CNN-ISO Rolling Bearing Fault Diagnosis Method Based on Dual Feature Selection

Abstract

This invention provides a CNN-ISO rolling bearing fault diagnosis method based on dual feature selection, belonging to the field of rotating machinery equipment fault diagnosis. The first stage utilizes a convolutional neural network (CNN) to automatically extract high-dimensional deep features from the original vibration signal, avoiding the subjectivity and complexity of manual feature construction and ensuring a more objective and comprehensive feature extraction process. The second stage combines an improved snake optimization (ISO) algorithm with feature removal and feature activation to achieve simultaneous optimization of feature subset selection and support vector machine (SVM) classifier parameters. It performs efficient global search in both the feature space and parameter space, fully exploring the potential for synergistic improvement in optimal feature combinations and classifier performance. This invention outperforms traditional methods in terms of accuracy, search efficiency, and model lightweighting, providing a new approach for real-time and accurate diagnosis of early bearing faults in industrial settings.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis of rotating machinery and equipment, and specifically relates to a CNN-ISO rolling bearing fault diagnosis method based on dual feature selection. Background Technology

[0002] Rolling bearings are among the most common components in rotating machinery systems and have a high failure probability. Their degradation process is usually accompanied by complex impact and frictional vibration signals. Currently, there are two main data-driven fault diagnosis paradigms for rolling bearing condition monitoring: One approach is based on shallow features and traditional machine learning. This involves first extracting manual features such as statistics, envelope spectrum indices, and Empirical Mode Decomposition (EEMD) energy from manually defined time, frequency, or time-frequency domains. Then, methods like Principal Component Analysis (PCA), Genetic Algorithm (GA), or Particle Swarm Optimization (PSO) are used for feature selection or dimensionality reduction. Finally, classifiers such as Support Vector Machine (SVM) or Random Forest (RF) are used to complete fault mode identification. However, this method has the following drawbacks: feature design is highly dependent on expert experience, requiring repeated parameter adjustments and experiments for different operating conditions; many manual features are redundant and highly correlated, leading to high computational costs in subsequent feature optimization; and in the early stages of faults, weak impact fault features are easily masked by background noise, potentially resulting in missed detections. On the other hand, end-to-end diagnostic processes based on deep learning, such as convolutional neural networks (CNNs), autoencoders (AEs), and long short-term memory networks (LSTMs), can directly model the original vibration sequences, reducing the need for manual intervention to some extent. However, to maintain high expressive power, these models typically contain a large number of parameters; on resource-constrained edge nodes, the storage requirements and computational load of such models increase significantly, leading to excessive resource consumption. Furthermore, under conditions such as high noise, imbalanced samples, or early micro-cracks, the generalization ability of deep models may decrease, and the inference time may be insufficient to meet the real-time early warning requirements.

[0003] In summary, existing methods have not yet achieved an ideal balance between adaptive feature selection and model efficiency. How to combine the discriminative power of deep features with adaptive feature selection mechanisms to fully utilize the high-discriminative features extracted by deep learning while reducing model complexity and computational overhead remains a pressing issue in this field. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a CNN-ISO rolling bearing fault diagnosis method based on dual feature selection. This method comprises two stages: the first stage utilizes a convolutional neural network (CNN) to automatically extract high-dimensional deep features from the original vibration signal, avoiding the subjectivity and complexity of manual feature construction and ensuring a more objective and comprehensive feature extraction process; the second stage combines an improved snake optimization (ISO) algorithm with feature removal and feature activation to achieve simultaneous optimization of feature subset selection and support vector machine (SVM) classifier parameters, performing efficient global search in both the feature space and parameter space to fully exploit the potential for synergistic improvement in optimal feature combinations and classifier performance.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

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

[0007] Data Import and Segmentation: Obtain the raw signal data of bearing vibration, which can be acquired through field sensors or imported from publicly available vibration signal datasets. The obtained raw vibration sequence is segmented into sample segments using a sliding window with a window length of 1024 or 2048 points, with a window overlap rate of 50%. No denoising / detrending filtering preprocessing is performed during segmentation; to facilitate training, the amplitude is min-max normalized to [-1, 1].

[0008] Phase 1: 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), with kernel sizes of 7, 5, and 3, and convolutional channel numbers of 32, 64, and 128, respectively. The stride of the first convolutional layer 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 using Global Average Pooling (GAP) and Batch Normalization (BN) to output a preliminary feature vector, thereby achieving automatic extraction of vibration signal features and reducing reliance on manual features.

[0009] Phase Two: Feature Selection and Joint Optimization of SVM Parameters (ISO): In this step, the snake swarm optimization algorithm is combined with feature culling and feature activation to achieve collaborative selection and optimization of feature subsets and classifier parameters. Specifically, it includes the following sub-steps:

[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 used to indicate whether a corresponding depth 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 feature-parameter individuals, a joint fitness function is constructed, with diagnostic performance as the primary objective, while also incorporating compression ratio and regularization penalty terms. The fitness function is defined as follows: .in, For the model ; The currently activated feature dimension; Original feature dimension; The two parameters of SVM Will Normalized to [0.1] on a logarithmic scale. For the trade-off coefficient (preferred) ).

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

[0013] Elimination Phase: Each currently selected feature is iterated through, and the individual's fitness is recalculated after temporarily removing that feature. If the fitness does not worsen after removing a feature (i.e., the fitness value does not increase), then the feature is confirmed for deletion (i.e., redundant features are removed). The feature is only ultimately accepted for deletion if the classification accuracy decreases by no more than 0.2% of the current accuracy after deletion, to avoid removing features that have an excessive impact on accuracy all at once.

[0014] Activation Phase: Previously unselected features are randomly activated (i.e., some 0s in the mask are set to 1s), while the parameters in the SVM parameter vector are simultaneously fine-tuned, and the fitness is recalculated. If the mutation operation improves the fitness (i.e., the fitness value decreases), the mutation is accepted, the corresponding feature is added to the feature subset, and the adjusted classifier parameters are retained. Feature activation is only accepted when the fitness value decreases by at least 0.5% relative to the pre-mutation fitness, to ensure that the introduced new features and parameter adjustments have a significant performance gain.

[0015] The aforementioned elimination and random activation phases are executed alternately in a loop, constituting a complete feature selection process. This strategy avoids premature convergence to local optima and performs a collaborative search in both the feature selection space and the classifier parameter space, thus balancing feature dimensionality 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 to 150 in this invention. The support vector machine classifier uses a radial basis function (RBF) kernel, whose parameters are searched in logarithmic space: , .

[0016] Model Training and Validation: A radial basis function kernel (RBM) SVM fault diagnosis model is trained using the optimal feature subset obtained through CNN-ISO optimization. Specifically, the extracted optimal feature subset is used as input to train the SVM model on the training set, and 5-fold cross-validation is used on the validation set to evaluate the model's classification accuracy and performance. At the same time, the time cost of the model to identify and classify faults on the test set is recorded.

[0017] The beneficial effects of this invention are:

[0018] Automated feature extraction reduces reliance on manual intervention: By introducing CNNs, automatic deep feature extraction of the original vibration signal is achieved, avoiding the subjectivity of manual feature design and eliminating the reliance on expert experience-driven feature construction. This makes feature extraction more objective and comprehensive, and can capture potential important patterns in complex signals.

[0019] Feature and parameter co-optimization leads to a highly efficient and accurate model: By co-optimizing feature subset selection and classifier parameters, the model parameter size and computational overhead are significantly reduced while maintaining fault classification and discrimination capabilities. An improved snake swarm optimization algorithm is utilized to fully explore the synergistic improvement between optimal feature combinations and classifier performance. Even in challenging scenarios such as uneven fault category distribution and weak early fault signals, the method of this invention maintains a high fault detection rate, reduces missed detections, and ensures that the model's inference efficiency meets the requirements of real-time monitoring.

[0020] In summary, this invention combines the advantages of deep learning and intelligent optimization algorithms to achieve high-precision and high-efficiency diagnosis of rolling bearing faults, and has significant engineering application value. Attached Figure Description

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

[0022] Figure 2 The number of selected features and A diagram illustrating the relationship between them;

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

[0024] Figure 4 For CNN-ISO and LSTM, CNN, CNN-LSTM methods Comparison chart;

[0025] Figure 5 This is a comparison chart of the test times for CNN-ISO, LSTM, CNN, and CNN-LSTM methods. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and specific experimental data. It should be understood that the following embodiments are intended to help understand the principle and beneficial effects of the method of the present invention, and are not intended to limit the scope of protection of the present invention.

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

[0028] Data Import and Segmentation: Obtain the raw signal data of bearing vibration, which can be acquired through field sensors or imported from publicly available vibration signal datasets. The obtained raw vibration sequence is segmented into sample segments using a sliding window with a window length of 1024 or 2048 points, with a window overlap rate of 50%. No denoising / detrending filtering preprocessing is performed during segmentation; to facilitate training, the amplitude is min-max normalized to [-1, 1].

[0029] Phase 1: 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), with kernel sizes of 7, 5, and 3, and convolutional channel numbers of 32, 64, and 128, respectively. The stride of the first convolutional layer 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 using Global Average Pooling (GAP) and Batch Normalization (BN) to output a preliminary feature vector, thereby achieving automatic extraction of vibration signal features and reducing reliance on manual features.

[0030] Phase Two: Feature Selection and Joint Optimization of SVM Parameters (ISO): In this step, the snake swarm optimization algorithm is combined with feature removal and feature activation to achieve collaborative selection and optimization of feature subsets and classifier parameters. This includes the following sub-steps.

[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 used to indicate whether a corresponding depth 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: To measure the quality of feature-parameter individuals, a joint fitness function is constructed, with diagnostic performance as the primary objective, while also incorporating compression ratio and regularization penalty terms. The fitness function is defined as follows: .in, For the model , The currently activated feature dimension; Original feature dimension; The two parameters of SVM Will Normalized to [0.1] on a logarithmic scale. For the trade-off coefficient (preferred) ).

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

[0034] Elimination Phase: Each currently selected feature is iterated through, and the individual's fitness is recalculated after temporarily removing the feature. If the fitness does not worsen after removing a feature (i.e., the fitness value does not increase), the feature is confirmed for deletion (i.e., redundant features are removed). The feature is only ultimately accepted for deletion if the classification accuracy decreases by no more than 0.2% after deletion, to avoid removing features that have a significant impact on accuracy all at once.

[0035] Activation Phase: Previously unselected features are randomly activated (i.e., some 0s in the mask are set to 1s), while the parameters in the SVM parameter vector are simultaneously fine-tuned, and the fitness is recalculated. If the mutation operation improves the fitness (i.e., the fitness value decreases), the mutation is accepted, the corresponding feature is added to the feature subset, and the adjusted classifier parameters are retained. Feature activation is only accepted when the fitness value decreases by at least 0.5% relative to before the mutation, to ensure that the introduced new features and parameter adjustments have a significant performance gain.

[0036] The aforementioned elimination phase and random activation phase are executed alternately and cyclically, constituting a complete feature selection strategy. This strategy avoids the algorithm from prematurely converging to local optima and performs a collaborative search in the feature selection space and classifier parameter space, thus balancing feature dimensionality 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 to 150 in this invention. The support vector machine classifier uses a radial basis function (RBF kernel), whose parameters are searched in logarithmic space: , .

[0037] Model Training and Validation: A radial basis function kernel (RBM) SVM fault diagnosis model is trained using the optimal feature subset obtained through CNN-ISO optimization. Specifically, the extracted optimal feature subset is used as input to train the SVM model on the training set, and 5-fold cross-validation is used on the validation set to evaluate the model's classification accuracy and performance. At the same time, the time cost of the model to identify and classify faults on the test set is recorded.

[0038] Example 1: Experimental Validation using the CWRU dataset

[0039] Data Import and Segmentation: This embodiment uses the Case Western Reserve University (CWRU) bearing fault dataset to experimentally verify the method of the present invention. The CWRU dataset includes vibration signals of rolling bearings under normal conditions and different fault types, covering normal, inner ring, rolling element, and outer ring faults. The sampling rate of the drive end vibration signal is 48kHz. The dataset is processed as follows: First, the original vibration sequence is segmented into a sliding window with a window length of 2048 points, and the window overlap rate is set to 50%. The data is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The number of samples in each dataset after segmentation is shown in Table 1. No denoising / detrending filtering preprocessing is performed during the segmentation process; to facilitate training, the amplitude is min-max normalized to [-1,1].

[0040] Table 1. Different Fault Types and Dataset Division (Number of Samples per Class)

[0041]

[0042] Phase 1: 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), with kernel sizes of 7, 5, and 3, and convolutional channel numbers of 32, 64, and 128, respectively. The stride of the first convolutional layer 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 using Global Average Pooling (GAP) and Batch Normalization (BN) to output a preliminary feature vector, thereby achieving automatic extraction of vibration signal features and reducing reliance on manual features.

[0043] Phase Two: Feature Selection and Joint Optimization of SVM Parameters (ISO): The snake swarm optimization algorithm population size was set to 40, and the maximum number of iterations was set to 150. The CNN-ISO algorithm was used to select a subset of features from the extracted 256-dimensional feature vectors, while simultaneously optimizing the parameters of the SVM classifier. After iterative search, the algorithm output the optimal feature subset with a dimension of 48, accounting for only 18.75% of the original feature dimension, significantly reducing the feature dimensionality. Simultaneously, the optimal combination of SVM parameters was obtained. , .

[0044] Model training and validation: Based on the optimal 48-dimensional feature subset and corresponding optimal parameters output by CNN-ISO. A radial basis function kernel SVM classifier model was trained on the training set. Then, a fault classification test was performed on the test set. Five-fold cross-validation was used to evaluate the model's classification accuracy and performance. It also records the average time cost of the model performing inference (fault diagnosis) on the test data. Figure 2 The number of features selected in 5-fold cross-validation and The relationship between them. Figure 3 , Figure 4 and Figure 5 The results of accuracy, macro-average F1 score, and test time of the proposed method (CNN-ISO) and the comparison method on the test set are presented respectively. Compared with traditional LSTM, pure CNN models, and hybrid models combining CNN and LSTM (CNN-LSTM), the proposed method achieves higher classification accuracy, It achieved optimal performance in terms of metrics such as model testing time overhead.

[0045] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A CNN-ISO rolling bearing fault diagnosis method based on dual feature selection, characterized in that, Includes: Step 1: Data import and segmentation: Obtain the original bearing vibration signal, perform sliding window segmentation according to the preset window length, and perform amplitude normalization; Step 2: First stage: 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 joint optimization. An improved snake swarm optimization algorithm is used to perform feature subset selection on the deep features in step 2 and optimize them in sync with the SVM parameters. 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, train a radial basis function SVM classifier and output the fault category; In step 3, a joint fitness function is constructed, with classification performance as the main term, and a penalty term related to feature compression rate and a regularization term for parameter magnitude are introduced to achieve the synergistic optimization of diagnostic performance and feature dimension compression. The joint fitness function is defined as follows: ,in, For the model ; The currently activated feature dimension; Original feature dimension; The two parameters of SVM Will Normalized to [0,1] on a logarithmic scale. These are the weighting coefficients; the weighting coefficients satisfy... .

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

3. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1, characterized in that, The convolutional neural network contains three residual blocks with kernel sizes of 7, 5, and 3, and convolutional channels of 32, 64, and 128, respectively. The first layer has a stride of 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, characterized in that, The length of the binary feature mask in step 3 is consistent with the dimension of the output feature in step 2; the SVM parameters include penalty coefficients. With kernel function parameters .

5. The CNN-ISO rolling bearing fault diagnosis method based on dual feature selection according to claim 1, characterized in that, The feature selection includes two types of mutation operations: feature removal and feature activation. The selected features are temporarily removed one by one, and their fitness is recalculated to perform removal. At least one feature is randomly activated from the unselected features, and fine-tuning is performed simultaneously. and Activate by performing the action; 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, characterized in that, The criteria for feature removal are: the removal of a single feature will only be accepted if the decrease in classification accuracy caused by deleting a single feature does not exceed 0.2% of the current accuracy, in order to avoid removing features that have too great an impact on accuracy at once.

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

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

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

Images