Bearing fault diagnosis method based on IGRO-LSSVM

Abstract

The invention discloses a bearing fault diagnosis method based on an IGRO-LSSVM (Inverse Generalized Register-Least Square Support The method comprises the following steps: firstly, extracting time-frequency domain features of original signals and features of SVD noise reduction signals based on information entropy to construct an original feature set; for irrelevant or redundant features existing in an original feature set, a feature evaluation method is adopted for feature selection, and a feature subset with good sensitivity and high stability is obtained. In view of insufficient global and local search capabilities of a gold washing optimizer and easy falling into local optimum, an elite reverse learning strategy, a dynamic boundary strategy of a cosine control factor and adaptive t distribution variation are introduced to form a multi-strategy enhanced gold washing optimizer (IGRO). And because the parameter selection of the LSSVM can influence the classification effect, the IGRO is adopted to optimize the parameters, and finally, a fault diagnosis model based on the IGRO-LSSVM is constructed to realize fault diagnosis. Experiments show that the method is high in fault diagnosis accuracy, and can accurately classify faults of bearings of different types and different damage diameters.

Description

Technical Field

[0001] The invention belongs to the technical field of bearing fault diagnosis, and relates to a bearing fault diagnosis method based on IGRO-LSSVM. Technical Background

[0002] Bearings are important parts of mechanical equipment and are crucial in operation. Due to long-term use and factors such as weather and temperature, bearings are prone to failure, which leads to reduced equipment efficiency or even shutdown, causing serious impact on production safety and economic benefits. Therefore, effective diagnosis of bearing failures and timely maintenance of equipment are of great significance to protecting people's lives and property and effectively preventing and avoiding safety accidents. At present, fault diagnosis technology mainly studies whether a fault occurs and locates the location and type of the fault. Rolling bearing fault diagnosis technology mainly consists of three parts: vibration signal preprocessing, fault feature extraction and fault type identification.

[0003] With the in-depth development of signal processing technology, many representative signal preprocessing methods have emerged, such as empirical mode decomposition (EMD), local mean decomposition (LMD) and variational mode decomposition (VMD).

[0004] In terms of fault feature extraction, when a rolling bearing fails, the representation of some characteristic parameters in the time domain and frequency domain will also change accordingly. Therefore, the fault features under different working conditions can be extracted based on the differences in the characteristic parameters in the time domain and frequency domain. However, for some signals with a low signal-to-noise ratio, it is difficult to obtain their fault features through simple analysis in the time domain and frequency domain. For this type of signal, it is necessary to combine it with other mathematical analysis methods to achieve the purpose of better extraction of fault features. Therefore, a fault feature extraction method based on information entropy is used.

[0005] With the continuous advancement of artificial intelligence technology, a large number of excellent machine learning models have emerged, including support vector machines, random forests, artificial neural networks, etc. Such models need to combine signal processing knowledge and use the extracted features as input to train the model. In addition, there are some hyperparameters that affect the performance of the model. Therefore, a large number of scholars combine signal processing, machine learning models and optimization algorithms for fault diagnosis. Least Squares Support Vector Machines (LS-SVM) is an extension of support vector machines (SVM). Least Squares Support Vector Machines transform the inequality constraints of the traditional support vector machine decision function into equality constraints, transform the convex quadratic programming problem in the support vector machine into a linear programming problem, transform the parameter search range from three dimensions to two dimensions, and use the quadratic loss function as the optimization indicator. While ensuring accuracy, the computational complexity is greatly reduced, which makes the least squares support vector machine have faster solution accuracy and operation speed when processing large data samples, and enhances the generalization ability. From the current status of rolling bearing fault diagnosis, it can be seen that the bearing fault diagnosis technology based on fault characteristic frequency is relatively mature, and there is a certain amount of technical accumulation in signal decomposition, feature extraction and type identification.

[0006] Traditional bearing fault diagnosis methods have the following defects: 1. Low speed causes the impact signal of rolling bearing fault to be weak, and the resonance frequency band response of the system fault caused is weak, which brings many difficulties to the selection of appropriate fault signals for analysis; 2. Heavy equipment is accompanied by relatively serious environmental noise during operation, which will make it difficult to extract and select fault features. Summary of the invention

[0007] The purpose of the present invention is to propose a bearing fault diagnosis method based on IGRO-LSSVM, to overcome the defects of traditional bearing fault diagnosis methods and to improve the accuracy of fault type identification, and to establish a bearing fault model based on IGRO-LSSVM.

[0008] The technical solution of the present invention is a bearing fault diagnosis method based on IGRO-LSSVM, and the specific steps are as follows:

[0009] Step 1: Extract the time domain and frequency domain features of the original vibration signal to construct feature set 1; perform SVD denoising on the original signal, extract features based on information entropy from the denoised signal, and construct feature set 2. Merge feature set 1 and feature set 2 to construct the original feature set.

[0010] Step 2: Use the feature evaluation method to evaluate the original feature set, obtain the evaluation score of the original feature, and select significant features from the original feature set according to the evaluation score, discard insensitive features, and finally obtain a low-dimensional sensitive feature set.

[0011] Step 3: Use IGRO to optimize the kernel function parameters and penalty factors of LSSVM, select the optimal parameter combination, and build a bearing fault diagnosis model based on IGRO-LSSVM.

[0012] The low-dimensional sensitive feature set obtained by feature selection is input into the bearing fault diagnosis model based on IGRO-LSSVM for training to realize the fault type diagnosis of the bearing.

[0013] As described above, the present invention relates to a bearing fault diagnosis method, which first extracts the time-frequency domain features of the bearing vibration signal, and performs SVD (Singular Value Decomposition) on the original signal to achieve noise reduction, extracts the information entropy-based features of the noise-reduced signal, and the time-domain features, frequency-domain features, and features based on information entropy together constitute the original feature set, and then the original feature vector set is divided into a training set and a test set according to a certain ratio. Secondly, a feature evaluation method is used to evaluate the training set to obtain an evaluation score for each feature, and according to the size of the original feature evaluation score, the first four features with larger evaluation scores are selected from the training set, and similarly, features with the same attributes as the training set are selected in the test set to construct the final training set and test set.

[0014] Then, in order to effectively improve the initial population quality of GRO, the elite reverse learning strategy is used for population initialization; the dynamic boundary strategy of the cosine control factor is introduced to balance the global and local search capabilities of the algorithm; the adaptive t distribution variation is introduced to enhance the algorithm's optimization performance and avoid local optimality. The elite reverse learning strategy, the dynamic boundary strategy of the cosine control factor and the adaptive t distribution variation together constitute IGRO.

[0015] In addition, the present invention also proposes a fault diagnosis model based on IGRO-LSSVM. Aiming at the problem that the LSSVM kernel function parameters and penalty factors are variables, the present invention uses IGRO to optimize the parameters of LSSVM, which greatly improves the recognition efficiency and recognition accuracy of bearing faults. Finally, the training set finally obtained is input into the bearing fault diagnosis model based on IGRO-LSSVM for training to realize the diagnosis of bearing fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Schematic diagram of rolling bearing fault diagnosis process.

[0017] Figure 2Flow chart of a rolling bearing fault diagnosis method in an embodiment of the present invention.

[0018] Figure 3 The time domain waveforms of the original signal and noisy signal of the outer race fault.

[0019] Figure 4 A line graph of evaluation scores according to an embodiment of the present invention.

[0020] Figure 5 This is the diagnostic result diagram of the GRO-LSSVM model.

[0021] Figure 6 This is the diagnosis result diagram of the IGRO-LSSVM model.

[0022] Figure 7 To improve the gold mining optimizer flow chart. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0024] like Figure 1 As shown, this embodiment describes a bearing fault diagnosis method, wherein the bearing fault diagnosis method mainly consists of four parts: vibration signal preprocessing, feature extraction, feature selection and fault diagnosis.

[0025] like Figure 2 As shown, a bearing fault diagnosis method based on IGRO-LSSVM includes the following steps:

[0026] Step 1: Extract the time domain and frequency domain features of the original vibration signal to construct feature set 1; perform SVD denoising on the original signal, extract features based on information entropy from the denoised signal, and construct feature set 2. Merge feature set 1 and feature set 2 to construct the original feature set.

[0027] SVD denoising is a nonlinear filtering method with good stability and invariance. After SVD denoising, the signal has a small phase shift and no time delay. Therefore, SVD is widely used in signal denoising, weak signal separation and periodic component extraction in fault diagnosis.

[0028] Time domain features, frequency domain features, and other domain features can reflect the relevant state information of the bearing from different aspects. However, for some signals with a low signal-to-noise ratio, it is difficult to obtain their fault characteristics through simple analysis of the time domain and frequency domain. For this type of signal, it is necessary to combine it with other mathematical analysis methods to achieve the purpose of better extraction of fault features. Therefore, a fault feature extraction method based on information entropy is established.

[0029] Step 1.1. Extract the time domain and frequency domain features of the original vibration signal and construct a feature set 1 based on the time domain and frequency domain. The specific content of the feature set 1 based on the time domain and frequency domain is shown in Table 1.

[0030] Table 1 Time-frequency domain characteristics

[0031]

[0032] Step 1.2. Use SVD to perform noise reduction on the original vibration signal. The specific process includes:

[0033] The vibration signal y(n) = [y 1 ,y 2 , ..., y N ] is constructed into a Hanke matrix (noise matrix A), expressed as:

[0034]

[0035] In the formula, when N is an even number, n=N / 2; when N is an odd number, n=(N+1) / 2, N=m+n-1, and the singular value decomposition of the noisy matrix A can be obtained:

[0036] A=USV T (2)

[0037] S=diag(σ 1 , σ 2 , ..., σ q , O), q = min (m, n) (3)

[0038] In the formula, σ i is the singular value of matrix A, σ 1 >σ 2 >...>σ q , O is a zero matrix, U and V are orthogonal matrices.

[0039] Select the number of valid singular values ​​p, retain the first p larger singular values ​​of the main diagonal matrix S, set all other singular values ​​to 0, and obtain a new diagonal matrix:

[0040] S′=diag(σ 1 , σ 2 ,...,σ p ,O) (4)

[0041] Then, a new signal matrix A′ is obtained through the orthogonal matrices U and V, and the reconstructed matrix A′ is restored to a one-dimensional signal.

[0042] Step 1.3. After obtaining the denoised signal, extract features based on information entropy and construct feature set 2 based on information entropy, including power spectrum entropy, singular spectrum entropy, approximate entropy, fuzzy entropy, permutation entropy and sample entropy.

[0043] The feature extraction based on information entropy is specifically calculated as follows:

[0044] (1) Calculate the power spectrum entropy H f :

[0045]

[0046] in, p k Indicates the proportion of the kth power spectrum component of the signal in the entire power spectrum, S k Represents the energy partition of a signal in the frequency domain.

[0047] (2) Calculate the singular spectrum entropy H t :

[0048]

[0049] in, p i represents the i-th singular value λ i The proportion of the whole singular value spectrum.

[0050] (3) Calculate the approximate entropy:

[0051] For a time series x(n)=x(1), x(2), x(3), ..., x(N) consisting of N data, they are sequentially organized into a vector sequence of dimension m. in

[0052] Defining vectors and Distance is the Chebyshev distance, that is, the maximum absolute value of the difference between each element:

[0053]

[0054] For a given statistics and The distance between The number of j (1≤j≤N-m+1) less than or equal to r, that is, the approximate number, denoted by c i For 1≤i≤N-m+1, define It is the ratio of the approximate quantity to the total quantity, that is, the approximate proportion:

[0055]

[0056] Definition φ m (r) is:

[0057]

[0058] According to the above definition, φ m The essence of (r) is the average value under the logarithmic representation of the approximate proportion.

[0059] Increase the dimension to m+1 and repeat the above steps to get φ m+1 (r).

[0060] The approximate entropy ApEn(m, r) of the signal sequence is calculated as:

[0061] ApEn(m,r)=φ m (r)-φ m+1 (r) (10)

[0062] (4) Calculate fuzzy entropy:

[0063] For a time series x(n)=x(1), x(2), x(3), ..., x(N) consisting of N data, they are sequentially organized into a vector sequence of dimension m. in express The mean value of

[0064] Defining vectors and Distance is the Chebyshev distance, that is, the maximum absolute value of the difference between each element:

[0065]

[0066] Introducing fuzzy membership n measure Similarity. For 1≤j≤Nm, j≠i:

[0067]

[0068] Definition φ m (n, r) is:

[0069]

[0070] Increase the dimension to m+1 and repeat the above steps to get φ m+1 (n, r).

[0071] The fuzzy entropy FuzzyEn(m, r, n) of the calculated signal sequence is as follows:

[0072] FuzzyEn(m, r, n) = lnφ m (n,r) - lnφ m+1 (n, r) (14)

[0073] (5) Calculate the permutation entropy:

[0074] For the signal sequence x(n) = x(1), x(2), x(3),..., x(N) with length N, perform phase space reconstruction to obtain the reconstructed vector sequence. where The parameter m is the embedding dimension, and the parameter τ is called the delay time, taking positive integers.

[0075] For Sort each element in the sequence to obtain the sorting index π.

[0076] For an m-dimensional sequence, there are m! possible sorting indices π, and the actual occurrence mode is denoted as J (J < m!). Define the j-th sorting index π j The number of occurrences v(π j ), then the occurrence probability p j is:

[0077]

[0078] Calculate the permutation entropy PermEn(m, τ) of the signal sequence as:

[0079]

[0080] The maximum value of the permutation entropy theory is Normalize the above permutation entropy:

[0081]

[0082] (6) Calculate the sample entropy:

[0083] For the time series x(n) = x(1), x(2), x(3),..., x(N) composed of N data, form a vector sequence with dimension m in order. where

[0084] Define the vector and The distance is the Chebyshev distance, that is, the maximum value of the absolute value of the difference of each element:

[0085]

[0086] For a given statistics and The distance between The number of j (1≤j≤NN-m, j≠i) less than or equal to r, that is, the approximate number, is denoted as

[0087] For 1≤j≤Nm, define It is the ratio of the approximate quantity to the total quantity, that is, the approximate proportion:

[0088]

[0089] Definition φ m (r) is:

[0090]

[0091] Increase the dimension to m+1 and repeat the above steps to get φ m+1 (r).

[0092] Calculate the sample entropy SampEn(m, r) of the signal sequence:

[0093]

[0094] When N is a finite value, use the following estimate:

[0095]

[0096] Step 1.4. Merge feature set 1 and feature set 2 to construct the original feature set, as shown in the formula:

[0097]

[0098] In the formula, T represents the time domain features, F represents the frequency domain features, En represents the features based on information entropy, and q represents the total number of samples; the order of the original feature vector set is [q×(11+13+6)].

[0099] Step 2: Use the feature evaluation method to evaluate the original feature set, obtain the evaluation score of the original feature, and select significant features from the original feature set according to the evaluation score, discard insensitive features, and finally obtain a low-dimensional sensitive feature set.

[0100] Step 2.1. Divide the original feature vector set into a training set and a test set according to a certain ratio. Then use the feature evaluation method to evaluate the features of the training set and obtain the evaluation score of the original features. The specific process is:

[0101] For a feature set with C pattern classes:

[0102] {q m,c,j , m=1,2,...,M c ;c=1,2,...,C;j=1,2,...,J} (24)

[0103] In the formula, q m，c，j represents the jth feature of the mth sample of the cth class, M c represents the total number of samples in the cth class, and J represents the total number of features for each sample.

[0104] (1) Calculate the average intra-class distance of all samples in the same class:

[0105]

[0106] Then calculate the average of the C intra-class distances:

[0107]

[0108] (2) Define and calculate the difference factor of intra-class distance:

[0109]

[0110] (3) Calculate the average value of all samples in the same class

[0111]

[0112] Then calculate the average distance between different classes:

[0113]

[0114] (4) Define and calculate the difference factor between class distances:

[0115]

[0116] (5) Define and calculate weighting factors:

[0117]

[0118] (6) Calculate the ratio of inter-class and intra-class distances with weighted factors:

[0119]

[0120] Finally, the maximum value regularization method is used to obtain the distance evaluation score:

[0121]

[0122] From the definition of formula (38), it can be seen that the smaller the average intra-class distance and the larger the average inter-class distance, the better the separability.

[0123] Step 2.2. According to the size of the feature evaluation score, select the first four features with larger evaluation scores from the training set. Similarly, select features with the same attributes as the training set in the test set to construct the final training set and test set.

[0124] Step 3: Use IGRO to optimize the kernel function parameters and penalty factors of the least squares support vector machine, select the optimal parameter combination, and build a bearing fault diagnosis model based on IGRO-LSSVM.

[0125] The low-dimensional sensitive feature set of feature selection is input into the bearing fault diagnosis model based on IGRO-LSSVM for training to realize the fault type diagnosis of the bearing.

[0126] Compared with traditional SVM, the convergence speed and recognition accuracy of least squares support vector machine are faster and higher, but the setting of penalty parameter and kernel function parameter will have a significant impact on the classification performance and accuracy of the classifier. Therefore, the present invention proposes a bearing fault diagnosis model based on IGRO-LSSVM, which can effectively improve the efficiency and accuracy of rolling bearing fault diagnosis on the basis of ensuring the globality and individual diversity of the search.

[0127] Step 3.1. The specific process of the least squares support vector machine is as follows:

[0128] Assume that the training set samples are {(x i ,y i ), i = 1, 2, ..., n}, x i ∈R n is the n-dimensional input variable, y i ∈R is a one-dimensional output variable.

[0129] Using nonlinear mapping function The input samples are mapped to a high-dimensional feature space, and a regression function is constructed in the high-dimensional feature space. According to the principle of structured risk minimization, the evaluation problem is described as an optimization problem:

[0130]

[0131] In the formula, ω is the weight vector; b is the bias term; ξ i is the fitting error, and γ is the penalty factor.

[0132] The Lagrangian function is used to solve the loss function of the optimization problem. Its mathematical expression is:

[0133]

[0134] In the formula, α i is the Lagrange multiplier. According to the KKT condition, ω, b, ξ in formula (18) are i , α i Find the partial derivative and make it 0. After solving it, the final LSSVM model is:

[0135]

[0136] Among them, y out (x) is the model prediction output, K(x, x) i ) is the radial basis kernel function selected by the present invention, σ is the kernel function parameter, x is a fixed data in the training data, x i is the i-th data in the training data.

[0137] Step 3.2. The optimization performance of GRO is greatly affected by the quality of the initial solution. A high-quality initial population can accelerate the convergence speed of the algorithm and help find the global optimal solution. Therefore, the present invention applies elite reverse learning to the GRO initialization stage (exploration stage), and uses the characteristic that elite individuals contain more effective information to construct its reverse population, and selects better individuals from the current population and the reverse population as the initial solution, thereby improving the quality and diversity of the initial population. The specific process of IGRO is as follows:

[0138] I. Gold exploration stage

[0139] The location of the gold diggers is stored in a file called M GP (N×D dimensional) matrix.

[0140]

[0141] In the formula, x ij represents the position of gold digger i in the jth dimension, D is the dimension size, and N is the number of gold diggers.

[0142] The basic steps for initializing a prospector using the elite reverse learning strategy are as follows:

[0143] (1) Randomly initialize the population S:

[0144] S N×D =Lb+rand(N,D)*(Ub-Lb) (39)

[0145] Where rand(N, D) represents the randomly initialized population; Ub and Lb are the upper and lower bounds, D is the dimension size, and N is the number of gold diggers.

[0146] (2) Select the first N / 2 individuals with better fitness to form an elite population X, and generate the reverse population X′ of the elite population X:

[0147] X′=rand*(Ub-Lb)-X (40)

[0148] (3) Combine populations S and X′ to obtain a new population {SUX'}, from which N individuals with better fitness are selected to form the initial gold prospector M GP .

[0149] II. Migration of Prospectors

[0150] The formula for simulating the migration of gold diggers to the gold mine is:

[0151]

[0152] In the formula, represents the best gold mine position in the tth iteration, represents the position of the gold digger i in the tth iteration, t is the current iteration number, is the new position of the gold digger i in the t+1th iteration, and is the vector coefficient, which is calculated as follows:

[0153]

[0154] In the formula, and is a random vector with a value range of [0, 1]. Convergence component l 1 It can be expressed as:

[0155]

[0156] In the formula, e is the exponential power. If e takes the value 1, then formula (51) is a linear decreasing function. At this time, l e It can be expressed as l 1 .

[0157] III. Gold Mining

[0158] The relevant mathematical expressions for gold mining are:

[0159]

[0160] In the formula, represents the position of the randomly selected gold digger r at the tth iteration, The specific calculation is as follows:

[0161]

[0162] IV. Cooperation among Prospectors

[0163] The modeling formula for collaboration between gold diggers is:

[0164]

[0165] In the formula, They represent the positions of gold diggers i, g1, and g2 at the tth iteration, respectively. g1 and g2 are two randomly selected gold diggers. is the distance vector between gold digger g2 and gold digger g1, The position of the next iteration updated by gold digger i according to the distance vector between g1 and g2 and the current position.

[0166] V. Relocation of Prospectors

[0167] The modeling formula for gold prospector relocation is:

[0168]

[0169] In the formula, represents the position of gold digger i at the tth iteration, Indicates the position of gold digger i in the next iteration.

[0170] Step 3.3. To balance the global and local search capabilities of the algorithm, a cosine control factor is introduced in each iteration of the gold panning optimization algorithm to dynamically adjust the search boundary:

[0171]

[0172] Where CMOA(iter) is the specific function value of the decision boundary calculated by the current iteration, iter is the number of current iterations; max iter is the maximum number of iterations; max and min represent the maximum and minimum values ​​of the corresponding function value of the decision boundary, respectively.

[0173] r is a random number that follows a uniform distribution between [0, 1]. When When the cooperation phase begins, When the cosine control factor is large, more emphasis is placed on broad global search; when the cosine control factor is small, more refined local search is preferred within a smaller range; when the cosine control factor is moderate in size, it helps promote information exchange and sharing between individuals.

[0174] At the same time, the t-distribution mutation operator with the iteration number i as the degree of freedom parameter of the t-distribution is used to perturb the position of the solution, so that the algorithm has better global development ability in the early stage of iteration and good local exploration ability in the later stage of iteration, and improves the convergence speed of the algorithm. The specific position update method is as follows:

[0175]

[0176] in, is the position of the optimal solution in the jth dimension after the adaptive t-distribution mutation disturbance; is the position of the optimal solution in the,th dimension before the mutation disturbance, and t(i) represents the t-distribution with the iteration number i as the degree of freedom parameter.

[0177] Step 3.4. The kernel function parameter σ and penalty factor y in the LSSVM algorithm will affect the generalization ability of the model, and thus affect its classification performance. Therefore, IGRO is used to iteratively find the kernel function parameter σ and penalty factor y of LSSVM.

[0178] In an embodiment of the present invention, the experimental data adopts the experimental data of the Bearing Data Center of Case Western Reserve University in the United States. The bearing model is SKF6205, the speed is 1730r / min, the load is 3HP, the sampling frequency is 12Khz, and the fault sizes are 0.1778mm, 0.3556mm and 0.5334mm respectively. The data includes normal data, inner ring fault, outer ring fault, and rolling element fault. Considering that the noise content of the bearing data set of Case Western Reserve University in the United States is relatively small, it is inevitably disturbed by noise in actual working conditions. Therefore, -6db Gaussian white noise is added as experimental data on the basis of the bearing data set of Case Western Reserve University in the United States. The original signal of the outer ring fault with a fault size of 0.1778mm and the time domain of the noisy signal are as follows Figure 3 shown.

[0179] Normal bearings are classified into one category, and 3 labels are defined for each type of faulty bearing. In this way, a total of 10 labels are determined. 100 groups of bearings of each label are collected, and each group contains 1024 data points. 75 groups are randomly selected as training sets and 25 groups are used as test sets. The original features of the training set are evaluated using the feature evaluation method to obtain the feature evaluation score, and its line graph is shown as follows: Figure 4 shown.

[0180] In order to verify the classification accuracy of the IGRO-LSSVM bearing fault diagnosis model, the experiment inputs the test sample set into the trained GRO-LSSVM model and the IGRO-LSSVM model for diagnosis and classification, and compares the classification results with the actual categories to obtain the GRO-LSSVM model diagnosis result diagram, as shown in Figure 5 As shown, the IGRO-LSSVM model diagnosis result diagram is obtained, as shown in Figure 6 As shown, the results show that the fault recognition rate of the improved and optimized rolling bearing diagnosis model is increased by 1.6% to 83.2%.

Claims

1. A bearing fault diagnosis method based on IGRO-LSSVM, characterized in that: The method comprises the following steps: Step 1: Extract the time domain and frequency domain features of the original vibration signal to construct feature set 1; perform SVD denoising on the original signal, extract features based on information entropy from the denoised signal, and construct feature set 2. Merge feature set 1 and feature set 2 to construct the original feature set. Step 2: Use the feature evaluation method to evaluate the original feature set, obtain the evaluation score of the original feature, and select significant features from the original feature set according to the evaluation score, discard insensitive features, and finally obtain a low-dimensional sensitive feature set. Step 3: Use IGRO to optimize the kernel function parameters and penalty factors of LSSVM, select the optimal parameter combination, and build a bearing fault diagnosis model based on IGRO-LSSVM. The low-dimensional sensitive feature set obtained by feature selection is input into the bearing fault diagnosis model based on IGRO-LSSVM for training to realize the fault type diagnosis of the bearing.

2. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: Construct the original feature set. The specific steps include: Step 1.

1. According to the bearing fault diagnosis index, the time domain and frequency domain features of the original vibration signal are extracted to construct a feature set 1 based on the time domain features and the frequency domain features; the original vibration signal is denoised using SVD, and then features based on information entropy are extracted to construct a feature set 2 based on information entropy. Step 1.

2. Merge feature set 1 and feature set 2 to construct the original feature set, as shown in formula (1): In the formula, T represents the time domain features, F represents the frequency domain features, En represents the features based on information entropy, and q represents the total number of samples; the order of the original feature vector set is [q×(11+13+6)].

3. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: Since the original feature set may contain irrelevant or redundant features, feature selection is performed using the feature evaluation method. The specific steps are as follows: Step 2.

1. Divide the original feature vector set into a training set and a test set according to a certain ratio. Then use the feature evaluation method to evaluate the features of the training set and obtain the evaluation score of the original features. The specific process is: For a feature set with C pattern classes: {q m,c,j ,m=1,2,...,M c ;c=1,2,...,C;j=1,2,...,J} (2) In the formula, q m,c,j represents the jth feature of the mth sample of the cth class, M c represents the total number of samples in the cth class, and J represents the total number of features for each sample. (1) Calculate the average intra-class distance of all samples in the same class: Then calculate the average of the C intra-class distances: (2) Define and calculate the difference factor of intra-class distance: (3) Calculate the average value of all samples in the same class: Then calculate the average distance between different classes: (4) Define and calculate the difference factor between class distances: (5) Define and calculate weighting factors: (6) Calculate the ratio of inter-class and intra-class distances with weighted factors: Finally, the maximum normalization method is used to obtain the distance-based feature evaluation score: From the definition of formula (10), it can be seen that the smaller the average intra-class distance and the larger the average inter-class distance, the better the separability; Step 2.

2. According to the size of the feature evaluation score, select the first four features with larger evaluation scores from the training set. Similarly, select features with the same attributes as the training set in the test set to construct the final training set and test set.

4. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: The specific implementation process of LSSVM is as follows: Assume that the training set samples are {(x i ,y i ), i = 1, 2, ..., n}, x i ∈R n is the n-dimensional input variable, y i ∈R is a one-dimensional output variable. Using nonlinear mapping function The input samples are mapped to a high-dimensional feature space, and a regression function is constructed in the high-dimensional feature space. According to the principle of structured risk minimization, the evaluation problem is described as an optimization problem: In the formula, ω is the weight vector; b is the bias term; ξ i is the fitting error, and γ is the penalty factor. The Lagrangian function is used to solve the loss function of the optimization problem. Its mathematical expression is: In the formula, α i is the Lagrange multiplier. According to the KKT condition, ω, b, ξ in formula (13) i , α i Find the partial derivative and make it 0, and finally get the LSSVM model after solving it: Among them, y out (x) is the model prediction output, K(x, x) i ) is the radial basis kernel function selected by the present invention, σ is the kernel function parameter, x is a fixed data in the training data, x i is the i-th data in the training data.

5. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: In order to prevent GRO from falling into the local optimal solution too early, the elite reverse learning strategy is introduced to increase the diversity of the search space. The specific process of IGRO is as follows: I. Gold exploration stage The basic steps for initializing a prospector using the elite reverse learning strategy are as follows: (1) Randomly initialize the population S: S N×D =Lb+rand(N,D)*(Ub-Lb) (16) Where rand(N, D) represents the randomly initialized population; Ub and Lb are the upper and lower bounds, D is the dimension size, and IV is the number of gold diggers. (2) Select the first N / 2 individuals with better fitness to form an elite population X, and generate the reverse population X′ of the elite population X: X′=rand*(Ub-Lb)-X (17) (3) Merge populations S and X′ to obtain a new population {S∪X′}, from which N individuals with better fitness are selected to form the initial gold prospectors. II. Migration of Prospectors The formula for simulating the migration of gold diggers to the gold mine is: In the formula, represents the best gold mine position in the tth iteration, represents the position of the gold digger i in the tth iteration, t is the current iteration number, is the new position of the gold digger i in the t+1th iteration, and is the vector coefficient, which is calculated as follows: In the formula, and is a random vector with a value range of [0, 1]. The convergence component l1 can be expressed as: In the formula, e is the exponential power. If e takes the value of 1, then formula (22) is a linear decreasing function. At this time, l e It can be expressed as l1. III. Gold Mining The relevant mathematical expressions for gold mining are: In the formula, represents the position of the randomly selected gold digger r at the tth iteration, The specific calculation is as follows: IV. Cooperation among Prospectors The modeling formula for collaboration between gold diggers is: In the formula, They represent the positions of gold diggers i, g1, and g2 at the tth iteration, respectively. g1 and g2 are two randomly selected gold diggers. is the distance vector between gold digger g2 and gold digger g1, The position of the next iteration updated by gold digger i according to the distance vector between g1 and g2 and the current position. V. Relocation of Prospectors The modeling formula for gold prospector relocation is: In the formula, represents the position of gold digger i at the tth iteration, Indicates the position of gold digger i in the next iteration.

6. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: In order to balance the global and local search capabilities of the algorithm, a cosine control factor is introduced in each iteration of the gold panning optimization algorithm to dynamically adjust the search boundary: Where CMOA(iter) is the specific function value of the decision boundary calculated by the current iteration, iter is the number of current iterations; max iter is the maximum number of iterations; max and min represent the maximum and minimum values ​​of the corresponding function value of the decision boundary, respectively. r is a random number that follows a uniform distribution between [0, 1]. When When the cooperation phase begins, When the cosine control factor is large, more emphasis is placed on broad global search; when the cosine control factor is small, more refined local search is preferred within a smaller range; when the cosine control factor is moderate in size, it helps promote information exchange and sharing between individuals.

7. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: The t-distribution mutation operator with the iteration number i as the degree of freedom parameter is used to perturb the position of the solution, so that the algorithm has better global development ability in the early stage of iteration and good local exploration ability in the later stage of iteration, and improves the convergence speed of the algorithm. The specific position update method is as follows: in, is the position of the optimal solution in the jth dimension after the adaptive t-distribution mutation disturbance; is the position of the optimal solution in the jth dimension before the mutation disturbance, and t(i) represents the t distribution with the iteration number i as the degree of freedom parameter.

8. The fault diagnosis method based on IGRO-LSSVM according to claim 1, characterized in that: Since the kernel function parameter σ and penalty factor γ of LSSVM are difficult to select and will affect the recognition rate and generalization ability of the model, IGRO is used to optimize the two parameters of LSSVM to obtain the optimal parameter combination (γ,σ). The training set finally obtained in step 2 is input into the bearing fault diagnosis model based on IGRO-LSSVM for training to realize the fault type diagnosis of the bearing.

Images