Small sample fault diagnosis method for rolling bearing

Abstract

This invention discloses a method for small-sample fault diagnosis of rolling bearings, comprising the following steps: S10 Data acquisition and preprocessing: for acquiring preprocessed data; S20 Gradient regularization generative adversarial network data enhancement: inputting the preprocessed data into a gradient regularization generative adversarial network to acquire enhanced samples. The discriminator introduces a gradient penalty mechanism to constrain its own gradient norm when distinguishing between real samples and generated samples, thereby expanding the small-sample dataset; S30 Multi-scale feature extraction and temporal modeling: inputting the enhanced samples into a one-dimensional convolutional neural network to extract local high-frequency impact features and low-frequency resonance features of the vibration signal, and then inputting the features output by the one-dimensional convolutional neural network into a bidirectional long short-term memory network to capture the long-range temporal dependencies of the features, and outputting a fused feature vector; S40 Fault classification decision: inputting the fused feature vector into a fully connected classification layer, and combining a confidence check mechanism to determine the fault type of the rolling bearing.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and specifically to a small-sample fault diagnosis method for rolling bearings based on a hybrid model. Background Technology

[0002] Rolling bearings, as core transmission components of rotating machinery, are widely used in high-end equipment such as wind turbine generators, rail transit power plants, precision industrial robots, and cigarette packaging machines. Statistics show that bearing-related failures account for nearly 40% of all failure cases in rotating machinery, resulting in unplanned downtime causing global industrial economic losses exceeding tens of billions of US dollars annually. Therefore, establishing an online monitoring and fault diagnosis system for the operating status of rolling bearings is of great significance for ensuring reliable equipment operation and optimizing full life-cycle management.

[0003] Existing rolling bearing fault diagnosis technologies are mainly divided into two categories: signal analysis-based and data-driven. Signal analysis-based methods (such as Fourier transform, wavelet transform, and empirical mode decomposition) rely on manually designed features, which suffer from insufficient time-frequency resolution and weak noise resistance in non-stationary signal processing. For example, empirical mode decomposition (EMD) is susceptible to mode aliasing and endpoint effects, and its signal-to-noise ratio improvement in fault feature extraction is less than 10% in noisy environments. In data-driven methods, traditional machine learning models (such as support vector machines and decision trees) require large amounts of labeled data and have poor adaptability to small sample sizes and class imbalance scenarios. While deep learning models (such as CNN and LSTM) can automatically extract features, they are prone to overfitting under small sample conditions and have limited generalization ability across different operating conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a small-sample fault diagnosis method for rolling bearings based on a hybrid model, so as to solve the technical problems of existing bearing fault diagnosis technology being prone to overfitting under small sample conditions, having weak generalization ability across operating conditions, and having insufficient noise resistance.

[0005] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0006] A method for small-sample fault diagnosis of rolling bearings based on a hybrid model includes the following steps: S10: Data Acquisition and Preprocessing The original vibration signal of the rolling bearing was collected by an accelerometer, high-frequency noise was removed by bandpass filtering, a fixed-length time domain segment was extracted, and then the preprocessed data was obtained by standardization. S20: Gradient Regularization for Generative Adversarial Network Data Augmentation The preprocessed data is input into the gradient regularization generative adversarial network. The generator of the adversarial network generates enhanced samples that are consistent with the distribution of real fault signals. When the discriminator distinguishes between real samples and generated samples, a gradient penalty mechanism is introduced to constrain its own gradient norm, thereby expanding the small sample dataset. S30: Multi-scale Feature Extraction and Temporal Modeling The enhanced samples are input into a one-dimensional convolutional neural network to extract local high-frequency impact features and low-frequency resonance features of the vibration signal. The features output by the one-dimensional convolutional neural network are then input into a bidirectional long short-term memory network to capture the long-range temporal dependencies of the features and output a fused feature vector. S40: Fault Classification Decision The fused feature vectors are input into a fully connected classification layer, and the failure type of the rolling bearing is determined by combining the confidence verification mechanism.

[0007] The beneficial effects of this invention are as follows: This invention addresses the technical problems of existing bearing fault diagnosis technologies, such as overfitting under small sample conditions, weak generalization ability across operating conditions, and insufficient noise resistance, by integrating the data augmentation capabilities of gradient-regularized generative adversarial networks (DRAGAN), the multi-scale feature extraction capabilities of one-dimensional convolutional neural networks (CNN), and the temporal dependency modeling capabilities of bidirectional long short-term memory networks (Bi-LSTM). It demonstrates the following technical advantages: 1. Excellent diagnostic performance with small sample size: The present invention achieves a classification accuracy of 78.9% for rolling bearing fault types with an extremely small sample size (5% training set), which is 23.1%, 20.8%, and 16.6% higher than traditional CNN (55.8%), Bi-LSTM (58.1%), and CNN-LSTM (62.3%), respectively. The accuracy reaches 86.2% with a 10% training set (small sample size), 92.7% with a 20% training set (medium sample size), and 95.6% with a 100% training set (full sample size). It can approach saturation performance with a medium sample size, and the data efficiency is significantly better than existing models.

[0008] 2. Strong generalization ability across operating conditions: In the cross-load condition migration task, the accuracy reaches 89.7±0.8% when operating condition A→B, which is 8.3% higher than CNN-LSTM (81.4±1.1%); when operating condition A→C, it reaches 85.1±0.9%, which is 10.9% higher than CNN-LSTM (74.2±1.3%); and when operating condition B→C, it reaches 87.6±0.7%. The fault characteristics maintain high cohesion and low coupling under the target operating condition, which is suitable for complex industrial scenarios with variable speed and variable load.

[0009] 3. Outstanding noise resistance: Even in a high-noise environment (SNR=-4dB), the accuracy reaches 68.7% with an extremely small sample size (5% of the training set), and the noise attenuation rate (NAR) is 87.1%, which is 26.6% higher than that of traditional CNN (42.1%). In a medium-noise environment (SNR=0dB), the NAR reaches 95.7% for the entire sample, which is close to the performance under noise-free conditions (98.4%). When SNR=10dB (weak noise), the accuracy reaches 93.8% for an extremely small sample size and 98.4% for the entire sample size, indicating that the present invention has strong adaptability to dynamic noise environments.

[0010] 4. High-quality generated samples: The gradient penalty mechanism of DRAGAN (penalty coefficient 10) used results in an FID score as low as 23.5 for the generated samples, and the model training loss variance is only 1.4 × 10⁻⁶. -3 This approach ensures both sample diversity and distribution accuracy, effectively mitigating overfitting issues in small sample scenarios and providing high-quality data support for subsequent feature extraction and classification. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the structure of the hybrid model established in an embodiment of the present invention; Figure 2 This is a flowchart of a small-sample fault diagnosis method for rolling bearings according to an embodiment of the present invention; Figure 3 This refers to the average classification accuracy of each model under different training sample sizes in the embodiments of the present invention. Figure 4 This is the confusion matrix of the fault classification accuracy of each model under extremely small sample conditions in the embodiments of the present invention; Figure 5 This is a schematic diagram comparing the diagnostic performance of a small sample across different operating conditions according to an embodiment of the present invention; Figure 6 This is the confusion matrix of the fault diagnosis accuracy of each model under operating conditions A→B in the embodiments of the present invention; Figure 7 This is a schematic diagram of cross-condition fault feature distribution visualization (t-SNE) according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the NAR curves of the mixture model under different sample sizes in an embodiment of the present invention; Figure 9This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Furthermore, the following description is for illustrative purposes and not for limitation, and sets forth specific details such as particular system architectures and techniques to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems and methods are omitted to avoid unnecessary detail that could obscure the description of the invention.

[0015] This invention provides a method for small-sample fault diagnosis of rolling bearings based on a hybrid model. The established DRAGAN-CNN-LSTM hybrid model is a combination of Gradient Regularized Generative Adversarial Network (DRAGAN), Convolutional Neural Network (CNN), and Bidirectional Long Short-Term Memory Network (Bi-LSTM). The core structure of this hybrid model consists of a DRAGAN data augmentation module, a multi-scale CNN feature extraction module, a Bidirectional LSTM temporal modeling module, and a classification decision module, as detailed below. Figure 1 As shown.

[0016] The DRAGAN data augmentation module achieves multi-scale feature extraction and temporal dependency modeling through a cascaded one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (Bi-LSTM). The one-dimensional convolutional neural network (1D-CNN) consists of a first convolutional layer (64 channels, kernel size 7, stride 1, using ReLU activation function, output feature dimension 2000×64), a max-pooling layer (window size 3, stride 3, downsampled to 666×64), and a second convolutional layer (128 channels, kernel size 5, output feature dimension 666×128). The bidirectional long short-term memory network (Bi-LSTM) layer adopts a bidirectional structure, with 128 LSTM units in each direction and a dropout rate of 0.3 between layers. Finally, a 256-dimensional temporal feature vector is output through bidirectional feature concatenation. The parameter design of each module is shown in Table 1. Table 1. Parameter Configuration Table for DRAGAN-CNN-LSTM Hybrid Model

[0017] Based on this, refer to Figure 2 The rolling bearing small-sample fault diagnosis method based on a hybrid model described in this embodiment specifically includes the following steps: S10: Data Acquisition and Preprocessing: The original vibration signal of the rolling bearing is acquired using an accelerometer, high-frequency noise is removed by bandpass filtering, a fixed-length time domain segment is extracted, and then the preprocessed data is obtained through standardization. S20: Gradient Regularized Generative Adversarial Network Data Augmentation: Preprocessed data is input into a gradient regularized generative adversarial network. The generator of the adversarial network generates augmented samples that are consistent with the distribution of real fault signals. When the discriminator distinguishes between real samples and generated samples, a gradient penalty mechanism is introduced to constrain its own gradient norm, thereby expanding the small sample dataset. S30: Multi-scale feature extraction and temporal modeling: The enhanced sample is input into a one-dimensional convolutional neural network to extract the local high-frequency impact features and low-frequency resonance features of the vibration signal. The features output by the one-dimensional convolutional neural network are then input into a bidirectional long short-term memory network to capture the long-range temporal dependencies of the features and output a fused feature vector. S40: Fault Classification Decision: Input the fused feature vector into the fully connected classification layer and combine it with the confidence check mechanism to determine the fault type of the rolling bearing: The Top-2 confidence check mechanism is adopted. If the maximum probability is >0.9 or the difference between the first two probabilities is >0.3, it is determined to be the corresponding fault category; otherwise, manual review is triggered.

[0018] To verify the rationality of the key hyperparameter settings in the DRAGAN-CNN-LSTM hybrid model, this embodiment uses the cigarette packaging machine paper feed roller bearing dataset (see Table 2) to conduct optimization experiments on the gradient penalty coefficient (λ), generator learning rate (lrg), and Dropout rate (p) using the control variable method.

[0019] Table 2 Composition and Division of Bearing Dataset

[0020] By comparing classification accuracy, generated sample quality, and training stability under different parameter combinations, the optimality of the model parameters is verified. The parameter optimization experiment is set as follows: Gradient penalty coefficient (λ): Choose λ=5, 10, 15, and fix other parameters (lrg=2e). 4 (p=0.3), to evaluate the impact of generated sample diversity on classification performance; Generator learning rate (lrg): Compared to lrg=1e 4 ,2e 4 5e 4With λ=10 and p=0.3 fixed, the stability of the training was analyzed. Dropout rate (p): Tested 4p=0.2, 0.3, 0.4, fixed λ=10, lrg=2e 4 To verify the effect of regularization.

[0021] The verification results are shown in Table 3.

[0022] Table 3 shows the experimental results of hyperparameter optimization for the DRAGAN-CNN-LSTM hybrid model.

[0023] Table 3 shows that when λ=10, the classification accuracy reaches 95.6%, an improvement of 6.4% and 4.2% compared to λ=5 and λ=15, respectively. Simultaneously, the FID score decreases to 23.5, and the loss variance is 1.4 × 10⁻⁶. -3 This indicates that a moderate gradient constraint can balance the diversity of generated samples with the realism of the distribution, avoiding overfitting or underfitting; lrg=2e 4 The model performs optimally at that time, with accuracy comparable to lrg=1e. 4 This resulted in a 3.5% improvement and a 48.1% reduction in loss variance. The excessively high learning rate (lrg=5e) was the cause. 4 This leads to the generator updating too quickly, resulting in a decrease in the quality of generated samples (FID Score=52.1); the model has the best generalization ability when p=0.3, but excessive regularization (p=0.4) inhibits the feature expression ability and reduces the accuracy by 4.9%; while excessive regularization (p=0.2) increases the risk of overfitting due to underregulation.

[0024] Subsequently, this embodiment verifies the fault diagnosis performance of rolling bearings in a small sample scenario based on the optimized DRAGAN-CNN-LSTM model.

[0025] Using a rolling bearing fault dataset collected by a certain institution as the subject, this study tests the fault diagnosis performance of the aforementioned DRAGAN-CNN-LSTM hybrid model under small sample conditions. The dataset covers four typical fault modes: normal, inner ring fault (IF), outer ring fault (OF), and rolling element fault (BF). The experimental data was collected from an SKF6205 bearing at the drive end, with a sampling frequency set to 12kHz. Single-point defects were preset using electrical discharge machining (EDM), with defect diameters ranging from 0.1778 to 0.5334 mm. Vibration signals under four load conditions (0HP, 1HP, 2HP, and 3HP) were selected for the experiment. The training set division for each type of fault sample was performed according to Table 4.

[0026] Table 4. Division by different training sample ratios

[0027] The test set is fixed at 20% of the remaining samples for each working condition to ensure the fairness of the model evaluation. The original vibration signals are preprocessed as follows: a 4th-order Butterworth filter (1-4kHz) is used to remove high-frequency noise; time-domain segments are truncated at 2048 points / sample (0.17 seconds) with an overlap rate of 50%; Z-score normalization is used to eliminate dimensional differences.

[0028] To verify the superiority of the DRAGAN-CNN-LSTM hybrid model, the following baseline models were selected for comparison: 1) Traditional CNN model: 4-layer one-dimensional convolutional network, number of channels 64→128→256→512, kernel size 64→32→16→8, stride 2, global average pooling at the end followed by a fully connected layer; 2) Bi-LSTM model: Bidirectional LSTM network, 128 hidden units, Dropout rate 0.3, fully connected classification layer at the end; 3) CNN-LSTM model: cascaded structure, the CNN part is the same as the traditional CNN model, and the LSTM part is the same as the Bi-LSTM model; The DRAGAN-CNN-LSTM hybrid model consists of a generator with 5 fully connected layers (256→512→1024→512→2048), a discriminator with 4 one-dimensional convolutional layers (64→32→16→8 kernels, 64→128→256→512 channels), a gradient penalty coefficient λ=10, and a generator learning rate of 2×10⁻⁶. -4 Discriminator learning rate 1×10 -5 The parameters of CNN and LSTM are the same as the baseline model.

[0029] All models used the Adam optimizer (learning rate 1×10⁻⁶). -4 The dataset was set to β1=0.9, β2=0.98, with a batch size of 64 and 200 training epochs. Early stopping (Patience=15) was used. The generative adversarial network generated 9 times the number of samples of the original data (585 samples for extremely small samples and 16200 samples for the full dataset). The experiments were implemented using the PyTorch framework and an NVIDIA RTX 3090 GPU. Each experiment was repeated 10 times to eliminate randomness. The results are shown in Table 5. Table 5. Average classification accuracy of each model under different training sample sizes.

[0030] Experimental data analysis shows, as shown in Table 5, Figure 3 as well as Figure 4As shown, the DRAGAN-CNN-LSTM hybrid model in this embodiment exhibits significant performance advantages in few-shot learning scenarios. Under the extreme condition of a training set ratio of only 5%, the model achieves a classification accuracy of 80.0%, which is 60, 40, and 20 percentage points higher than the benchmark models of traditional CNN (20.0%), Bi-LSTM (40.0%), and CNN-LSTM (60.0%), respectively. This verifies the dual advantages of the hybrid model architecture: the data augmentation method based on DRAGAN generative adversarial networks effectively alleviates the limitation of insufficient training samples, while the fusion of spatial feature encoding of convolutional neural networks and temporal dependency modeling of long short-term memory networks significantly improves the model's representation learning efficiency for sparse data.

[0031] As the proportion of training samples increases, the performance of all models shows an upward trend, but the magnitude of the improvement varies. In this embodiment, the DRAGAN-CNN-LSTM hybrid model improves accuracy by 20% when the sample proportion increases from 5% to 10%, while the traditional CNN model only improves by 40%. When the sample proportion reaches 20%, the DRAGAN-CNN-LSTM hybrid model achieves 100% accuracy in validation, while the comparison model requires the full dataset (100% samples) to approach this level. This indicates that the established model has stronger data efficiency, especially in reaching saturation performance with a moderate sample size (20%).

[0032] Next, to verify the generalization ability of the DRAGAN-CNN-LSTM hybrid model across working conditions in complex industrial scenarios, a small-sample cross-load diagnostic experiment was designed based on the rolling bearing fault dataset collected by a certain organization.

[0033] This experiment focuses on analyzing the fault identification performance of the model under different load conditions (differences between training and testing conditions), and compares it with the same three existing models used in the diagnostic performance verification under the small sample scenario. The experiment selects three typical load conditions as shown in Table 6 below: Table 6 Dataset Load Conditions

[0034] For each fault type (IF / OF / BF), 20 samples (each sample with 2048 points and 50% overlap) are randomly selected from the source condition (e.g., condition A), and 100 samples are selected from the normal state to form the training set (160 samples in total). The full dataset of the target conditions (e.g., conditions B / C) is used as the test set to verify the model's ability to transfer data across different conditions. The experimental results are shown in Table 7. Table 7 Comparison of Diagnostic Performance Across Operating Conditions in Small Samples (Unit: %)

[0035] Experimental data analysis shows that, as shown in Table 7 above, and Figure 5 and Figure 6 As shown, in the scenario of migrating from condition A to B, the DRAGAN-CNN-LSTM hybrid model achieved an average recognition accuracy of 89.7 ± 0.8%, an improvement of 8.3% compared to the traditional CNN-LSTM model. When migrating to condition A→C, the model's recognition accuracy improved by 10.9 percentage points to 85.1% compared to the baseline model. In the more challenging task of migrating from condition B to C, the model maintained an accuracy of 87.6% and a CDR value of 95.4%, with performance improvements exceeding 10 percentage points. The advantage of this model stems from the gradient constraint mechanism of the generative adversarial network, which effectively separates load-related features and reduces the interference intensity of frequency domain modulation on fault features. The adversarial enhancement module significantly improves the capture efficiency of time-frequency features under small sample conditions (20 samples per class), solving the feature distortion problem that occurs in traditional models when data is insufficient. The model constructed in this experiment exhibits good stability and transfer capability in complex operating conditions such as variable speed and variable load, providing an effective technical path for cross-domain fault diagnosis of industrial equipment.

[0036] Figure 7 The distribution of fault features in the t-SNE dimensionality reduction space of DRAGAN-CNN-LSTM in the A→B transfer task is shown. Different fault categories maintain high cohesion and low coupling under the target condition, and their feature distributions highly overlap with those of the source condition, verifying the model's strong representation ability for cross-condition fault modes.

[0037] Furthermore, this embodiment also verifies the noise immunity performance by applying Gaussian white noise interference of different intensities to the test set samples.

[0038] Based on a rolling bearing dataset collected by a certain institution, comparative validation was designed for different training sample sizes and noise interference scenarios. The training set was divided into the following proportions: extremely small samples (5%, total sample size 65, including 50 normal samples and 5 samples of each type of failure), small samples (10%, total sample size 140), medium samples (20%, total sample size 280), and full samples (100%, total sample size 1800). The test set was injected with Gaussian white noise at different signal-to-noise ratios (SNR), ranging from -4dB (strong noise) to 10dB (weak noise), while the training set remained unchanged. The comparison models included traditional CNN, Bi-LSTM, CNN-LSTM, and DRAGAN-CNN-LSTM, with network parameters consistent with the diagnostic performance validation in the small sample scenario. Evaluation metrics included classification accuracy and noise attenuation rate (NAR), where NAR is defined as the percentage ratio of model accuracy at a certain SNR to accuracy without noise (SNR=∞), used to quantify the degree of performance degradation under noise interference.

[0039] The noise power is dynamically adjusted based on the signal energy, using the following formula:

[0040] In the formula, The original signal power is represented by , and SNR is the signal-to-noise ratio. The experimental results are shown in Table 8. Table 8. Classification accuracy of the model under different signal-to-noise ratios (unit: %)

[0041] Experimental results show that in a high-noise environment (SNR=-4dB), the DRAGAN-CNN-LSTM hybrid model in this embodiment achieves the best recognition accuracy of 68.7%, with a corresponding noise attenuation rate (NAR) of 87.1%. In comparison, traditional CNN and Bi-LSTM models, due to the lack of data augmentation techniques and limited global feature extraction capabilities, both achieve accuracies below 50% under low signal-to-noise ratio conditions, reflecting their insufficient noise robustness. When the training sample size is increased from 5% to 100%, the model's accuracy improves by 24.8 percentage points to 92.5% under the same noise intensity, confirming the synergistic optimization effect of adversarial data augmentation strategies and a sufficient training sample size.

[0042] Figure 8 The NAR curves of the DRAGAN-CNN-LSTM hybrid model under different sample sizes are shown. With an extremely small sample size (5%), the SNR drops from 10dB to -4dB, but the NAR only decreases by 25.1% (93.8%→68.7%), while the traditional CNN decreases by 39.4% (81.5%→42.1%), highlighting the model's strong noise suppression capability. Under moderate noise (SNR=0dB), the model achieves a full-sample NAR of 95.7%, approaching the performance under noise-free conditions (98.4%), indicating its suitability for dynamic noise environments in real-world industrial scenarios.

[0043] Another embodiment of the present invention relates to a small-sample fault diagnosis system for rolling bearings based on the above-described hybrid model, comprising: The data acquisition and preprocessing module is used to acquire the original vibration signal of the rolling bearing through the accelerometer, remove high-frequency noise through bandpass filtering, extract a fixed-length time domain segment, and then obtain preprocessed data through standardization. The gradient regularization generative adversarial network data augmentation module is used to input preprocessed data into the gradient regularization generative adversarial network. The generator of the adversarial network generates augmented samples that are consistent with the distribution of real fault signals. When the discriminator distinguishes between real samples and generated samples, it introduces a gradient penalty mechanism to constrain its own gradient norm, thereby expanding the small sample dataset. The multi-scale feature extraction and temporal modeling module is used to input the enhanced samples into a one-dimensional convolutional neural network to extract the local high-frequency impact features and low-frequency resonance features of the vibration signal. Then, the features output by the one-dimensional convolutional neural network are input into a bidirectional long short-term memory network to capture the long-range temporal dependencies of the features and output a fused feature vector. The fault classification decision module is used to input the fused feature vector into the fully connected classification layer and combine it with the confidence verification mechanism to determine the fault type of the rolling bearing.

[0044] Figure 9 This is a schematic diagram of an electronic device 10 provided in another embodiment of the present invention. (See diagram below.) Figure 9 As shown, the electronic device 10 of this embodiment includes: a processor 11, a memory 12, and a computer program 13 stored in the memory 12 and executable on the processor 11, such as a program for a small-sample fault diagnosis method for rolling bearings based on a hybrid model. When the processor 11 executes the computer program 13, it implements the steps in the above embodiment of the small-sample fault diagnosis method for rolling bearings based on a hybrid model, for example... Figure 2 The steps are shown.

[0045] For example, computer program 13 may be divided into one or more modules / units, one or more of which are stored in memory 12 and executed by processor 11 to complete the present invention. One or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 13 in electronic device 10.

[0046] Electronic device 10 can be a desktop computer, laptop, handheld computer, cloud server, or other computing device. Electronic device 10 may include, but is not limited to, a processor 11 and a memory 12. Those skilled in the art will understand that... Figure 9 This is merely an example of electronic device 10 and does not constitute a limitation on electronic device 10. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device 10 may also include input / output devices, network access devices, buses, etc.

[0047] The processor 11 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0048] The memory 12 can be an internal storage unit of the electronic device 10, such as a hard disk or RAM of the electronic device 10. The memory 12 can also be an external storage device of the electronic device 10, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the electronic device 10. Furthermore, the memory 12 can include both internal and external storage units of the electronic device 10. The memory 12 is used to store computer programs and other programs and data required by the electronic device 10. The memory 12 can also be used to temporarily store data that has been output or will be output.

[0049] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0050] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0051] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0052] In the embodiments provided by this invention, it should be understood that the disclosed devices / electronic devices and methods can be implemented in other ways. For example, the device / electronic device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0053] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0054] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0055] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for small-sample fault diagnosis of rolling bearings based on a hybrid model, characterized in that: Including steps S10: Data Acquisition and Preprocessing The original vibration signal of the rolling bearing was collected, high-frequency noise was removed by bandpass filtering, a fixed-length time domain segment was extracted, and then the preprocessed data was obtained by standardization. S20: Gradient Regularization for Generative Adversarial Network Data Augmentation The preprocessed data is input into the gradient regularization generative adversarial network. The generator of the adversarial network generates enhanced samples that are consistent with the distribution of real fault signals. When the discriminator distinguishes between real samples and generated samples, a gradient penalty mechanism is introduced to constrain its own gradient norm, thereby expanding the small sample dataset. S30: Multi-scale Feature Extraction and Temporal Modeling The enhanced samples are input into a one-dimensional convolutional neural network to extract local high-frequency impact features and low-frequency resonance features of the vibration signal. The features output by the one-dimensional convolutional neural network are then input into a bidirectional long short-term memory network to capture the long-range temporal dependencies of the features and output a fused feature vector. S40: Fault Classification Decision The fused feature vectors are input into a fully connected classification layer, and the failure type of the rolling bearing is determined by combining the confidence verification mechanism.

2. The method for diagnosing small-sample faults in rolling bearings according to claim 1, characterized in that: The generator of the adversarial network in step S20 is a 5-layer fully connected network with the number of hidden layer neurons changing sequentially from 256 to 512 to 1024 to 512 to 2048. The input is a Gaussian noise vector of dimension 100, and the generator learning rate is set to 2×10⁻⁶. -4 The discriminator is a 4-layer one-dimensional convolutional network with kernel sizes of 64→32→16→8 and channel numbers of 64→128→256→512. The discriminator learning rate is set to 1×10⁻⁶. -5 .

3. The method for diagnosing small-sample faults in rolling bearings according to claim 1, characterized in that: The penalty coefficient of the gradient penalty mechanism in step S20 is set to 10.

4. The method for diagnosing small-sample faults in rolling bearings according to claim 1, characterized in that: The one-dimensional convolutional neural network in step S30 includes a two-stage "convolution-activation-pooling" structure: the first convolutional layer has 64 feature channels, a kernel size of 7, a stride of 1, and uses an activation function, with an output feature dimension of 2000×64; the max pooling layer has a window size of 3 and a stride of 3, downsampling the output features of the first convolutional layer to 666×64; the second convolutional layer has 128 feature channels, a kernel size of 5, a stride of 1, and an output feature dimension of 666×128.

5. The method for diagnosing small-sample faults in rolling bearings according to claim 1, characterized in that: The bidirectional long short-term memory network in step S30 adopts a bidirectional structure.

6. The method for diagnosing small-sample faults in rolling bearings according to claim 1, characterized in that: The bandpass filter in step S10 is a fourth-order Butterworth filter with a filter frequency band of 1-4kHz; the length of the time domain segment is 2048 points, corresponding to a duration of 0.17 seconds, and the segment overlap rate is set to 50%.

7. The method for diagnosing small-sample faults in rolling bearings according to claim 1, characterized in that: The confidence verification mechanism in step S40 is Top-2 confidence verification: if the probability of the maximum fault category is greater than 0.9, the fault type corresponding to the maximum probability is directly output; if the probability difference between the first two fault categories is greater than 0.3, the corresponding fault type is output; if none of the above conditions are met, manual review is triggered.

8. A small-sample fault diagnosis system for rolling bearings based on a hybrid model, characterized in that: The system is used to implement the steps of the rolling bearing small sample fault diagnosis method according to any one of claims 1 to 7, comprising: The data acquisition and preprocessing module is used to acquire the original vibration signal of the rolling bearing, remove high-frequency noise through bandpass filtering, extract a fixed-length time domain segment, and then obtain preprocessed data through standardization. The gradient regularization generative adversarial network data augmentation module is used to input preprocessed data into the gradient regularization generative adversarial network. The generator of the adversarial network generates augmented samples that are consistent with the distribution of real fault signals. When the discriminator distinguishes between real samples and generated samples, it introduces a gradient penalty mechanism to constrain its own gradient norm, thereby expanding the small sample dataset. The multi-scale feature extraction and temporal modeling module is used to input the enhanced samples into a one-dimensional convolutional neural network to extract the local high-frequency impact features and low-frequency resonance features of the vibration signal. Then, the features output by the one-dimensional convolutional neural network are input into a bidirectional long short-term memory network to capture the long-range temporal dependencies of the features and output a fused feature vector. The fault classification decision module is used to input the fused feature vector into the fully connected classification layer and combine it with the confidence verification mechanism to determine the fault type of the rolling bearing.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the rolling bearing small sample fault diagnosis method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the rolling bearing small sample fault diagnosis method as described in any one of claims 1 to 7.

Images