Rolling bearing fault diagnosis method and system oriented to complex working conditions

Abstract

The invention provides a rolling bearing fault diagnosis method and system oriented to complex working conditions, and the method comprises the steps: firstly carrying out the preprocessing of a collected rolling bearing vibration signal through employing a rolling bearing fault signal noise reduction method oriented to optimization SSA-SGMD, and carrying out the preprocessing of the vibration signal; then, screening the preprocessed bearing vibration signals, performing synchronous extrusion wavelet transform on the screened vibration signals, and converting one-dimensional vibration signals into two-dimensional images; and finally, performing image expansion by using a novel integrated generative adversarial network, extracting and classifying two-dimensional features by using an improved VGG network model, extracting feature differences among different fault types, and finally identifying and classifying the two-dimensional features to finally achieve the purpose of fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a rolling bearing fault diagnosis method and system under complex working conditions. Background Art

[0002] As an important part of mechanical transmission, rolling bearings are widely used in industrial manufacturing, transportation, energy industry, aerospace and other fields. If the mechanical system is overloaded or improperly used during use, the rolling bearing will be damaged to a certain extent. At the same time, harsh working environments, such as high temperature, high humidity, dust, chemical corrosion, dust, etc., also affect the normal operation of rolling bearings and cause failures. Therefore, it is necessary to diagnose them in real time to ensure personal safety and economic benefits. The diagnosis of rolling bearings usually relies on the analysis of vibration signals, but in actual production, the production environment is noisy and there is a lot of noise interference. The collected vibration signals usually contain a lot of noise that interferes with feature extraction. In addition, one-dimensional vibration signals cannot intuitively show the spatial distribution and changes of fault modes. Two-dimensional features can more comprehensively reflect the fault information of rolling bearings. Using two-dimensional images for fault diagnosis can better analyze the fault types of rolling bearings. Therefore, it is of great significance to propose solutions to the problems of signal noise reduction and two-dimensional feature extraction of rolling bearings.

[0003] In the actual production environment, the friction, impact, looseness between the components of the production machinery itself, and the noise generated during the processing will be introduced into the vibration signal. At the same time, the operation of other machines and equipment in the production environment will also generate vibrations. These vibrations may be transmitted through the air or structure and become noise sources. As a result, the collected vibration signal contains some noise signals, resulting in a decrease in the performance of the diagnostic model. At present, there are two noise reduction methods for vibration signals: filter-based and signal processing-based methods. The filter-based method is a relatively traditional method for vibration signal noise reduction, among which the filter is a commonly used tool in noise reduction. It mainly eliminates or weakens certain frequency components by selecting a certain frequency band in the spectrum, thereby achieving a noise reduction effect. For the signal processing-based method, common methods include wavelet transform and empirical mode decomposition, and the key lies in how to offset the noise in the original vibration signal. For the method of converting one-dimensional vibration signals into two-dimensional images, the main method is to convert the time domain signal into a frequency domain signal, and then combine it to draw the time-frequency image. Common methods such as short-time Fourier transform and continuous wavelet transform have certain defects in the processing of high-frequency signals and cannot fully extract the two-dimensional features of vibration signals. At present, neural networks mainly process two-dimensional images through convolution and pooling, extract features from two-dimensional images of different fault types, input the extracted features into the classifier for classification, and finally output the recognition results. At present, the advantages of converting one-dimensional vibration signals into two-dimensional images after noise reduction in fault diagnosis have not been fully utilized, resulting in an overall low diagnostic accuracy in complex production environments. Summary of the invention

[0004] In order to overcome the deficiencies of the prior art, an object of the present invention is to provide a rolling bearing fault diagnosis method and system for complex working conditions.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A rolling bearing fault diagnosis method under complex working conditions, comprising:

[0007] The optimization parameters of symplectic geometric mode decomposition are optimized using the sparrow search algorithm to obtain the optimized SSA-SGMD algorithm;

[0008] The SSA-SGMD algorithm is used to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal;

[0009] The preprocessed signal is screened, and the screened preprocessed signal is subjected to synchronous squeezing wavelet transformation to convert the one-dimensional screened preprocessed signal into a two-dimensional image;

[0010] Performing sample expansion on the two-dimensional image to obtain a first data set, and inputting the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set;

[0011] The improved VGG network model is used to extract and classify the two-dimensional images in the first data set and the second data set to obtain the characteristic differences between different fault types, so as to realize the fault diagnosis of rolling bearings.

[0012] Preferably, the rolling bearing vibration signal is collected through a multi-condition rolling bearing fault simulation experimental platform; the fault types of the data acquired by the multi-condition rolling bearing fault simulation experimental platform include: normal state, outer ring crack, inner ring crack, roller crack and mixed fault.

[0013] Preferably, the optimization parameters of the symplectic geometric mode decomposition are optimized using a sparrow search algorithm, including:

[0014] Initialization parameters K and α, the range of K and α is: 0≤K≤0.1; 0.8≤α≤1; where K is the first optimization parameter and α is the second optimization parameter;

[0015] Initialize the population; the formula for initializing the population is: t is the number of iterations, a and b are adjustable parameters, y1 represents the initial position of the population individual, and y2 represents the next position of the population individual;

[0016] Calculate the optimal chaotic sequence; the formula of the optimal chaotic sequence is: z(t)=abs[ωy1(t)], i=1,2,…n; wherein z(t) is the optimal chaotic sequence, ω is the scale factor, n is the total number of signals, and abs represents the absolute value;

[0017] According to the optimal chaotic sequence, the inverse mapping is performed to obtain the sparrow position; the formula of the sparrow position is: i t =lb i +z(t)(ub i -lb i );where ub i lb i are the upper and lower limits of the search space, x i t is the position of the sparrow;

[0018] The minimum sample entropy is set as the fitness function; the formula of the fitness function is: Among them, X is a sample set consisting of the positions of sparrows, n' is the number of different categories, x i is the i-th category in the sample, P(x i) is the probability of the i-th category in the sample, H(X) is the fitness function;

[0019] A fitness function is set according to the fitness function; the formula of the fitness function is: Wherein, F(X) is the fitness function;

[0020] Calculate the nonlinear adaptive weight factor; the formula of the nonlinear adaptive weight factor is: Wherein, T is the maximum number of iterations, and λ is the nonlinear adaptive weight factor;

[0021] Update the finder's position; the formula for updating the finder's position is: Among them, R2 is the warning value, R2∈[0,1]; ST is the safety value, ST∈[0.5,1]; Q is a random number that obeys the normal distribution, and are the current position and next position of the finder respectively;

[0022] The formula for calculating the value of Levy in the Levy flight strategy is: Among them, Γ(x)=(x-1)!; r3 and r4 are calculation parameters, r3, r4∈[0,1]; ξ is an adjustable constant, the value is 1.5,

[0023] The global optimal position is calculated according to the Cauchy strategy; the calculation formula of the global optimal position is: newbest =x best +x best ×Cauchy(0,1); where Cauchy(0,1) is the standard Cauchy distribution function, x best The current optimal position, x newbest is the global optimal position;

[0024] Update the follower position; the calculation formula of the follower position is: Among them, x worst is the worst position of the finder, d is the vector dimension, when i>n / 2, the hungry i-th follower flies to other positions to get food;

[0025] Update the alerter position; the formula for the alerter position is: Among them, the first optimization parameter K represents the moving direction, K∈[-1,1]; β is the step size parameter; f g 、f w are the global optimal and worst fitness respectively, f i is the individual fitness, ο is the minimum constant, is the position with the largest fitness value, It is the position where the moderation value is minimum to prevent the denominator from being zero.

[0026] Repeat steps "Set the minimum sample entropy as the fitness function" to step "Update the position of the alerter" until the maximum number of iterations is reached to obtain the optimal first optimization parameter K and second optimization parameter α.

[0027] Preferably, the SSA-SGMD algorithm is used to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal, including:

[0028] Calculate the PSD of the original rolling bearing vibration signal x and find the frequency f of the maximum peak max , let the frequency f max With sampling frequency F s When the ratio is less than K obtained by the sparrow search algorithm, the embedding dimension d'=n / 3 is set, otherwise, d=1.2×(F S / f max ), for the delay time, usually τ = 1, to obtain the trajectory matrix X g The rolling bearing vibration signal x is expressed as: x = x1, x2, ..., x n" , n" represents the signal data length, and the formula of the trajectory matrix is: Where d' represents the embedding dimension, and m = n"-(d'-1)τ;

[0029] For the matrix X g Perform autocorrelation analysis to obtain the covariance matrix A, A = X g T X g ;

[0030] According to the symplectic geometry theorem, the Hamilton matrix M is constructed using the covariance matrix A;

[0031] Processing the Hamilton matrix M yields another Hamilton matrix W; W = M 2 ;

[0032] Establish equations involving symplectic geometry matrix G based on Hamilton matrix W; Among them, the matrices G and B are symplectic orthogonal matrices and upper triangular matrices respectively;

[0033] Using QR decomposition, we can decompose the matrix A 2 Decompose to obtain the eigenvector matrix Q; matrix A 2 Indicates the multiplication of matrix A and matrix A;

[0034] Using the eigenvector matrix Q and the original trajectory matrix X gConstruct the coefficient matrix S; S i =Q i T X g ;

[0035] The reconstruction matrix Z is obtained by using the eigenvector matrix Q and the coefficient matrix S; Z i =Q i S i ; Among them, Z i (i=1,2,…d") is the initial single-component reconstruction matrix; the reconstruction matrix Z is composed of d" groups of initial single-component reconstruction matrices Z i Composition: Z = Z1 + Z2 + ... + Z d" ;

[0036] Reconstruct the matrix Z from the single component using diagonal averaging i (1≤k≤d) is transformed to obtain a one-dimensional initial single-component signal with a length of l, and finally d" one-dimensional initial single-component signals are obtained. The sum of all initial single components is the original signal; let Z i The elements in are defined as z ij (1≤i≤m,1≤j≤d"); set d * =min(m,d"), m * =max(m,d") and 1≤i≤m,1≤j≤d", if m<d", then z * ij =z ij , otherwise z * ij =z ji ; The expression to be converted is: Here, the original signal is the signal used in the subsequent steps after the original rolling vibration bearing vibration signal is processed;

[0037] Calculate the period of d" one-dimensional initial single component signals, select the first one-dimensional initial single component signal Y1, and compare Y1 with the period of Y1 plus d"-1 initial single components, reorganize the initial single components within the range of the optimal first optimization parameter α, obtain the first symplectic geometric component SGC1, remove the constituent initial single component signals of SGC1, and sum the remaining initial single components after removal, denoted as g1; finally, select the first component of the remaining initial single components and perform period similarity analysis with other components to obtain SGC2; after h iterations, obtain h symplectic geometric components, and the sum of the remaining initial single components is denoted as g h ;

[0038] Calculate the normalized mean square error NMSE of the residual signal; Among them, g h (e) is g hThe normalized mean square error of , x(e) is the normalized mean square error of x;

[0039] When the NMSE h When it is less than the set threshold, the decomposition ends; otherwise, the iterative decomposition continues, and the expression when the decomposition ends is expressed as: Where N is the number of symplectic geometry components obtained;

[0040] Preferably, the SSA-SGMD algorithm is used to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal, and further includes:

[0041] The vibration signal after noise reduction processing is segmented, and the length of each segment of the vibration signal is determined, and overlapping sampling is used to obtain the pre-processed signal converted into a two-dimensional image.

[0042] Preferably, performing synchronous squeezing wavelet transform on the filtered preprocessed signal to convert the one-dimensional filtered preprocessed signal into a two-dimensional image comprises:

[0043] The CWT of the preprocessed signal after transformation and screening in the time domain is: Where: W s (a1, b1) is the complex wavelet coefficient spectrum of the signal; a1 is the scale factor; b1 is the time shift factor; ψ(t) is the mother wavelet function; s(t) is the original signal;

[0044] Convert the time domain to the frequency domain, the conversion formula is: Where: ξ is the angular frequency; is the Fourier transform of the original signal; is a complex exponential function, representing a rotation factor used to adjust the position of the wavelet on the time axis; is the complex conjugate of the mother wavelet;

[0045] By resolving the phase of the wavelet coefficients in the time domain, the instantaneous frequency of the signal is obtained:

[0046] In the discrete case, for any center frequency ω1 The inner wavelet coefficients are squeezed to synchronously squeeze the wavelet transform values. The mathematical expression is: Among them, T s (ω1,b1) represents the result of synchronous squeezing wavelet transform when the scale is ω1 and the position is b1; W s (a k ,b1) indicates the scale is a k Continuous wavelet transform coefficient when position is b1; (Δa) k For different scales a k The interval between.

[0047] When the variables are continuous, the mathematical expression is:

[0048] The synchronous squeezing wavelet transformed data are used for graphing, and the two-dimensional images of synchronous squeezing wavelet transformed time-frequency of various bearing fault types are obtained.

[0049] Preferably, the two-dimensional image is sample expanded to obtain a first data set, and the first data set is input into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set, including:

[0050] Flipping, rotating, cropping, color dithering, and adding noise to the two-dimensional image to obtain the first data set;

[0051] Based on the first data set, images are generated using a deep convolutional generative adversarial network framework layer based on DCGAN to obtain the second data set.

[0052] A rolling bearing fault diagnosis system for complex working conditions, comprising:

[0053] An optimization module is used to optimize the optimization parameters of symplectic geometric mode decomposition using a sparrow search algorithm to obtain an optimized SSA-SGMD algorithm;

[0054] A noise reduction module, used for performing noise reduction preprocessing on the collected rolling bearing vibration signal by using the SSA-SGMD algorithm to obtain a preprocessed signal;

[0055] A screening module, used for screening the pre-processed signal, and performing synchronous extrusion wavelet transform on the screened pre-processed signal, so as to convert the one-dimensional screened pre-processed signal into a two-dimensional image;

[0056] An expansion module, used to perform sample expansion on the two-dimensional image to obtain a first data set, and input the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set;

[0057] The diagnosis module is used to extract and classify the two-dimensional images in the first data set and the second data set using the improved VGG network model to obtain the characteristic differences between different fault types to realize the fault diagnosis of the rolling bearing.

[0058] Preferably, the rolling bearing vibration signal is collected through a multi-condition rolling bearing fault simulation experimental platform; the fault types of the data acquired by the multi-condition rolling bearing fault simulation experimental platform include: normal state, outer ring crack, inner ring crack, roller crack and mixed fault.

[0059] Preferably, the optimization module comprises:

[0060] A parameter initialization unit, used for initializing parameters K and α, wherein the ranges of K and α are: 0≤K≤0.1; 0.8≤α≤1; wherein K is the first optimization parameter, and α is the second optimization parameter;

[0061] The population initialization unit is used to initialize the population; the formula for initializing the population is: t is the number of iterations, a and b are adjustable parameters, y1 represents the initial position of the population individual, and y2 represents the next position of the population individual;

[0062] A sequence calculation unit is used to calculate an optimal chaotic sequence; the formula of the optimal chaotic sequence is: z(t)=abs[ωy1(t)], i=1,2,…n; wherein z(t) is the optimal chaotic sequence, ω is a proportional factor, n is the total number of signals, and abs represents an absolute value;

[0063] The inverse mapping unit is used to perform inverse mapping according to the optimal chaotic sequence to obtain the sparrow position; the formula of the sparrow position is: i t =lb i +z(t)(ub i -lb i );where ub i lb i are the upper and lower limits of the search space, x i t is the position of the sparrow;

[0064] The first function determination unit is used to set the minimum sample entropy as the fitness function; the formula of the fitness function is: Among them, X is a sample set consisting of the positions of sparrows, n' is the number of different categories, x i is the i-th category in the sample, P(x i ) is the probability of the i-th category in the sample, H(X) is the fitness function;

[0065] The second function determination unit is used to set a fitness function according to the fitness function; the formula of the fitness function is: Wherein, F(X) is the fitness function;

[0066] The factor calculation unit is used to calculate the nonlinear adaptive weight factor; the formula of the nonlinear adaptive weight factor is: Wherein, T is the maximum number of iterations, and λ is the nonlinear adaptive weight factor;

[0067] The first position updating unit is used to update the finder position; the formula for updating the finder position is: Among them, R2 is the warning value, R2∈[0,1]; ST is the safety value, ST∈[0.5,1]; Q is a random number that obeys the normal distribution, and are the current position and next position of the finder respectively;

[0068] The flight strategy unit is used to calculate the Levy value in the Levy flight strategy; the calculation formula for the Levy value is: Among them, Γ(x)=(x-1)!; r3 and r4 are calculation parameters, r3, r4∈[0,1]; ξ is an adjustable constant, the value is 1.5,

[0069] The optimal position determination unit is used to calculate the global optimal position according to the Cauchy strategy; the calculation formula of the global optimal position is: newbest =x best +x best ×Cauchy(0,1); where Cauchy(0,1) is the standard Cauchy distribution function, x best The current optimal position, x newbest is the global optimal position;

[0070] The second position updating unit is used to update the follower position; the calculation formula of the follower position is: Among them, x worst is the worst position of the finder, d is the vector dimension, when i>n / 2, the hungry i-th follower flies to other positions to get food;

[0071] The third position updating unit is used to update the position of the alerter; the formula of the position of the alerter is: Among them, the first optimization parameter K represents the moving direction, K∈[-1,1]; β is the step size parameter; f g 、f w are the global optimal and worst fitness respectively, f i is the individual fitness, ο is the minimum constant, is the position with the largest fitness value, It is the position where the moderation value is minimum to prevent the denominator from being zero.

[0072] The repeating unit is used to repeat the steps of "setting the minimum sample entropy as the fitness function" to "updating the position of the alerter" until the maximum number of iterations is reached, so as to obtain the optimal first optimization parameter K and the second optimization parameter α.

[0073] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0074] The present invention provides a rolling bearing fault diagnosis method and system for complex working conditions, the method comprising: optimizing the optimization parameters of symplectic geometric modal decomposition using a sparrow search algorithm to obtain an optimized SSA-SGMD algorithm; using the SSA-SGMD algorithm to perform denoising preprocessing on the collected rolling bearing vibration signal to obtain a preprocessing signal; screening the preprocessing signal, and performing synchronous squeezing wavelet transform on the screened preprocessing signal to convert the one-dimensional screened preprocessing signal into a two-dimensional image; performing sample expansion on the two-dimensional image to obtain a first data set, and inputting the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set; using an improved VGG network model to extract and classify the two-dimensional images in the first data set and the second data set to obtain feature differences between different fault types to achieve rolling bearing fault diagnosis. The present invention. The present invention can improve the accuracy of rolling bearing fault diagnosis and has good generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0076] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0077] Figure 2 A schematic diagram of a technical route provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0078] 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.

[0079] The purpose of the present invention is to provide a rolling bearing fault diagnosis method under complex working conditions, which can improve the accuracy of rolling bearing fault diagnosis and has good generalization ability.

[0080] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0081] Figure 1A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a rolling bearing fault diagnosis method under complex working conditions, comprising:

[0082] Step 100: Optimizing the optimization parameters of symplectic geometric mode decomposition using a sparrow search algorithm to obtain an optimized SSA-SGMD algorithm;

[0083] Step 200: using the SSA-SGMD algorithm to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal;

[0084] Step 300: screening the preprocessed signal, and performing synchronous squeezing wavelet transform on the screened preprocessed signal to convert the one-dimensional screened preprocessed signal into a two-dimensional image;

[0085] Step 400: performing sample expansion on the two-dimensional image to obtain a first data set, and inputting the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set;

[0086] Step 500: Use the improved VGG network model to extract and classify the two-dimensional images in the first data set and the second data set to obtain feature differences between different fault types to achieve rolling bearing fault diagnosis.

[0087] Specifically, this embodiment provides a rolling bearing fault diagnosis method, comprising the following steps:

[0088] (1) Design and build a multi-condition rolling bearing fault simulation experimental platform to obtain a data set, and obtain the Case Western Reserve University bearing data set (CWRU);

[0089] (2) Optimize the symplectic geometric mode decomposition (SGMD) parameters and use the sparrow search algorithm (SSA);

[0090] (3) Optimized SSA-SGMD denoising of the Western Reserve University bearing data;

[0091] (4) segmenting the vibration signal after noise reduction processing, determining the length of each segment of the vibration signal, and using overlapping sampling to determine the original data converted into a two-dimensional image;

[0092] (5) Perform continuous wavelet transform on the selected vibration signal to convert the time domain signal into the time-frequency domain signal, obtain new signal features, and analyze the energy distribution of the signal at different frequencies and times;

[0093] (6) Calculate the instantaneous frequency of the bearing vibration signal, squeeze and reorganize the signal after continuous wavelet transform to obtain the synchronous squeezed wavelet transform value, and draw a two-dimensional feature image;

[0094] (7) Use basic image processing methods to expand samples and obtain the data set T-set. Then use DCGAN to generate an adversarial network to generate the data set G-set. Finally, use the improved VGG for classification.

[0095] Furthermore, the multi-condition rolling bearing fault simulation experimental platform in step (1) is composed of a first-stage vibration isolation platform, a drive motor, a magnetic powder brake loading device, a three-stage reduction gearbox, and an input and output shaft rolling bearing fault simulation kit. The data set obtained by the multi-condition rolling bearing fault simulation experimental platform includes normal state, outer ring crack, inner ring crack, roller crack and mixed fault, a total of 5 types.

[0096] Furthermore, the step (2) of optimizing the symplectic geometric mode decomposition parameters and using the sparrow search algorithm is to optimize the symplectic geometric mode decomposition parameters and using the sparrow search algorithm.

[0097] Furthermore, in step (4), the length of each vibration signal is 1024, and an overlapping sampling method is adopted, wherein the overlapping rate is 0.5, so as to expand the sample capacity.

[0098] Furthermore, in step (5), the vibration signal is subjected to a continuous wavelet transform, which decomposes the signal into information in the time domain and the frequency domain, and provides the ability to locate fault features on the time-frequency plane, which can more accurately determine the time and frequency of the occurrence of the fault, and helps to accurately diagnose the health status of the bearing.

[0099] Furthermore, in step (6), by analyzing the phase of the wavelet coefficients in the time domain, the instantaneous frequency of the signal can be obtained, and the wavelet coefficients at any center frequency can be squeezed, and the wavelet transform values ​​can be squeezed synchronously, and finally the image is drawn using the synchronously squeezed wavelet transform values.

[0100] Furthermore, in step (7), the generated two-dimensional image is expanded by using basic image processing methods, and the two-dimensional image is flipped, rotated, cropped, color-dithered, and noise is added to obtain the data set T-set. A deep convolutional generative adversarial network framework based on DCGAN is used to generate an image of size 256×256 pixels2 to obtain the data set G-set.

[0101] Furthermore, the specific steps of performing fault diagnosis and identification using the improved VGG neural network constructed in step (7) are as follows:

[0102] (7-1) Construct an improved VGG neural network;

[0103] (7-2) Using samples for training, wherein each fault signal is provided with 300 sets of samples, and the samples are divided into a training set, a test set, and a validation set in a ratio of 7:2:1;

[0104] (7-3) Visualize the t-SNE clustering distribution of the early maximum pooling layer activations, the final convolutional layer activations, and the final softmax layer activations of the neural network;

[0105] (7-4) Use the confusion matrix to summarize the prediction results of the classification model.

[0106] Furthermore, the optimization of the VGG neural network in step (4) includes: in order to reduce parameters, using global average pooling (GAP) to replace the FC layer; improving the normalization method by normalizing each input data by making the expected value 0 and the variance 1; optimizing the filter size to increase the training speed of the network model.

[0107] like Figure 1 As shown, the rolling bearing fault diagnosis method of the present invention comprises the following specific steps:

[0108] Step 1: The multi-condition rolling bearing fault simulation experimental platform consists of a first-level vibration isolation platform, a drive motor, a magnetic powder brake loading device, a three-stage reduction gearbox, and an input and output shaft rolling bearing fault simulation kit. The data set obtained by the multi-condition rolling bearing fault simulation experimental platform includes 5 types: normal state, outer ring crack, inner ring crack, roller crack, and mixed fault.

[0109] Step 2: Optimize the symplectic geometric mode decomposition (SGMD) parameters K and α using the sparrow search algorithm (SSA), as follows:

[0110] (1) Initialize the parameters K and α and their range, namely:

[0111] 0≤K≤0.1

[0112] 0.8≤α≤1

[0113] (2) Hénon chaotic map initializes population parameters, namely:

[0114] Initialize the population, that is:

[0115]

[0116] Among them, t is the number of iterations, a and b are adjustable parameters, and the optimal chaotic sequence is calculated, that is:

[0117] z(t)=abs[ωy1(t)]i=1,2,…n

[0118] The inverse mapping obtains the sparrow position, which determines the position of the sparrow in the search space. The sparrow position is continuously optimized during the algorithm iteration process to find the optimal solution:

[0119] x i t=lb i +z(t)(ub i -lb i )

[0120] Among them, z(t) is the optimal chaotic sequence; ub i lb i are the upper and lower limits of the search space,

[0121] (3) Calculate the fitness value of the initialized population, namely:

[0122] Set the minimum sample entropy as the fitness function, that is:

[0123]

[0124] Among them, X i is the sample set, n' is the number of different categories, x i is the i-th category in the sample, P(x i ) is the probability of the i-th category in the sample, setting the moderation function,

[0125]

[0126] (4) Update the discoverer through nonlinear adaptive weight factors, namely:

[0127] Calculate the nonlinear adaptive weight factor, namely:

[0128]

[0129] Where t is the current iteration number; T is the maximum iteration number, and the discoverer position is updated:

[0130]

[0131] Among them, α is a random number within (0,1]; R2 is the warning value, R2∈[0,1]; ST is the safety value, ST∈[0.5,1]; Q is a random number that obeys the normal distribution;

[0132] (5) Use Levy flight and Cauchy mutation strategies to update the follower position and the alerter position, that is:

[0133] Calculate the value of Levy, that is,

[0134]

[0135] Among them, Γ(x)=(x-1)!, Q is a random number that obeys the normal distribution; r3, r4∈[0,1]; ξ is an adjustable constant with a value of 1.5. The global optimal position is calculated according to the Cauchy strategy.

[0136] x newbest =x best +x best ×Cauchy(0,1)

[0137] Among them, Cauchy (0, 1) is the standard Cauchy distribution function, and the updated follower position is:

[0138]

[0139] Among them, x worst is the worst position of the finder, d is the vector dimension, when i>n / 2, the hungry i-th follower flies to other locations to get food and updates the position of the alerter.

[0140]

[0141] Among them, K is the moving direction, which is also the control parameter, K∈[-1,1]; β is the step size parameter; f g 、f w are the global optimal and worst fitness respectively, f i is the individual fitness, ο is the minimum constant to prevent the denominator from being 0,

[0142] (6) Repeat steps (3)-(5) to the maximum number of iterations to obtain the optimal parameters K and α. The input of the sparrow search algorithm is a two-dimensional array, and the optimized result is also a two-dimensional array, so the output is the optimal K and α.

[0143] Step 3: Perform optimized SSA-SGMD denoising on the Western Reserve University bearing data. The steps are as follows:

[0144] (1) The power spectral density (PSD) method is used to obtain the signal trajectory matrix X, namely:

[0145] For the time series signal x=x1,x2,…,x n (n represents the signal data length), calculate the PSD of the original signal x and find the frequency f with the maximum peak max , let the frequency f max With sampling frequency F s When the ratio is less than K obtained by the sparrow search algorithm, the embedding dimension d = n / 3 is set; otherwise, d = 1.2 × (F S / f max), for the delay time, usually τ = 1, to obtain the trajectory matrix X g .

[0146]

[0147] Where d' represents the embedding dimension, and m = n"-(d'-1)τ;

[0148] (2) Obtain the covariance matrix A, namely:

[0149] For the matrix X g Perform autocorrelation analysis to obtain the covariance matrix A, which is:

[0150] A=X g T X g ;

[0151] (3) Establish the symplectic geometric orthogonal matrix Q, namely:

[0152] According to the symplectic geometry theorem, the Hamilton matrix M is constructed, namely:

[0153]

[0154] Further processing of the matrix M yields another Hamilton matrix W, namely:

[0155] W=M 2 ;

[0156] Establish an equation involving the symplectic geometry matrix G, namely:

[0157]

[0158] Among them, the matrices G and B are symplectic orthogonal matrices and upper triangular matrices respectively. Set a Householder matrix Q and construct the matrix H = [Q 0; 0 Q], which can be proved that H is also a Householder matrix and the matrix G can be replaced by the matrix H. Assume is the eigenvalue of matrix A, then λ1,λ2,…λ d is the eigenvalue of matrix B, Q i (i=1,2,…,d) is the matrix A 2 The eigenvectors corresponding to the eigenvalues, the eigenvalues ​​of matrix A and matrix B are correlated. Through symplectic geometric similarity transformation, matrix Q maintains the essential characteristics of the original signal, and QR decomposition is used to transform matrix A. 2 After decomposition, we can get the eigenvector matrix Q.

[0159] (4) Reconstruct the initial single-component reconstruction matrix to obtain the reconstruction matrix Z, which is:

[0160] The coefficient matrix S is constructed using the eigenvector matrix Q and the original trajectory matrix X:

[0161] S i =Q i T X g

[0162] Use the eigenvector matrix Q and the coefficient matrix S to obtain the reconstruction matrix Z:

[0163] Z i =Q i S i

[0164] Among them, Z i (i=1,2,…d") is the initial single-component reconstruction matrix; the reconstruction matrix Z is composed of d" groups of initial single-component reconstruction matrices Z i Composition: Z = Z1 + Z2 + ... + Z d" .

[0165] (5) Construct the initial symplectic geometric component matrix, namely:

[0166] Reconstruct the matrix Z from the single component using diagonal averaging i (1≤k≤d) is transformed to obtain an initial single-component signal of length l, and finally d" one-dimensional initial single-component signals are obtained. The sum of all initial single components is the original signal; let Z i The elements in are defined as z ij (1≤i≤m,1≤j≤d"); set d * =min(m,d"), m * =max(m,d") and 1≤i≤m,1≤j≤d", if m<d", then z * ij =z ij , otherwise z * ij =z ji The conversion expression is:

[0167]

[0168] This allows the initial single-component reconstruction matrix Z i Convert to a one-dimensional time series Y i (the obtained y1,y2,…,y n is a one-dimensional time series Y i Therefore, a one-dimensional time series can be obtained by diagonally averaging each initial single-component reconstruction matrix, and then d groups of initial single-component signals can be obtained, namely Y1, Y2, …Y d .

[0169] (6) Reorganize the initial symplectic geometric component matrix, that is:

[0170] First, the period of the d" initial single components is calculated; then the first initial single component Y1 is selected and compared with the period of the remaining d"-1 initial single components, and the initial single components are reorganized within the allowable error range (α obtained by the sparrow search algorithm) to obtain the first symplectic geometric component (SGC1); then the constituent initial single component signals of SGC1 are removed, and the remaining initial single components are summed, denoted as g1; finally, the first component of the remaining initial single components is selected to perform period similarity analysis with other components to obtain SGC2. After h iterations, h symplectic geometric components can be obtained, and the sum of the remaining initial single components can be denoted as g h ,

[0171] (7) Calculate the normalized mean square error NMSE, that is:

[0172] The normalized mean square error NMSE of the residual signal can be calculated as follows:

[0173]

[0174] Among them, g h (e) is g h The normalized mean square error of , x(e) is the normalized mean square error of x,

[0175] (8) Obtain the symplectic geometric component SGC, namely:

[0176] When the NMSE h If the value is less than the set threshold, the decomposition ends; otherwise, the iterative decomposition continues until the decomposition ends.

[0177] The expression for the stop time can be expressed as:

[0178]

[0179] Where N is the number of symplectic geometric components obtained, x(n) is the original signal, SGC h (n) hth decomposition component, g (N+1) (n) is the remaining component after N times of decomposition.

[0180] Step 4: First, perform preprocessing operations such as denoising and filtering on the one-dimensional vibration data to eliminate noise or interference that may affect signal analysis. Overlapping sampling is used to reduce the interval between sampling windows, thereby increasing the sampling density. This means that more data can be acquired in a shorter time, improving the efficiency of sampling. In addition, overlapping sampling can reduce the information loss caused by the interval between sampling windows, which helps to improve the accuracy and reliability of sampling while improving the frequency resolution, which is particularly useful for analyzing high-frequency parts or signals that require more refined frequency analysis.

[0181] Step 5. First, perform continuous wavelet transform on the vibration signal. Continuous wavelet transform decomposes the signal into information in the time domain and frequency domain, and provides the ability to locate fault characteristics on the time-frequency plane, which can more accurately determine the time and frequency of the fault, and help to accurately diagnose the health of the bearing. It can also convert the time domain signal to the time-frequency domain, analyze the energy distribution of the signal at different frequencies and times, and effectively remove noise, improving the accuracy and reliability of fault diagnosis. Generally speaking, a non-stationary signal can be represented by the sum of a series of harmonics of different frequencies as follows:

[0182]

[0183] Among them: A k (t) is the instantaneous amplitude of the kth harmonic; is the instantaneous phase of the kth harmonic; r(t) is the noise. Taking the derivative of the instantaneous image phase in time, we get the instantaneous frequency:

[0184]

[0185] Perform continuous wavelet analysis on the filtered vibration signal. The steps are as follows:

[0186] (1) The CWT of the transformed vibration signal in the time domain is:

[0187]

[0188] Where: W s (a1, b1) is the complex wavelet coefficient spectrum of the signal; s(t) is the original signal; a1 is the scale factor; b1 is the time shift factor; ψ(t) is the wavelet basis function.

[0189] (2) Convert the time domain into the frequency domain. The conversion formula is:

[0190]

[0191] in: is the Fourier transform of s(ξ); is the complex conjugate of the Fourier transform of ψ(aξ); ξ is the angular frequency.

[0192] Step 6. In actual signal analysis, the wavelet coefficients are distributed close to the frequency domain, which will cause the signal's time-frequency curve to appear blurred and have low resolution, especially when the signal components are complex. Although the wavelet coefficients are diffused in the frequency direction, the signal phase remains constant. Find the instantaneous frequency of the signal and perform synchronous squeezed wavelet transform. The steps are as follows:

[0193] (1) By analyzing the phase of the wavelet coefficients in the time domain, the instantaneous frequency of the signal can be obtained:

[0194]

[0195] (2) The wavelet coefficients are transformed from the time scale domain to the time-frequency domain. Based on this concept, the instantaneous frequency of the signal can be obtained by analyzing the phase of the wavelet coefficients in the time domain, and on this basis, the frequency is squeezed and reconstructed to concentrate its energy, thereby improving the fuzzy phenomenon in the frequency domain. In the discrete case, for any center frequency ω1 The inner wavelet coefficients are squeezed, and the wavelet transform values ​​can be squeezed synchronously. The mathematical expression is as follows:

[0196]

[0197] When the variables are continuous, the mathematical expression is:

[0198] T s (ω,b1)=∫ A(b) W s (a1,b1)a -3 / 2 δ(W(a1,b1)-ω)da1

[0199] Finally, the synchronous squeezing wavelet transformed data is used for graph drawing to obtain the synchronous squeezing wavelet transform time-frequency diagram of various bearing fault types.

[0200] Step 7: Use basic image processing methods to expand samples and obtain the data set T-set. Then use DCGAN to generate adversarial networks to generate data sets G-set. Finally, use improved VGG for classification, including:

[0201] (1) The generated two-dimensional image is expanded using basic image processing methods. The two-dimensional image is flipped, rotated, cropped, color-dithered, and noise is added to obtain the data set T-set.

[0202] (2) The deep convolutional generative adversarial network (APCDCGAN) framework based on DCGAN is used. DCGAN generates images of size 64×64 pixels2. To avoid low resolution, APCDCGAN is designed to generate images of size 256×256 pixels2. Compared with DCGAN, the deconvolutional layers and convolutional layers of the generator and discriminator of APCDCGAN are deconvolutional layers, respectively.

[0203] (3) DCGAN was trained on three datasets: LSUN, Imagenet-1k, and a Faces dataset, which contain at least tens of thousands of images. The kernel size of 4×4 was replaced with a kernel size of 3×3.

[0204] For example, the three datasets are: LSUN, a picture dataset from the Computer Vision Laboratory of Stanford University, Imagenet-1k, a picture dataset from the Imagenet official website, and Faces, a picture dataset from Flickr.

[0205] (4) Then the generated dataset G-set is input into the improved VGG network for training. In order to reduce parameters, global average pooling (GAP) is used to replace the FC layer.

[0206] (5) First, the two-dimensional image data is input into the network for batch normalization processing. Each input data is normalized by making the expected value 0 and the variance 1:

[0207]

[0208] Among them, x (k) , E and Var are the normalized activation value, mean value and variance respectively. For each n-dimensional input x = (x (1) …x (n) ), k represents the kth dimension.

[0209] (6) To achieve the same transformation, change the normalized value to:

[0210]

[0211] in, and β (k) =E[x (k) ]. These two settings can restore the feature distribution learned by a specific layer.

[0212] (7) The normalized image is input into the convolutional layer of the network model for convolution operation. First, the original image is 128×128. After being input into the first layer of the network model, convolution and pooling operation is performed. Then, after passing through the first convolution and pooling layer, the size is reduced to 64×64. After passing through another convolution and pooling layer, the size is reduced to 32×32.

[0213] (8) Optimize the filter size to improve the training speed of the network model. VGG16 involves a convolution kernel of size 3×3 and a maximum pooling of size 2×2. In each convolution block, the stride of the convolution layer is set to 1, and the same padding is used in the maximum pooling layer with a stride of 2.

[0214] (9) Then the improved VGG16 network model is classified using the softmax classifier. In this model, the classification is reduced to 5 categories. The loss function combining the softmax function and the cross entropy is defined as:

[0215]

[0216] Among them, p(x i ) is the true distribution, q(x i ) is the predicted distribution.

[0217] Corresponding to the above method, this embodiment further provides a rolling bearing fault diagnosis system for complex working conditions, including:

[0218] An optimization module is used to optimize the optimization parameters of symplectic geometric mode decomposition using a sparrow search algorithm to obtain an optimized SSA-SGMD algorithm;

[0219] A noise reduction module, used for performing noise reduction preprocessing on the collected rolling bearing vibration signal by using the SSA-SGMD algorithm to obtain a preprocessed signal;

[0220] A screening module, used for screening the pre-processed signal, and performing synchronous extrusion wavelet transform on the screened pre-processed signal, so as to convert the one-dimensional screened pre-processed signal into a two-dimensional image;

[0221] An expansion module, used to perform sample expansion on the two-dimensional image to obtain a first data set, and input the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set;

[0222] The diagnosis module is used to extract and classify the two-dimensional images in the first data set and the second data set using the improved VGG network model to obtain the characteristic differences between different fault types to realize the fault diagnosis of the rolling bearing.

[0223] Preferably, the rolling bearing vibration signal is collected through a multi-condition rolling bearing fault simulation experimental platform; the fault types of the data acquired by the multi-condition rolling bearing fault simulation experimental platform include: normal state, outer ring crack, inner ring crack, roller crack and mixed fault.

[0224] Preferably, the optimization module comprises:

[0225] A parameter initialization unit, used for initializing parameters K and α, wherein the ranges of K and α are: 0≤K≤0.1; 0.8≤α≤1; wherein K is the first optimization parameter, and α is the second optimization parameter;

[0226] The population initialization unit is used to initialize the population; the formula for initializing the population is: t is the number of iterations, a and b are adjustable parameters, y1 represents the initial position of the population individual, and y2 represents the next position of the population individual;

[0227] A sequence calculation unit is used to calculate an optimal chaotic sequence; the formula of the optimal chaotic sequence is: z(t)=abs[ωy1(t)], i=1,2,…n; wherein z(t) is the optimal chaotic sequence, ω is a proportional factor, n is the total number of signals, and abs represents an absolute value;

[0228] The inverse mapping unit is used to perform inverse mapping according to the optimal chaotic sequence to obtain the sparrow position; the formula of the sparrow position is: i t =lb i +z(t)(ub i -lb i );where ub i lb i are the upper and lower limits of the search space, x i t is the position of the sparrow;

[0229] The first function determination unit is used to set the minimum sample entropy as the fitness function; the formula of the fitness function is: Among them, X is a sample set consisting of the positions of sparrows, n' is the number of different categories, x i is the i-th category in the sample, P(x i ) is the probability of the i-th category in the sample, H(X) is the fitness function;

[0230] The second function determination unit is used to set a fitness function according to the fitness function; the formula of the fitness function is: Wherein, F(X) is the fitness function;

[0231] The factor calculation unit is used to calculate the nonlinear adaptive weight factor; the formula of the nonlinear adaptive weight factor is: Wherein, T is the maximum number of iterations, and λ is the nonlinear adaptive weight factor;

[0232] The first position updating unit is used to update the finder position; the formula for updating the finder position is: Among them, R2 is the warning value, R2∈[0,1]; ST is the safety value, ST∈[0.5,1]; Q is a random number that obeys the normal distribution, and are the current position and next position of the finder respectively;

[0233] The flight strategy unit is used to calculate the Levy value in the Levy flight strategy; the calculation formula for the Levy value is: Among them, Γ(x)=(x-1)!; r3 and r4 are calculation parameters, r3, r4∈[0,1]; ξ is an adjustable constant, the value is 1.5,

[0234] The optimal position determination unit is used to calculate the global optimal position according to the Cauchy strategy; the calculation formula of the global optimal position is: newbest =x best +x best ×Cauchy(0,1); where Cauchy(0,1) is the standard Cauchy distribution function, x best The current optimal position, x newbest is the global optimal position;

[0235] The second position updating unit is used to update the follower position; the calculation formula of the follower position is: Among them, x worst is the worst position of the finder, d is the vector dimension, when i>n / 2, the hungry i-th follower flies to other positions to get food;

[0236] The third position updating unit is used to update the position of the alerter; the formula of the position of the alerter is: Among them, the first optimization parameter K represents the moving direction, K∈[-1,1]; β is the step size parameter; f g 、f w are the global optimal and worst fitness respectively, f i is the individual fitness, ο is the minimum constant, is the position with the largest fitness value, It is the position where the moderation value is minimum to prevent the denominator from being zero.

[0237] A repeating unit is used to repeat the steps of "setting the minimum sample entropy as the fitness function" to "updating the position of the alerter" until the maximum number of iterations is reached to obtain the optimal first optimization parameter. K The second optimization parameter α.

[0238] The beneficial effects of the present invention are as follows:

[0239] The present invention is applicable to the situations where parts in a mechanical system operate in coordination, the working environment is disturbed and contains a large amount of noise. The present invention can effectively reduce the noise of the signal and extract distinctive two-dimensional features. The improved VGG neural network is used to identify and classify the two-dimensional features. The comprehensive accuracy rate is as high as over 97%. In addition, the method of the present invention can achieve high-accuracy classification in different working environments, and has strong generalization ability and adaptability to working conditions.

[0240] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0241] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A rolling bearing fault diagnosis method for complex working conditions. It is characterized in that include: The optimization parameters of symplectic geometric mode decomposition are optimized using the sparrow search algorithm to obtain the optimized SSA-SGMD algorithm; The SSA-SGMD algorithm is used to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal; The preprocessed signal is screened, and the screened preprocessed signal is subjected to synchronous squeezing wavelet transformation to convert the one-dimensional screened preprocessed signal into a two-dimensional image; Performing sample expansion on the two-dimensional image to obtain a first data set, and inputting the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set; The improved VGG network model is used to extract and classify the two-dimensional images in the first data set and the second data set to obtain the characteristic differences between different fault types, so as to realize the fault diagnosis of rolling bearings.

2. The rolling bearing fault diagnosis method under complex working conditions according to claim 1, It is characterized in that The rolling bearing vibration signal is collected through a multi-operating condition rolling bearing fault simulation experimental platform; The fault types of the data acquired by the multi-operating condition rolling bearing fault simulation experimental platform include: normal state, outer ring crack, inner ring crack, roller crack and mixed fault.

3. The rolling bearing fault diagnosis method under complex working conditions according to claim 1, It is characterized in that The optimization parameters of symplectic geometric mode decomposition are optimized using the sparrow search algorithm, including: Initialization parameters K and α, the range of K and α is: 0≤K≤0.1; 0.8≤α≤1; where K is the first optimization parameter and α is the second optimization parameter; Initialize the population; the formula for initializing the population is: t is the number of iterations, a and b are adjustable parameters, y 1 represents the initial position of the individuals in the population, y 2 represents the next position of the individuals in the population; Calculate the optimal chaotic sequence; the formula for the optimal chaotic sequence is: z(t) = abs[ωy 1 (t)], i = 1, 2, ... n; wherein z(t) is the optimal chaotic sequence, ω is the proportional factor, n is the total number of signals, and abs represents the absolute value; According to the optimal chaotic sequence, inverse mapping is performed to obtain the sparrow position; the formula for the sparrow position is: Among them, ub i lb i are the upper and lower limits of the search space, x i t is the position of the sparrow; The minimum sample entropy is set as the fitness function; the formula of the fitness function is: Among them, X is a sample set consisting of the positions of sparrows, n' is the number of different categories, x i is the i-th category in the sample, P(x i ) is the probability of the i-th category in the sample, H(X) is the fitness function; A fitness function is set according to the fitness function; the formula of the fitness function is: Wherein, F(X) is the fitness function; Calculate the nonlinear adaptive weight factor; the formula of the nonlinear adaptive weight factor is: Wherein, T is the maximum number of iterations, and λ is the nonlinear adaptive weight factor; Update the finder's position; the formula for updating the finder's position is: Among them, R 2 is the warning value, R 2 ∈[0,1]; ST is a safety value, ST∈[0.5,1]; Q is a random number that follows a normal distribution, and are the current position and next position of the finder respectively; The formula for calculating the value of Levy in the Levy flight strategy is: Where Γ(x) = (x-1)!; r 3 and r 4 are all calculation parameters, r 3 、r 4 ∈[0,1]; ξ is an adjustable constant, with a value of 1.5, The global optimal position is calculated according to the Cauchy strategy; the calculation formula of the global optimal position is: newbest =x best +x best ×Cauchy(0,1); where Cauchy(0,1) is the standard Cauchy distribution function, x best The current optimal position, x newbest is the global optimal position; Update the follower position; the calculation formula of the follower position is: Among them, x worst is the worst position of the finder, d is the vector dimension, when i>n / 2, the hungry i-th follower flies to other positions to get food; Update the alerter position; the formula for the alerter position is: Among them, the first optimization parameter K represents the moving direction, K∈[-1,1]; β is the step size parameter; f g 、f w are the global optimal and worst fitness respectively, f i is the individual fitness, ο is the minimum constant, is the position with the largest fitness value, It is the position where the moderation value is minimum to prevent the denominator from being zero. Repeat steps "Set minimum sample entropy as fitness function" to step "Update alerter position" until the maximum number of iterations is reached to obtain the optimal first optimization parameter K and second optimization parameter α.

4. The rolling bearing fault diagnosis method under complex working conditions according to claim 3, It is characterized in that The SSA-SGMD algorithm is used to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal, including: Calculate the PSD of the original rolling bearing vibration signal x and find the frequency f of the maximum peak max , let the frequency f max With sampling frequency F s When the ratio is less than K obtained by the sparrow search algorithm, the embedding dimension d'=n / 3 is set, otherwise, d=1.2×(F S / f max ), for the delay time, usually τ = 1, to obtain the trajectory matrix X g The rolling bearing vibration signal x is expressed as: x = x 1 ,x 2 ,…,x n" , n" represents the signal data length, and the formula of the trajectory matrix is: Where d' represents the embedding dimension, and m = n"-(d'-1)τ; For the matrix X g Perform autocorrelation analysis to obtain the covariance matrix A, A = X g T X g ; According to the symplectic geometry theorem, the Hamilton matrix M is constructed using the covariance matrix A; Processing the Hamilton matrix M yields another Hamilton matrix W; W = M 2 ; Establish equations involving symplectic geometry matrix G based on Hamilton matrix W; Among them, the matrices G and B are symplectic orthogonal matrices and upper triangular matrices respectively; Using QR decomposition, we can decompose the matrix A 2 Decompose and obtain the eigenvector matrix Q; Using the eigenvector matrix Q and the original trajectory matrix X g Construct the coefficient matrix S; S i =Q i T X g ; The reconstruction matrix Z is obtained by using the eigenvector matrix Q and the coefficient matrix S; Z i =Q i S i ; Among them, Z i (i=1,2,…d") is the initial single-component reconstruction matrix; the reconstruction matrix Z is composed of d" groups of initial single-component reconstruction matrices Z i Composition: Z = Z 1 +Z 2 +…+Z d" ; Reconstruct the matrix Z from the single component using diagonal averaging i (1≤k≤d) is transformed to obtain a one-dimensional initial single-component signal with a length of l, and finally d" one-dimensional initial single-component signals are obtained. The sum of all initial single components is the original signal; let Z i The elements in are defined as z ij (1≤i≤m,1≤j≤d"); set d * =min(m,d"), m * =max(m,d") and 1≤i≤m,1≤j≤d", if m<d", then z * ij =z ij , otherwise z * ij =z ji ;d * Represents the number of one-dimensional initial unit component signals obtained after conversion; m * It is used to define Z i The maximum index of Represents the converted Z i Elements in the matrix; the expression to be converted is: y k Represents the kth element of the converted signal; k index variable, indicating the position of the current signal processing; Calculate the period of d" one-dimensional initial single-component signals and select the first one-dimensional initial single-component signal Y 1 , and let Y 1 With Y 1 The periods of the remaining d"-1 initial single components are compared, and the initial single components are reorganized within the optimal range of the first optimization parameter α to obtain the first symplectic geometric component SGC1. The initial single component signal that constitutes SGC1 is removed, and the remaining initial single components after removal are summed up and recorded as g1; finally, the first component of the remaining initial single component is selected to perform period similarity analysis with other components to obtain SGC2; after h iterations, h symplectic geometric components are obtained, and the sum of the remaining initial single components is recorded as g h ; Calculate the normalized mean square error NMSE of the residual signal; Among them, g h (e) is g h The normalized mean square error of , x(e) is the normalized mean square error of x; When the NMSE h When it is less than the set threshold, the decomposition ends; otherwise, the iterative decomposition continues, and the expression when the decomposition ends is expressed as: Where N is the number of symplectic geometry components obtained.

5. The rolling bearing fault diagnosis method under complex working conditions according to claim 4, It is characterized in that The SSA-SGMD algorithm is used to perform noise reduction preprocessing on the collected rolling bearing vibration signal to obtain a preprocessed signal, and further includes: The vibration signal after noise reduction processing is segmented, and the length of each segment of the vibration signal is determined, and overlapping sampling is used to obtain the pre-processed signal converted into a two-dimensional image.

6. The rolling bearing fault diagnosis method under complex working conditions according to claim 5, It is characterized in that Performing synchronous squeezing wavelet transform on the filtered preprocessed signal to convert the one-dimensional filtered preprocessed signal into a two-dimensional image, including: The CWT of the preprocessed signal after transformation and screening in the time domain is: Where: W s (a 1 ,b 1 ) is the complex wavelet coefficient spectrum of the signal; a 1 is the scale factor; b 1 is the time shift factor; ψ(t) is the mother wavelet function; s(t) is the original signal; Convert the time domain to the frequency domain, the conversion formula is: Where: ξ is the angular frequency; is the Fourier transform of the original signal; is a complex exponential function, representing a rotation factor used to adjust the position of the wavelet on the time axis; is the complex conjugate of the mother wavelet; By resolving the phase of the wavelet coefficients in the time domain, the instantaneous frequency of the signal is obtained: In the discrete case, for any center frequency ω 1 Department The inner wavelet coefficients are squeezed to synchronously squeeze the wavelet transform values. The mathematical expression is: Δω represents the frequency interval; T s (ω 1 ,b 1 ) means when the scale is ω 1 Position b 1 The result of synchronous squeezing wavelet transform when W s (a k ,b 1 ) indicates the scale is a k Position b 1 Continuous wavelet transform coefficient at time ; (Δa) k For different scales a k The interval between When the variables are continuous, the mathematical expression is: The synchronous squeezing wavelet transformed data are used for graphing, and the two-dimensional images of synchronous squeezing wavelet transformed time-frequency of various bearing fault types are obtained.

7. The rolling bearing fault diagnosis method under complex working conditions according to claim 1, It is characterized in that Performing sample expansion on the two-dimensional image to obtain a first data set, and inputting the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set, including: Flipping, rotating, cropping, color dithering, and adding noise to the two-dimensional image to obtain the first data set; Based on the first data set, images are generated using a deep convolutional generative adversarial network framework layer based on DCGAN to obtain the second data set.

8. A rolling bearing fault diagnosis system for complex working conditions. It is characterized in that include: An optimization module is used to optimize the optimization parameters of symplectic geometric mode decomposition using a sparrow search algorithm to obtain an optimized SSA-SGMD algorithm; A noise reduction module, used for performing noise reduction preprocessing on the collected rolling bearing vibration signal by using the SSA-SGMD algorithm to obtain a preprocessed signal; A screening module, used for screening the pre-processed signal, and performing synchronous extrusion wavelet transform on the screened pre-processed signal, so as to convert the one-dimensional screened pre-processed signal into a two-dimensional image; An expansion module, used to perform sample expansion on the two-dimensional image to obtain a first data set, and input the first data set into a deep convolutional generative adversarial network framework based on DCGAN to obtain a second data set; The diagnosis module is used to extract and classify the two-dimensional images in the first data set and the second data set using the improved VGG network model to obtain the characteristic differences between different fault types to realize the fault diagnosis of the rolling bearing.

9. The rolling bearing fault diagnosis system for complex working conditions according to claim 8, It is characterized in that The rolling bearing vibration signal is collected through a multi-operating condition rolling bearing fault simulation experimental platform; The fault types of the data acquired by the multi-operating condition rolling bearing fault simulation experimental platform include: normal state, outer ring crack, inner ring crack, roller crack and mixed fault.

10. The rolling bearing fault diagnosis system for complex working conditions according to claim 8, It is characterized in that The optimization module includes: A parameter initialization unit, used for initializing parameters K and α, wherein the ranges of K and α are: 0≤K≤0.1; 0.8≤α≤1; wherein K is the first optimization parameter, and α is the second optimization parameter; The population initialization unit is used to initialize the population; the formula for initializing the population is: t is the number of iterations, a and b are adjustable parameters, y 1 represents the initial position of the individuals in the population, y 2 represents the next position of the individuals in the population; The sequence calculation unit is used to calculate the optimal chaotic sequence; the formula of the optimal chaotic sequence is: z(t) = abs[ωy 1 (t)], i = 1, 2, ... n; wherein z(t) is the optimal chaotic sequence, ω is the proportional factor, n is the total number of signals, and abs represents the absolute value; The inverse mapping unit is used to perform inverse mapping according to the optimal chaotic sequence to obtain the sparrow position; the formula of the sparrow position is: Among them, ub i lb i are the upper and lower limits of the search space, is the position of the sparrow; The first function determination unit is used to set the minimum sample entropy as the fitness function; the formula of the fitness function is: Among them, X is a sample set consisting of the positions of sparrows, n' is the number of different categories, x i is the i-th category in the sample, P(x i ) is the probability of the i-th category in the sample, H(X) is the fitness function; The second function determination unit is used to set a fitness function according to the fitness function; the formula of the fitness function is: Wherein, F(X) is the fitness function; The factor calculation unit is used to calculate the nonlinear adaptive weight factor; the formula of the nonlinear adaptive weight factor is: Wherein, T is the maximum number of iterations, and λ is the nonlinear adaptive weight factor; The first position updating unit is used to update the finder position; the formula for updating the finder position is: Among them, R 2 is the warning value, R 2 ∈[0,1]; ST is a safety value, ST∈[0.5,1]; Q is a random number that follows a normal distribution, and are the current position and next position of the finder respectively; The flight strategy unit is used to calculate the Levy value in the Levy flight strategy; the calculation formula for the Levy value is: Where Γ(x) = (x-1)!; r 3 and r 4 are all calculation parameters, r 3 、r 4 ∈[0,1]; ξ is an adjustable constant, with a value of 1.5, The optimal position determination unit is used to calculate the global optimal position according to the Cauchy strategy; the calculation formula of the global optimal position is: newbest =x best +x best ×Cauchy(0,1); where Cauchy(0,1) is the standard Cauchy distribution function, x best The current optimal position, x newbest is the global optimal position; The second position updating unit is used to update the follower position; the calculation formula of the follower position is: Among them, x worst is the worst position of the finder, d is the vector dimension, when i>n / 2, the hungry i-th follower flies to other positions to get food; The third position updating unit is used to update the position of the alerter; the formula of the position of the alerter is: Among them, the first optimization parameter K represents the moving direction, K∈[-1,1]; β is the step size parameter; f g 、f w are the global optimal and worst fitness respectively, f i is the individual fitness, ο is the minimum constant, is the position with the largest fitness value, It is the position where the moderation value is minimum to prevent the denominator from being zero. The repeating unit is used to repeat the steps of "setting the minimum sample entropy as the fitness function" to "updating the position of the alerter" until the maximum number of iterations is reached to obtain the optimal first optimization parameter K and the second optimization parameter α.

Images