Application research of improved classification algorithm based on multi-feature signal fusion in aviation bearing fault diagnosis
By adopting multi-eigen signal fusion and improving atomic search optimization support vector machine (IASO-SVM) algorithm in aviation bearing fault diagnosis, the problem of low fault diagnosis accuracy and model susceptible to hyperparameters in the existing technology is solved, and a more efficient and robust fault diagnosis effect is achieved.
Patent Information
- Application Number
- CN202510104632.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-16
- Publication Date
- 2025-06-03
AI Technical Summary
The existing technology is difficult to effectively integrate multi-feature signals, resulting in low accuracy in aerospace bearing fault diagnosis, and traditional models such as SVM are easily affected by hyperparameters, and atomic search optimization algorithms are easily trapped in local optimization.
The improved atomic search optimization support vector machine (IASO-SVM) algorithm based on multi-feature signal fusion is adopted to extract frequency domain and time-frequency characteristic signals through fast Fourier transform, ICEEMDAN and VMD, perform feature fusion and screening, and combine Levy flight variation and parameter optimization to optimize the SVM model.
The accuracy of aviation bearing fault diagnosis is improved, the robustness and optimization capabilities of the model are enhanced, the dependence on punishment factors and core parameters is reduced, and the local optimal problem is avoided.
Smart Images
Figure CN120086725A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of aero-bearing fault diagnosis, and particularly to an improved atom search optimization classification fault diagnosis method based on multi-feature signal fusion. Background Technique
[0002] Rolling bearings are key supporting components of aero-engines. However, under complex working conditions such as heavy loads, strong impacts, and high rotational speeds, rolling bearings are prone to failure. Once a bearing fails, it may cause damage to the equipment, and even result in significant economic losses and casualties. In addition, since the measured vibration signals of rolling bearings usually exhibit non-linearity and non-stationarity, it is difficult for traditional methods to effectively capture the fault feature information therein. Therefore, it is particularly important to study an effective feature extraction method and pattern recognition method to timely detect potential faults of aero-bearings. Current research mainly focuses on the extraction of time-frequency domain feature signals, but multi-feature signals are not effectively fused. Therefore, the extracted features are difficult to comprehensively reflect the operating state of the bearing from multiple angles. In the case of relatively severe noise interference, these features are difficult to effectively characterize the fault information, thereby resulting in low accuracy of fault recognition. In addition, the performance of the support vector machine (SVM) model commonly used for fault diagnosis is easily affected by the penalty factor (C) and the kernel parameter (g). Due to the problem of being easily trapped in local optima, the atom search optimization algorithm (ASO) cannot find suitable SVM hyperparameters, and the fault diagnosis effect is not ideal. Summary of the Invention
[0003] The purpose of the present invention is to provide an improved atom search algorithm optimized support vector machine (IASO-SVM) fault diagnosis algorithm based on multi-feature signal fusion to improve the accuracy of aero-bearing fault diagnosis.
[0004] To achieve the above purpose, the present invention provides the following solution:
[0005] An IASO-SVM fault diagnosis algorithm based on multi-feature signal fusion, comprising:
[0006] Obtain the vibration signal of the aero-bearing;
[0007] Use the fast Fourier transform (FFT) to calculate the centroid frequency, average frequency, root mean square frequency, and frequency variance of the vibration signal of the aero-bearing respectively to extract the frequency domain feature signal;
[0008] Perform modal decomposition on the vibration signal of the aviation bearing, that is, use the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) and variational mode decomposition (VMD) to decompose the above signal respectively to obtain the decomposed vibration signal of the aviation bearing; calculate the approximate entropy (ApEn) and fuzzy entropy (FuzzyEn) for the decomposed vibration signal of the aviation bearing respectively to obtain two time-frequency feature signals;
[0009] Fuse the frequency domain feature and the two time-frequency domain features to obtain a fused feature signal.
[0010] Use the random forest feature screening method for the fused feature signal to obtain the importance of each feature signal, arrange them in descending order according to the feature importance, and then select effective features according to the accuracy of different numbers of features to filter out redundant features;
[0011] Use principal component analysis (PCA) to reduce the dimension of the effective feature signal to obtain a low-dimensional fused feature signal to improve the calculation efficiency;
[0012] Subsequently, input the low-dimensional fused feature signal into the IASO-SVM fault diagnosis algorithm for fault diagnosis. The specific process is as follows:
[0013] First, initialize the population using the Logistic-Tent mapping, that is, randomly generate the penalty factor C and kernel parameter g of the support vector machine with a uniform distribution;
[0014] Secondly, use the IASO algorithm to calculate the standard deviation to measure the data volatility, and judge whether the population needs to mutate. Set the mutation coefficient threshold to 0.1. When the standard deviation (Std) of the optimal fitness in the last 3 iterations is less than 0.1, perform Levy flight mutation on the atom positions to help the algorithm jump out of the local extreme value. The position mutation formula after introducing Levy flight is as follows: where, is the position after mutation in the t-th iteration, rand(t) is a random number uniformly distributed between (0, 1) in the t-th iteration, x(t) is the position before mutation in the t-th iteration, Levy(t) is the flight function, and x best (t) is the position of the best fitness in the t-th iteration.
[0015] Then, according to the fault diagnosis accuracy rate, the improved atomic search algorithm is used to optimize the penalty factor C and the kernel parameter g of the support vector machine, so as to obtain the penalty factor C and the kernel parameter g under the optimal fault diagnosis accuracy rate;
[0016] Finally, the above low-dimensional fusion feature signal data is used to train the support vector machine with optimized parameters, and then this model is used for the fault diagnosis of aviation bearings. Description of the Drawings
[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required to be used in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0018] Figure 1 Flowchart of the IASO-SVM fault diagnosis algorithm;
[0019] Figure 2 Relationship diagram of the number of features, accuracy rate and running time of the RF feature screening method;
[0020] Figure 3 Diagram of the number of features and accuracy rate of different feature selection methods:
[0021] Figure 4 Relationship diagram of the number of PCA features, accuracy rate and running time;
[0022] Figure 5 Test function diagram;
[0023] Figure 6 Optimization curve of the test function. Detailed Embodiments
[0024] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the protection scope of the present invention.
[0025] To make the above objects, features and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.
[0026] As Figure 1 shown, the present invention provides an IASO-SVM fault diagnosis algorithm based on multi-feature signal fusion, including:
[0027] Step 201: Obtain the vibration signal of the aviation bearing. In practice, the vibration signal of the aviation bearing is collected by a vibration accelerometer.
[0028] Step 202: Calculate the centroid frequency, average frequency, root mean square frequency, and frequency variance of the above bearing vibration signal through fast Fourier transform (FFT) to extract the frequency domain characteristic signal; then, decompose the bearing vibration signal using ICEEMDAN and VMD, and calculate the fuzzy entropy and approximate entropy of the decomposed signal to obtain the time-frequency characteristic signal. Then, fuse the above frequency domain characteristics with the two time-frequency domain characteristics to obtain the fused characteristic signal.
[0029] Step 203: Use the random forest feature screening method for the above fused characteristic signal to obtain the importance of each feature, and arrange them in descending order according to the feature importance. Subsequently, select the effective features according to the accuracy rate of different numbers of features, and filter out the redundant features.
[0030] Step 204: Use principal component analysis (PCA) to reduce the dimension of the effective features to obtain a low-dimensional fused characteristic signal to improve the calculation efficiency.
[0031] Step 205: Input the dimension-reduced fused characteristic signal into the IASO-SVM fault diagnosis algorithm for model training and fault diagnosis respectively.
[0032] Step 205 specifically includes:
[0033] Initialize the population. Initialize the population using the Logistic-Tent mapping, that is, randomly generate the penalty factor C and kernel parameter g of the support vector machine with a uniform distribution.
[0034] Levy flight mutation. Use the IASO algorithm to calculate the standard deviation to measure the data volatility, judge whether the population needs to mutate, and set the mutation coefficient threshold to 0.1. When the standard deviation of the optimal fitness in the last 3 iterations is less than 0.1, perform Levy flight mutation on the atomic position to help the algorithm jump out of the local extreme value. The position mutation formula after introducing Levy flight is as follows: Among them, is the position after mutation in the t-th iteration, rand(t) is a random number uniformly distributed between (0, 1) in the t-th iteration, x(t) is the position before mutation in the t-th iteration, Levy(t) is the flight function, and xbest(t) is the position of the best fitness in the t-th iteration.
[0035] Parameter optimization. According to the fault diagnosis accuracy rate, the IASO algorithm is used to optimize the penalty factor C and the kernel parameter g to obtain the penalty factor C and the kernel parameter g under the optimal fault diagnosis accuracy rate;
[0036] Model training and fault diagnosis. Use the above low-dimensional fusion feature signal data to train the parameter-optimized SVM, and then use this model for fault diagnosis.
[0037] The time-frequency feature signal extraction process usually involves the decomposition of vibration signals and the calculation of entropy values for each sub-signal. Therefore, the present invention combines ICEEMDAN decomposition, local mean decomposition (LMD), VMD decomposition, and wavelet packet decomposition (WPD) with approximate entropy, fuzzy entropy, and sample entropy, and evaluates various time-frequency features by accuracy, as shown in Table 1. Table 1 Comparison of recognition accuracy rates of time-frequency feature signals with different combinations
[0038] As can be seen from Table 1, the time-frequency feature signal extraction method of combining ICEEMDAN decomposition with fuzzy entropy and the extraction method of combining VMD decomposition with approximate entropy have better recognition accuracy rates than other methods. Therefore, the present invention intends to combine these two time-frequency feature signal extraction methods with other domain features to find the optimal feature fusion scheme, and the recognition accuracy rates of the combined features are shown in Table 2. Table 2 Accuracy rates of two signal feature combinations
[0039] As can be seen from Table 2, the best result is the time-frequency fusion feature signal combined with the frequency domain and VMD decomposition, with an accuracy rate of 97.3061%, which is 24.4081% higher than that of the single-time domain feature signal and 6.6939% higher than that of the frequency domain feature signal. In addition, in order to more comprehensively characterize the signal features, the present invention also studied the effects of combinations of two or more features, and the results are shown in Table 3. Table 3 Training time and recognition accuracy rates of multiple signal feature combinations ① Time domain feature, ② Frequency domain feature, ③ ICEEMDAN_FuzzyEn, ④ VMD_ApEn
[0040] As can be seen from Table 3, after fusing three or more features, the highest fault recognition accuracy rate is achieved by the fusion feature signal combining the frequency-domain feature signal with ICEEMDAN_FuzzyEn and VMD_ApEn (②+③+④). When the number of features is further increased, the recognition accuracy rate of SVM does not improve. Instead, due to the introduction of too many redundant features, the accuracy rate slightly decreases, and at the same time, the training time also increases. In summary, it is relatively scientific for the present invention to select the above-mentioned fusion time-frequency domain feature signal as the feature parameter of the acquisition signal.
[0041] In practical engineering applications, it is necessary to strike a balance between diagnostic accuracy and calculation time. To better play the role of feature selection, the present invention studies the relationship between the diagnostic accuracy rate and calculation time of the random forest (RF) feature screening method (referred to as the RF method) under different feature dimensions, as Figure 2 shown. It can be seen from the figure that when this method selects 13-dimensional features, the classification accuracy rate reaches the highest 98.2857%, and the calculation time is 76.548 seconds, which is reduced by 21.7469% compared with before, achieving a good balance between diagnostic accuracy and calculation time. This shows that it is more in line with the actual engineering requirements for the RF method to select 13-dimensional features.
[0042] To further verify the effectiveness of the RF method, this application compares it with the chi-square test feature screening method and the Laplacian feature screening method, as Figure 3 shown. It can be observed from the figure that when the number of selected features is less than 5, the accuracy rate of the RF method is slightly lower than that of the chi-square test feature screening method, but significantly higher than that of the Laplacian feature screening method. When the feature dimension is between 5 and 10 dimensions, the accuracy rate of the RF method rises rapidly and approaches the maximum value at 7 dimensions, significantly higher than the other two methods; while between 10 and 15 dimensions, the performance of the RF method is almost equivalent to that of the chi-square test feature screening method, and it still maintains a high classification accuracy rate at 13 dimensions and is better than the Laplacian feature screening method. When the feature dimension is between 16 and 20 dimensions, the performance of the Laplacian feature screening method is similar to the accuracy rates of the other two methods. However, since the more features are selected, the more calculation time is required, therefore, overall, the RF method is slightly better than the chi-square test feature screening method, verifying the effectiveness of the RF method. Finally, this application selects the RF method for feature selection and determines 13-dimensional features for subsequent testing.
[0043] In order to further reduce the complexity of the data, reduce the feature dimension, retain the main features and improve the data quality, the present invention performs principal component analysis (PCA) on the data after feature screening and studies the relationship between the number of principal components and the running time. The results are as Figure 4As shown in the figure. It shows that after principal component analysis, when the dimension is 12, the accuracy of the model reaches 98.4898%, and the calculation time at this time is 73.724 seconds. This indicates that while the accuracy is improved, the calculation time is also reduced. Therefore, it is very necessary to perform PCA.
[0044] Subsequently, the above low-dimensional fusion feature signal is input into the IASO-SVM algorithm for fault diagnosis. As mentioned above, the performance of SVM is easily affected by the penalty factor (C) and the kernel parameter (g). Based on this, the present invention proposes an improved atomic search algorithm (IASO) to solve the problem that SVM falls into local optimum. The process is as Figure 1 shown.
[0045] First, use the Logistic-Tent mapping to randomly generate the SVM penalty factor C and the kernel parameter g with a uniform distribution; then, use the IASO algorithm to calculate the standard deviation to measure the data volatility. When the standard deviation of the optimal fitness in the last 3 iterations is less than 0.1, perform Levy flight mutation on the atomic positions to help the algorithm jump out of the local extreme value; then, according to the fault diagnosis accuracy, use the IASO algorithm to find the penalty factor C and the kernel parameter g under the optimal fault diagnosis accuracy; finally, use the above low-dimensional fusion feature signal data to train the SVM model with optimized parameters, and then perform fault diagnosis according to this model.
[0046] To verify the effectiveness of the improved algorithm, the present invention also makes a comparative study on the optimization results of the IASO algorithm, the Coral Reefs Optimization Algorithm (CRO), the Human Conception Optimizer (HCO), and the standard ASO algorithm for the Schwefel 2.22 (F1), Quartic (F2), Ackley (F3), and Griewank (F4) functions. Among the above test functions, F1 and F2 are unimodal test functions, while F3 and F4 are multimodal test functions. Their expressions are shown in Table 4, and the function graphs are as Figure 5 shown.
[0047] During the comparison process, the population size is uniformly set to 30, and the maximum number of iterations is 500. To reduce the error impact caused by chance, each algorithm is executed 10 times during independent testing. The optimal value (Best), the average value (Mean), and the standard deviation (Std) are used as evaluation indicators. The optimization results are shown in Table 5 in detail.
[0048] As can be seen from Table 5, under the conditions of the same population size and the maximum number of iterations, the optimization results of the IASO algorithm are better than those of other algorithms. For the unimodal test functions F1 and F2, IASO can respectively find the theoretical optimal values of 4.1168×10-16 and 2.9726×10 -6 , showing its good optimization performance. In the multimodal test function F3, both the ASO and IASO algorithms are closer to the optimal value, highlighting the superiority of these algorithms. For the F4 function, although the optimization result of the ASO algorithm is just so-so, the improved IASO algorithm can stably find the optimal solution. To sum up, the IASO algorithm is superior to other algorithms in terms of optimization accuracy and robustness, indicating that it has both strong global optimization ability and local exploration ability. Table 4 Test function expressions Table 5 Comparison of the effects of different algorithms in test functions
[0049] To further prove the advantages of the IASO algorithm in terms of optimization accuracy and convergence speed, the present invention analyzed the performance of the above four algorithms on test functions and plotted the corresponding convergence curves, as Figure 6 shown.
[0050] It can be Figure 6 seen that in both unimodal and multimodal test functions, the convergence accuracy of the IASO algorithm and the ASO algorithm is significantly higher than that of other algorithms. At the same time, the number of iterations required for the IASO algorithm to reach the same accuracy is less than that of the ASO algorithm, indicating that its convergence speed has been improved.
[0051] To verify the effect of the above improved algorithm in measured data, the present invention took 875 groups of detection data from a certain aviation bearing test bench as an example and analyzed its improvement effect on fault identification.
[0052] Research shows that the fault categories of aviation bearings can be inner ring indentation 150μm, inner ring indentation 250μm, inner ring indentation 450μm, normal, roller indentation 150μm, roller indentation 250μm, roller indentation 450μm, which are represented by numbers 1-7 respectively. The data sample distribution is shown in Table 6. Table 6 Sample data distribution
[0053] The present invention comparatively verifies the classification accuracies, precisions, and recall rates of IASO-SVM, CRO-SVM, HCO-SVM, ASO-SVM, as well as grid search, random forest (RF), logistic regression (LR), gradient boosting method (GBM), and the method in the paper (H.E H, Salama D A, N.H H, et al. A Hybrid Barnacles Mating Optimizer Algorithm With Support Vector Machines for Gene Selection of Microarray Cancer Classification[J]. IEEE ACCESS, 2021, 964895-64905). The results are shown in Table 7. Table 7 Comparison of the effects of different algorithms
[0054] From the comparative analysis of Table 7, it can be seen that the accuracy, precision, and recall rate of IASO-SVM are all higher than those of other algorithms. This indicates that this algorithm can more reliably monitor the operating state of aeroengine bearings, thereby providing more accurate fault diagnosis.
[0055] In this specification, each embodiment is described in a progressive manner. The key point of each embodiment is to illustrate the differences from other embodiments. For the same or similar parts among the embodiments, reference can be made to each other. For the system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and reference can be made to the description in the method part for the relevant parts.
[0056] In this paper, specific examples are used to elaborate on the principle and implementation mode of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, there will be changes in the specific implementation mode and application scope according to the idea of the present invention. In summary, the content of this specification should not be construed as a limitation to the present invention.
Claims
1. An improved atomic search algorithm based on multi-feature signal fusion to optimize support vector machine fault diagnosis algorithm, characterized in that: Extract and fuse characteristic signals from aviation bearing vibration signals, including: The center of gravity frequency, average frequency, root mean square frequency and frequency variance of the collected aviation bearing vibration signal are calculated by fast Fourier transform to extract the frequency domain characteristic signal. Performing modal decomposition on the above-mentioned aviation bearing vibration signal, that is, using improved fully adaptive noise set empirical mode and variational mode to decompose the above-mentioned signal to obtain a decomposed aviation bearing vibration signal; calculating fuzzy entropy and approximate entropy of the decomposed aviation bearing vibration signal to obtain two time-frequency characteristic signals; The frequency domain characteristic signal is fused with two time-frequency domain characteristic signals to obtain a fused characteristic signal.
2. According to the fusion feature signal extracted according to claim 1, feature screening is performed using a random forest feature screening method, and dimension reduction of the screened feature signal is performed using principal component analysis, specifically comprising: A random forest feature screening method is used on the fused feature signal to obtain the importance of each feature, and the features are arranged from large to small according to their importance, and then effective features are selected according to the accuracy of different feature numbers, and redundant features are filtered out; The principal component analysis is used to reduce the dimension of the screening features to obtain low-dimensional feature signals to improve computational efficiency.
3. According to claim 2, the low-dimensional fusion feature signal is input into a support vector machine optimized by an improved atomic search algorithm to perform model training and fault diagnosis, specifically including: First, the population is initialized using the Logistic-Tent mapping, that is, the penalty factor C and kernel parameter g of the support vector machine with uniform distribution are randomly generated; Secondly, the improved atomic search optimization algorithm is used to calculate the standard deviation to measure data volatility and determine whether the population needs to mutate. The coefficient of variation threshold is set to 0.
1. When the standard deviation of the optimal fitness of the last three iterations is less than 0.1, the atomic position is subjected to Levy flight mutation to help the algorithm jump out of the local extreme value. The position mutation formula after the introduction of Levy flight is as follows: in, is the position after mutation in the tth iteration, rand(t) is a random number uniformly distributed between (0, 1) in the tth iteration, x(t) is the position before mutation in the tth iteration, Levy(t) is the flight function, X best (t) is the best fitness position in the tth iteration. Then, according to the fault diagnosis accuracy, the penalty factor C and kernel parameter g of the support vector machine are optimized using the improved atomic search algorithm to obtain the penalty factor C and kernel parameter g under the optimal fault diagnosis accuracy; Finally, the low-dimensional fusion feature signal data is used to train a support vector machine with optimized parameters, and then the model is used for aviation bearing fault diagnosis.