained in the previous step, V' represents the approximate matrix after decomposition, W represents the basis, H represents the linear combination coefficient, and V' n-1 represents the approximate matrix of the n-1th iteration, W n-1 represents the basis of the n-1th iteration, H n-1 represents the linear combination coefficient of the n-1th iteration, Q H Indicates H n-1 The correction coefficient matrix, Q W W n-1 The correction coefficient matrix, H n represents the linear combination coefficient of the nth iteration, W n represents the basis of the nth iteration, V' n represents the approximate matrix of the nth iteration; ε represents the preset target of the difference before and after matrix decomposition.

[0024] Furthermore, the formula for calculating the pixel intensity of the grayscale image is:

[0025]

[0026] Among them, P represents the pixel intensity of the grayscale image, δ represents the sample selection range, L(.) represents the length value of the segmented sample, MinL represents the minimum length value of the segmented sample, and Max L represents the maximum length value of the segmented sample.

[0027] Preferably, the process of processing the images in the training set using the MP-CNN network includes:

[0028] Step 1: Use the optimized grayscale image in the training set as the input of MP-CNN, use the convolution layer in BLOCK1 to perform the second feature extraction, and perform batch normalization processing in the BN layer, activation by the Relu activation function and pooling in the pooling layer. The pooled result is subjected to the Dropout layer, the fully connected layer and the SVM to achieve binary classification, and the first fault classification result is obtained. The first fault classification result is used as the input of the convolution layer in BLOCK2;

[0029] Step 2: Use the convolution layer in BLOCK2 to extract features from the output of BLOCK1 again, and use the same process to finally achieve four-classification of the output, and obtain the second fault classification result; use the second fault classification result as the input of the convolution layer in BLOCK3;

[0030] Step 3: Use the convolutional layer in BLOCK3 to extract features from the output of BLOCK2 again. After the same process as steps 1 and 2, the third fault classification result is obtained.

[0031] Preferably, the process of weighting the loss functions of the three feature maps using the attention mechanism optimized by the wolf pack algorithm includes:

[0032] Step 1: Initialize the relevant parameters of the wolf pack algorithm; randomly generate S artificial scout wolves, and calculate the fitness of the prey searched by each scout wolf;

[0033] X m =(x m1 ,x m2 ,…x md )

[0034] Y lead =Loss f

[0035] Among them, X m represents the initialized wolf pack, x md represents the position of the mth wolf in the d-dimensional space variable to be optimized, Y lead represents the prey fitness of the alpha wolf;

[0036] Step 2: The scout wolves perform wandering behavior in h directions according to the wandering step length, and when the fitness of a scout wolf is better than the current alpha wolf or the number of wandering times is greater than the predetermined maximum number, the scout wolf becomes the new alpha wolf; the fitness of each scout wolf is compared with the current alpha wolf, if the scout wolf has better fitness, it replaces the current alpha wolf as the new alpha wolf, otherwise the alpha wolf is not updated; the scout wolf position update formula is:

[0037]

[0038] in, represents the updated position of the mth wolf in the pth direction in the d-dimensional space variable to be optimized, p represents the pth direction, and h represents all directions in which the wolf can roam. represents the walking step length;

[0039] Step 3: After the wandering behavior ends, the alpha wolf howls to initiate a summoning behavior, quickly summoning the surrounding wolves to its location. The wolves quickly approach the alpha wolf with a running step length and search for prey. If there is a wolf with a higher fitness during the running process, it will replace the current alpha wolf. The position update formula of the wolves is:

[0040]

[0041] in, It represents the updated position of the nth wolf in the d-dimensional space variable to be optimized. It indicates the position of the nth wolf in the d-dimensional space variable to be optimized. Indicates the running stride length, Indicates the current position of the wolf leader. Indicates the direction in which the wolf is currently rushing;

[0042] Step 4: When the distance between the fierce wolf and the leader wolf is less than the set threshold, it switches to siege behavior; it runs towards the leader wolf according to the siege step length; if the fierce wolf has a higher fitness during the running process, it will replace the current leader wolf; the position update formula of the fierce wolf is:

[0043]

[0044]

[0045] in, It represents the updated position of the rth wolf in the d-dimensional space variable to be optimized. represents the position of the rth wolf in the d-dimensional space variable to be optimized, λ represents the step length coefficient, represents the siege step length, Indicates the current position of the leader wolf, that is, the current target position. Indicates the siege direction, max d and min d They represent the maximum and minimum values ​​in the value range of the dth variable to be found respectively;

[0046] Step 5: Update the wolf packs at different locations according to the above mechanism and determine the final position of the leader wolf, that is, output to the target value Loss f The smallest three weights.

[0047] Furthermore, the conditions that the leader wolf must satisfy for the update are:

[0048] Y m >Ylead

[0049] T m >T max

[0050] Among them, Y m represents the fitness of the mth wolf, T m represents the number of times the mth exploratory wolf wanders, T max The preset maximum number of walks.

[0051] Furthermore, the threshold is set as:

[0052] D ns <σ

[0053]

[0054] Among them, D ns represents the distance between the nth wolf and the current leader, σ represents the set threshold, D represents the total number of variables to be optimized, and ω represents the determination factor.

[0055] Furthermore, the attention mechanism formula is:

[0056]

[0057] Among them, Loss f is the weighted final loss value, a i is the weight, Loss i is the loss value of each layer.

[0058] Preferably, the final loss function expression is:

[0059]

[0060] Among them, N represents the amount of data in the training set, i represents the sample number, j represents the category number, k represents the number of weight rows, l represents the number of weight columns, λ represents the regularization parameter, and x represents the number of weights. i represents the i-th sample data, y i represents the correct label, W represents the weight set, f(x i ; W) represents the scores of different classification categories, f(x i ; W) yi represents the score of the jth category, and Δ represents the boundary value.

[0061] The present invention combines the strong feature extraction ability of the hierarchical idea in the field of mechanical fault diagnosis with the efficient classification function of SVM, and solves the problem of low diagnostic accuracy caused by the large number and types of samples when only CNN-SVM is used as mentioned in the background, and the problem of low diagnostic accuracy caused by only considering the hierarchical idea and using the Softmax classifier without reasonably allocating loss weights, thereby effectively solving the problem of low accuracy in mechanical fault diagnosis caused by noise interference and variable working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is the overall flow chart of the present invention;

[0063] Figure 2 It is a framework diagram of the fault diagnosis method of the present invention;

[0064] Figure 3 It is a structural diagram of the rotating machinery equipment fault diagnosis model based on MP-CNN and SVM of the present invention;

[0065] Figure 4 Flowchart for determining weights for optimizing the attention mechanism for the wolfpack algorithm. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] A method for fault diagnosis of rotating machinery equipment based on a multi-level progressive classification convolutional neural network (MP-CNN) combined with a support vector machine (SVM). By combining the application of hierarchical thinking in fault diagnosis and the high efficiency of SVM in solving multi-classification problems, the designed MP-CNN can realize three-level progressive fault classification of rotating machinery equipment from coarse to fine, and SVM outputs different types of classification respectively. At the same time, the attention mechanism is introduced to reasonably distribute the weights of the loss values ​​at the three-level classification, so as to obtain the final loss value of the model.

[0068] A method for fault diagnosis of rotating mechanical equipment based on MP-CNN and SVM, the method comprising: acquiring an image of the rotating mechanical equipment to be diagnosed, preprocessing the image of the rotating mechanical equipment to be diagnosed, inputting the preprocessed image into a fault diagnosis model of the rotating mechanical equipment based on MP-CNN and SVM, and obtaining a fault diagnosis result of the mechanical equipment.

[0069] like Figure 1As shown, the training of the rotating machinery equipment fault diagnosis model based on MP-CNN and SVM includes:

[0070] S1: Obtain the original data set, preprocess the data in the original data set, and divide the preprocessed data into a training set and a test set;

[0071] S2: Input the images in the training set into the MP-CNN network for feature extraction to obtain three progressive feature maps;

[0072] S3: Input the three progressive feature maps into SVM respectively to obtain three progressive fault classification results;

[0073] S4: Calculate the loss function of each result according to the three progressive fault classification results, and use the Wolf Pack Algorithm (WPA) to optimize the attention mechanism to find the optimal weight distribution. Then, weight the three loss functions to obtain the final loss function.

[0074] S5: Use back propagation to update the parameters of the model;

[0075] S6: When the final loss function value is the smallest, the model training is completed.

[0076] The process of preprocessing the data in the original data set includes: filling the pixels of the image with the time domain signal of the middle bearing fault in the original data set in sequence; randomly selecting a pixel with a length of K from the time domain signal; 2 The segmented sample is used as an input signal to obtain a K×K pixel image; the K×K image is converted into a grayscale image, and the pixel intensity of the grayscale image is calculated. The non-negative matrix factorization (NMF) embedded in the expectation-maximization (EM) iterative algorithm is used to perform preliminary feature extraction on the grayscale image, and the decomposed form is reconstructed, and the obtained optimized grayscale image is used as the input information for the subsequent steps. The formula for calculating the pixel intensity of the grayscale image is:

[0077]

[0078] Among them, P represents the pixel intensity of the grayscale image, δ represents the sample selection range, that is, δ=1,2,…K 2 , L(.) represents the length value of the segmented sample, Min L represents the minimum length value of the segmented sample, and MaxL represents the maximum length value of the segmented sample.

[0079] The NMF process formula embedded in the EM iterative algorithm is:

[0080] V≈V'=WH

[0081] V' n-1 =W n-1 H n-1

[0082]

[0083]

[0084] Q H ×H n-1 →H n

[0085] Q W ×W n-1 →W n

[0086] V' n =W n H n

[0087] Among them, V represents the matrix corresponding to the grayscale image obtained in the previous step, V' represents the approximate matrix after decomposition, W represents the basis, H represents the linear combination coefficient, and V' n-1 represents the approximate matrix of the n-1th iteration, W n-1 represents the basis of the n-1th iteration, H n-1 represents the linear combination coefficient of the n-1th iteration, Q H Indicates H n-1 The correction coefficient matrix, Q W W n-1 The correction coefficient matrix, H n represents the linear combination coefficient of the nth iteration, W n represents the basis of the nth iteration, V' n Represents the approximate matrix of iteration n.

[0088] The iteration stops when the following objective function is satisfied:

[0089] ||VV′ n ||<ε

[0090] Among them, ε represents the preset target of the difference before and after matrix decomposition.

[0091] like Figure 2 As shown in , the MP-CNN network includes three two-dimensional convolutional building blocks, each of which is used to extract features from the input image. Figure 3As shown in the figure, the two-dimensional convolution building block consists of a convolution layer, a batch normalization layer (BN), an activation layer, and a pooling layer. A Dropout layer, a fully connected layer, and a SVM are added after each building block to finally achieve multi-level progressive classification, in which the structure of the selected convolution kernel and the added bias structure are also shown.

[0092] 1024 data points in the original signal are selected as a group, and the pixels are converted into a 32×32 grayscale image. The third dimension with a depth of 1 is added to it, which can be regarded as a 32×32×1 three-dimensional input signal. The specific network parameters of CNN are shown in Table 1, where (3,2,16,16) of convolution layer 1 means: the convolution kernel size is 3×3, the moving step is 2, the number of convolution kernels is 16, and the bias is 16; (10,0.01) means: the penalty factor C in the Gaussian radial basis function (RBF) SVM is 10, and the kernel parameter γ is 0.01. The network uses the Relu activation function, the number of training iterations is 100, the minimum batch size is 100, the learning rate lr is 0.001, the Dropout is 0.75, and the training algorithm is Adam.

[0093] Table 1 MP-CNN related network parameters

[0094]

[0095]

[0096] For SVM multi-classification, the leaf loss function is introduced to evaluate the consistency between the output value of the input data after the model and the expected value, and the regularization penalty term is introduced to suppress the weights with large values, thereby improving the generalization of the model. The Loss function formula is as follows:

[0097]

[0098] Among them, N represents the amount of data in the training set, i represents the sample number, j represents the category number, k represents the number of weight rows, l represents the number of weight columns, λ represents the regularization parameter, and x represents the number of weights. i represents the i-th sample data, y i represents the correct label, W represents the weight set, f(x i ; W) represents the scores of different classification categories, f(x i ; W) yi represents the score of the jth category, and Δ represents the boundary value.

[0099] Since each layer of classification will cause different degrees of loss, and the loss of the entire model needs to be weighted, the attention mechanism optimized by the wolf pack algorithm is introduced to combine the losses of the three layers of classification with different weights to obtain the final loss value. The relevant parameters are also optimized and updated during the training process to improve the accuracy of the model. Figure 4 The figure shows the flow chart of the weight determination of the wolf pack algorithm optimization attention mechanism. The specific process is as follows:

[0100] Step 1: Initialize the parameters related to the wolf pack algorithm. Randomly generate S artificial scout wolves, calculate the fitness of the prey searched by each scout wolf, that is, obtain the three weight values ​​with the smallest final loss value, and select the scout with the best fitness as the current leader wolf;

[0101] X m =(x m1 ,x m2 ,…x md )

[0102] Y lead =Loss f

[0103] Among them, X m represents the initialized wolf pack, x md represents the position of the mth wolf in the d-dimensional space variable to be optimized, Y lead It indicates the prey fitness (the target value sought) of the leader wolf at this time.

[0104] Step 2: The scout wolf performs wandering behavior in h directions according to the wandering step length, and when the fitness of a scout wolf is better than the current alpha wolf or the number of wandering times is greater than the predetermined maximum number, the scout wolf becomes the new alpha wolf. The scout wolf position update formula is:

[0105]

[0106] in, represents the updated position of the mth wolf in the pth direction in the d-dimensional space variable to be optimized, p represents the pth direction, and h represents all directions in which the wolf can roam. Indicates the walking step length.

[0107] Then, the fitness of each current scout is compared with that of the current leader. If a scout has better fitness, it will replace the current leader as the new leader. Otherwise, the leader will not be updated. The leader update must meet the following conditions (any one):

[0108] Y m >Y lead

[0109] T m >T max

[0110] Among them, Y m represents the fitness of the mth wolf, T m represents the number of times the mth exploratory wolf wanders, T max The preset maximum number of walks.

[0111] Step 3: After the wandering behavior ends, the alpha wolf howls to initiate a summoning behavior, quickly summoning the surrounding wolves to their location. The wolves quickly approach the alpha wolf at a running pace and search for prey.

[0112]

[0113] in, It represents the updated position of the nth wolf in the d-dimensional space variable to be optimized. It indicates the position of the nth wolf in the d-dimensional space variable to be optimized. Indicates the running stride length, Indicates the current position of the leader wolf, that is, the current target position. Indicates the direction the wolf is currently charging. If there is a wolf with higher fitness during the charging process, it will replace the current leader.

[0114] Step 4: When the distance between the fierce wolf and the leader wolf is less than a certain threshold, they switch to siege behavior:

[0115] D ns <σ

[0116]

[0117] Among them, D ns represents the distance between the nth wolf and the current leader, σ represents the set threshold, D represents the total number of variables to be optimized, and ω represents the determination factor.

[0118] Run towards the leader wolf according to the siege step length. If a higher fitness is found during the run, the current leader wolf will be replaced.

[0119]

[0120]

[0121] in, It represents the updated position of the rth wolf in the d-dimensional space variable to be optimized. represents the position of the rth wolf in the d-dimensional space variable to be optimized, λ represents the step length coefficient, represents the siege step length, Indicates the current position of the leader wolf, that is, the current target position. Indicates the siege direction, max d and min d They respectively represent the maximum and minimum values ​​in the range of values ​​of the dth variable to be found.

[0122] Step 5: Update the wolf packs at different locations according to the above mechanism to determine the final position of the alpha wolf, that is, output to the target value Loss fThe smallest three weights.

[0123] The attention mechanism formula is as follows:

[0124]

[0125] Among them, Loss f is the weighted final loss value, a i is the weight, Loss i is the loss value of each layer.

[0126] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A rotating machinery equipment fault diagnosis method based on MP-CNN and SVM, characterized in that: include: Acquire an image of a rotating mechanical device to be diagnosed, preprocess the image of the rotating mechanical device to be diagnosed, input the preprocessed image into a rotating mechanical device fault diagnosis model based on MP-CNN and SVM, and obtain a fault diagnosis result of the mechanical device; Training the rotating machinery fault diagnosis model based on MP-CNN and SVM includes: S1: Obtain the original data set, preprocess the data in the original data set, and divide the preprocessed data into a training set and a test set; S2: Input the images in the training set into the MP-CNN network for feature extraction to obtain three progressive feature maps; specifically including: Step 1: Use the optimized grayscale image in the training set as the input of MP-CNN, use the convolution layer in BLOCK1 to perform the second feature extraction, and perform batch normalization processing in the BN layer, activation by the Relu activation function and pooling in the pooling layer. The pooled result is subjected to the Dropout layer, the fully connected layer and the SVM to achieve binary classification, and the first fault classification result is obtained. The first fault classification result is used as the input of the convolution layer in BLOCK2; Step 2: Use the convolution layer in BLOCK2 to extract features from the output of BLOCK1 again, and use the same process to finally achieve four-classification of the output, and obtain the second fault classification result; use the second fault classification result as the input of the convolution layer in BLOCK3; Step 3: Use the convolutional layer in BLOCK3 to extract features from the output of BLOCK2 again. After the same process as steps 1 and 2, the third fault classification result is obtained. S3: Input the three progressive feature maps into SVM respectively to obtain three progressive fault classification results; S4: Calculate the loss function of each result according to the three progressive fault classification results, and use the wolf pack algorithm to optimize the attention mechanism to obtain the optimal weight distribution. Weight the three loss functions according to the optimal weight distribution to obtain the final loss function. S5: Use back propagation to update the parameters of the model; S6: When the final loss function value is the smallest, the model training is completed.

2. A rotating machinery equipment fault diagnosis method based on MP-CNN and SVM according to claim 1, characterized in that: The process of preprocessing the data in the original data set includes: filling the pixels of the image with the time domain signal of the middle bearing fault in the original data set in sequence; randomly selecting a pixel with a length of K from the time domain signal; 2 The segmented sample is used as an input signal to obtain a picture with K×K pixels; the K×K picture is converted into a grayscale image; a non-negative matrix decomposition method embedded in an expectation-maximization iterative algorithm is used to perform preliminary feature extraction on the grayscale image, the preliminary features after decomposition are reconstructed to obtain an optimized grayscale image, and the pixel intensity of the optimized grayscale image is calculated.

3. A rotating machinery equipment fault diagnosis method based on MP-CNN and SVM according to claim 2, characterized in that: The formula for obtaining the optimized grayscale image is: V≈V'=WH V' n-1 =W n-1 H n-1 Q H ×H n-1 →H n Q W ×W n-1 →W n V' n =W n H n The iteration stops when the following objective function is satisfied: ||V-V′ n ||<ε Among them, V represents the matrix corresponding to the grayscale image obtained in the previous step, V' represents the approximate matrix after decomposition, W represents the basis, H represents the linear combination coefficient, and V' n-1 represents the approximate matrix of the n-1th iteration, W n-1 represents the basis of the n-1th iteration, H n-1 represents the linear combination coefficient of the n-1th iteration, Q H Indicates H n-1 The correction coefficient matrix, Q W W n-1 The correction coefficient matrix, H n represents the linear combination coefficient of the nth iteration, W n represents the basis of the nth iteration, V' n represents the approximate matrix of the nth iteration; ε represents the preset target of the difference before and after matrix decomposition.

4. The rotating machinery equipment fault diagnosis method based on MP-CNN and SVM according to claim 2 is characterized in that: The formula for calculating the pixel intensity of the grayscale image is: Among them, P represents the pixel intensity of the grayscale image, δ represents the sample selection range, L(.) represents the length value of the segmented sample, MinL represents the minimum length value of the segmented sample, and Max L represents the maximum length value of the segmented sample.

5. The method for fault diagnosis of rotating machinery equipment based on MP-CNN and SVM according to claim 1, characterized in that: The process of weighting the loss functions of the three feature maps using the attention mechanism optimized by the wolf pack algorithm includes: Step 1: Initialize the relevant parameters of the wolf pack algorithm; randomly generate S artificial scout wolves, and calculate the fitness of the prey searched by each scout wolf; X m =(x m1 ,x m2 ,…x md ) Y lead =Loss f Among them, X m represents the initialized wolf pack, x md represents the position of the mth wolf in the d-dimensional space variable to be optimized, Y lead represents the prey fitness of the alpha wolf; Step 2: The scout wolves perform wandering behavior in h directions according to the wandering step length, and when the fitness of a scout wolf is better than the current alpha wolf or the number of wandering times is greater than the predetermined maximum number, the scout wolf becomes the new alpha wolf; the fitness of each scout wolf is compared with the current alpha wolf, if the scout wolf has better fitness, it replaces the current alpha wolf as the new alpha wolf, otherwise the alpha wolf is not updated; the scout wolf position update formula is: in, represents the updated position of the mth wolf in the pth direction in the d-dimensional space variable to be optimized, p represents the pth direction, and h represents all directions in which the wolf can roam. represents the walking step length; Step 3: After the wandering behavior ends, the alpha wolf howls to initiate a summoning behavior, quickly summoning the surrounding wolves to its location. The wolves quickly approach the alpha wolf with a running step length and search for prey. If there is a wolf with a higher fitness during the running process, it will replace the current alpha wolf. The position update formula of the wolves is: in, It represents the updated position of the nth wolf in the d-dimensional space variable to be optimized. It indicates the position of the nth wolf in the d-dimensional space variable to be optimized. Indicates the running stride length, Indicates the current position of the wolf leader. Indicates the direction in which the wolf is currently rushing; Step 4: When the distance between the fierce wolf and the leader wolf is less than the set threshold, it switches to siege behavior; it runs towards the leader wolf according to the siege step length; if the fierce wolf has a higher fitness during the running process, it will replace the current leader wolf; the position update formula of the fierce wolf is: in, It represents the updated position of the rth wolf in the d-dimensional space variable to be optimized. represents the position of the rth wolf in the d-dimensional space variable to be optimized, λ represents the step length coefficient, represents the siege step length, Indicates the current position of the leader wolf, that is, the current target position. Indicates the siege direction, max d and min d They represent the maximum and minimum values ​​in the value range of the dth variable to be found respectively; Step 5: Update the wolf packs at different locations according to the above mechanism and determine the final position of the leader wolf, that is, output to the target value Loss f The smallest three weights.

6. A rotating machinery equipment fault diagnosis method based on MP-CNN and SVM according to claim 5, characterized in that: The conditions for the Wolf Leader update are: AND m -And lead T m >T max Among them, Y m represents the fitness of the mth wolf, T m represents the number of times the mth wolf wanders, T max The preset maximum number of walks.

7. The method for fault diagnosis of rotating machinery equipment based on MP-CNN and SVM according to claim 5, characterized in that: The thresholds are set as: D ns <s Among them, D ns represents the distance between the nth wolf and the current leader, σ represents the set threshold, D represents the total number of variables to be optimized, and ω represents the determination factor.

8. The method for fault diagnosis of rotating machinery equipment based on MP-CNN and SVM according to claim 5, characterized in that: The attention mechanism formula is: Among them, Loss f is the weighted final loss value, a i is the weight, Loss i is the loss value of each layer.

9. The method for fault diagnosis of rotating machinery equipment based on MP-CNN and SVM according to claim 1, characterized in that: The final loss function expression is: Among them, N represents the amount of data in the training set, i represents the sample number, j represents the category number, k represents the number of weight rows, l represents the number of weight columns, λ represents the regularization parameter, and x represents the number of weights. i represents the i-th sample data, y i represents the correct label, W represents the weight set, f(x i ; W) represents the scores of different classification categories, f(x i ; W) yi represents the score of the jth category, and Δ represents the boundary value.

Images