Bearing Fault Diagnosis Model Training Method and Training Device

Abstract

The present invention provides a method and a device for training a bearing fault diagnosis model. The training method includes the following steps: performing data segmentation on the fault vibration signals of different types of bearing faults to obtain a training set, where the training set includes a number of training samples; calculating the silhouette coefficients of the training samples in the training set under different parameters to select the optimal parameters of the enhanced composite multi-scale slope entropy; using the enhanced composite multi-scale slope entropy after parameter selection to extract fault features to obtain a training feature set; and using the training feature set to train the honey badger algorithm optimized kernel extreme learning machine to obtain a trained composite multi-scale fault diagnosis model. The present invention can effectively identify bearing faults of different types and damage degrees, and the average recognition rate reaches 99.9%.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis, and particularly relates to a method and a device for training a bearing fault diagnosis model. Background Art

[0002] Mechanical equipment is developing towards the direction of being large-scale and precision, and the intelligence is also increasing day by day, which puts forward higher requirements for the reliability during the operation of the equipment. As the core rotary support component in rotating machinery, slight defects on the surface of the bearing may lead to operation failures of the entire device system, causing huge casualties and property losses.

[0003] Traditional methods for detecting rolling bearings require workers to regularly remove the bearings from the mechanical equipment for safety inspection, which consumes a large amount of manpower and material resources. Therefore, intelligent detection methods are particularly important. In recent years, many nonlinear dynamics feature extraction methods have emerged, such as sample entropy, permutation entropy, dispersion entropy, symbolic dynamics entropy, etc. These entropy algorithms can effectively extract information features, but they only contain information of a single scale. Therefore, some scholars have proposed multi-scale permutation entropy, which enhances the performance of the permutation entropy algorithm by extracting the entropy value information of multi-scale sequences. However, the multi-scale sequence cannot reflect all the constituent patterns of the time series, and the permutation entropy ignores the amplitude information of the time series. Therefore, some people have proposed composite multi-scale weighted permutation entropy to improve the deficiencies of multi-scale permutation entropy. However, the composite multi-scale weighted permutation entropy only reflects the low-frequency information of the time series and cannot reflect the high-frequency information, which may lead to large errors in the collected fault feature information.

[0004] Therefore, finding an entropy algorithm that can simultaneously reflect the low-frequency information and high-frequency information of the extracted time series is of crucial significance for improving the accuracy of bearing fault diagnosis. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention proposes a method and a device for training a bearing fault diagnosis model, which use the difference method to extract the high-frequency information of the time series, use the average method to extract the low-frequency information of the sequence, use slope entropy to extract the complex features of the two multi-scale sequences, and average the obtained high and low frequency features, thereby constituting an enhanced composite multi-scale slope entropy, and input the extracted fault features into the honey badger algorithm optimized kernel extreme learning machine classifier to realize the fault diagnosis of the bearing.

[0006] The present invention proposes a method for training a bearing fault diagnosis model, including the following steps:

[0007] Perform data segmentation on the fault vibration signals of different types of faults of the bearing to obtain a training set, and the training set includes a number of training samples;

[0008] Calculate the silhouette coefficient of the training samples in the training set under different parameters to select the optimal parameters for enhancing the composite multi-scale slope entropy;

[0009] Use the enhanced composite multi-scale slope entropy after parameter selection to extract fault features and obtain a training feature set;

[0010] Use the training feature set to train the honey badger algorithm optimized kernel extreme learning machine to obtain a trained composite multi-scale fault diagnosis model.

[0011] In an embodiment of the present invention, the different types of faults include one or a combination of more of normal state, inner ring single point fault, outer ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, inner ring multi-point fault, outer ring multi-point fault, and roller multi-point fault.

[0012] In an embodiment of the present invention, the different parameters include scale factor, delay time, embedding dimension, low threshold, and high threshold.

[0013] In an embodiment of the present invention, the selection of the optimal parameters of the enhanced composite multi-scale slope entropy includes the following steps:

[0014] Set the scale factor, delay time, and embedding dimension as fixed values, and observe the change trend of the silhouette coefficient of the training samples in the training set under different low thresholds and different high thresholds to determine the optimal low threshold;

[0015] After determining the optimal low threshold, observe the silhouette coefficient of the training samples in the training set under different embedding dimensions and different high thresholds to determine the optimal embedding dimension and high threshold.

[0016] In an embodiment of the present invention, the optimal parameter combination is: the maximum scale factor is 8, the delay time is 1, the embedding dimension is 3, the low threshold is 0.001°, and the high threshold is 60°.

[0017] In an embodiment of the present invention, the enhanced composite multi-scale slope entropy is obtained by taking the average of the composite multi-scale slope entropy and the composite multi-scale slope entropy based on difference under each scale factor.

[0018] In an embodiment of the present invention, the composite multi-scale slope entropy is obtained in the following manner:

[0019] Calculate the slope entropy value of each coarse-grained sequence under the scale factor s, and average the s slope entropy values to obtain the entropy value of the composite multi-scale slope entropy under the scale factor s.

[0020] In an embodiment of the present invention, the composite multi-scale slope entropy based on difference is obtained in the following manner:

[0021] Calculate the slope entropy value of each differential coarse-grained sequence under the scale factor s, and average the s slope entropy values to obtain the entropy value of the composite multi-scale slope entropy based on the difference under the scale factor s.

[0022] In an embodiment of the present invention, the honey badger algorithm is used to optimize the kernel extreme learning machine with the training feature set to obtain a trained composite multi-scale fault diagnosis model. The steps include:

[0023] Input the training feature set, initialize the population size and position, determine the population size, the maximum number of iterations w max , the constant C and ρ, and use the regularization optimization parameter A and the kernel function parameter α of the kernel extreme learning machine as the position information of the honey badger;

[0024] Establish the classification error rate of the training set as the fitness function, calculate the fitness of each honey badger, and record the position of the best fitness as x prey ;

[0025] If the number of iterations w ≤ w max , then update the positions of the honey badger population using the predefined formula and update the optimal position for each iteration;

[0026] Output the final honey position, that is, the optimal parameters A and α, and use them to construct a kernel extreme learning machine model.

[0027] To achieve the above and other related purposes, the present invention also proposes a training device for a bearing fault diagnosis model. The training device includes:

[0028] A training set acquisition module for performing data segmentation on the fault vibration signals of different types of faults of the bearing to obtain a training set, where the training set includes a number of training samples;

[0029] A parameter selection module for calculating the silhouette coefficients of the training samples in the training set under different parameters to select the optimal parameters for enhancing the composite multi-scale slope entropy;

[0030] A feature set acquisition module for performing fault feature extraction using the enhanced composite multi-scale slope entropy after parameter selection to obtain a training feature set;

[0031] A model training module for using the training feature set to train the honey badger algorithm to optimize the kernel extreme learning machine to obtain a trained composite multi-scale fault diagnosis model.

[0032] The present invention proposes a method and a training device for training a bearing fault diagnosis model, which uses the difference method to extract the high-frequency information of the time series, uses the average method to extract the low-frequency information of the sequence, uses the slope entropy to extract the complex features of the two multi-scale sequences, and averages the obtained high- and low-frequency features, thereby constituting an enhanced composite multi-scale slope entropy. The extracted fault features are input into a honey badger algorithm-optimized kernel extreme learning machine classifier to achieve bearing fault diagnosis.

[0033] The enhanced composite multi-scale slope entropy proposed by the present invention combines the high- and low-frequency information of the time series. Compared with general entropy algorithms, the extracted time series and scale information are more comprehensive and rich. At the same time, the regularization coefficient and kernel function parameters of the kernel extreme learning machine are optimized by the honey badger algorithm to achieve parameter self-adaptation.

[0034] The method of the present invention can effectively extract the bearing fault features in different health states, has good recognition ability for single and composite faults of bearings, can distinguish bearing faults of different types and damage degrees, and the average recognition rate reaches 99.9%. Brief Description of the Drawings

[0035] Figure 1 It is a flowchart of the bearing fault diagnosis method proposed by the present invention.

[0036] Figure 2 It is a flowchart of the bearing fault diagnosis method proposed by the present invention.

[0037] Figure 3 It is a diagram showing the relationship between time series difference and symbols.

[0038] Figure 4 It is a structural block diagram of a training device for a bearing fault diagnosis model.

[0039] Figure 5 It is a structural block diagram of a bearing test device.

[0040] Figure 6 Bar chart of silhouette coefficients at three high thresholds for different low thresholds δ.

[0041] Figure 7 Bar chart of silhouette coefficients for different combinations of embedding dimensions and high thresholds.

[0042] Figure 8-1 It is a t-sne visualization and dimensionality reduction result diagram for extracting the features of 9 fault signal samples using the enhanced composite multi-scale slope entropy algorithm.

[0043] Figure 8-2 It is a t-sne visualization and dimensionality reduction result diagram for extracting the features of 9 fault signal samples using the composite multi-scale slope entropy algorithm.

[0044] Figure 8-3To extract the features of 9 kinds of fault signal samples using the differential-based composite multi-scale slope entropy algorithm, the t-sne visualization dimensionality reduction result graph.

[0045] Figure 8-4 To extract the features of 9 kinds of fault signal samples using the refined composite multi-scale walk entropy algorithm, the t-sne visualization dimensionality reduction result graph.

[0046] Figure 8-5 To extract the features of 9 kinds of fault signal samples using the composite multi-scale weighted permutation entropy algorithm, the t-sne visualization dimensionality reduction result graph.

[0047] Figure 9-1 The diagnostic result graph of one experiment using the enhanced composite multi-scale slope entropy algorithm.

[0048] Figure 9-2 The diagnostic result graph of one experiment using the composite multi-scale slope entropy algorithm.

[0049] Figure 9-3 The diagnostic result graph of one experiment using the differential-based composite multi-scale slope entropy algorithm.

[0050] Figure 9-4 The diagnostic result graph of one experiment using the refined composite multi-scale walk algorithm.

[0051] Figure 9-5 The diagnostic result graph of one experiment using the composite multi-scale weighted permutation entropy algorithm.

[0052] Figure 10 The line graph of the recognition accuracy rate for 30 fault experiments using five algorithms.

[0053] Figure 11 The bar graph of the average recognition accuracy rate for fault experiments using five algorithms. Specific implementation manner

[0054] The following uses specific specific examples to illustrate the implementation manner of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. The details in this specification can also be variously modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0055] It should be noted that the diagrams provided in this embodiment only illustrate the basic concept of the present invention in a schematic manner. Therefore, only the components related to the present invention are shown in the diagrams, rather than being drawn according to the number, shape, and size of the components in actual implementation. The type, quantity, and ratio of each component in actual implementation can be arbitrarily changed, and the component layout type may also be more complex.

[0056] The present invention proposes a bearing fault diagnosis model training method, which uses a difference method to extract high-frequency information of a time series, uses an average method to extract low-frequency information of a sequence, uses slope entropy to extract complex features of two multi-scale sequences, and averages the acquired high and low frequency features to form an enhanced composite multi-scale slope entropy, which is used to solve the problem that existing algorithms contain incomplete information when extracting effective information features.

[0057] Now combined Figure 1 and Figure 2 The training method for the bearing fault diagnosis model of this embodiment is described below. The training method for the bearing fault diagnosis model includes the following steps:

[0058] S1. performing data segmentation on fault vibration signals of different types of bearing faults to obtain a training set, wherein the training set includes a plurality of training samples;

[0059] S2, calculating the silhouette coefficients of the training samples in the training set under different parameters to select the optimal parameters for enhancing the composite multi-scale slope entropy;

[0060] S3, using the enhanced composite multi-scale slope entropy after parameter selection to extract fault features and obtain a training feature set;

[0061] S4. Using the training feature set to train the Honey Badger algorithm optimization kernel extreme learning machine to obtain a trained composite multi-scale fault diagnosis model.

[0062] In step S1, data segmentation is performed on the collected fault vibration signals of different types of bearings, N samples are selected for each fault signal, M samples are randomly selected as training sets, and the remaining NM samples are used as test sets.

[0063] In step S1, the different types of faults include one or more combinations of normal state, inner ring single point fault, outer ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, inner ring multi-point fault, outer ring multi-point fault and roller multi-point fault.

[0064] In step S2, the different parameters include a scale factor s, a delay time τ, an embedding dimension m, a low threshold δ, and a high threshold γ.

[0065] The silhouette coefficient SC measures the distance between data by defining the cohesion and dispersion, and its value range is distributed between [0, 1]. The closer the value is to 1, the smaller the within-class dispersion and the larger the between-class dispersion. At this time, the feature extraction effect is better. Moreover, the numerical selection of the scale factor s, delay time τ, embedding dimension m, low threshold δ, and high threshold γ of the enhanced composite multi-scale slope entropy will affect the result of the silhouette coefficient SC, specifically affecting the values of a(i) and b(i) in the calculation formula of the silhouette coefficient SC described below. The calculation formula of the silhouette coefficient SC is as follows:

[0066]

[0067] In the formula: N is the total number of samples, a(i) is the average distance from sample i to other samples within the cluster, and b(i) is the minimum average distance from sample i to other clusters.

[0068] In this embodiment, step S2 further includes: first, setting the scale factor s, delay time τ, and embedding dimension m as fixed values, observing the change trend of the silhouette coefficient of the training samples in the training set under different low thresholds and different high thresholds to determine the optimal low threshold; after determining the optimal low threshold, observing the silhouette coefficient of the training samples in the training set under different embedding dimensions and different high thresholds to determine the optimal embedding dimension and high threshold.

[0069] In step 3, the enhanced composite multi-scale slope entropy ECMSE is obtained by taking the average of the composite multi-scale slope entropy CMSE and the differential-based composite multi-scale slope entropy DBCMSE under each scale factor, that is:

[0070]

[0071] The main parameters of ECMSE include the embedding dimension m, delay time τ, scale factor s, low threshold δ, and high threshold γ. The delay time τ has little effect on the entropy value, and usually τ = 1 can be set. There is no specific requirement for the selection of the scale factor, and a moderate value can be selected. In a specific example, the maximum scale factor s max = 8. The embedding dimension m and the thresholds (low threshold δ and high threshold γ) have a certain impact on the feature extraction effect of ECMSE. Usually, m is between [2, 7]. There is no specific standard for setting the thresholds, but the high threshold δ needs to be in a reasonable range. If γ is too small or too large, it cannot reflect the true change of the slope. To facilitate the description of the threshold range, this embodiment uses angles as the measurement unit of the threshold. The selection range of δ is set as: {10 -1° , 10 -2° , 10 -3° , 10 -4° , 10 -5° , 10-6°}, and the range of γ is {30°, 45°, 60°}.

[0072] In this embodiment, the composite multi-scale slope entropy CMSE is obtained in the following manner:

[0073] Calculate the slope entropy value SE of each coarse-grained sequence at scale factor s, and average the s slope entropy values SE to obtain the entropy value of the composite multi-scale slope entropy at scale factor s, that is:

[0074]

[0075] where the coarse-grained sequence can be defined according to x i . It should be noted that 1 ≤ k ≤ s, s = 1, 2…, s max , when s = 1, the coarse-grained sequence is the original sequence. [·] represents rounding of numbers, represents the k-th average coarse-grained sequence at the s-th scale factor, and j represents the j-th point of the k-th coarse-grained sequence .

[0076] Now, further explanation is given for the slope entropy value SE mentioned above:

[0077] First, assume a time series x = {x i , i = 1, 2,…,}, and after performing phase space reconstruction on it, a subspace sequence can be obtained: where m is the embedding dimension, τ is the delay time, and t = 1, 2,…, L - τ(m - 1). Define the vertical increment threshold γ and the horizontal increment threshold δ; γ is a relatively large quantity used to measure the significant difference between vector sequences to distinguish different fluctuation amplitudes; δ is a very small value used to classify approximately equal amplitude cases. Secondly, take the difference of adjacent elements of the time series, and define subsequences as different symbols through the threshold, Figure 3 showing the relationship between the time series difference and symbols.

[0078] Through the above definitions, a symbol subspace can be obtained. Count the total number Z of permutation patterns that appear in all symbol subspaces, and the number of occurrences of each permutation is h i , where i = 1, 2, L, Z, and the corresponding probability can be obtained as p i = h i / Z, and the slope entropy can be defined as:

[0079]

[0080] In this embodiment, the differential-based composite multi-scale slope entropy DBCMSE is obtained in the following manner:

[0081] Calculate the slope entropy value of each differential coarse-grained sequence at the scale factor s, and average the s slope entropy values to obtain the entropy value of the differential-based composite multi-scale slope entropy at the scale factor s, that is:

[0082]

[0083] where the differential-based coarse-grained sequence can be defined according to where a = i - [(j - 1)s + k], 1 ≤ k ≤ s, s = 1, 2…, s max , represents the combination of choosing a numbers from s - 1 numbers, [·] represents rounding the number, represents the k-th differential coarse-grained sequence at the s-th scale factor, and j represents the j-th point of the k-th coarse-grained sequence .

[0084] In step S4, the honey badger algorithm-optimized kernel extreme learning machine is trained using the training feature set to obtain a trained composite multi-scale fault diagnosis model. The specific steps include:

[0085] Input the training set features, initialize the population size and position, determine the population size, the maximum number of iterations w max , determine the constants C and ρ, and use A and α as the position information of the honey badger.

[0086] Establish the classification error rate of the training set as the fitness function, calculate the fitness of each honey badger, and record the position of the best fitness as x prey .

[0087] If the number of iterations w ≤ w max , then update the position of the honey badger population using the following formula and update the optimal position for each iteration.

[0088]

[0089]

[0090] where ρ is a constant ≥ 1, defaulting to 6, representing the ability of the honey badger to find food. r 3 , r 4 , r 5 are all random numbers between [0 1].

[0091] When \(0.5\leq r\leq1\), the honey badger follows the honeyguide to the beehive, and the final position is determined by \(x\). new2 Determined.

[0092] Where \(r\) 7 is a random number between \([0, 1]\).

[0093] Output the final honey position, that is, the optimal parameters \(A\) and \(\alpha\), and use them to construct the Kernel Extreme Learning Machine (KELM) model.

[0094] Now, further explanations are given for the Kernel Extreme Learning Machine and the Honey Badger Optimization Algorithm mentioned above:

[0095] Kernel Extreme Learning Machine

[0096] The kernel function maps the feature space to a higher-dimensional feature space, thereby obtaining better feature discrimination effects. The Kernel Extreme Learning Machine is obtained by combining the kernel function and the Extreme Learning Machine. Compared with the Extreme Learning Machine, the Kernel Extreme Learning Machine usually has better stability and stronger feature fitting ability. According to Mercer's theorem, the kernel function \(\Omega\) can be defined as:

[0097] \(\Omega = HH^T\), \(\Omega(i, j)=h(x_i)h(x_j)=K(x_i, x_j)\), T , where \(H\) is the output matrix of the hidden layer, and \(h(x_i)\), \(h(x_j)\) are the output row vectors corresponding to the input vectors \(x_i\), \(x_j\). i \(x_i\) j \(x_j\) i , \(x_j\) j )

[0098] where \(H\) is the output matrix of the hidden layer, and \(h(x_i)\), \(h(x_j)\) are the output row vectors corresponding to the input vectors \(x_i\), \(x_j\). i \(h(x_i)\) j \(h(x_j)\) i \(x_i\) j \(x_j\)

[0099] The single-hidden layer feedforward network model can be expressed as: \(g(x)=h(x)\beta = H\beta\), where \(\beta\) is the weight matrix between the hidden layer and the output layer. By solving the dual optimization problem of the multi-output ELM, we can obtain: where \(I\) is the identity matrix, \(A\) is the regularization coefficient, and \(T\) is the expected output matrix. Combining the expressions of the kernel function \(\Omega\), the single-hidden layer feedforward network model, and the weight matrix \(\beta\) between the hidden layer and the output layer, the actual output can be obtained as:

[0100]

[0101] In the present invention, the selected kernel function is the radial basis function, that is:

[0102]

[0103] where \(\alpha\) is the undetermined kernel parameter, and the performance of the kernel function can be improved by adjusting \(\alpha\).

[0104] Honey Badger Optimization Algorithm

[0105] The Honey Badger Algorithm is a bio-inspired algorithm proposed by Fatma A. Hashima et al. in 2021. The algorithm simulates the process of honey badgers searching for honey. During the search phase, the honey badger locates the hive by the smell of honey; if the smell is strong, the movement of the honey pot will intensify, otherwise it will slow down. The movement trajectory of the honey badger is a heart-shaped curve, and the direction of movement has a certain degree of randomness. When led by a honey bird, the honey badger will follow the honey bird directly to the hive. This foraging behavior can prevent the algorithm from falling into a local optimum. The mathematical theory of the Honey Badger Algorithm is as follows:

[0106] (1) Initialize the honey badger group

[0107] x t =lb t +r 1 ×(ub t -lb t )

[0108] where x t represents the position of the tth honey badger, r 1 is a random number between [0,1], ub t and lb t for search upper bound and search lower bound.

[0109] (2) Defining odor intensity

[0110] The intensity of the smell is determined by the concentration S of the honey and the distance between the honey badger and the honey. If the smell intensity is high, the honey badger will move faster, otherwise the movement will slow down.

[0111] S=(x t -x t+1 ) 2

[0112] d i =x prey -x t

[0113]

[0114] where d t represents the distance between the honey and the tth honey badger, x prey represents the best prey location found so far, I t represents the odor intensity between honey and the i-th honey badger, r 2 is a random number between [0 1].

[0115] (3) Update density factor

[0116] Density Factor The change over time is gradually reduced to ensure the smoothness of the search process.

[0117]

[0118] Where C is a constant ≥ 1, the default value is 2, and w max is the maximum number of iterations.

[0119] (4) Search and honey bird leading stage

[0120] The action pattern of the honey badger is determined by r, which is a random number between [0 1]. When 0≤r<0.5, the honey badger searches based on the smell, and its movement trajectory is heart-shaped. By setting the search direction F, the honey location can be better found. new1 It indicates the best position updated by the honey badger through autonomous search.

[0121]

[0122]

[0123] Where ρ is a constant ≥ 1, with a default value of 6, representing the ability of the honey badger to find food. 3 、r 4 、r 5 are all random numbers between [0 1].

[0124] When 0.5≤r≤1, the honey badger follows the honey bird to the beehive, and the final position is x new2 Decide.

[0125] where r 7 is a random number between [0 1].

[0126] like Figure 4As shown, the present invention also proposes a training device 100 for a bearing fault diagnosis model, which applies the bearing fault diagnosis method described in the above embodiment. Specifically, the bearing fault diagnosis device 100 includes a training set acquisition module 1, a parameter selection module 2, a feature set acquisition module 3, and a model training module 4. The training set acquisition module 1 is used to perform data segmentation on the fault vibration signals of different types of bearing faults to obtain a training set, and the training set includes a number of training samples; the parameter selection module 2 is used to calculate the contour coefficients of the training samples in the training set under different parameters to select the optimal parameters of the enhanced composite multi-scale slope entropy; the feature set acquisition module 3 is used to extract fault features using the enhanced composite multi-scale slope entropy after parameter selection to obtain a training feature set; the model training module 4 is used to train the Honey Badger algorithm optimization kernel extreme learning machine using the training feature set to obtain a trained composite multi-scale fault diagnosis model.

[0127] The training process of the bearing fault diagnosis model of the present invention will be described below with reference to a specific example.

[0128] In order to demonstrate the effectiveness and feasibility of the method of the present invention, data collected on-site on the test bench are used for verification. The data is collected from the aircraft engine bearing test bench (i.e., bearing experimental device) of our unit. Figure 5 As shown, the bearing test device 2 includes a bearing test machine 21 , a loading station 22 , a lubrication system 23 , and a cooling system 24 .

[0129] Bearing damage types include healthy, inner ring, outer ring, roller single-point and multi-point faults, as well as outer ring / roller and inner ring / roller composite faults, a total of 9 states, with numbers 1 to 9 as the category labels of different faults. Table 1 shows their corresponding relationships.

[0130] Table 1 Correspondence table of bearing fault types, quantities and locations

[0131]

[0132] The experimental data were collected using LMS Test.lab software, with a sampling frequency of 20480Hz. The bearing operating conditions were axial load of 2kn and speed of 2000rpm. All fault signals were segmented into 1024 sample points per segment, and 100 samples were taken for each fault. 20 samples were randomly selected as the sample set, and the remaining 80 samples were used as the test set. There were 180 samples in the sample set and 720 samples in the test set. The test set was used to test the trained diagnostic model.

[0133] In this embodiment, in order to simplify the selection of parameters, we first observe the variation law of the low threshold δ. Set the embedding dimension m = 3, the delay time τ = 1, and the maximum scale factor smax =8, the silhouette coefficients of different low thresholds δ under high thresholds γ are as follows Figure 6 As shown in Figure 2, it can be seen that under the three high thresholds γ, as δ decreases, the corresponding silhouette coefficient gradually increases, and all of them are around 10 -3 When δ continues to decrease, the silhouette coefficient no longer changes, so δ=10 -3 As the optimal low threshold. Secondly, observe the corresponding relationship between the embedding dimension m = 2 to 7 and the contour coefficients of the three high thresholds. Figure 7 As shown, it can be seen that m = 3, γ = 60° is the optimal combination. The final parameters are determined to be τ = 1, s max =8, m=3, δ=0.001°, γ=60°, where, s max is the maximum scale factor.

[0134] At the same time, in this example, in order to verify the feature extraction performance of the enhanced composite multi-scale slope entropy algorithm, it is compared with the composite multi-scale slope entropy (CMSE), the difference-based composite multi-scale slope entropy (DBCMSE), the refined composite multi-scale dispersion entropy (RCMDE), and the composite multi-scale weighted permutation entropy (CMWPE). In order to ensure the fairness of the experiment, all embedding dimensions are set to m = 3, the delay time is τ = 1, and the maximum scale factor is set to s max =8,. The number of categories of the fine composite multi-scale walk entropy is selected as c = 5. Five algorithms are used to extract the features of 9 fault signal samples. The t-SNE visualization dimension reduction results are shown in the figure. Figures 8-1 to 8-5 shown.

[0135] Depend on Figures 8-1 to 8-5It can be seen that, except for the short distances between the outer ring / roller composite fault and the inner ring / roller composite fault, the other fault types have obvious distinction. Overall, the distances within the bearing fault feature classes of the enhanced composite multi-scale slope entropy ECMSE are compact, and there is no crossover between the fault features. The composite multi-scale slope entropy CMSE has the phenomenon of crossover of roller single-point fault, outer ring / roller composite fault and inner ring / roller composite fault features, and some inner ring single-point faults are classified as normal categories. The visualization effect of the composite multi-scale slope entropy DBCMSE based on difference is better than that of the composite multi-scale slope entropy CMSE, but there are shortcomings such as the short distances between multiple fault classes and the crossover of minor faults. It can be seen that the enhanced composite multi-scale slope entropy ECMSE, which has both high and low frequency features, has better performance than the composite multi-scale slope entropy CMSE and the composite multi-scale slope entropy DBCMSE based on difference. Both the refined composite multi-scale slope entropy RCMDE and the composite multi-scale slope entropy CMWPE have the problem that the intra-class distance is not compact enough, and there is a serious overlap between faults, and the overall discrimination is relatively poor. In summary, the composite multi-scale slope entropy ECMSE proposed in the present invention has the best discrimination characteristics and the best feature extraction ability.

[0136] In addition, in this embodiment, in order to verify the effect of the rolling bearing fault diagnosis method in this paper, the five algorithms are input into the Honey Badger algorithm optimization kernel extreme learning machine classifier to verify the classification performance. In this embodiment, the population number of the HBA algorithm is initialized to 30, and the maximum number of iterations is 50. Figures 9-1 to 9-5 , which are the diagnostic results of one experiment of five algorithms.

[0137] Depend on Figures 9-1 to 9-5 It can be seen that the diagnostic accuracy of ECMSE is 100%, and all 720 faults are correctly classified. DBCMSE has 6 misdiagnosed faults. Compared with CMSE, DBCMSE has higher diagnostic performance, which reflects the superiority of the differential method. However, RCMDE and CMWPE have more fault classification errors and poor fault identification performance. Experiments show that ECMSE+HBA-KELM has the highest diagnostic accuracy and the best diagnostic ability.

[0138] At the same time, in order to further explore the stability of ECMSE, the five algorithms were subjected to 30 experiments. The training set and test set of each experiment were randomly selected, and the number of training sets and test sets was still 180 and 720. The experimental results are as follows: Figure 10 As shown. Figure 10It can be seen that in 30 experiments, the accuracy of ECMSE is always the highest, and the fluctuation is very low. DBMSE shows an accuracy similar to CMSE in some experiments, but the overall effect is better than CMSE. The recognition results of RCMDE and CMWPE fluctuate greatly and have poor diagnostic ability. The average accuracy of the five methods is shown in Figure 11 As shown, the average accuracy of the ECMSE proposed by the present invention reaches 99.9%, which is an ideal bearing fault diagnosis method.

[0139] In summary, the present invention proposes a bearing fault diagnosis model training method and training device, which utilizes the difference method to extract the high-frequency information of the time series, utilizes the average method to extract the low-frequency information of the sequence, uses the slope entropy to extract the complex features of the two multi-scale sequences, and averages the acquired high and low frequency features to form an enhanced composite multi-scale slope entropy. The extracted fault features are input into the Honey Badger algorithm optimized kernel extreme learning machine classifier to realize bearing fault diagnosis.

[0140] Compared with the general entropy algorithm, the enhanced composite multi-scale slope entropy proposed in the present invention extracts more comprehensive and rich time series and scale information; at the same time, the global optimization capability of the honey badger algorithm is used to realize the automatic determination of the parameters of the kernel extreme learning machine.

[0141] The method of the present invention can effectively extract the fault characteristics of bearings in different health states, has good recognition capabilities for both single and compound bearing faults, and can effectively distinguish bearing faults of different types and damage degrees, with an average recognition rate of 99.9%.

[0142] The above description is only a preferred embodiment of the present application and an explanation of the technical principles used. Those skilled in the art should understand that the scope involved in the present application is not limited to the technical solution formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the inventive concept, such as a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

[0143] Except for the technical features described in the specification, the remaining technical features are known technologies to those skilled in the art. In order to highlight the innovative features of the present invention, the remaining technical features will not be described here in detail.

Claims

1. A method for training a bearing fault diagnosis model, characterized in that, it includes the following steps: Perform data segmentation on the fault vibration signals of different types of bearing faults to obtain a training set, where the training set includes a number of training samples; Calculate the silhouette coefficients of the training samples in the training set under different parameters to select the optimal parameters of the enhanced composite multi-scale slope entropy, where the different parameters include scale factor, delay time, embedding dimension, low threshold, and high threshold; Use the enhanced composite multi-scale slope entropy after parameter selection to extract fault features and obtain a training feature set; Use the training feature set to train the honey badger algorithm optimized kernel extreme learning machine to obtain a trained composite multi-scale fault diagnosis model; Among them, the enhanced composite multi-scale slope entropy ECMSE is obtained by taking the average of the composite multi-scale slope entropy CMSE and the differential-based composite multi-scale slope entropy DBCMSE under each scale factor, and its formula is as follows: where \(x\) represents the time series, represents the embedding dimension, represents the delay time, represents the scaling factor, represents the lower threshold, represents the upper threshold, represents the \(k\) -th average coarse - grained sequence under the \(s\) -th scaling factor, represents the \(k\) -th differential coarse - grained sequence under the \(s\) -th scaling factor; According to the definition, According to the definition, where , represents rounding the number, represents choosing numbers from numbers, and \(N\) is the total number of samples.

2. The method for training a bearing fault diagnosis model according to claim 1, characterized in that, the different types of faults include one or more combinations of normal state, inner ring single point fault, outer ring single point fault, roller single point fault, outer ring roller composite fault, inner ring roller composite fault, inner ring multi-point fault, outer ring multi-point fault, and roller multi-point fault.

3. The method for training a bearing fault diagnosis model according to claim 1, characterized in that, the selection of the optimal parameters of the enhanced composite multi-scale slope entropy includes the following steps: Set the scale factor, delay time, and embedding dimension as fixed values, and observe the change trend of the silhouette coefficients of the training samples in the training set under different low thresholds and different high thresholds to determine the optimal low threshold; After determining the optimal low threshold, observe the silhouette coefficients of the training samples in the training set under different embedding dimensions and different high thresholds to determine the optimal embedding dimension and high threshold.

4. The method for training a bearing fault diagnosis model according to claim 3, characterized in that, the optimal parameter combination is: the maximum scale factor is 8, the delay time is 1, the embedding dimension is 3, the low threshold is 0.001°, and the high threshold is 60°.

5. The method for training a bearing fault diagnosis model according to claim 1, characterized in that, the steps of using the training feature set to train the honey badger algorithm optimized kernel extreme learning machine to obtain a trained composite multi-scale fault diagnosis model include: Input the training feature set, initialize the population size and positions, determine the population size, and the maximum number of iterations , the constant C and ρ, as the regularization optimization parameters of the kernel extreme learning machine and the kernel function parameters as the position information of the honey badgers; Establish the classification error rate of the training set as the fitness function, calculate the fitness of each honey badger, and record the position of the best fitness as ; If the number of iterations , then update the positions of the honey badger population using a predefined formula and update the optimal position for each iteration; Output the final honey location, i.e., the optimal parameters and , and use them to construct a kernel extreme learning machine model.

6. A bearing fault diagnosis model training device, characterized in that, it includes: A training set acquisition module for performing data segmentation on the fault vibration signals of different types of bearing faults to obtain a training set, where the training set includes a number of training samples; A parameter selection module for calculating the silhouette coefficients of the training samples in the training set under different parameters to select the optimal parameters of the enhanced composite multi-scale slope entropy, where the different parameters include scale factor, delay time, embedding dimension, low threshold, and high threshold; A feature set acquisition module for using the enhanced composite multi-scale slope entropy after parameter selection to extract fault features and obtain a training feature set; The model training module is used to train the honey badger algorithm optimized kernel extreme learning machine by using the training feature set to obtain a trained composite multi-scale fault diagnosis model; Among them, the enhanced composite multi-scale slope entropy ECMSE is obtained by taking the average of the composite multi-scale slope entropy CMSE and the differential-based composite multi-scale slope entropy DBCMSE under each scale factor. The formula is as follows: where \(x\) represents the time series, represents the embedding dimension, represents the delay time, represents the scaling factor, represents the low threshold, represents the high threshold, represents the \(k\)-th average coarse-grained sequence under the \(s\)-th scaling factor, represents the \(k\)-th differential coarse-grained sequence under the \(s\)-th scaling factor; According to the definition, According to the definition, where , represents rounding of a number, represents choosing numbers from numbers, and \(N\) is the total number of samples.

Images