Method for diagnosing faults in a transmission system based on hybrid feature selection

Abstract

The application discloses a kind of transmission system fault diagnosis methods based on hybrid feature selection, z-score mode is used to standardize all fault features in the holographic feature library of transmission system;Irrelevant and redundant features are removed from the standardized holographic feature library using the maximum correlation minimum redundancy filtering feature selection method, to obtain the feature preliminary selection library;Class separability quantification index based on cross entropy is used to evaluate the feature subset under different selection, and on this basis, fitness function is designed, and the optimal feature subset is further selected from the feature preliminary selection library using binary particle swarm optimization algorithm;According to the feature selection result of training set, the test set after feature optimization is input into fault diagnosis system, to realize the identification of transmission system fault type.The application can effectively improve transmission system fault diagnosis efficiency.

Description

Technical Field

[0001] This invention relates to a technology in the field of mechanical fault detection, specifically a transmission system fault diagnosis method based on hybrid feature selection. Background Technology

[0002] Existing hybrid feature selection models typically employ classifiers to evaluate the quality of feature subsets in their fitness function design, aiming to maximize classification accuracy. Multi-fold cross-validation is also used to improve the reliability of feature selection. However, this fitness function design has the following drawbacks: 1. Using classification accuracy as a metric makes it difficult to quantify subtle differences between different selected feature subsets, thus increasing the difficulty of searching for the optimal feature subset; 2. Multi-fold cross-validation significantly increases time complexity. Therefore, it is necessary to improve the design of the fitness function in existing hybrid feature selection algorithms to enhance the efficiency of transmission system fault diagnosis. Summary of the Invention

[0003] This invention addresses the problems of low sensitivity to feature selection in existing technologies, inability to quantify subtle differences between different selected feature subsets, and difficulty in effectively searching for the optimal feature subset. It proposes a transmission system fault diagnosis method based on hybrid feature selection, which effectively improves the efficiency of transmission system fault diagnosis by constructing an optimal feature library and fault diagnosis system for the transmission system.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a transmission system fault diagnosis method based on hybrid feature selection. The method employs a z-score approach to standardize all fault features in the holographic feature library of the transmission system. A maximum correlation minimum redundancy filtering feature selection method is used to remove irrelevant and redundant features from the standardized holographic feature library, resulting in an initial feature selection library. A cross-entropy-based class separability quantification index is used to evaluate the feature subsets under different selections. Based on this, a fitness function is designed, and a binary particle swarm optimization algorithm is used to further select the optimal feature subset from the initial feature selection library. According to the feature selection results of the training set, the optimized test set is input into the fault diagnosis system to achieve the identification of transmission system fault types.

[0006] The evaluation of feature subsets under different choices using a cross-entropy-based categorical separability quantification index specifically includes:

[0007] 1) Using the K-Nearest Neighbors (KNN) clustering approach, and following the similarity metric, calculate the feature selection for each training sample X. n (S *The similarity between the training sample and all training samples is used to find the K most similar samples in the training sample set. The number of each category in the K nearest neighbors is counted to estimate the posterior probability p(ξ) of each category of the training sample. t |X n (S * ), where: the selected feature subset is S * ξ t (t = 1, 2, L, T) represents the categories, and T represents the number of categories.

[0008] 2) Perform one-hot encoding on the target label c of all training samples to convert it into one-hot label form.

[0009] 3) Based on cross-entropy, the classification ability of the selected feature subset is measured using the class separability evaluation criterion, specifically: Where: b n,t X represents the nth sample. n (S * The t-th category ξ t The true probability.

[0010] 4) The class separability of the selected feature subset is quantified by calculating the average of the cross-entropy between the estimated class posterior probability of the training sample subset of the selected feature subset and the actual one-hot label.

[0011] The fitness function described above is specifically: Wherein: S * The preferred feature subset, This represents the remaining unselected feature subset.

[0012] This invention relates to a system for implementing the above-mentioned method, comprising: a feature extraction unit, a feature initial selection unit, a feature optimization unit, and a fault diagnosis unit, wherein: the feature extraction unit extracts features from single-channel vibration data from the time domain, frequency domain, and time-frequency domain to construct a holographic feature library of the transmission system; the feature initial selection unit uses a maximum correlation minimum redundancy feature selection algorithm to remove irrelevant and redundant features from the holographic feature library to obtain a preliminary feature library of the transmission system; the feature optimization unit uses an improved encapsulated feature selection algorithm based on binary particle swarm optimization to further filter the preliminary feature library to obtain the optimal feature subset of the transmission system; the fault diagnosis unit inputs the feature-optimized test set into a trained SVM model to diagnose and identify faults in the transmission system. Technical effect

[0013] This invention quantifies the classification performance of selected feature subsets using a class separability metric based on cross-entropy and KNN. Through the design of a fitness function in binary particle swarm optimization, it balances the strong class discriminative power of preferred features with the relatively weak class discriminative power of remaining features, maximizing the selection of the optimal feature subset. Compared to existing feature subset evaluation criteria, this invention can quantify subtle differences between different selected feature subsets in less time, while the fitness function fully considers the relatively weak class discriminative power of non-preferred feature subsets, further improving feature selection performance. Attached Figure Description

[0014] Figure 1 This is a flowchart of the present invention;

[0015] Figure 2 This is a schematic diagram of the system of the present invention;

[0016] Figure 3 This is a schematic diagram of the initial selection library of transmission system features in the embodiment;

[0017] Figure 4 This is a schematic diagram of the fitness function variation curve in the embodiment;

[0018] Figure 5 This is a schematic diagram of the preferred feature subset results in the embodiment;

[0019] Figure 6 This is a schematic diagram of the fault identification results in the embodiment;

[0020] Figure 7 This is a schematic diagram of the fault identification results in the embodiment without considering the weak class discriminativeness of the remaining feature subset;

[0021] Figure 8 The fitness function variation curve for existing technologies;

[0022] Figure 9 This is the result of fault identification for existing technology. Detailed Implementation

[0023] like Figure 1As shown, this embodiment relates to a transmission system fault diagnosis method based on hybrid feature selection. The data was collected on a fault simulation test bench of a scaled-down model of a planetary gearbox for a certain type of engine, with a sampling frequency of 20kHz. To simulate the complexity of actual working conditions, vibration signal acquisition was conducted under the following conditions: speed 600 r / min, load 400 Nm. Seven fault types were simulated: planetary gear cracking (PGC), planetary gear severe pitting (PGP), gear ring minor cracking (GRC), gear ring severe pitting (GRP), sun gear moderate cracking (SGC), sun gear severe pitting (SGP), and sun gear moderate wear (SGW). Data sets collected under these seven fault states and normal conditions were used as analysis data. Under each working condition, there were 238 samples for each fault state, each sample size being 4096. This embodiment specifically includes the following steps:

[0024] Step 1: Extract fifteen time-domain statistical features Δ1:Δ from the collected vibration signals. 15 Twelve frequency domain statistical characteristics Λ1:Λ 12 Multiscale permutation entropy MPE1:MPE 25 Using the hierarchical entropy HPE1:HPE8, a holographic feature library of the transmission system is constructed.

[0025] The fifteen time-domain statistical features specifically include: Δ1 = max(X), Δ2 = max(X), Δ3 = min(X), Δ5 = max(X) - min(X), Δ 12 =Δ2 / Δ8, Δ 13 =Δ7 / Δ6, Δ 14 =Δ2 / Δ6, Δ 15 =Δ2 / Δ7, where: X is the vibration data to be processed, and L is the data length.

[0026] The twelve frequency domain statistical features specifically include: Λ 10 =Λ6 / Λ5, Where: c(s) is the spectrum of signal X, s = 1, 2, K, S, S is the number of spectral lines, f s Let be the frequency value corresponding to the s-th spectral line.

[0027] In the aforementioned holographic feature library, the time-domain statistical feature Δ1:Δ 15 Numbered 1-15 respectively; Frequency domain statistical characteristics Λ1:Λ 12Numbered 16-27 respectively, multi-scale permutation entropy MPE1:MPE 25 They are numbered 28-52 respectively; the hierarchical entropy HPE1:HPE8 are numbered 53-60 respectively.

[0028] The hyperparameters in the multi-scale permutation entropy and hierarchical permutation entropy are set as follows: maximum scale factor is 25, number of levels is 3, embedding dimension is 4, and delay time is 1.

[0029] The second step is to use z-score to standardize the holographic feature library, and then randomly select 24 samples from each fault state sample after standardization as the training set, and the remaining 214×8 standardized samples as the test set.

[0030] The third step involves using a maximum correlation minimum redundancy filtering feature selection algorithm to sort the 60 fault features in the holographic feature library S according to their importance from highest to lowest, selecting the top 30 fault features as the initial feature selection library for the transmission system. Specifically, this includes:

[0031] 3.1) A relevance evaluation criterion based on mutual information is used to measure the relevance of each feature, i.e. Where: I(S) i c) is the i-th feature S in the holographic feature library S. i The mutual information between the target category c and the target category c is calculated using the above formula. i The average value of the mutual information between the feature and the target category c is used to quantitatively assess the correlation between each feature and the target category.

[0032] 3.2) A redundancy evaluation criterion based on mutual information is used to measure the redundancy between features, i.e. Where: I(S) i ,S j ) represents the i-th feature S i The mutual information between the j-th feature Sj and the j-th feature Sj is calculated using the above formula. i And another feature S j The average value of mutual information between features is used to quantitatively assess the redundancy between them.

[0033] 3.3) Simultaneously optimize U and R based on the maximum correlation and minimum redundancy evaluation criterion, i.e. The incremental search method is used to sort all features in the holographic feature library S, specifically as follows: in: A subset of features has been selected based on the maximum relevance and minimum redundancy evaluation criteria, with m-1 features chosen first. This process continues until all features in the original feature set are sorted according to the maximum relevance and minimum redundancy criteria.

[0034] 3.4) Based on the sorting results, select the top 30 fault features as the initial feature selection library. like Figure 3 As shown.

[0035] The fourth step involves using an improved encapsulated feature selection algorithm based on binary particle swarm optimization to further refine the initial feature library and obtain the optimal feature subset for the transmission system. This includes:

[0036] 4.1) Each particle is encoded in binary form, and the particle swarm is randomly initialized;

[0037] 4.2) Based on the particle encoding scheme, the binary representation of each particle is transformed into a preferred feature subset;

[0038] 4.3) Calculate the fitness function value of each particle, evaluate the class separability of the selected feature subset represented by each particle, and update the optimal feature position of each particle and the global optimal feature position, specifically: Where: Ω(S) * () is a quantitative index of the class separability of the preferred feature subset. It serves as a quantitative indicator of the class separability of the remaining unselected feature subsets.

[0039] The aforementioned categorical separability quantification index is obtained in the following way:

[0040] a) Perform one-hot encoding on the target label c of all samples to convert it into one-hot label form.

[0041] b) Using the KNN clustering approach, with k set to 11, calculate the clustering results for each training sample X after feature selection. n (S * The similarity between the sample and all training samples is used to find the K most similar samples in the training sample set. The number of samples in each category among the neighboring samples is counted to estimate the posterior probability p(ξ) of each category for each sample. t |X n (S * )), where ξ t (t = 1, 2, L, T) represents the categories, and T represents the number of categories.

[0042] c) Class separability assessment based on cross-entropy, i.e. Where: b n,t X represents the nth sample. n (S * The t-th category ξ tThe ideal probability is calculated. The class separability of the selected feature subset is comprehensively evaluated by calculating the average of the cross-entropy between the estimated class posterior probability of the sample set of the selected feature subset and the actual one-hot label.

[0043] like Figure 4 , 5 As shown, this embodiment presents the fitness function change curve and feature optimization results during the binary group optimization process.

[0044] 4.4) Update the particle's velocity and binary position;

[0045] 4.5) Repeat steps 4.2 to 4.4 until the number of iterations reaches the set value;

[0046] 4.6) Convert the binary representation of the global optimal particle position into the selected optimal feature subset.

[0047] The hyperparameter settings for the binary particle swarm optimization are as follows: particle swarm size is 20, maximum number of iterations is 50, and the two learning factors are 2 and 4 respectively.

[0048] Step 5: Train the linear SVM model using the feature-optimized training set. Based on the feature selection results of the training set, input the feature-optimized test set into the trained linear SVM to obtain the fault identification results, such as... Figure 6 As shown.

[0049] In contrast, with other parameters remaining unchanged, the fault identification results are given when the fitness function does not consider the weak discriminative power of the remaining feature subset, such as... Figure 7 As shown. Comparison Figure 6 and Figure 7 As can be seen, the present invention fully considers the strong class discriminativeness of the preferred features and the weak class discriminativeness of the remaining features in the fitness function design, effectively avoiding misdiagnosis of some test samples, thereby improving fault identification performance.

[0050] In contrast, with other parameters remaining constant, the fitness function variation curves and fault identification results for the existing hybrid feature selection (mRMR+BPSO) case are presented, as shown below. Figure 8 , 9 As shown. Existing hybrid feature selection (mRMR+BPSO) uses a nearest neighbor classifier as the learner in its fitness function design, aiming to maximize accuracy, while employing five-fold cross-validation to improve the reliability of feature selection. (From...) Figure 8 It is known that the fitness function under the existing hybrid feature selection (mRMR+BPSO) uses test accuracy as a metric, which has low sensitivity to feature selection, cannot quantify the subtle differences between different selected feature subsets, and is difficult to effectively search for the optimal feature subset. (Comparison) Figure 7 and Figure 9 Compared to existing fitness functions, the designed fitness function, based on the cross-entropy class separability criterion, finds a more suitable feature subset by quantifying the subtle differences between different preferred feature subsets, thereby improving the fault identification rate to some extent.

[0051] Since this invention does not require multi-fold cross-validation, the group optimization time in the above embodiment is 7.54s. If the weak discriminative power of the remaining feature subset is not considered, the group optimization time of this invention is even less, only 3.46s. In contrast, the group optimization time of existing hybrid feature selection methods is 14.59s under the same conditions. This indicates that compared to existing hybrid feature selection (mRMR+BPSO) methods, this invention has higher computational efficiency and is more suitable for practical engineering applications.

[0052] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for fault diagnosis of a transmission system based on hybrid feature selection, characterized in that, The z-score method is used to standardize all fault features in the holographic feature library of the transmission system; the maximum correlation minimum redundancy filtering feature selection method is used to remove irrelevant and redundant features from the standardized holographic feature library to obtain the initial feature selection library. The feature subsets under different selections are evaluated using a cross-entropy-based category separability quantification index. Based on this, a fitness function is designed, and the optimal feature subset is further selected from the initial feature selection library using a binary particle swarm optimization algorithm. According to the feature selection results of the training set, the test set after feature optimization is input into the fault diagnosis system to realize the identification of the fault type of the transmission system. The evaluation of feature subsets under different choices using a cross-entropy-based categorical separability quantification index specifically includes: 1) Calculate the similarity of each training sample after feature selection according to the similarity measurement criteria. The similarity with all training sample sets is used to find [the similarity] within the training sample set. The sample most similar to the training sample was statistically analyzed. The number of each category in the nearest neighbor samples is used to estimate the posterior probability of each category in the training sample. Where: the selected feature subset is , As a category, , Number of categories; 2) Target labels for all training samples Perform one-hot encoding to convert it into one-hot tag format. ; 3) Based on cross-entropy, the classification ability of the selected feature subset is measured through the category separability evaluation criterion, specifically: category separability quantification index. ,in: Indicates the first Sample No. Categories The true probability; 4) The class separability of the selected feature subset is quantified by calculating the average cross-entropy between the estimated class posterior probability of the training sample subset of the selected feature subset and the actual one-hot label; The fitness function described above is specifically: ,in: The preferred feature subset, This represents the remaining unselected feature subset.

2. The transmission system fault diagnosis method based on hybrid feature selection according to claim 1, characterized in that, The aforementioned holographic feature library of the transmission system refers to the extraction of fifteen time-domain statistical features from the collected vibration signals. Twelve frequency domain statistical characteristics Multiscale permutation entropy and hierarchical entropy ; The fifteen time-domain statistical features specifically include: , , , , , , , , , , , , , , ,in: For the vibration data that needs to be processed, For data length; The twelve frequency domain statistical features specifically include: , , , , , , , , , , , ,in: For signal The spectrum, , The number of spectral lines. For the first The frequency value corresponding to each spectral line.

3. The transmission system fault diagnosis method based on hybrid feature selection according to claim 1, characterized in that, The aforementioned initial feature selection library utilizes a maximum correlation minimum redundancy filtering feature selection algorithm to select the holographic feature library. The 60 fault features were selected after being sorted from most important to least important, and specifically include: 1) A relevance evaluation criterion based on mutual information is used to measure the relevance of each feature, i.e. ,in: For holographic feature library Middle Features and target category The mutual information between features is calculated using the above formula in the feature space. and target category The average value of the mutual information between features is used to quantitatively assess the correlation between each feature and the target category; 2) A redundancy evaluation criterion based on mutual information is used to measure the redundancy between features, i.e. ,in: For the first Features and the Special The mutual information between features is calculated using the above formula in the feature space. And another feature The average value of mutual information between features is used to quantitatively assess the redundancy between them; 3) Simultaneous optimization based on the maximum correlation and minimum redundancy evaluation criteria. and ,Right now An incremental search method is used to analyze the holographic feature library. Sort all features as follows: ,in: This involves selecting a subset of features that have been evaluated according to the maximum relevance and minimum redundancy criterion, using m-1 features, until all features in the original feature set are sorted according to the maximum relevance and minimum redundancy criterion. 4) Based on the size and sorting results of the initial feature selection library, select the top 30 sorted fault features as the initial feature selection library. .

4. The transmission system fault diagnosis method based on hybrid feature selection according to claim 1, characterized in that, The optimal feature subset is obtained in the following way: 1) Each particle is encoded in binary form, and the particle swarm is randomly initialized; 2) Based on the particle encoding scheme, the binary representation of each particle is transformed into a preferred feature subset; 3) Calculate the fitness function value of each particle, evaluate the class separability of each particle representing the selected feature subset, and update the optimal feature position of each particle and the global optimal feature position, specifically: ,in: The quantification index of the class separability of the preferred feature subset. It serves as a quantitative indicator of the class separability of the remaining unselected feature subsets; 4) Update the particle's velocity and binary position; 5) Repeat steps 2 to 4 until the number of iterations reaches the set value; 6) Transform the binary representation of the global optimal particle position into the selected optimal feature subset.

5. The transmission system fault diagnosis method based on hybrid feature selection according to claim 4, characterized in that, The aforementioned categorical separability quantification index is obtained in the following way: a) Target labels for all samples Perform one-hot encoding to convert it into one-hot tag format. ; b) Calculate the features selected for each training sample The similarity with all training sample sets is used to find [the similarity] within the training sample set. Find the sample most similar to the training sample, count the number of each category in the neighboring samples, and thus estimate the posterior probability of each category for each sample. ,in As a category, For the number of categories, ; c) Class separability assessment based on cross-entropy, i.e. ,in: Indicates the first Sample No. Categories The ideal probability is obtained by calculating the average of the cross-entropy between the estimated class posterior probability of the sample set of the selected feature subset and the actual one-hot label, and then comprehensively evaluating the class separability of the selected feature subset.

6. The transmission system fault diagnosis method based on hybrid feature selection according to claim 1, characterized in that, The identification of transmission system fault types refers to: training a linear SVM model using a training set with optimized features, obtaining optimized features for the corresponding test set based on the feature selection results of the training set, and inputting the processed test set into the trained linear SVM to obtain the fault identification result.

7. A transmission system fault diagnosis system based on hybrid feature selection that implements the method of any one of claims 1-6, characterized in that, include: The system comprises a feature extraction unit, a preliminary feature selection unit, a feature optimization unit, and a fault diagnosis unit. Specifically: the feature extraction unit extracts features from single-channel vibration data in the time domain, frequency domain, and time-frequency domain to construct a holographic feature library for the transmission system; the preliminary feature selection unit uses a maximum correlation minimum redundancy feature selection algorithm to remove irrelevant and redundant features from the holographic feature library, resulting in a preliminary feature library for the transmission system; the feature optimization unit uses an improved encapsulated feature selection algorithm based on binary particle swarm optimization to further filter the preliminary feature library, obtaining the optimal feature subset for the transmission system; and the fault diagnosis unit inputs the feature-optimized test set into a trained SVM model to diagnose and identify faults in the transmission system.

Images