Bearing fault feature extraction method for chemical fiber winding machine based on GEP and mutual information

Abstract

The present invention proposes a method for extracting fault features of chemical fiber winding machine bearings that combines GEP and mutual information. An improved gene expression programming with evolutionary search performance is used for feature extraction of rolling bearings. In order to improve the availability of gene expression programming fitting features and the degree of integration of bearing features in industrial scenarios, a feature selection method based on mutual information is proposed, which analyzes the correlation between the current gene expression programming features and the fault label sequence, and the redundancy with the current candidate gene expression programming feature set, and on this basis, analyzes the complementarity between the current candidate subset and the discrete feature set in the time-frequency domain. Compared with the classification accuracy of the original feature pair, the gene expression programming training process can better retain the dominant features. This method can also be applied to similar industrial scenarios with changing working conditions to help extract fault features and achieve efficient fault diagnosis.

Description

Technical Field

[0001] The invention relates to a method for extracting bearing fault features of a chemical fiber winding machine by combining gene expression programming (GEP for short) and mutual information, and belongs to the fields of fault diagnosis and machine learning. Background Art

[0002] The normal operation of chemical fiber equipment plays a vital role in the quality of chemical fiber products, the safety of chemical fiber production and the normal use of chemical fiber equipment. Chemical fiber production includes multiple steps such as melting, extrusion, spinning, bundling, winding, and texturing. Chemical fiber winding is a very important and complex process in the chemical fiber production process. The equipment used for chemical fiber winding, the chemical fiber winding machine, has the characteristics of complex structure, many moving parts, and continuous operation. Once a fault occurs, it will directly affect the quality of chemical fiber yarn forming. Therefore, the fault diagnosis problem of chemical fiber winding machines has received more and more attention from enterprises and relevant researchers.

[0003] As a key component of the chemical fiber winding machine, the winding chuck has a complex structure, including multiple groups of bearings, transmission shafts, sleeves, chuck couplings, sleeves and other parts. The probability of chuck bearing failure is the highest. Extracting the fault features of the bearing is an important prerequisite for bearing fault diagnosis. At present, the fault feature extraction methods used for bearings mainly include time-frequency domain analysis, empirical mode decomposition, spectral kurtosis analysis, classical signal processing and mechanical dynamics methods. However, due to the special position of the chuck, one end is fixed on the rotating wheel and the other end is suspended. Multiple bearings are wrapped inside the chuck, which brings challenges to the fault diagnosis of the chuck bearing. At the same time, during the winding process, the chuck speed continues to decrease, the silk cake load continues to increase, and the chuck position continues to change, making the traditional bearing feature extraction method unsustainable in the fault diagnosis of the winding machine chuck bearing. The extraction of chuck bearing fault features using support vector machines, fuzzy theory, neural networks and other methods can skip the influence of the complex structure of the chuck, but the effectiveness of its features is overly dependent on data accuracy and is prone to underfitting.

[0004] Gene expression programming (GEP) can effectively fit the relationship between variables and target quantities. In particular, GEP is a combination of genetic algorithms and genetic programming, integrating the advantages of optimization and symbolic coding, and can be well used for function discovery and feature mining. However, GEP uses common selection methods such as roulette and tournaments when selecting offspring, and fails to fully combine with fault information, and cannot be directly used to extract fault features of winding chuck bearings. If the relationship between GEP expression and feature classification accuracy is established, the effectiveness of feature extraction will be greatly increased. Summary of the invention

[0005] The purpose of the present invention is to achieve accurate and effective extraction of winder chuck bearing fault features, by using improved gene expression programming with evolutionary search performance for rolling bearing feature extraction, thereby improving the fit and adaptability of fault features.

[0006] In order to achieve the above object, the technical solution of the present invention is to provide a method for extracting bearing fault features of a chemical fiber winding machine by combining GEP and mutual information, which is characterized by comprising the following steps:

[0007] Step 1: Use a vibration acceleration sensor and a data acquisition device to collect vibration acceleration signals of the winding machine in normal and fault states;

[0008] Step 2: Remove the outliers in the original sampled signal obtained in step 1, and fill in the blanks by using the neighborhood mean completion method;

[0009] Step 3: Use empirical mode decomposition to divide the signal processed in step 2 into multiple IMF components as shown in formula (1), calculate the period of each IMF component, calculate the fault frequency of the inner ring, outer ring, rolling element, and cage of each bearing of the winding machine, match it with the period of each IMF component, and determine the IMF component corresponding to each part of the winding machine:

[0010]

[0011]

[0012]

[0013]

[0014]

[0015] In formula (1), x(t) is the original sampled signal after processing in step 2, c i (t) is the i-th IMF component after empirical mode decomposition, n is the total number of IMF components, r n (t) is the residual that does not meet the decomposition requirements; in formula (2), fo is the characteristic frequency of the bearing outer ring fault, fr is the natural frequency of the rotor during rotation, z is the number of rolling elements in the rolling bearing, d is the rolling element diameter, D is the rolling bearing pitch diameter, and α is the contact angle; in formula (3), fi is the characteristic frequency of the bearing inner ring fault; in formula (4), fb is the characteristic frequency of the bearing rolling element fault; in formula (5), fc is the characteristic frequency of the bearing cage fault;

[0016] Step 4: Convert the IMF components into frequency domain representation;

[0017] Step 5, subtract the frequency domain amplitude of the fault signal processed by steps 2 to 4 from the normal signal to obtain the frequency domain amplitude difference;

[0018] Step 6: Establish an amplitude difference-frequency value dictionary, sort by amplitude difference from large to small, sort in multiple cycles and cluster the corresponding frequency values, then select multiple regions, and determine the corresponding number of marking positions according to the cluster centers of these regions;

[0019] Step 7: Input the amplitude difference corresponding to the marked position into the constructed gene expression programming and decision tree model, and output the feature expression, wherein the gene expression programming and decision tree model includes encoding, decoding, selection, and evolution parts constituted by gene expression programming and an evaluation part constituted by a decision tree.

[0020] Preferably, in step 1, four types of data are recorded, namely normal, inner ring fault, outer ring fault, and inner and outer ring mixed fault, and each type of data takes 4 seconds to obtain.

[0021] Preferably, in step 7, the encoding process consisting of gene expression programming includes the following steps:

[0022] The encoding is determined by the four arithmetic operators +, -, *, and / , and the variable symbols use 26 English letters to replace the pre-processed rolling bearing vibration signal data; when encoding, the encoding head is set to the operator, the middle part of the encoding is the operator + variable symbol, and the end of the encoding is the variable symbol. Then a gene sequence starts with an operator, followed by a string of operators and variable symbols, and finally ends with a variable symbol.

[0023] Preferably, in step 7, the decoding process consisting of gene expression programming includes the following steps:

[0024] The variable symbol is used as the operation unit, and the operation results are stacked from bottom to top according to the tree structure, and finally the characteristic value of the starting variable symbol position is calculated as the rolling bearing fault characteristic value.

[0025] Preferably, in step 7, the evolutionary process constituted by gene expression programming is to take evolutionary measures unique to gene expression programming for genes, including replication, mutation, string insertion, and recombination, wherein the string insertion methods adopted are divided into string insertion IS and root string insertion RIS: the string insertion IS process is to randomly select a substring in the original gene sequence, randomly insert it into any non-head position in the original sequence, and the subsequent gene sequences are sequentially extended, and the excess part is truncated; while the root string insertion RIS is to put the substring into the head of the gene sequence, and then sequentially extend it, and truncate the excess gene fragment.

[0026] Preferably, in step 7, three types of recombination methods are used: single-point recombination, two-point recombination, and gene recombination, wherein: single-point recombination is to set a certain recombination point on the two parent gene sequences, and cross-exchange the gene sequence after the recombination point; two-point recombination is to set two recombination points on the two parent gene sequences, and cross-exchange the gene sequence between the two recombination points; gene recombination is aimed at chromosomes composed of multiple genes, and the connection position of the two chromosomes in the parent individuals is used as the recombination point, and the multiple gene sequences are cross-exchanged.

[0027] Preferably, in step 7, a fitness function based on classification accuracy is designed in the gene expression programming and decision tree model:

[0028] The classification accuracy is constructed as the fitness function, and the optimization direction is combined with the final goal, as shown in the following formula (6):

[0029]

[0030] In formula (6), Score (i) Indicates the classification accuracy of the current i-th feature, using fitness (i) Represents the fitness function of the current i-th feature.

[0031] Preferably, in step 7, in the gene expression programming and decision tree model, for the genes after the evolutionary process, multiple groups of feature formulas with local optimal characteristics are generated through multiple cycles of gene expression programming, and a feature selection method based on mutual information is designed, which specifically includes the following steps:

[0032] Step 701: Use mutual information to measure the correlation between candidate features and fault labels. The i-th candidate feature f i Correlation between known fault labels Tg As shown in formula (7), the i-th candidate feature f i The mutual information I(f i ; Tg) is calculated as shown in formula (8);

[0033]

[0034] I(f i ; Tg)=H(f i )-H(f i |Tg) (8)

[0035] In formula (8), H(f i ) represents the candidate feature f i The information entropy, H(f i |Tg) represents the candidate feature f iand the conditional entropy of the fault label Tg;

[0036] Step 702: measure the redundancy between features by the amount of shared information between the current GEP feature and the selected GEP dominant feature subset, where the i-th candidate feature f i and the selected jth GEP advantage feature set S j The redundancy definition of is shown in formula (9);

[0037]

[0038] Step 703: Based on the measurement of correlation and redundancy, the complementarity is measured by similarity analysis between the candidate features and the time-frequency domain formula feature set, where the i-th candidate feature f i and the kth time-frequency domain formula feature set S k Each feature of has a small similarity to form feature complementarity. The complementary measure definition is shown in formula (10):

[0039] Com(f i ,Tg,S k )=I(f i ,Tg|S k ) (10)

[0040] In formula (10), Com(f i ,Tg,S k ) represents the i-th candidate feature f i and the kth time-frequency domain formula feature set S k complementarity;

[0041] Step 704: After steps 701, 702, and 703, consider the correlation, redundancy, and complementarity between the features and design a key parameter selection measure Select (f i ), in the known fault label Tg, the currently selected GEP advantage feature set S j and the time-frequency domain formula feature set S calculated by the time-frequency domain feature formula k In the case of i The correlation with the fault label Tg is measured, and the calculation method is shown in formula (11):

[0042]

[0043] In formula (10), α and β represent weight variables, and 0≤α, 0≤β, (α+β)≤1;

[0044] Step 705: Based on the principle of “strong correlation, low redundancy, and high complementarity”, the candidate features are screened, and the final feature expression is shown in formula (12):

[0045]

[0046] In formula (12), P [i] Represents the final feature expression selected as GEP.

[0047] The present invention uses an improved gene expression programming (GEP) with evolutionary search performance for feature extraction of rolling bearings. In order to improve the availability of gene expression programming fitting features and the degree of integration of bearing features in industrial scenarios, a feature selection method based on mutual information is proposed, which analyzes the correlation between the current gene expression programming features and the fault label sequence, the redundancy with the current candidate gene expression programming feature set, and on this basis, analyzes the complementarity between the current candidate subset and the discrete feature set in the time-frequency domain. Compared with the classification accuracy of the original feature pair, the gene expression programming training process can better retain the dominant features. This method can also be applied to similar industrial scenarios with variable working conditions to help extract fault features and achieve efficient fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a coding structure diagram;

[0049] Figure 2 It is a schematic diagram of the decoding calculation process;

[0050] Figure 3 It is a method implementation flow chart;

[0051] Figure 4 yes Figure 3 Schematic diagram of the middle plug string

[0052] Figure 5 yes Figure 3 Schematic diagram of the reorganization. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only intended to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art may make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0054] A method for extracting bearing fault features of a chemical fiber winding machine combined with GEP and mutual information provided by the present invention specifically comprises the following steps:

[0055] Step 1: Use vibration acceleration sensor and data acquisition equipment to collect the vibration acceleration signal of the winding machine. Refer to the experience of standard bearing vibration signal acquisition, set the sampling frequency to 51200 and the sampling sensitivity to 5.11m / s. 2, each recorded four types of data: normal, inner circle fault, outer circle fault, and inner and outer circle mixed fault. Each type of data was collected for 4 seconds. The relevant parameters of the single-circuit data set are shown in Table 1. Before the timing domain and frequency domain formulas, the original data was segmented into 20 points.

[0056]

[0057] Table 1 Bearing test bench bearing test related parameters

[0058] Step 2: Remove the outliers in the original sampling signal and fill in the blanks by using the neighborhood mean completion method.

[0059] Step 3: Use empirical mode decomposition to divide the signal processed in step 2 into multiple IMF components as shown in formula (1). Calculate the period of each IMF component, calculate the fault frequency of the inner ring, outer ring, rolling element, and cage of each bearing of the winding machine, match them with the period of each IMF component, and determine the IMF components corresponding to each part of the winding machine.

[0060]

[0061]

[0062]

[0063]

[0064]

[0065] In formula (1), x(t) is the original sampled signal after processing in step 2, c i (t) is the i-th IMF component after empirical mode decomposition, n is the total number of IMF components, r n (t) is the residual that does not meet the decomposition requirements; in formula (2), fo is the characteristic frequency of the bearing outer ring fault, fr is the natural frequency of the rotor during rotation, z is the number of rolling elements in the rolling bearing, d is the rolling element diameter, D is the rolling bearing pitch diameter, and α is the contact angle; in formula (3), fi is the characteristic frequency of the bearing inner ring fault; in formula (4), fb is the characteristic frequency of the bearing rolling element fault; in formula (5), fc is the characteristic frequency of the bearing retainer fault.

[0066] Step 4: Convert the IMF components into frequency domain representation.

[0067] Step 5: Subtract the frequency domain amplitudes of the various types of fault signals processed in steps 2 to 4 from the normal signals to obtain the frequency domain amplitude difference.

[0068] Step 6: Establish an amplitude difference-frequency value dictionary, sort by amplitude difference from large to small, sort in multiple cycles and cluster the corresponding frequency values, then select 52 areas, and determine 52 marking positions based on the cluster centers of the 52 areas.

[0069] Step 7: Input the amplitude difference corresponding to the marker position into the constructed Gene Expression Program (GEP) and Decision Tree (DT) model to output the feature expression. The Gene Expression Program and Decision Tree model consists of two parts: the encoding, decoding, selection, and evolution part composed of GEP and the evaluation part composed of DT.

[0070] Step 7 specifically includes the following steps:

[0071] Step 701, the encoding process, including gene character representation and gene structure construction, further includes the following steps:

[0072] Step 7011: The commonly used symbol set for gene character representation is {+,-,×,÷,exp,sqrt,ln,tan,sin,cos}; the terminator is {A,B,C,D,……,a,b,c,d,……,?}, where uppercase and lowercase letters represent variables and “?” represents constants;

[0073] Step 7012: The gene structure consists of a gene head, a gene middle, and a gene tail. Only symbols are allowed to appear in the gene head, the gene middle consists of symbols and terminators, and only terminators are allowed to appear in the gene tail. The middle of the gene is added to improve the coding effectiveness and avoid the situation where only the starting symbol of the gene fragment is an operation symbol and all the subsequent symbols are variables.

[0074] Step 702: The decoding part is determined by the symbol and the number of factors involved in the operation of the symbol. The variable symbol is used as the operation unit, and the operation results are stacked from bottom to top in a tree structure. Finally, the characteristic value of the starting variable symbol position is calculated as the rolling bearing fault characteristic value, which specifically includes the following steps:

[0075] Step 7021: Match the point data in the clusters selected after clustering with the codes one by one, and reserve the sign bit of the corresponding coding area in the decoding result;

[0076] Step 7022: Combine each symbol in the coding area, calculate the total number of factors m involved in the operation, and select m bits from left to right in the cluster midpoint data;

[0077] Step 7023: Select the operation factors from right to left in the selected m positions, calculate them in sequence according to the symbols in the coding area, and fill the reserved sign positions;

[0078] Step 703: construct the classification accuracy as the fitness function, and combine the optimization direction with the final goal. The fitness function designed based on the classification accuracy is shown in formula (6):

[0079]

[0080] In formula (6), Score (i) Indicates the classification accuracy of the current i-th feature, using fitness (i) Represents the fitness function of the current i-th feature;

[0081] Step 704: GEP-specific evolutionary measures are taken on the gene, including duplication, mutation, insertion, and recombination, which specifically include the following steps:

[0082] Step 7041: randomly select individuals from the parent gene and replicate them as new genes. When the replication threshold is exceeded, replication is performed.

[0083] Step 7042: randomly select individual genes from the parent gene, mutate to other symbols or variables, and perform mutation when the mutation threshold is exceeded;

[0084] Step 7043: For the gene processed by step 7042, the insertion string method is divided into insertion string (IS) and root insertion string (RIS), wherein: the insertion string process is to randomly select a substring in the original gene sequence, randomly insert it into any position other than the head of the original sequence, and the subsequent gene sequence is sequentially extended, and the excess part is cut off; while the root insertion string is to put the substring into the head of the gene sequence, and then sequentially extend it, and cut off the excess gene fragment;

[0085] Step 7044: for the gene after step 7043, the gene is intercepted according to the recombination principle, and the intercepted gene fragments are exchanged to form a recombinant gene;

[0086] Step 705: for the genes after step 7044, multiple sets of feature formulas with local optimal characteristics are generated through gene expression programming cycles for multiple times, and a feature selection method based on mutual information is designed, which specifically includes the following steps:

[0087] Step 7051: Use mutual information to measure the correlation between the candidate feature and the fault label. The i-th candidate feature f i Correlation between known fault labels Tg As shown in formula (7), the i-th candidate feature f i The mutual information I(f i ; Tg) is calculated as shown in formula (8);

[0088]

[0089] I(f i ; Tg)=H(f i )-H(f i |Tg) (8)

[0090] In formula (8), H(f i ) represents the candidate feature f i The information entropy, H(f i |Tg) represents the candidate feature f i and the conditional entropy of the fault label Tg;

[0091] Step 7052: measure the redundancy between features by the amount of shared information between the current GEP feature and the selected GEP advantage feature subset, where the i-th candidate feature f i and the selected jth GEP advantage feature set S j The redundancy definition of is shown in formula (9);

[0092]

[0093] Step 7053: Based on the measurement of correlation and redundancy, the complementarity is measured by similarity analysis between the candidate features and the time-frequency domain formula feature set, where the i-th candidate feature f i and the kth time-frequency domain formula feature set S k Each feature of has a small similarity to form feature complementarity. The complementary measure definition is shown in formula (10):

[0094] Com(f i ,Tg,S k )=I(f i ,Tg|S k ) (10)

[0095] In formula (10), Com(f i ,Tg,S k ) represents the i-th candidate feature f i and the kth time-frequency domain formula feature set S k complementarity;

[0096] Step 7054: After steps 7051, 7052, and 7053, consider the correlation, redundancy, and complementarity between the features and design the key parameter selection measure Select (f i ), in the known fault label Tg, the currently selected GEP advantage feature set S j and the time-frequency domain formula feature set S calculated by the time-frequency domain feature formula k In the case of iThe correlation with the fault label Tg is measured, and the calculation method is shown in formula (11):

[0097]

[0098] In formula (10), α and β represent weight variables, and 0≤α, 0≤β, (α+β)≤1;

[0099] Step 7055: Based on the principle of “strong correlation, low redundancy, and high complementarity”, the candidate features are screened, and the final feature expression is shown in formula (12):

[0100]

[0101] In formula (12), P [i] Represents the final feature expression selected as GEP.

[0102] A specific implementation process of the present invention comprises the following steps:

[0103] Step 1: Data Collection

[0104] The target dataset was collected from a newly built spinning mill in the format shown in Table 1. The input of the data is the energy efficiency time series recorded every 5 seconds from the smart meter connected to the spinning frame. The length of the time series is 40, and the output of the data is 3 abnormal categories, except for 1 normal category. The source domain data was collected from the historical database of the old mill, which is consistent with the format of the target dataset. The vibration data during the rotation of the winding head was collected using a vibration acceleration sensor, and feature extraction experiments were performed at instantaneous linear speeds of chemical fiber winding yarns of 1000, 2000, and 3000 m / min. The sampling frequency is 51200 Hz, the sampling points are 716800, and the collection time for each category is 1 second, with a total of 14 categories. When using the winding head data as verification, the multi-round single-loop data set is also used for verification, and the original data is also segmented according to 20 points per segment.

[0105] Step 2: Build the model

[0106] Built using Python programming language Figure 3 For the model shown, the firmware should be above the following configuration: Core i7 CPU @ 2.20GHz and 8GB RAM.

[0107] Step 3: Model training and validation

[0108] For GEP and improved GEP, the number of individuals in the population is set to 30, the number of iterations is set to 50, the outermost loop is 3 rounds, and the optimal individual retention mechanism is set. The standard C4.5 decision tree classification algorithm is used to classify the test cases based on the extracted features and the original time domain and frequency domain features, and the performance of the feature extraction method is evaluated based on the classification accuracy.

Claims

1. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information, characterized in that: The following steps are involved: Step 1: Use a vibration acceleration sensor and a data acquisition device to collect vibration acceleration signals of the winding machine in normal and fault states; Step 2: Remove the outliers in the original sampled signal obtained in step 1, and fill in the blanks by using the neighborhood mean completion method; Step 3: Use empirical mode decomposition to divide the signal processed in step 2 into multiple IMF components as shown in formula (1), calculate the period of each IMF component, calculate the fault frequency of the inner ring, outer ring, rolling element, and cage of each bearing of the winding machine, match it with the period of each IMF component, and determine the IMF component corresponding to each part of the winding machine: In formula (1), x(t) is the original sampled signal after processing in step 2, c i (t) is the i-th IMF component after empirical mode decomposition, n is the total number of IMF components, r n (t) is the residual that does not meet the decomposition requirements; in formula (2), fo is the characteristic frequency of the bearing outer ring fault, fr is the natural frequency of the rotor during rotation, z is the number of rolling elements in the rolling bearing, d is the rolling element diameter, D is the rolling bearing pitch diameter, and α is the contact angle; in formula (3), fi is the characteristic frequency of the bearing inner ring fault; in formula (4), fb is the characteristic frequency of the bearing rolling element fault; in formula (5), fc is the characteristic frequency of the bearing cage fault; Step 4: Convert the IMF components into frequency domain representation; Step 5, subtract the frequency domain amplitude of the fault signal processed by steps 2 to 4 from the normal signal to obtain the frequency domain amplitude difference; Step 6: Establish an amplitude difference-frequency value dictionary, sort by amplitude difference from large to small, sort in multiple cycles and cluster the corresponding frequency values, then select multiple regions, and determine the corresponding number of marking positions according to the cluster centers of these regions; Step 7: Input the amplitude difference corresponding to the marked position into the constructed gene expression programming and decision tree model, and output the feature expression, wherein the gene expression programming and decision tree model includes encoding, decoding, selection, and evolution parts constituted by gene expression programming and an evaluation part constituted by a decision tree.

2. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 1, characterized in that: In step 1, four types of data are recorded: normal, inner ring fault, outer ring fault, and inner and outer ring mixed fault. Each type of data takes 4 seconds to obtain.

3. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 1, characterized in that: In step 7, the coding process consisting of gene expression programming includes the following steps: The encoding is determined by the four arithmetic operators +, -, *, and / , and the variable symbols use 26 English letters to replace the pre-processed rolling bearing vibration signal data; when encoding, the encoding head is set to the operator, the middle part of the encoding is the operator + variable symbol, and the end of the encoding is the variable symbol. Then a gene sequence starts with an operator, followed by a string of operators and variable symbols, and finally ends with a variable symbol.

4. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 1, characterized in that: In step 7, the decoding process consisting of gene expression programming includes the following steps: The variable symbol is used as the operation unit, and the operation results are stacked from bottom to top according to the tree structure, and finally the characteristic value of the starting variable symbol position is calculated as the rolling bearing fault characteristic value.

5. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 1, characterized in that: In step 7, the evolutionary process constituted by gene expression programming is to take evolutionary measures unique to gene expression programming for genes, including replication, mutation, string insertion, and recombination. Among them, the string insertion methods adopted are divided into string insertion IS and root string insertion RIS: the string insertion IS process is to randomly select a substring in the original gene sequence, randomly insert it into any non-head position in the original sequence, and the subsequent gene sequences are sequentially extended, and the excess part is truncated; while the root string insertion RIS is to put the substring to the head of the gene sequence, and then sequentially extend it, and truncate the excess gene fragment.

6. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 5, characterized in that: In step 7, three types of recombination methods are used: single-point recombination, two-point recombination, and gene recombination, among which: single-point recombination is to set a recombination point on the two parent gene sequences, and cross-exchange the gene sequence after the recombination point; two-point recombination is to set two recombination points on the two parent gene sequences, and cross-exchange the gene sequence between the two recombination points; gene recombination is aimed at chromosomes composed of multiple genes, with the connection position of the two chromosomes in the parent individuals as the recombination point, and cross-exchange the multiple gene sequences.

7. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 1, characterized in that: In step 7, a fitness function based on classification accuracy is designed in the gene expression programming and decision tree model: The classification accuracy is constructed as the fitness function, and the optimization direction is combined with the final goal, as shown in the following formula (6): In formula (6), Score (i) Indicates the classification accuracy of the current i-th feature, using fitness (i) Represents the fitness function of the current i-th feature.

8. A method for extracting bearing fault features of a chemical fiber winding machine combining GEP and mutual information as claimed in claim 1, characterized in that: In step 7, in the gene expression programming and decision tree model, for the genes after evolution processing, multiple groups of feature formulas with local optimal characteristics are generated through gene expression programming cycles for multiple times, and a feature selection method based on mutual information is designed, which specifically includes the following steps: Step 701: Use mutual information to measure the correlation between candidate features and fault labels. The i-th candidate feature f i Correlation between known fault labels Tg As shown in formula (7), the i-th candidate feature f i The mutual information I(f i ; Tg) is calculated as shown in formula (8); I(f i ;Tg)=H(f i )-H(f i |Tg) (8) In formula (8), H(f i ) represents the candidate feature f i The information entropy, H(f i |Tg) represents the candidate feature f i and the conditional entropy of the fault label Tg; Step 702: measure the redundancy between features by the amount of shared information between the current GEP feature and the selected GEP dominant feature subset, where the i-th candidate feature f i and the selected jth GEP advantage feature set S j The redundancy definition of is shown in formula (9); Step 703: Based on the measurement of correlation and redundancy, the complementarity is measured by similarity analysis between the candidate features and the time-frequency domain formula feature set, where the i-th candidate feature f i and the kth time-frequency domain formula feature set S k Each feature of has a small similarity to form feature complementarity. The complementary measure definition is shown in formula (10): Com(f i ,Tg,S k )=I(f i ,Tg|S k ) (10) In formula (10), Com(f i , Tg, S k ) represents the i-th candidate feature f i and the kth time-frequency domain formula feature set S k complementarity; Step 704: After steps 701, 702, and 703, consider the correlation, redundancy, and complementarity between the features and design a key parameter selection measure Select (f i ), in the known fault label Tg, the currently selected GEP advantage feature set S j and the time-frequency domain formula feature set S calculated by the time-frequency domain feature formula k In the case of i The correlation with the fault label Tg is measured, and the calculation method is shown in formula (11): In formula (11), α and β represent weight variables, and 0≤α, 0≤β, (α+β)≤1; Step 705: Based on the principle of "strong correlation, low redundancy, and high complementarity", the candidate features are screened, and the final feature expression is shown in formula (12): In formula (12), P [i] Represents the final feature expression selected as GEP.

Images