CNN-ISO rolling bearing fault diagnosis method based on double feature selection

Abstract

The invention provides a CNN-ISO rolling bearing fault diagnosis method based on double feature selection. Belongs to the field of rotary mechanical equipment fault diagnosis. In the first stage, a convolutional neural network (CNN) is utilized to automatically extract high-dimensional depth features from original vibration signals, subjectivity and complexity of artificial feature construction are avoided, and it is ensured that the feature extraction process is more objective and comprehensive; in the second stage, an improved snake swarm optimization (ISO) algorithm is combined with feature elimination and feature activation, synchronous optimization of feature subset selection and support vector machine (SVM) classifier parameters is realized, efficient global search is carried out in a feature space and a parameter space, and the collaborative improvement potential of optimal feature combination and classifier performance is fully mined. The method is superior to a traditional method in the aspects of accuracy, search efficiency and model lightweight, and can provide a thought for real-time and accurate diagnosis of the early failure of the industrial field bearing.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fault diagnosis of rotating machinery equipment, and particularly relates to a CNN-ISO rolling bearing fault diagnosis method based on double feature selection. BACKGROUND

[0002] Rolling bearings are the most common and high-failure-probability components in rotating machinery systems, and their degradation process is usually accompanied by complex impact and friction vibration signals. For the state monitoring of rolling bearings, there are currently two data-driven fault diagnosis paradigms: On the one hand, the diagnosis process based on shallow features and traditional machine learning first extracts statistical quantities, envelope spectrum indicators and empirical mode decomposition (EEMD) energy and other hand-crafted features in the time domain, frequency domain or time-frequency domain divided by humans, then performs feature selection or dimensionality reduction using principal component analysis (PCA), genetic algorithm (GA) or particle swarm optimization (PSO), and finally completes fault pattern recognition using a support vector machine (SVM), random forest (RF) or other classifier. However, this method has the following disadvantages: The feature design is highly dependent on expert experience, and different working conditions require repeated parameter adjustment for testing; there is redundancy and high correlation between a large number of hand-crafted features, and the subsequent feature optimization process has high computational overhead; in the early fault stage, weak impact fault features are easily masked by background noise, which may lead to missed detection. On the other hand, the end-to-end diagnosis process based on deep learning, such as convolutional neural networks (CNN), autoencoders (AE) and long short-term memory networks (LSTM), can directly model the original vibration sequence, to some extent, reducing human intervention. However, to maintain high expressiveness, such models usually contain a large number of parameters; on resource-constrained edge nodes, the storage requirements and computational load of such models increase significantly, resulting in high resource occupation. In addition, under high noise, sample imbalance or early subtle crack conditions, the generalization ability of deep models may decrease, and the inference time is difficult to meet the real-time warning requirements.

[0003] In summary, the existing methods have not achieved an ideal balance between feature adaptive selection and model efficiency. How to combine the discriminative ability of deep features with an adaptive feature selection mechanism, to fully utilize the high discriminative features extracted by deep learning, and to reduce model complexity and computational overhead, is still a problem to be solved in the field. SUMMARY

[0004] To solve the above problems, the application proposes a CNN-ISO rolling bearing fault diagnosis method based on double feature selection. The method includes two stages: in the first stage, a convolutional neural network (CNN) is used to automatically extract high-dimensional deep features from the original vibration signal, avoiding the subjectivity and complexity of manual feature construction, ensuring that the feature extraction process is more objective and comprehensive; in the second stage, an improved snake optimization algorithm (Improved Snake Optimization, ISO) is combined with feature elimination and feature activation to realize simultaneous optimization of feature subset selection and support vector machine (SVM) classifier parameters, and efficient global search is performed in the feature space and parameter space to fully exploit the collaborative improvement potential of the optimal feature combination and classifier performance.

[0005] To achieve the above purpose, the technical scheme adopted by the application is:

[0006] The CNN-ISO rolling bearing fault diagnosis method based on double feature selection includes the following steps:

[0007] Data import and segmentation: obtain the original signal data of bearing vibration, which can be collected by field sensors or imported from public vibration signal data sets. The obtained original vibration sequence is segmented into sample segments by sliding window with window length of 1024 or 2048 points, and the window overlap rate is 50%. No denoising / detrending filtering preprocessing is performed during the segmentation process; for ease of training, the amplitude is normalized to [-1, 1] by min-max normalization.

[0008] First stage: feature construction (CNN extraction): a convolutional neural network (CNN) model is constructed to automatically extract deep features. The CNN uses 3-5 residual blocks (preferably 4 residual blocks), and the convolution kernel size is 7, 5, and 3 in turn, and the convolution channel number is 32, 64, and 128 in turn. The step length of the first layer convolution is set to 2. Each sample segment is input into the CNN model to obtain a 256-dimensional high-level feature tensor; the feature tensor is compressed by global average pooling (GAP) and batch normalization (BN), and a preliminary feature vector is output, thereby realizing automatic extraction of vibration signal features and reducing dependence on manual features.

[0009] Second stage: feature selection and SVM parameter joint optimization (ISO): in this step, the snake optimization algorithm is combined with feature elimination and feature activation to realize collaborative selection and optimization of feature subsets and classifier parameters. Specifically, the following sub-steps are included:

[0010] a) Individual encoding: In the snake swarm optimization algorithm, each candidate solution individual is encoded into two parts: a feature selection mask and an SVM parameter vector. The feature selection mask is a 256-bit binary string that indicates whether the corresponding deep feature is selected; the SVM parameter vector contains the parameters of the support vector machine, such as the penalty coefficient. and kernel function parameters .

[0011] b) Fitness function: To measure the quality of individual features and parameters, a joint fitness function is constructed with diagnostic performance as the primary objective, while also introducing compression rate and regularization penalty terms. The fitness function is defined as follows: .in, For the model ; : The currently activated feature dimension; : original feature dimension; : Two parameters of SVM, Will Normalized to [0.1] on a logarithmic scale, is the trade-off coefficient (preferably ).

[0012] c) Feature elimination and feature activation: In each iteration of the snake swarm algorithm, two mutation operations, elimination phase and random activation phase, are performed on each candidate solution individual:

[0013] Elimination phase: Each currently selected feature is traversed, temporarily removed, and the individual's fitness is recalculated. If fitness does not deteriorate after removing a feature (i.e., the fitness value does not increase), the feature is confirmed for deletion (i.e., redundant features are eliminated). Only when the classification accuracy drops by no more than 0.2% after removing a feature is the feature finally eliminated, to avoid removing features that have a significant impact on accuracy.

[0014] Activation phase: Randomly enable previously unselected features (i.e., set some 0s in the mask to 1s), simultaneously fine-tune the parameters in the SVM parameter vector, and then recalculate the fitness. If the mutation results in an improvement in fitness (i.e., a decrease in fitness), the mutation is accepted, the corresponding feature is added to the feature subset, and the adjusted classifier parameters are retained. Feature activation is only performed if the fitness value decreases by at least 0.5% relative to the current fitness before the mutation, ensuring that the introduced new features and parameter adjustments have a significant performance gain.

[0015] The above-mentioned elimination stage and random activation stage are alternately executed, constituting a complete feature selection process. This strategy can avoid premature convergence of the algorithm to a local optimum, and perform a collaborative search in the feature selection space and the classifier parameter space, thereby balancing feature dimension compression and classification performance improvement. For the improved snake swarm optimization algorithm, the population size is preferably set to 40, and the maximum number of iterations is 150. The support vector machine classifier adopts a radial basis kernel function (RBF kernel), and the parameters are searched in the logarithmic space: , .

[0016] Model training and verification: according to the optimal feature subset obtained by CNN-ISO optimization, a radial basis kernel SVM fault diagnosis model is trained. Specifically, the extracted optimal feature subset is used as input, the SVM model is trained on the training set, and the classification accuracy and of the model are evaluated on the validation set using 5-fold cross-validation, and the time overhead of the model for fault identification classification on the test set is recorded.

[0017] Advantages of the present application:

[0018] Automatic feature extraction, reducing dependence on manual work: by introducing CNN, automatic deep feature extraction of the original vibration signal is realized, avoiding the subjectivity of manual feature design, eliminating the dependence on expert experience-driven feature construction, making feature extraction more objective and comprehensive, and being able to capture potential important patterns in complex signals.

[0019] Feature and parameter collaborative optimization, efficient and accurate model: through collaborative optimization of feature subset selection and classifier parameters, the model parameter size and computational overhead are greatly reduced while ensuring fault classification and discrimination ability. By using the improved snake swarm optimization algorithm, the synergistic improvement of optimal feature combination and classifier performance is fully explored. In the case of uneven distribution of fault categories, weak early fault signals and other difficult scenarios, the method of the present application can still maintain a high fault detection rate, reduce the occurrence of missed detection, and the model reasoning efficiency meets the requirements of real-time monitoring.

[0020] In summary, the present application combines the advantages of deep learning and intelligent optimization algorithm, realizes high-precision and high-efficiency diagnosis of rolling bearing faults, and has important engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a schematic diagram of the overall process of the method of the present application;

[0022] Figure 2 is a schematic diagram of the relationship between the number of selected features and

[0023] Figure 3 ​A comparison chart of the classification accuracy of the CNN-ISO and the LSTM, CNN, and CNN-LSTM methods;

[0024] Figure 4 A comparison chart of the classification accuracy of the CNN-ISO and the LSTM, CNN, and CNN-LSTM methods;

[0025] Figure 5 A comparison chart of the test time of the CNN-ISO and the LSTM, CNN, and CNN-LSTM methods. DETAILED DESCRIPTION

[0026] The application will be further described below in conjunction with the drawings and specific experimental data. It should be understood that the following examples are intended to help understand the principles and beneficial effects of the method of the application and are not intended to limit the protection scope of the application.

[0027] The CNN-ISO rolling bearing fault diagnosis method based on double feature selection comprises the following steps:

[0028] Data import and segmentation: Obtain the original signal data of bearing vibration, which can be collected by field sensors or imported from a public vibration signal dataset. The obtained original vibration sequence is segmented into sample segments by sliding window with a window length of 1024 or 2048 points, and the window overlap rate is 50%. No denoising / detrending filtering preprocessing is performed during the segmentation process; for ease of training, the amplitude is normalized to [-1, 1] by min-max normalization.

[0029] First stage: feature construction (CNN extraction): a convolutional neural network (CNN) model is constructed for automatic extraction of deep features. The CNN uses 3-5 residual blocks (preferably 4 residual blocks), and the convolution kernel size is 7, 5, and 3 in turn, and the convolution channel number is 32, 64, and 128 in turn. The step length of the first layer convolution is set to 2. Each sample segment is input into the CNN model to obtain a 256-dimensional high-level feature tensor; the feature tensor is compressed by global average pooling (GAP) and batch normalization (BN), and a preliminary feature vector is output, thereby realizing automatic extraction of vibration signal features and reducing dependence on artificial features.

[0030] Second stage: feature selection and SVM parameter joint optimization (ISO): in this step, the combination of the swarm optimization algorithm and feature elimination and feature activation is realized to achieve collaborative selection and optimization of the feature subset and the classifier parameters. Specifically, the following sub-steps are included.

[0031] ​a) Individual encoding: In the snake swarm optimization algorithm, each candidate solution individual is encoded into two parts: a feature selection mask and an SVM parameter vector. The feature selection mask is a 256-bit binary string that indicates whether the corresponding deep feature is selected; the SVM parameter vector contains the parameters of the support vector machine, such as the penalty coefficient. and kernel function parameters .

[0032] b) Fitness function: In order to measure the quality of individual features and parameters, a joint fitness function is constructed with diagnostic performance as the main goal, while introducing compression rate and regularization penalty terms. The fitness function is defined as follows: .in, For the model , : The currently activated feature dimension; : original feature dimension; : Two parameters of SVM, Will Normalized to [0.1] on a logarithmic scale, is the trade-off coefficient (preferably ).

[0033] c) Feature elimination and feature activation: In each iteration of the snake swarm algorithm, two mutation operations, elimination phase and random activation phase, are performed on each candidate solution individual:

[0034] Elimination phase: Each currently selected feature is traversed, temporarily removed, and the individual's fitness is recalculated. If fitness does not deteriorate after removing a feature (i.e., the fitness value does not increase), the feature is confirmed for deletion (i.e., redundant features are eliminated). Only when the classification accuracy drops by no more than 0.2% after removing a feature is the feature finally eliminated, to avoid removing features that have a significant impact on accuracy.

[0035] Activation phase: Randomly enable previously unselected features (i.e., set some 0s in the mask to 1s), simultaneously fine-tune the parameters in the SVM parameter vector, and then recalculate the fitness. If the mutation results in an improvement in fitness (i.e., a decrease in fitness), the mutation is accepted, the corresponding feature is added to the feature subset, and the adjusted classifier parameters are retained. Feature activation is only performed if the fitness value decreases by at least 0.5% compared to before the mutation, ensuring that the introduced new features and parameter adjustments have a significant performance gain.

[0036] The above-mentioned elimination stage and random activation stage are alternately executed, constituting a complete feature selection strategy. This strategy can avoid premature convergence of the algorithm to a local optimum, and perform collaborative search in the feature selection space and the classifier parameter space, thereby balancing feature dimension compression and classification performance improvement. For the improved snake swarm optimization algorithm, the population size is preferably set to 40 and the maximum number of iterations is 150 in the present application. The support vector machine classifier adopts a radial basis kernel function (RBF kernel), and the parameters are searched in the logarithmic space: , .

[0037] Model training and verification: According to the optimal feature subset obtained by CNN-ISO optimization, a radial basis kernel SVM fault diagnosis model is trained. Specifically, the extracted optimal feature subset is used as input, the SVM model is trained on the training set, and the classification accuracy and of the model are evaluated on the validation set using 5-fold cross-validation. At the same time, the time overhead of the model for fault identification classification on the test set is recorded.

[0038] Example 1: CWRU dataset experimental verification

[0039] Data import and segmentation: In this embodiment, the method of the present application is experimentally verified using the bearing fault dataset of the Case Western Reserve University (CWRU) in the United States. The CWRU dataset includes normal and different fault types of rolling bearing vibration signals, and the fault types cover normal, inner ring, rolling body and outer ring. The driving end vibration signal sampling rate is 48 kHz. The dataset is processed as follows: first, the original vibration sequence is cut into sliding windows with a window length of 2048 points, and the window overlap rate is set to 50%. The data is divided into training set, validation set and test set, with a ratio of 6:2:2. The sample size of each dataset after segmentation is shown in Table 1. No filtering preprocessing such as denoising / detrending is performed during the segmentation process; for ease of training, the amplitude is normalized to [-1, 1] by min-max normalization.

[0040] Table 1: Different fault types and dataset division (number of samples per type)

[0041] The first stage is feature construction (CNN extraction): a convolutional neural network (CNN) model is constructed for automatically extracting deep features. The CNN adopts 3-5 residual blocks (preferably 4 residual blocks), and the convolution kernel size is 7, 5 and 3 in sequence, and the convolution channel number is 32, 64 and 128 in sequence. The step length of the first layer convolution is set to 2. Each sample segment is input into the CNN model to obtain a 256-dimensional high-level feature tensor; the feature tensor is compressed through global average pooling (GAP) and batch normalization (BN) to output a preliminary feature vector, so as to realize automatic extraction of vibration signal features and reduce the dependence on artificial features.

[0042] The second stage is feature selection and SVM parameter joint optimization (ISO): the population size of the snake swarm optimization algorithm is set to 40, and the maximum iteration number is 150. The 256-dimensional feature vector extracted above is subjected to feature subset selection by using the CNN-ISO algorithm, and the parameters of the SVM classifier are simultaneously optimized. After iterative search, the algorithm outputs an optimal feature subset with a dimension of 48, which accounts for only 18.75% of the original feature dimension, and significantly reduces the feature dimension. At the same time, the optimal parameter combination of the SVM is obtained, and the parameter , .

[0043] Model training and verification: according to the optimal 48-dimensional feature subset and the corresponding optimal parameters output by the CNN-ISO, a radial basis kernel function SVM classifier model is trained on the training set. Then, the test set is subjected to fault classification test. The classification accuracy and of the model are evaluated by using 5-fold cross-validation, and the average time overhead of the model for reasoning (fault diagnosis) on the test data is recorded. Among them, Figure 2 is the relationship between the number of selected features and in 5-fold cross-validation. Figure 3 、 Figure 4 and Figure 5 respectively give the accuracy, macro average F1 value and test time of the method (CNN-ISO) of the present application and the comparative method on the test set. Compared with the traditional LSTM, pure CNN model and hybrid model (CNN-LSTM) combining CNN and LSTM, the method of the present application has the optimal performance in the classification accuracy, and model test time overhead and other indicators.

[0044] It should be emphasized that the embodiments of the present application are illustrative rather than limiting, and therefore the present application includes but is not limited to the embodiments described in the specific embodiments, and any other embodiments derived from the technical solutions of the present application by those skilled in the art also belong to the scope of protection of the present application.

Claims

1. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection is characterized by: The method includes: Step 1: Data import and segmentation: obtaining the original bearing vibration signal, performing sliding window segmentation according to a preset window length, and performing amplitude normalization; Step 2: Phase 1: Feature Construction: Input the sample segment into the convolutional neural network and output a fixed-dimensional deep feature vector; Step 3: Second stage: Feature selection and SVM parameter optimization. The improved snake swarm optimization algorithm is used to perform feature subset selection on the deep features in step 2 and optimize the SVM parameters simultaneously. The candidate solution consists of a binary feature mask and a parameter vector. Step 4: Training and diagnosis: Based on the optimal feature subset and optimal parameters obtained in step 3, the radial basis kernel function SVM classifier is trained and the fault category is output; Among them, in step 3, a joint fitness function is constructed with classification performance as the main term, and a penalty term related to the feature compression rate and a regularization term for the parameter amplitude are introduced to achieve the coordinated optimization of diagnostic performance and feature dimension compression.

2. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The sliding window length is 2048, the window overlap rate is 50%, and the amplitude is normalized to [-1, 1] through min-max.

3. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The convolutional neural network contains 3 residual blocks, the convolution kernel sizes are 7, 5, and 3 respectively, the number of convolution channels are 32, 64, and 128 respectively, the first layer convolution step size is 2, and the depth feature vector is obtained by global average pooling and batch normalization in the last layer.

4. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The length of the binary feature mask in step 3 is consistent with the output feature dimension of step 2; the SVM parameters include the penalty coefficient and kernel function parameters .

5. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The feature selection includes two types of mutation operations: feature elimination and feature activation: temporarily removing the currently selected features one by one and recalculating the fitness to perform elimination; randomly activating at least one feature in the unselected features and synchronously fine-tuning and to perform activation; accept the corresponding mutation when the preset criteria are met.

6. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The criterion for feature elimination is that the classification accuracy will not drop by more than 0.2% of the current accuracy if a single feature is deleted, and the feature will be eliminated to avoid removing features that have a large impact on the accuracy at one time.

7. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The criterion for feature activation is: the feature activation operation is accepted only when the fitness value decreases by at least 0.5% of the current fitness value relative to the value before the mutation, so as to ensure that the introduced new features and parameter adjustments have obvious performance gains.

8. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The joint fitness function is defined as: ,in, For the model ; : The currently activated feature dimension; : original feature dimension; : Two parameters of SVM, Will Normalized to [0,1] on a logarithmic scale, is a trade-off coefficient; the trade-off coefficient satisfies .

9. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: right The search is performed in the logarithmic scale space and normalized, and the SVM uses the radial basis kernel.

10. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1 is characterized in that: The population size of the ISO is 40 and the maximum number of iterations is 150.

Images