An efficient method for constructing rolling bearing health indicators

Abstract

This invention discloses an efficient method for constructing health indicators for rolling bearings, comprising: sampling monitoring data of the entire life cycle of rolling bearings; extracting primary features of the entire life cycle of rolling bearings and using them as input to generate an initial chromosome population; optimizing the output of higher-level features of the chromosomes based on evaluation function one; obtaining a higher-level feature population by outputting higher-level features of the chromosomes; evaluating the higher-level feature population constructed by the current chromosome population based on evaluation function two; if a termination condition is met, proceeding to the next step; otherwise, performing genetic operations on the current population according to evaluation function two, repeating the above steps until the termination condition is met; after meeting the termination condition, outputting the chromosome with the highest fitness in the final generation population, and using its corresponding higher-level feature as a health indicator reflecting the degradation trend of rolling bearings. This invention can solve the problem that traditional genetic programming algorithms cannot efficiently search for health indicators that effectively characterize the degradation trend of rolling bearings.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of rolling bearing fault diagnosis and residual life prediction, and particularly relates to a high-efficiency rolling bearing health index construction method. BACKGROUND

[0002] The health state of rotating machinery equipment is closely related to rolling bearings, and accurate estimation of the residual service life of rolling bearings can greatly ensure the stability of the operation of rotating machinery equipment and the safety of personnel accompanying. The construction of health indicators is the main content of the prediction of the residual service life of rolling bearings, and the purpose is to represent the key information of the degradation process of rolling bearings. Therefore, the health index construction process of rolling bearings is crucial to the prediction accuracy of the residual service life of rolling bearings.

[0003] According to whether the performance degradation information of the equipment has a physical meaning, it can be divided into physical health indicators and virtual health indicators. Physical health indicators are usually obtained through statistical learning or signal processing. However, the physical health indicators identified from the performance degradation information of the analyzed equipment and their respective construction methods have a strong correlation with the analyzed equipment, and therefore cannot be directly applied to other similar equipment. In addition to physical health indicators, the fusion model for constructing virtual health indicators can also be divided into two categories: mathematical model-based methods and deep learning-based methods. The research on mathematical model-based methods is based on assumptions about the form of degradation over time or professional knowledge of signal processing techniques. However, in practice, it is often difficult to obtain such information, especially for complex systems. Deep learning models can automatically generate health indicators using rich data without requiring much expert knowledge of the system. However, the deep features created by deep learning models have weak interpretability and cannot be explained as physical features of the equipment. Genetic programming algorithm is an effective method for mining the non-linear relationship between multiple features of rolling bearings, and it can construct health indicators that represent the degradation trend of rolling bearings with good performance according to requirements. And so far, genetic programming evolutionary algorithms represent a key enabler for discovering interpretable machine learning models. However, the genetic programming algorithms currently developed for constructing health indicators only rely on one layer of evaluation functions to evaluate the chromosomes of each generation, which cannot effectively control the randomness of the algorithm, resulting in low search efficiency of genetic programming algorithms for effective health indicators.

[0004] Therefore, it is urgent to solve the above problems. SUMMARY

[0005] The purpose of the present application is to provide a high-efficiency rolling bearing health index construction method, which can solve the problem that traditional genetic programming algorithms cannot efficiently search for health indicators that represent the degradation trend of rolling bearings with good performance.

[0006] Technical solution: To achieve the above object, the application discloses a kind of efficient construction method of rolling bearing health index, comprising the following steps:

[0007] (1) the full life cycle monitoring data of rolling bearing is sampled at equal interval time T times, and the sample point quantity of each sampling is R, and is recorded as sequence

[0008] (2) the S time-frequency domain primary feature extraction of each sampling data is carried out, and the i th primary feature corresponding to the sampling data of full life cycle is recorded as Wherein z i,k It indicates the characteristic value of the i th primary feature at the t k Time, and k=1,2,…,T, i=1,2,…,S, further, the primary feature set corresponding to the full life cycle sampling data can be recorded as

[0009] (3) the individual of each generation population in genetic programming algorithm is the hierarchical structure tree composed of function and primary feature, input primary feature set as terminal set, and "+" "," "-" "," "*" and "sqrt" and "cos" are as function set, and the population quantity of each generation is set to E, and the maximum depth of chromosome is set to m layer, to generate initial chromosome population;

[0010] (4) based on "+" node, each chromosome is decomposed, and the function item represented by all sub-trees of chromosome is respectively as middle-level feature p, and the number of middle-level features output in this stage is recorded as λ, and further, the i th middle-level feature corresponding to the sampling data of full life cycle of rolling bearing is recorded as Wherein the characteristic value of the i th middle-level feature p i At the t k Time is recorded as p i,k , and i=1,2,…,λ, k=1,2,…,T, further, the set of middle-level features in this stage can be recorded as

[0011] (5) based on evaluation function, c The middle-level features available in each chromosome are screened, and c≤λ, then the middle-level feature set screened after each chromosome corresponding can be recorded as

[0012] (6) based on exponential weighted moving average method, the middle-level feature set P c×T Of each chromosome is handled, then the middle-level feature set handled after each chromosome corresponding can be recorded as V

[0013] (7) the middle-level feature set V c×TDimensionality reduction is performed, and a first principal component is taken to construct a high-level feature, denoted as wherein is an eigenvalue of the high-level feature at the t k time point;

[0014] (8) The high-level feature population constructed by the current chromosome population is evaluated based on the evaluation function two. If the termination condition of the algorithm is met, the next step is performed. If not, the current population is subjected to genetic operation according to the evaluation function two, and the above steps are repeated until the termination condition is met.

[0015] (9) After the termination condition is met, the chromosome with the highest fitness in the final generation population is output, and the corresponding high-level feature is taken as the health index reflecting the degradation trend of the rolling bearing.

[0016] In step (2), S time-frequency domain primary features are extracted from each sampling data, wherein the features refer to mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square, peak factor, shape factor, pulse factor, margin factor time domain eigenvalue, and corresponding spectral kurtosis features, and S>5.

[0017] Preferably, the decomposition manner in step (4) is specifically as follows: each chromosome is decomposed into a plurality of function items based on the '+' node; wherein the sub-tree representing the function item is determined by decomposing the chromosome from the root node to the nonlinear node, i.e. the node other than '+'.

[0018] Further, the specific evaluation process of the intermediate feature in each chromosome based on the evaluation function one in step (5) is as follows: first, the fitness value M of each intermediate feature is obtained, and the fitness threshold W is set in advance according to the requirement; if M≤W, the intermediate feature is removed; otherwise, it is retained; finally, c intermediate features available in each chromosome are screened out, and c≤λ, then the set of intermediate features screened out corresponding to each chromosome can be denoted as

[0019] Further, the process of processing the intermediate feature of each chromosome based on the exponential weighted moving average method in step (6) is as follows: the eigenvalue of the i k th intermediate feature processed by the exponential weighted moving average method at the t i,k time point is denoted as v i,0 , and the i i,k th intermediate feature corresponding to the sampling data of the whole life cycle of the rolling bearing is denoted as v =0, wherein p i,k The calculation formula of the exponential weighted moving average method is as follows:

[0020] v i,k-1 =α×v i,k +(1-α)×p

[0021] wherein, a is the weight value of the exponential weighted moving average method, and 0 < a < 1, thus obtaining the intermediate feature set processed by the exponential weighted average method

[0022] Further, in step (7), the dimensionality of the screened intermediate feature set V c×T is reduced by principal component analysis method, and the first principal component is selected to construct the advanced feature, and the specific steps are as follows:

[0023] Firstly, the c intermediate features are sequentially decentered, as shown in the following formula:

[0024]

[0025] wherein, denotes the decentered i-th intermediate feature, mean(.) denotes the average value function, and the average value of the output sequence, and the intermediate feature set corresponding to each chromosome at this time is denoted as Secondly, the covariance matrix of the decentered intermediate feature set is calculated, and the covariance matrix of the rolling bearing is denoted as s c×c ; further, the eigenvalues and eigenvectors of the covariance matrix are obtained based on singular value decomposition, and are arranged in descending order according to the eigenvalue size; further, the eigenvector corresponding to the maximum eigenvalue is selected, and the unit orthogonal eigenvector u 1×c of the first principal component is obtained, which is the coefficient of the first principal component about the original variable; further, the intermediate feature set of the rolling bearing is reconstructed based on the vector; thus, the original c-dimensional feature is reduced to 1-dimensional, that is, wherein is the output advanced feature, is the eigenvalue of the advanced feature at the t k time.

[0026] Preferably, the function formula of the evaluation function one in step (5) is the same as that of the evaluation function two in step (8).

[0027] Further, the function formula of the evaluation function one in step (5) is different from that of the evaluation function two in step (8).

[0028] Beneficial effects: compared with the prior art, the present application has the following remarkable advantages: in the case of rich input features, the present application comprehensively considers the evaluation standards of multiple health indicators by two-layer evaluation functions, progressively evaluates features at different levels, and further reduces the randomness of the algorithm, so as to efficiently and comprehensively search for suitable rolling bearing health indicators; the present application simplifies the output solution by changing the output mode of the chromosome. Attached Figure Description

[0029] Figure 1 This is a flowchart of the present invention;

[0030] Figure 2 This is a schematic diagram of high-level chromosome output features in this invention;

[0031] Figure 3 These are the E chromosomes randomly generated in the first generation of this invention;

[0032] Figure 4 This is a schematic diagram of chromosome decomposition in this invention. Detailed Implementation

[0033] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0034] like Figure 1 and Figure 2 As shown, the present invention provides an efficient method for constructing rolling bearing health indicators, comprising the following steps:

[0035] (1) The rolling bearing's full life cycle monitoring data is sampled at equal intervals for T times, with R being the number of sample points in each sampling, denoted as the sequence.

[0036] (2) For each sampled data, S primary features in the time-frequency domain are extracted. The i-th primary feature corresponding to the sampled data throughout the entire lifecycle is denoted as: Where z i,k This indicates that the i-th primary feature is at the t-th position. k The feature values ​​at each time point, where k = 1, 2, ..., T, i = 1, 2, ..., S, further, the primary feature set corresponding to the whole lifecycle sampling data can be denoted as...

[0037] For each sampled data, S primary time-frequency features are extracted. The features are the time-domain feature values ​​of mean, standard deviation skewness, kurtosis, peak-to-peak value, root mean square, peak factor, shape factor, impulse factor, margin factor and their corresponding spectral kurtosis features, and S>5.

[0038] (3) In the genetic programming algorithm, each generation of individuals is a hierarchical structure tree composed of functions and primary features. The primary feature set is input into it as the terminal set, and "+", "-", "×", "sqrt", and "cos" are used as the function set. The population size of each generation is set to E, and the maximum chromosome depth is set to m layers, where m can be 6. This generates the first generation chromosome population, such as... Figure 3 As shown;

[0039] (4) Decompose each chromosome based on the "+" node, and take the function terms represented by all subtrees of the chromosome as intermediate features p respectively. Record the number of intermediate features output in this stage as λ. Further, set the i-th intermediate feature corresponding to the sampling data of the rolling bearing's entire life cycle as λ. Where the i-th intermediate feature p i At the t k The feature value at time t is denoted as p. i,k And i = 1, 2, ..., λ, k = 1, 2, ..., T. Furthermore, the set of intermediate features in this stage can be denoted as...

[0040] The specific process of the decomposition method is as follows: each chromosome is decomposed into several function terms based on the "+" node, such as... Figure 4 As shown; where the subtree representing the function term is determined by decomposing the chromosome from the root node to nonlinear nodes, i.e., nodes that are not '+'; Figure 4 Taking the tree structure in the image as an example, the root node is a "+" operator, so the tree can be decomposed into two subtrees: tree A and tree B. The root node of tree A is also a "+" operator, so it can be decomposed into trees C and D. The root node of tree B is a non-linear node "×", so it cannot be decomposed. This decomposition process ultimately yields three subtrees: B, C, and D. Taking the function terms represented by all subtrees of the chromosome as intermediate features p, and denoting the number of intermediate features output in this stage as λ, the i-th intermediate feature corresponding to the sampling data of the rolling bearing's entire life cycle is denoted as p. Where the i-th intermediate feature p i At the t k The feature value at time t is denoted as p. i,k And i = 1, 2, ..., λ, k = 1, 2, ..., T. Furthermore, the set of intermediate features in this stage can be denoted as...

[0041] (5) Based on the evaluation function, select c usable intermediate features for each chromosome, where c ≤ λ. Then, the set of intermediate features selected for each chromosome can be denoted as:

[0042] The specific evaluation process for intermediate features in each chromosome based on the evaluation function is as follows: First, obtain the fitness value M for each intermediate feature, and pre-set the fitness threshold W according to the requirements; if M≤W, remove the intermediate feature; otherwise, retain it; finally, select c usable intermediate features in each chromosome, where c≤λ. The set of intermediate features selected for each chromosome can then be denoted as:

[0043] (6) Based on the Exponential Weighted Moving Average (EWMA) method, the intermediate feature set P of each chromosome is calculated. c×T After processing, the intermediate feature set corresponding to each chromosome can be denoted as .

[0044] The process of processing the intermediate features of each chromosome based on the exponentially weighted moving average method is as follows: Let the i-th intermediate feature, processed by the exponentially weighted moving average method, be at point t. k The feature value at time t is denoted as v. i,k And the i-th intermediate feature corresponding to the sampling data of the entire life cycle of the rolling bearing is denoted as v i,0 =0, where p i,k The calculation formula after processing with the exponentially weighted moving average method is as follows:

[0045] v i,k =α×v i,k-1 +(1-α)×p i,k

[0046] Where α is the weight value of the exponentially weighted moving average method, and 0 < α < 1, thus obtaining the intermediate feature set after processing by the exponentially weighted moving average method.

[0047] (7) The intermediate feature set V obtained after processing in each chromosome is processed by principal component analysis. c×T Dimensionality reduction is performed, and the first principal component is used to construct high-level features, denoted as . in It is the tth high-level feature k The feature values ​​at each time point;

[0048] Principal component analysis was used to screen the intermediate feature set V from each chromosome. c×T Dimensionality reduction is performed, and the first principal component is selected to construct high-level features. The specific steps are as follows:

[0049] First, the c intermediate features are decentralized sequentially, as shown in the following equation:

[0050]

[0051] in, Let represent the i-th intermediate feature after decentralization, and let mean(.) denote the average value function, outputting the average value of the sequence. Then, the intermediate feature set corresponding to each chromosome is denoted as . Secondly, for the decentralized intermediate feature set The covariance matrix is calculated, and the covariance matrix of the rolling bearing is denoted as s c×c ; further, eigenvalues and eigenvectors of the covariance matrix are obtained based on singular value decomposition, and the eigenvectors corresponding to the maximum eigenvalues are selected according to the size of the eigenvalues from large to small; further, the unit orthogonal eigenvector u 1×c of the first principal component is obtained based on the vector; further, the intermediate feature set of the rolling bearing is reconstructed based on the vector ; thus, the original c-dimensional features are reduced to 1-dimensional, that is, wherein is the output high-level feature, is the eigenvalue of the high-level feature at the t k time;

[0052] (8) The high-level feature population constructed by the current chromosome population based on the evaluation function two is evaluated, if the termination condition of the algorithm is met, the next step is performed; if not, the current population is genetically operated according to the evaluation function two, and the above steps are repeated until the termination condition is met;

[0053] (9) After the termination condition is met, the chromosome with the highest fitness in the final generation population is output, and the corresponding high-level feature is taken as the health index reflecting the degradation trend of the rolling bearing.

[0054] The function formula of the evaluation function one in step (5) and the evaluation function two in step (8) can be the same or different, and the evaluation function one and the evaluation function two are used to evaluate the performance of different levels of features.

[0055] Example 1

[0056] Data set introduction and primary feature extraction: Example 1 adopts the XJTU-SY rolling bearing accelerated life test data set to verify the method proposed in the present application, and the test rolling bearing is LDK UER204 rolling bearing. The acceleration vibration signal in the horizontal direction is selected as the source of signal analysis in the present application, the single sampling frequency is 25.6 kHz, the sampling duration is 1.28 s, and the interval between adjacent two samplings is 1 min. Example 1 selects the data set of the rolling bearing under the first type of working condition 1-2 to verify the method proposed in the present application. Example 1 extracts 20 signal features from the rolling bearing vibration signal, as shown in Table 1. The primary features cannot meet the performance requirement of accurately predicting the remaining useful life. Therefore, the primary features need to be selected and processed, and then the high-level features with good performance of reflecting the degradation trend of the rolling bearing are constructed.

[0057] Table 1: Extracted rolling bearing features

[0058]

[0059] Comparison of traditional and improved output operation in genetic programming: To prove the superiority of the improved chromosome output method of the present application, the characteristic data set of rolling bearing 1-2 of the XJTU data set is input into the improved genetic algorithm and the traditional genetic programming algorithm respectively in Example 1. Considering the principle of quantitative analysis, after setting the fitness function of the traditional genetic programming algorithm as a monotonic evaluation function, the evaluation functions one and two in the improved genetic programming algorithm are also set as monotonic evaluation functions, as shown in the following formula:

[0060]

[0061] wherein Mon(y i ) represents the monotonicity score of the i-th feature of the rolling bearing, sgn(.) is a sign function, indicating returning an integer variable, i.e. indicating the sign of the parameter; y i represents the i-th feature time series, y i,k represents the k-th feature value of the i-th feature, i.e. y i = [y i,1 , y i,2 , …, y i,k , …, y i,T ], and k = 1, 2, …, T, i = 1, 2, …, S. In order to reduce the randomness of the genetic programming algorithm, each method is run 5 times independently. And, the maximum number of iterations is set to 100, the population size of each generation is 1000, and the maximum depth of the chromosome is set to 6 layers. Among them, the fitness values of the optimal chromosomes obtained by the two methods in the final generation are shown in Table 2. Here, the optimal chromosome with a fitness value greater than 0.8 is regarded as a usable health indicator. As the algorithm iterates, as can be seen from Table 2, the number of usable health indicators generated by the improved method is higher than that of the traditional genetic programming algorithm. This shows that the method of the present application can improve the search efficiency and accuracy of the process of constructing usable health indicators by genetic programming algorithm.

[0062] Table 2 Fitness values of optimal chromosomes of the final generation of two methods

[0063]

[0064] Example 2

[0065] Dataset introduction and primary feature extraction: Example 2 also uses the XJTU-SY rolling bearing accelerated life test dataset to verify the method proposed in the application, and the test rolling bearing is LDK UER204 rolling bearing. Example 2 selects the data set of the first type of working condition 1-1 rolling bearing to verify the method proposed in the application. Example 2 extracts 20 signal features from the rolling bearing vibration signal, as shown in Table 1 in Example 1. The primary features cannot meet the performance requirements of accurate RUL prediction. Therefore, the primary features need to be selected and processed to construct high-level features that represent the performance of the rolling bearing degradation trend.

[0066] Performance comparison between the constructed health index and the original features: Example 2 constructs high-level features for the 1-1 rolling bearing data of the XJTU dataset, and the input primary features are 20 time-frequency domain features of the rolling bearing. The maximum number of iterations is set to 100, the number of populations in each generation is 1000, the maximum depth of the chromosome is set to 6 layers, the evaluation function one of Example 2 is set as a monotonicity evaluation function, and the evaluation function two is set as shown in the following formula, and the remaining settings are as described in the application.

[0067]

[0068] Wherein, H represents the ideal health index time series based on the least square method for the rolling bearing high-level feature, and Cov(.) represents the covariance of the two, Var(.) represents the variance; f represents the fitness of the high-level feature, L corresponds to the node number of the chromosome, and a1 and a2 represent the hyperparameters of the penalty function, which are 30 and 2 respectively; then the mathematical formula of the rolling bearing 1-1 health index obtained by Example 2 is as shown in the following formula:

[0069]

[0070] Wherein, z i represents the time series of the i-th feature of the rolling bearing, and EWMA(·) represents the exponential weighted moving average function.

[0071] The application compares the health index constructed by the method with the 20 original features based on fitness, as shown in Table 3. As shown in Table 3, compared with the original features, the fitness of the high-level features constructed according to the method of the application is higher. In summary, the high-level features constructed by the method proposed in the application have good fitness.

[0072] Table 3 Performance comparison between the health index constructed by the method proposed in the application and the original features

[0073]

[0074]

[0075] The application reduces the randomness existing in the process of generating chromosomes by genetic programming algorithm, improves the search efficiency of the algorithm on the health index of the rolling bearing, and is beneficial to accurately predict the remaining service life of the rolling bearing.

Claims

1. A method for efficient construction of rolling bearing health indicators, characterized in that, The method comprises the following steps: (1) The full life cycle monitoring data of the rolling bearing is sampled for T times at equal intervals, and the number of sample points of each sampling is R, denoted as sequence (2) S primary features are extracted from each sampling data, and the i-th primary feature corresponding to the whole life cycle sampling data is denoted as wherein z i,k represents the feature value of the i-th primary feature at the t k -th time, and k = 1, 2, …, T, i = 1, 2, …, S, and further, the primary feature set corresponding to the whole life cycle sampling data can be denoted as (3) The individuals of each generation population in the genetic programming algorithm are hierarchical structure trees composed of functions and primary features, the primary feature set is input as the terminal set, "+", "-", "x", "sqrt" and "cos" are set as the function set, the population number of each generation is set as E, and the maximum depth of the chromosome is set as m layers, thereby generating the initial chromosome population; (4) Decompose each chromosome based on the "+" node, and take the function terms represented by all subtrees of the chromosome as intermediate features p respectively. Record the number of intermediate features output in this stage as λ. Further, set the i-th intermediate feature corresponding to the sampling data of the rolling bearing's entire life cycle as λ. Where the i-th intermediate feature p i At the t k The feature value at time t is denoted as p. i,k And i = 1, 2, ..., λ, k = 1, 2, ..., T. Furthermore, the set of intermediate features in this stage can be denoted as... (5) Based on the evaluation function, c intermediate features in each chromosome are selected, c < λ, and the set of selected intermediate features corresponding to each chromosome can be denoted as (6) The medium-level feature set P of each chromosome is calculated based on the exponential weighted moving average method c×T After processing, the medium-level feature set of each chromosome after processing can be recorded as (7) The intermediate feature set V obtained by processing each chromosome is subjected to principal component analysis c×T to reduce the dimension and take the first principal component to construct a high-level feature, denoted as wherein is the eigenvalue of the high-level feature at the t k time (8) Based on the evaluation function two, the senior feature population constructed by the current chromosome population is evaluated, if the termination condition of the algorithm is met, the next step is performed, if not, the current population is genetically operated according to the evaluation function two, and the above steps are repeated until the termination condition is met; (9) After the termination condition is met, the chromosome with the highest fitness in the final generation population is output, and the corresponding senior feature is taken as the health index reflecting the degradation trend of the rolling bearing.

2. The method according to claim 1, wherein: In the step (2), S time-frequency domain primary features are extracted from each sampling data, wherein the features refer to mean, standard deviation, skewness, kurtosis, peak-to-peak value, root mean square, peak factor, shape factor, pulse factor, margin factor time domain characteristic value and corresponding spectral kurtosis characteristics, and S>5.

3. The method of claim 1, wherein: In the step (4), the specific process of the decomposition mode is that each chromosome is decomposed into a plurality of function items based on the "+" node; wherein the sub-tree representing the function item is determined by decomposing the chromosome from the root node to the nonlinear node, that is, the node other than '+'.

4. The method of claim 1, wherein: The specific evaluation process of the intermediate features in each chromosome based on the evaluation function in step (5) is as follows: first, the fitness value M of each intermediate feature is obtained, and the fitness threshold W is set in advance according to requirements; if M≤W, the intermediate feature is removed; otherwise, it is retained; finally, c intermediate features available in each chromosome are screened out, c≤λ, and the set of the screened intermediate features corresponding to each chromosome can be recorded as 5. The method of claim 1, wherein: The process of processing the intermediate features of each chromosome based on the exponentially weighted moving average method in step (6) is as follows: The i-th intermediate feature processed by the exponentially weighted moving average method is set at point t. k The feature value at time t is denoted as v. i,k And the i-th intermediate feature corresponding to the sampling data of the entire life cycle of the rolling bearing is denoted as v i,0 =0, where p i,k The calculation formula after processing with the exponentially weighted moving average method is as follows: v i,k = a x v i,k-1 + (1 - a) x p i,k wherein a is an index weight value of the exponential weighted moving average method, and 0 < a < 1, so as to obtain the intermediate feature set processed by the exponential weighted average method 6. The method of claim 1, wherein: The step (7) is to perform principal component analysis on the screened intermediate feature set V in each chromosome c×T The dimension is reduced, and the first principal component is selected to construct the high-level feature, and the specific steps are as follows: Firstly, the c intermediate features are sequentially decentered, as shown in the following formula: wherein, denotes the i-th intermediate feature after decentralization, mean(.) denotes the mean function, and the mean value of the output sequence, and the intermediate feature set corresponding to each chromosome at this time is denoted as Secondly, the covariance matrix is calculated based on the intermediate feature set after decentralization, and the covariance matrix of the rolling bearing is denoted as s c ×c ; Furthermore, the eigenvalues and eigenvectors of the covariance matrix are obtained based on singular value decomposition, and the eigenvectors corresponding to the maximum eigenvalues are selected according to the size of the eigenvalues from large to small; Further, the unit orthogonal eigenvector u 1×c of the first principal component is obtained based on the eigenvectors, and the vector is the coefficient of the first principal component with respect to the original variable; Further, the intermediate feature set of the rolling bearing is reconstructed based on the vector; At this time, the original c-dimensional feature is reduced to 1-dimensional, that is, wherein is the output high-level feature, is the eigenvalue of the high-level feature at the t k time.

7. The method of claim 1, wherein: The function formula of the evaluation function one in the step (5) is the same as that of the evaluation function two in the step (8).

8. The method of claim 1, wherein: The function formula of the evaluation function one in the step (5) is different from that of the evaluation function two in the step (8). The function formula of the evaluation function one in the step (5) is different from that of the evaluation function two in the step (8).

Images