Mechanical equipment fault diagnosis method based on approximate evaluation of multi-wavelet feature cascade
By employing the feature cascading technique of a three-branch convolutional neural network and a multi-wavelet kernel function filter, the problem of mechanical equipment fault diagnosis under noise imbalance samples was solved, achieving efficient fault identification and improved accuracy.
Patent Information
- Application Number
- CN202410956515.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-17
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-07-17
AI Technical Summary
Existing intelligent diagnostic methods are difficult to effectively handle mechanical equipment fault diagnosis under noise imbalance samples, especially in the fault diagnosis of high-speed rotating parts in mechanical equipment. Noise interference causes fault information to be submerged, and sample imbalance leads to insufficient diagnostic accuracy and robustness.
A three-branch convolutional neural network model is adopted, combined with multi-wavelet kernel function filters and feature cascading techniques. Signal denoising, high and low frequency analysis, and multi-scale analysis are performed through Gaussian difference, wavelet kernel function layers, and Laplacian wavelet kernel function layers. The optimal cascading result is adaptively selected for fault diagnosis using a self-supervised wavelet kernel weight matrix and a two-distribution approximation evaluation method.
It effectively reduces preprocessing steps, improves the diagnostic accuracy and robustness of noisy and imbalanced samples, enhances the network's representation ability, enables it to adapt to complex fault diagnosis, and improves the accuracy and precision of intelligent diagnosis between fault locations in rotating components such as bearings and gears.
Smart Images

Figure CN118503633B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to intelligent fault diagnosis technology and deep learning technology, and in particular to a mechanical equipment fault diagnosis method for approximate evaluation of multi-wavelet feature cascade. Background Art
[0002] High-speed rotating parts such as bearings and gears are widely used in mechanical equipment. These parts are in harsh working conditions such as high temperature, high speed, and high load for a long time, which directly determines the safety, stability, and reliability of the mechanical equipment. With the deepening of precision, integration, automation, and modularization of mechanical equipment, traditional fault diagnosis methods can no longer meet the requirements of diagnostic speed, accuracy, reliability, and comprehensive analysis. In order to avoid catastrophic accidents and improve the safety and reliability of mechanical equipment, fault diagnosis technology has quickly become a research hotspot in domestic industry and academia. Fault diagnosis based on deep learning does not rely on the expert knowledge and experience of the engineering object, but automatically processes signal features end-to-end, has strong comprehensive analysis capabilities, fast speed, and broad application prospects.
[0003] The fault information of mechanical equipment is distributed in many physical sources and has the characteristics of nonlinearity, non-stationarity and multi-coupling. On the one hand, the components in mechanical operation are interrelated and influence each other, and there is a complex mapping relationship between the cause of the fault and the fault mode, which causes noise interference in the signal and drowns out the fault information; on the other hand, mechanical equipment is rarely in a fault state, and it is difficult to collect enough available fault data in real operation scenarios, resulting in sample imbalance between healthy and faulty states. How to improve the accuracy, precision and robustness of intelligent diagnosis of noise sample imbalance of high-speed running bearings, gears and other components in mechanical equipment is still an engineering problem that needs to be solved urgently.
[0004] Among the existing technical solutions, there are few systematic studies on noise imbalance sample signals. For example, patent CN111751133A proposes an intelligent fault diagnosis method based on a deep convolutional neural network model with non-local mean embedding to analyze the type of mechanical fault, but does not test the diagnostic ability of noise signals; patent CN113158984A proposes a bearing fault diagnosis method based on complex Morlet wavelet and lightweight convolutional network, but it processes balanced samples and needs to convert the signal into a two-dimensional image in advance, which adds a preprocessing step; patent CN114383844A proposes a rolling bearing fault diagnosis method based on a distributed deep neural network. Although it realizes automatic feature extraction from the original vibration signal without manual feature selection and denoising, it does not describe the processing of fault imbalance; patent CN116256174A proposes a noise robust bearing fault diagnosis method based on interpretable multi-layer wavelet drive, which is also for fault diagnosis of noise samples, and fails to achieve systematic diagnosis of noise imbalance samples.
[0005] In summary, it is difficult for existing intelligent diagnosis methods to achieve fault diagnosis under noise imbalance samples. Summary of the invention
[0006] Purpose of the invention: The purpose of the present invention is to solve the deficiencies in the prior art and to provide a mechanical equipment fault diagnosis method that approximates the evaluation of multi-wavelet feature cascades.
[0007] Technical solution: A mechanical equipment fault diagnosis method of the present invention based on approximate evaluation of multi-wavelet feature cascade includes the following steps:
[0008] Step S1, obtaining vibration data of healthy and faulty states of different parts of target high-speed rotating components (such as bearings, gears, and rotors, etc.) under different working conditions, and adding noise to divide the sample unbalanced data set;
[0009] Step S2: construct a three-branch convolutional neural network model, and extract various features of the unbalanced signal of the noise sample from the unbalanced sample data set through the three-branch convolutional neural network model;
[0010] The three-branch convolutional neural network model includes a convolution layer, a multi-wavelet kernel function filter and a classification layer. The multi-wavelet kernel function filter includes three parallel branches, including a Gaussian difference wavelet kernel function layer, a Morlet wavelet kernel function layer, and a Laplace wavelet kernel function layer. The Gaussian difference wavelet kernel function layer, the Morlet wavelet kernel function layer, and the Laplace wavelet kernel function layer respectively perform fault signal denoising, high and low frequency analysis, and multi-scale analysis; a self-supervised wavelet kernel weight matrix is designed in the multi-wavelet kernel function filter, and the kernel of a learnable multi-wavelet kernel function operator is used to supervise and update the weight matrix of the filter, thereby realizing a filtering operation specific to noise imbalance samples;
[0011] Here, denoising, high and low frequency analysis, and multi-scale analysis are performed simultaneously, which reduces the preprocessing steps for noise removal and uses multiple wavelets to help extract different feature information of noise imbalance samples and extract various features of noise sample imbalance signals;
[0012] Step S3, performing feature cascading on the three different branches of step S2 to reduce global fusion redundancy, extracting feature information of the significant position of the unbalanced signal of the noise sample, and increasing the representation ability of the network. The feature cascading method is: performing feature cascading on the first branch and the second branch, performing feature cascading on the first branch and the third branch, and performing feature cascading on the second branch and the third branch; the first branch is a Morlet wavelet kernel function layer, the second branch is a Laplace wavelet kernel function layer, and the third branch is a Gaussian difference wavelet kernel function layer;
[0013] Step S4, using the two-distribution approximation evaluation method to measure the approximation degree of the branch feature cascade, accurately locate the differential expression of the multi-wavelet kernel function, and autonomously select the best cascade result to enter the decision layer of the model, which expands the expected feature information during the network training process, thereby achieving high-accuracy mechanical equipment fault diagnosis. Among them, the decision layer includes a flattening layer, a fully connected layer, and a softmax classification layer, and the best cascade result here refers to the maximum approximation degree.
[0014] Furthermore, step S1 is specifically as follows: adding noise to healthy samples of the target rotating parts and fault samples of different parts under different working conditions, setting different unbalanced sample sizes, and dividing them into training set, test set and validation set in proportion, and finally obtaining a training set with noise imbalance for different working condition combinations and a validation set and test set with noise balance for different working condition combinations. Before entering the model, the training set, test set and validation set are randomly shuffled.
[0015] Furthermore, the multi-wavelet kernel function filter embedded in step S2 is composed of a Morlet wavelet kernel function, a Laplace wavelet kernel function and a Gaussian difference wavelet kernel function, which respectively perform low-frequency analysis, high-frequency analysis and multi-scale analysis on the signal and have a denoising characteristic.
[0016] The formula of multiwavelet kernel function is expressed as:
[0017]
[0018] in, represents the Morlet wavelet kernel function, represents the Laplace wavelet kernel function, represents the Gaussian difference wavelet kernel function, A is the amplitude coefficient, f is the center frequency, τ is the time shift factor, ζ is the hysteresis factor, a is the scale factor, u For translation factors, s is the translation factor, is the difference coefficient ( is the difference coefficient of Gaussian difference wavelet. The form of Gaussian difference wavelet is the subtraction of two Gaussian wavelets with different difference coefficients. is the difference coefficient 1, is the difference coefficient 2.
[0019] In order to learn and update during the model training process, obtain a shape that better matches the wavelet and the fault signal, that is, a filter with a specific shape, and thus improve the accuracy of identifying fault modes, the kernel of the learnable multi-wavelet kernel function operator is further used to supervise and update the weight matrix of the filter. The formula of the self-supervised wavelet kernel weight matrix in the multi-wavelet kernel function filter is expressed as:
[0020]
[0021] in, W [,] represents the kernel matrix vector, c Represents the channel, is the kernel function, initialized to a random normal distribution.
[0022] Furthermore, in order to reduce global redundancy and extract feature information of the significant position of the unbalanced signal of the noise sample, the specific method of feature cascading in step S3 is: feature cascading of the first branch and the second branch, feature cascading of the first branch and the third branch, and feature cascading of the second branch and the third branch. The formula is expressed as:
[0023]
[0024] in, Represents the first branch and the second branch i The feature cascade on feature channels, Represents the first and third branches. i The feature cascade on feature channels, Represents the second and third branches. i The feature cascade on feature channels, Represents the first l The output value of the layer, n represents the sequence length, represents the width of the pooling kernel, Represents the convolutional layer l Tier i In the feature map h The activation value of a neuron.
[0025] Furthermore, the two-distribution approximation evaluation method described in step S4 is expressed as follows:
[0026]
[0027] in, is the maximum value of the approximate estimate obtained by the two distribution approximate evaluation method, E represents the expected form of relative entropy, It indicates the weights given to the first branch and the second branch by the two-distribution approximation evaluation method. It indicates the weights given to the second and third branches by the two-distribution approximation evaluation method. and All calculations Intermediate calculation data in the process, express Convert to Distribution An approximate estimate of the feature cascade of express Convert to Distribution An approximate estimate of the feature cascade of express Convert to Distribution An approximate estimate of the feature cascade of ;
[0028] The calculation formulas are:
[0029]
[0030] in, express , and The value in the channel, represents the weights given to the first and third branches by the two-distribution approximation evaluation method, Obeying a simple Gaussian distribution, , and express , and distribution.
[0031] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0032] (1) The present invention greatly reduces the preprocessing steps and effectively solves the problem that a single wavelet kernel function cannot take into account signal characteristics and that there is feature loss in the convolution process for the diagnosis of noise imbalance sample systems. It can adapt to complex fault diagnosis and classification problems.
[0033] (2) The multi-wavelet kernel function filter of the present invention has bandpass characteristics, locality in the time and frequency domains, and multi-scale analysis characteristics, and can be effectively used for high-frequency analysis and detection, while processing periodic low-frequency signals, revealing details and structures at different levels of the signal, and performing effective high-frequency analysis, low-frequency analysis, denoising and multi-scale analysis on fault signals.
[0034] (3) The present invention adopts a feature cascade method to make up for the gap between the data collected by the experimental platform and the actual working condition data, enhance the representation ability of the network, and can improve the accuracy, precision and robustness of intelligent diagnosis between different fault locations of rotating parts such as bearings and gears. The calculation efficiency is relatively high.
[0035] (4) The present invention uses a two-distribution approximation evaluation method to measure the approximation degree of the branch feature cascade, assigns different weights to the distribution according to the degree of discreteness, and adaptively selects the better result. It can accurately locate the differential expression of the multi-wavelet kernel function, thereby improving the accuracy of the diagnostic method. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a schematic diagram of the overall process of the present invention.
[0037] Figure 2 TSNE two-dimensional and three-dimensional visualization of the diagnosis results when there is no noise in the embodiment.
[0038] Figure 3 TSNE two-dimensional and three-dimensional visualization of the diagnosis results when the power ratio is 0.0316 in the embodiment.
[0039] Figure 4 TSNE two-dimensional and three-dimensional visualization of the diagnosis results when the power ratio is 0.1259 in the embodiment.
[0040] Figure 5 is the weight matrix under strong noise samples in the embodiment Visualization diagram.
[0041] Figure 6 is the weight matrix under strong noise samples in the embodiment Visualization diagram.
[0042] Figure 7 is the weight matrix under strong noise samples in the embodiment Visualization diagram.
[0043] Figure 8 It is a scatter matrix visualization result of the two distribution approximation evaluation method under strong noise sample imbalance of the diagnostic method in the embodiment.
[0044] in, Figure 2 (a) is a 2D visualization diagram of the diagnosis result TSNE without noise. Figure 2 (b) is a 3D visualization diagram of the diagnosis result TSNE without noise. Figure 3 (a) is a 2D visualization diagram of the diagnostic results TSNE when the power ratio is 0.0316. Figure 3 (b) is a 3D visualization diagram of the diagnosis result TSNE when the power ratio is 0.0316. Figure 4 (a) is a 2D visualization diagram of the diagnostic results TSNE when the power ratio is 0.1259. Figure 4 (b) is a 3D visualization diagram of the diagnosis result TSNE when the power ratio is 0.1259. DETAILED DESCRIPTION
[0045] The technical solution of the present invention is described in detail below, but the protection scope of the present invention is not limited to the embodiments.
[0046] The mechanical equipment fault diagnosis method of the present invention based on approximate evaluation of multi-wavelet feature cascade comprises the following steps:
[0047] Step S1, obtaining vibration data of healthy and faulty states of different parts of a target high-speed rotating component under different working conditions, and adding noise to divide the sample unbalanced data set;
[0048] Step S2, constructing a three-branch convolutional neural network model, and extracting various features of the unbalanced signal of the noise sample from the unbalanced sample data set through the three-branch convolutional neural network model; the three-branch convolutional neural network model includes a convolution layer, a multi-wavelet kernel function filter and a classification layer, and the multi-wavelet kernel function filter includes three parallel branches, including a Gaussian difference wavelet kernel function layer, a Morlet wavelet kernel function layer, and a Laplace wavelet kernel function layer, and the Gaussian difference wavelet kernel function layer, the Morlet wavelet kernel function layer and the Laplace wavelet kernel function layer respectively perform fault signal denoising, high and low frequency analysis and multi-scale analysis; a self-supervised wavelet kernel weight matrix is designed in the multi-wavelet kernel function filter;
[0049] Step S3, performing feature cascading on the three different branches of step S2, the method of feature cascading is: performing feature cascading on the first branch and the second branch, performing feature cascading on the first branch and the third branch, and performing feature cascading on the second branch and the third branch; the first branch is a Morlet wavelet kernel function layer, the second branch is a Laplace wavelet kernel function layer, and the third branch is a Gaussian difference wavelet kernel function layer;
[0050] Step S4, use the two-distribution approximation evaluation method to measure the approximation degree of the branch feature cascade, accurately locate the differential expression of the multi-wavelet kernel function, and autonomously select the best cascade result to enter the decision layer of the model, where the decision layer includes a flattening layer, a fully connected layer and a softmax classification layer, and the best cascade result here refers to the maximum approximation degree.
[0051] Embodiments: To facilitate understanding of the technical solutions of the present invention, further description is given below in conjunction with the accompanying drawings and embodiments.
[0052] like Figure 1 As shown, the mechanical equipment fault diagnosis method of the approximate evaluation multi-wavelet feature cascade in this embodiment specifically includes the following steps:
[0053] Step S1, obtaining vibration data of healthy and faulty states of different parts of target high-speed rotating components such as bearings, gears, rotors, etc. under different working conditions, and adding noise to divide the sample unbalanced data set.
[0054] Noise is added to healthy samples of the target rotating parts and faulty samples of different parts under different working conditions (as shown in Table 1, the ratio of added noise can be 0.3:0.2:0.2:0.1:0.1:0.05:0.05), and different unbalanced sample sizes are set, which are divided into training sets, test sets, and validation sets in proportion. Finally, the training sets with unbalanced noise under different working condition combinations and the validation sets and test sets with balanced noise under different working condition combinations are obtained, and the order is randomly shuffled before entering the model.
[0055] In the bearing data set of this example, the sampling frequency is 50kHZ, and the following are normal state (1000r / min), label 0; inner ring fault (600r / min), label 1; inner ring fault (1000r / min), label 2; rolling element fault (600r / min), label 3; rolling element fault (1000r / min), label 4; outer ring fault (600r / min), label 5; outer ring fault (1000r / min), label 6, a total of 7 categories for diagnosis. The ratio of training set, validation set, and test set is 0.7:0.15:0.15, the ratio of unbalanced samples in the training set is 0.3:0.2:0.2:0.1:0.1:0.05:0.05, the total sample size is 7000, each sample contains 700 data points, and the sampling is overlapped. Data samples of the training set divided according to Table 1.
[0056] surface Sample size of bearings under different working conditions
[0057]
[0058] Step S2, construct a three-branch convolutional neural network model, embed a multi-wavelet kernel function filter, and design a self-supervised wavelet kernel weight matrix in the multi-wavelet kernel function filter, use the learnable multi-wavelet kernel function operator to supervise and update the filter weight matrix, so as to achieve a specific filtering operation for the noise imbalance sample. At the same time, denoising, high and low frequency analysis and multi-scale analysis are performed to extract various features of the noise sample imbalance signal.
[0059] The multi-wavelet kernel function filter of this embodiment is composed of the Morlet wavelet kernel function, the Laplace wavelet kernel function and the Gaussian difference wavelet kernel function, which respectively perform low-frequency analysis, high-frequency analysis and multi-scale analysis on the signal, and also have denoising characteristics. The basic shape of the multi-wavelet is:
[0060]
[0061] in, represents the morlet wavelet, represents the Laplace wavelet, represents Gaussian difference wavelet, A represents amplitude coefficient, f represents center frequency, t represents time domain signal, is the difference coefficient of the high-resolution difference wavelet.
[0062] Considering that the time domain signal is shifted over time when passing through the filter, a bandwidth factor a is introduced to control the time shift factor and determine the degree of selection of a specific frequency range;
[0063] Wavelet basic dictionary based on time domain , the multi-wavelet kernel function can be further constructed as:
[0064]
[0065] in, u , s Both are wavelet basic dictionaries The dynamic adaptive parameters in u is the learned translation factor, s is the learned scale factor. Both are adaptive and adjustable parameters and are completely differentiable. Substitute them into the sigmoid activation function respectively, and their basic form is , the formula of multi-wavelet kernel function is expressed as:
[0066]
[0067] in, represents the Morlet wavelet kernel function, represents the Laplace wavelet kernel function, represents the Gaussian difference wavelet kernel function, A is the amplitude coefficient, f is the center frequency, τ is the time shift factor, ζ is the hysteresis factor, a is the scale factor, u For translation factors, s is the translation factor, is the difference coefficient.
[0068] The self-supervised wavelet kernel weight matrix uses the kernel of the learnable multi-wavelet kernel function operator to supervise and update the weight matrix of the filter, and the formula is expressed as:
[0069]
[0070] Among them, W[,] represents the kernel matrix vector, c Represents the channel, is the kernel function, initialized to a random normal distribution.
[0071] In this embodiment, the amplitude coefficient of the multiwavelet kernel function is 1.5, the center frequency is 0.5, the time shift factor is 0.1, the hysteresis factor is -0.3, and the scale factor is 0.5.
[0072] The three-branch convolutional neural network model is mainly composed of a convolution layer, a multi-wavelet kernel function layer, a feature cascade layer, a two-distribution approximation estimation layer and a classification layer. The multi-wavelet kernel function layer is composed of a Morlet wavelet kernel function layer, a Laplace wavelet kernel function layer and a Gaussian difference wavelet kernel function layer.
[0073] As shown in Table 2, Bcm represents Figure 1 The batch convolution shown in the figure consists of a convolution layer, a ReLU layer, and a batch normalization layer. M represents the maximum pooling layer, W represents the wavelet kernel and the designed self-supervised wavelet kernel weight matrix, and the shape is kernel size, step size, and channel. The fault signal passes through the convolution layer for network modeling. The first branch and the second branch, the first branch and the third branch, and the second branch and the third branch where the multi-wavelet kernel function is located are distributed and fused to form a feature cascade layer, and then the cascade features enter the two distribution approximation estimation layer and the classification layer.
[0074] Table 2 Three-branch convolutional neural network model
[0075]
[0076] In this embodiment, the workflow of the batch normalization layer is:
[0077] (1) For batch input , calculate the batch data mean: ;
[0078] (2) Use a small positive number to avoid division by 0 and calculate the variance of the batch data: ;
[0079] (3) Standardization: , which will be multiplied by the scale factor Adjust the value and add the translation factor After adding the offset we get : ;
[0080] (4) Return the parameters learned by the network and , output the normalized network response: .
[0081] In this embodiment, the noise added for testing model diagnosis is Gaussian white noise, and the ratios of Gaussian white noise power to signal power of no noise, weak noise and strong noise corresponding to the noise level and model training results are shown in Table 3.
[0082] Table 3 Different noise signal-to-noise ratios and corresponding diagnostic results
[0083]
[0084] Step S3: perform feature cascading on different branches to reduce global fusion redundancy, extract feature information of significant positions of unbalanced signals of noise samples, and increase the representation ability of the network.
[0085] The specific method of feature cascading is as follows: the first branch and the second branch perform feature cascading, the first branch and the third branch perform feature cascading, and the second branch and the third branch perform feature cascading. The formula is expressed as:
[0086] ;
[0087] ;
[0088] ;
[0089] in, Represents the first branch and the second branch i The feature cascade on feature channels, Represents the first and third branches. i The feature cascade on feature channels, Represents the second and third branches. i The feature cascade on feature channels, Represents the first l The output value of the layer, n represents the sequence length, represents the width of the pooling kernel, Represents the convolutional layer l Tier i In the feature map h The activation value of a neuron.
[0090] Step S4: Use the two-distribution approximation evaluation method to measure the approximation degree of the branch feature cascade, accurately locate the differential expression of the multi-wavelet kernel function, and autonomously select the best cascade result to enter the decision layer of the model.
[0091] The approximate evaluation formula of the two distributions in this embodiment is expressed as:
[0092]
[0093] in, is the maximum value of the approximate estimate obtained by the two distribution approximate evaluation method, E represents the expected form of relative entropy, It indicates the weights given to the first branch and the second branch by the two-distribution approximation evaluation method. It indicates the weights given to the second and third branches by the two-distribution approximation evaluation method. express Convert to Distribution An approximate estimate of the feature cascade of express Convert to Distribution An approximate estimate of the feature cascade of express Convert to Distribution An approximate estimate of the feature cascade of .
[0094] The calculation formulas are:
[0095]
[0096] in, express , and The value in the channel, represents the weights given to the first and third branches by the two-distribution approximation evaluation method, Obeying a simple Gaussian distribution, , and express , and distribution.
[0097] In this embodiment, Figures 2 to 4 It is a visualization diagram of the diagnosis results of the noise-free, weak noise and strong noise samples in Table 3. Among them, Figure 2 (a) is a 2D visualization diagram of the diagnosis result TSNE without noise. Figure 2 (b) is a 3D visualization diagram of the diagnosis result TSNE without noise. Figure 3 (a) is a 2D visualization diagram of the diagnostic results TSNE when the power ratio is 0.0316. Figure 3 (b) is a 3D visualization diagram of the diagnosis result TSNE when the power ratio is 0.0316. Figure 4 (a) is a 2D visualization diagram of the diagnostic results TSNE when the power ratio is 0.1259. Figure 4 (b) is a 3D visualization diagram of the diagnosis result TSNE when the power ratio is 0.1259.
[0098] like Figures 5 to 7 As shown in FIG. 1 , three schematic diagrams of self-supervised wavelet kernel weight matrices in this embodiment are shown. The arrows represent the convergence direction. It can be seen that the kernel weight values are adjusted during the training process and finally converged. When the kernel weight matrix gradually converges and maintains a stable shape, it indicates that the model has learned a set of effective filters for feature extraction.
[0099] In this embodiment, Figure 8 The scatter matrix visualization results of the dynamic weight selection for the two distributions approximate estimation, where [1] is the first branch and the second branch. i Feature fusion on feature channels , [2] is the first branch and the third branch i Feature fusion on feature channels , [3] is the first i Feature fusion on feature channels , [4] is the result of selecting the two distribution approximate evaluation method ,It can be seen that the wavelet kernel function group ,processes different experimental signal details differently, which effectively alleviates ,the fixed extraction mode of a single wavelet and increases the ,depth of attention to faults.
[0100] In summary, the present invention can diagnose the imbalance of noise samples under different working conditions, and at the same time adopts a visualization method to make the multi-wavelet kernel function embedded in the self-supervised wavelet kernel weight matrix and the two-distribution approximation evaluation method interpretable, thereby improving the intelligent diagnosis accuracy, precision and robustness of noise sample imbalances in different degrees of faults in the same part of the bearing and between different fault parts, and the calculation efficiency is high.
Claims
1. A mechanical equipment fault diagnosis method based on approximate evaluation of multi-wavelet feature cascade, characterized in that: The following steps are involved: Step S1, obtaining vibration data of healthy and faulty states of different parts of a target high-speed rotating component under different working conditions, and adding noise to divide the sample unbalanced data set; Step S2: construct a three-branch convolutional neural network model, and extract various features of the unbalanced signal of the noise sample from the unbalanced sample data set through the three-branch convolutional neural network model; The three-branch convolutional neural network model includes a convolution layer, a multi-wavelet kernel function filter and a classification layer. The multi-wavelet kernel function filter includes three parallel branches, including a Gaussian difference wavelet kernel function layer, a Morlet wavelet kernel function layer, and a Laplace wavelet kernel function layer. The Gaussian difference wavelet kernel function layer, the Morlet wavelet kernel function layer, and the Laplace wavelet kernel function layer respectively perform fault signal denoising, high and low frequency analysis, and multi-scale analysis; a self-supervised wavelet kernel weight matrix is designed in the multi-wavelet kernel function filter; Step S3, performing feature cascading on the three different branches of step S2, the method of feature cascading is: performing feature cascading on the first branch and the second branch, performing feature cascading on the first branch and the third branch, and performing feature cascading on the second branch and the third branch; the first branch is a Morlet wavelet kernel function layer, the second branch is a Laplace wavelet kernel function layer, and the third branch is a Gaussian difference wavelet kernel function layer; Step S4, using a two-distribution approximation evaluation method to measure the approximation degree of the branch feature cascade, locating the differential expression of the multi-wavelet kernel function through the approximation degree, and autonomously selecting the best cascade result to enter the decision layer of the model, the decision layer includes a flattening layer, a fully connected layer and a softmax classification layer, and the best cascade result refers to the maximum approximation degree; The specific contents of the two distribution approximation evaluation method in step S4 are as follows: ; in, is the maximum value of the approximate estimate obtained by the two distribution approximate evaluation method, E represents the expected form of relative entropy, It indicates the weights given to the first branch and the second branch by the two-distribution approximation evaluation method. It indicates the weights given to the second and third branches by the two-distribution approximation evaluation method. express Convert to Distribution An approximate estimate of the feature cascade of express Convert to Distribution An approximate estimate of the feature cascade of express Convert to Distribution An approximate estimate of the feature cascade of ; The calculation formulas are: ; in, express , and The value in the channel, It indicates the weights given to the first and third branches by the two-distribution approximation evaluation method. z Obeying a simple Gaussian distribution, , and express , and distribution.
2. The mechanical equipment fault diagnosis method of approximate evaluation multi-wavelet feature cascade according to claim 1 is characterized in that: The step S1 adds noise to healthy samples of the target high-speed rotating component and fault samples of different parts under different working conditions, sets different unbalanced sample sizes, and divides them into training sets, test sets and verification sets in proportion, and finally obtains noise-unbalanced training sets for different working condition combinations and noise-balanced verification sets and test sets for different working condition combinations.
3. The mechanical equipment fault diagnosis method of approximate evaluation multi-wavelet feature cascade according to claim 1 is characterized in that: The multi-wavelet kernel function filter in step S2 performs low-frequency analysis, high-frequency analysis, multi-scale analysis and denoising on the signal. The calculation formula of the multi-wavelet kernel function filter is as follows: ; in, represents the Morlet wavelet kernel function, represents the Laplace wavelet kernel function, represents the Gaussian difference wavelet kernel function, A is the amplitude coefficient, f is the center frequency, is the time shift factor, ζ is the hysteresis factor, a is the scale factor, u For translation factors, s is the translation factor, is the difference coefficient 1, is the difference coefficient 2.
4. The mechanical equipment fault diagnosis method of approximate evaluation multi-wavelet feature cascade according to claim 1 or 3, characterized in that: The formula of the self-supervised wavelet kernel weight matrix in the multi-wavelet kernel function filter is as follows: ; in, W [,] represents the kernel matrix vector, c Represents the channel, is the kernel function, initialized to a random normal distribution.
5. The mechanical equipment fault diagnosis method of approximate evaluation multi-wavelet feature cascade according to claim 1 is characterized in that: The detailed formula of the feature cascade in step S3 is as follows: ; ; ; in, Represents the first branch and the second branch i The feature cascade on feature channels, Represents the first and third branches. i The feature cascade on feature channels, Represents the second and third branches. i The feature cascade on feature channels, Represents the first l The output value of the layer, n represents the sequence length, represents the width of the pooling kernel, Represents the convolutional layer l Tier i In the feature map h The activation value of a neuron.
