A bearing fault diagnosis method and system for small samples

Abstract

The application discloses a bearing fault diagnosis method and system for small samples, relates to the technical field of bearing fault diagnosis, and obtains a bearing vibration signal; the bearing vibration signal is subjected to multidimensional decomposition to obtain an enhanced fault feature; the enhanced fault feature is subjected to adaptive fusion to obtain a unified feature representation after fusion; based on a meta-learning or small sample learning paradigm, the unified feature representation is used for fault classification to obtain a fault classification result. The application solves the technical problems of the existing bearing fault diagnosis technology, such as dependence on a large number of labeled samples, poor generalization ability in a small sample scene, insufficient feature mining and weak robustness in complex working conditions, and through a hierarchical architecture of feature construction and enhancement-adaptive fusion-small sample classification, in combination with two core technical routes of bimodal cross-modal fusion and VMD decomposition-component fusion, high-precision and high-stability fault diagnosis in a small sample scene is realized.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, and more specifically to a bearing fault diagnosis method and system for small sample sizes. Background Technology

[0002] In industrial production, bearings are core components of rotating machinery, and their operating status directly affects equipment reliability. Traditional bearing fault diagnosis relies on training models with a large number of labeled fault samples. However, in actual industrial scenarios, fault samples are scarce and class imbalanced, resulting in poor generalization ability and low classification accuracy of existing diagnostic models. Therefore, this is an urgent industry pain point.

[0003] Therefore, how to provide a diagnostic system that is adaptable to small sample scenarios and can fully explore fault characteristics, and solve the problems of limited samples and difficulty in diagnosis, is an urgent issue that needs to be addressed by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a bearing fault diagnosis method and system for small sample scenarios, which solves the technical problems of existing bearing fault diagnosis technology, such as reliance on a large number of labeled samples, poor generalization ability in small sample scenarios, insufficient feature mining, and weak robustness in complex working conditions. Through a hierarchical architecture of feature construction and enhancement-adaptive fusion-small sample classification, combined with two core technical routes of bimodal cross-modal fusion and VMD decomposition-component fusion, high-precision and high-stability fault diagnosis in small sample scenarios is achieved.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a bearing fault diagnosis method for small sample sizes, comprising: Acquire bearing vibration signals; The bearing vibration signal is decomposed in multiple dimensions to obtain enhanced fault characteristics; The enhanced fault features are adaptively fused to obtain a unified feature representation after fusion. Based on meta-learning or few-shot learning paradigms, fault classification is performed using the unified feature representation to obtain fault classification results.

[0006] Preferably, the bearing vibration signal is decomposed into multiple dimensions, including: The bearing vibration signal is synchronously converted into a two-dimensional time-domain waveform image and a manually generated feature vector; Alternatively, variational mode decomposition can be used to adaptively decompose the bearing vibration signal into multiple intrinsic mode function components.

[0007] Preferably, it further includes: during the training phase, a fixed number of samples are drawn evenly from each type of fault state, and the SMOTE algorithm is used to oversample in the feature space to alleviate class imbalance.

[0008] Preferably, the enhanced fault features are adaptively fused to obtain a unified feature representation after fusion, including: using an attention mechanism to fuse the features extracted from the two-dimensional time-domain waveform image and the hand-crafted feature vector respectively; Alternatively, features extracted from each intrinsic mode function component can be fused.

[0009] The handcrafted feature vector is 18-dimensional, including time-domain statistics, frequency-domain indices, envelope features, and wavelet energy features. The time-domain statistics include peak value, peak-to-peak value, mean, variance, kurtosis, and skewness. The frequency-domain indices include center frequency, mean square frequency, and frequency variance. The envelope features include envelope peak value and envelope kurtosis. The wavelet energy features are the energy proportions of each scale component after wavelet decomposition.

[0010] Preferably, an attention mechanism is used to fuse the features extracted from the two-dimensional time-domain waveform image and the hand-crafted feature vectors, including: an image feature extraction branch, which uses a VGG-style convolutional network to extract the spatial hierarchical features of the two-dimensional time-domain waveform image; The manual feature extraction branch uses a multilayer perceptron to perform a nonlinear transformation on the manual feature vector; Align the output features of the two branches to the same semantic space; Adaptive interactive fusion is performed on the aligned bi-branch features.

[0011] The VGG-style convolutional network includes 3 to 4 convolutional blocks, each containing 2 to 3 convolutional layers and 1 max pooling layer; the multilayer perceptron includes 2 to 3 hidden layers; and the multi-head attention module has 4 to 6 heads.

[0012] Preferably, a WDCNN-SENet composite network is used to extract fault-sensitive features of each intrinsic mode function component.

[0013] The WDCNN includes 5 convolutional layers, with the first layer being a wide convolution and the rest being narrow convolutions; the SENet module has a compression ratio of 16-32.

[0014] Preferably, the features extracted from each intrinsic mode function component are fused, including: Employing at least one of attention-score-based weighted fusion, feature splicing fusion, max pooling fusion, or adaptive fusion strategies, specific features of each intrinsic mode function component are dynamically aggregated.

[0015] Preferably, based on a meta-learning or few-shot learning paradigm, fault classification is performed using the unified feature representation to obtain fault classification results, including: A prototype-based classifier forms a category prototype by calculating the feature mean of each category sample in the support set, and classifies the query sample to the nearest prototype category by using a distance metric.

[0016] Preferably, a bearing fault diagnosis system for small sample sizes includes: Signal acquisition unit, used to acquire bearing vibration signals; The feature construction and enhancement unit is used to perform multi-dimensional decomposition on the bearing vibration signal to obtain enhanced fault features. An adaptive feature fusion unit is used to adaptively fuse the enhanced fault features to obtain a unified feature representation after fusion. The few-sample classification unit, based on meta-learning or few-sample learning paradigm, uses the unified feature representation to classify faults and obtain fault classification results.

[0017] As can be seen from the above technical solutions, compared with the prior art, the present invention discloses a bearing fault diagnosis method and system for small samples, which has the following beneficial effects: (1) It proposes two feature construction schemes: dual-modal cross-modal fusion and VMD decomposition-component fusion, which fully mine the fault information in the vibration signal from the perspectives of morphological statistics and frequency domain decoupling, respectively, effectively solving the problem of insufficient feature mining in traditional diagnostic methods.

[0018] (2) Through the collaborative design of small sample equalization module, multi-strategy component fusion and meta-learning training paradigm, high-precision diagnosis can be achieved in extreme small sample scenarios where only a small number of samples are available for each type of fault.

[0019] (3) Through the collaborative design of VGG-style convolutional networks, multilayer perceptrons, and multi-head attention, adaptive interactive fusion of dual-modal features is achieved. Compared with traditional fusion strategies such as simple splicing, it can more fully explore the complementary information between modalities and improve feature discrimination. The WDCNN-SENet composite network enhances the ability to capture fault-sensitive features by combining wide convolutions and channel attention, which is superior to traditional deep convolutional models without attention mechanisms. At the same time, training optimization mechanisms such as progressive unfreezing of branch weights, gradient pruning, and early stopping effectively avoid gradient explosion and overfitting problems during training, taking into account both training stability and diagnostic efficiency.

[0020] (4) Based on commonly used industrial vibration signal acquisition equipment and general deep learning framework, the hardware deployment cost is low and the portability is strong; the data augmentation strategy can be adapted to complex working conditions such as noise interference and signal amplitude fluctuation in industrial scenarios, and improve the robustness of the model in harsh environments; at the same time, the two technical routes can be flexibly selected according to the scarcity of samples, which significantly improves the system's scenario adaptability and facilitates its application in different industrial scenarios. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, 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 embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of a bearing fault diagnosis method for small sample sizes provided by the present invention.

[0023] Figure 2 This invention provides a schematic diagram of a bearing fault diagnosis system designed for small sample sizes. Detailed Implementation

[0024] 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.

[0025] This invention discloses a bearing fault diagnosis method for small sample sizes, such as... Figure 1 As shown, it includes: Acquire bearing vibration signals; The bearing vibration signal is decomposed in multiple dimensions to obtain enhanced fault characteristics; The enhanced fault features are adaptively fused to obtain a unified feature representation after fusion. Based on meta-learning or few-shot learning paradigms, fault classification is performed using the unified feature representation to obtain fault classification results.

[0026] Specifically, the bearing vibration signal is decomposed in multiple dimensions, including: The bearing vibration signal is synchronously converted into a two-dimensional time-domain waveform image and a manually generated feature vector; Alternatively, variational mode decomposition can be used to adaptively decompose the bearing vibration signal into multiple intrinsic mode function components.

[0027] Specifically, this also includes: during the training phase, a fixed number of samples are drawn evenly from each type of fault state, and the SMOTE algorithm is used to oversample in the feature space to alleviate class imbalance.

[0028] Specifically, the enhanced fault features are adaptively fused to obtain a unified feature representation, including: using an attention mechanism to fuse the features extracted from the two-dimensional time-domain waveform image and the hand-crafted feature vectors respectively; Alternatively, features extracted from each intrinsic mode function component can be fused.

[0029] The handcrafted feature vector is 18-dimensional, including time-domain statistics, frequency-domain indices, envelope features, and wavelet energy features. The time-domain statistics include peak value, peak-to-peak value, mean, variance, kurtosis, and skewness. The frequency-domain indices include center frequency, mean square frequency, and frequency variance. The envelope features include envelope peak value and envelope kurtosis. The wavelet energy features are the energy proportions of each scale component after wavelet decomposition.

[0030] Specifically, an attention mechanism is used to fuse the features extracted from the two-dimensional time-domain waveform image and the hand-crafted feature vectors, including: an image feature extraction branch, which uses a VGG-style convolutional network to extract the spatial hierarchical features of the two-dimensional time-domain waveform image; The manual feature extraction branch uses a multilayer perceptron to perform a nonlinear transformation on the manual feature vector; Align the output features of the two branches to the same semantic space; Adaptive interactive fusion is performed on the aligned bi-branch features.

[0031] The VGG-style convolutional network includes 3 to 4 convolutional blocks, each containing 2 to 3 convolutional layers and 1 max pooling layer; the multilayer perceptron includes 2 to 3 hidden layers; and the multi-head attention module has 4 to 6 heads.

[0032] Specifically, a WDCNN-SENet composite network is used to extract fault-sensitive features of each intrinsic mode function component.

[0033] The WDCNN includes 5 convolutional layers, with the first layer being a wide convolution and the rest being narrow convolutions; the SENet module has a compression ratio of 16-32.

[0034] Specifically, the features extracted from each intrinsic mode function component are fused, including: Employing at least one of attention-score-based weighted fusion, feature splicing fusion, max pooling fusion, or adaptive fusion strategies, specific features of each intrinsic mode function component are dynamically aggregated.

[0035] Specifically, based on meta-learning or few-shot learning paradigms, fault classification is performed using the unified feature representation to obtain fault classification results, including: A prototype-based classifier forms a category prototype by calculating the feature mean of each category sample in the support set, and classifies the query sample to the nearest prototype category by using a distance metric.

[0036] This invention provides a fault diagnosis method based on multi-dimensional feature construction, adaptive fusion, and few-shot learning techniques, solving the problems of existing technologies that rely on a large number of labeled samples, have poor generalization in few-shot scenarios, and lack sufficient feature mining. The system adopts a hierarchical architecture of feature construction and enhancement, adaptive fusion, and few-shot classification, providing two technical routes: dual-modal cross-modal fusion and VMD-WDCNN-SENet component fusion. Through multi-dimensional feature mining, adaptive fusion strategies, and meta-learning training paradigms, combined with optimization mechanisms such as SMOTE oversampling and data augmentation, it achieves high-precision diagnosis of bearing faults in few-shot scenarios. This invention is adaptable to complex industrial conditions, has low deployment costs, strong portability, and possesses significant technical innovation and practical value.

[0037] In one specific embodiment of the present invention, a bearing fault diagnosis system for small sample sizes, such as... Figure 2 As shown, it includes: Signal acquisition unit, used to acquire bearing vibration signals; The feature construction and enhancement unit is used to perform multi-dimensional decomposition on the bearing vibration signal to obtain enhanced fault features. An adaptive feature fusion unit is used to adaptively fuse the enhanced fault features to obtain a unified feature representation after fusion. The few-sample classification unit, based on meta-learning or few-sample learning paradigm, uses the unified feature representation to classify faults and obtain fault classification results.

[0038] Specifically, the feature construction and enhancement unit is used to decompose or represent the input one-dimensional bearing vibration signal in multiple dimensions to construct and enhance features containing fault information; the adaptive feature fusion unit is used to adaptively fuse the features output by the feature construction and enhancement unit to obtain a unified feature representation after fusion; and the few-shot classification unit is used to perform fault classification based on meta-learning or few-shot learning paradigms using the unified feature representation.

[0039] Specifically, the feature construction and enhancement unit includes: a multimodal data representation module, used to synchronously convert the one-dimensional vibration signal into a two-dimensional time-domain waveform image and a handcrafted feature vector; the adaptive feature fusion unit is a cross-modal fusion unit, which uses an attention mechanism to fuse the features extracted from the two-dimensional time-domain waveform image and the handcrafted feature vector respectively; the handcrafted feature vector is 18-dimensional, including time-domain statistics, frequency-domain indices, envelope features, and wavelet energy features; wherein, the time-domain statistics include peak value, peak-to-peak value, mean, variance, kurtosis, and skewness; the frequency-domain indices include center frequency, mean square frequency, and frequency variance; the envelope features include envelope peak value and envelope kurtosis; and the wavelet energy features are the energy proportions of each scale component after wavelet decomposition.

[0040] Specifically, it also includes a small sample balancing module, which is used to draw a fixed number of samples from each type of fault state in a balanced manner during the training phase, and to use the SMOTE algorithm to oversample in the feature space to alleviate class imbalance.

[0041] Specifically, after extracting the time-domain, frequency-domain, and wavelet features of the original signal, the program automatically triggers the SMOTE process when generating training set features. The algorithm first analyzes the sample distribution of each class in the current feature space, identifying the minority class. Then, for each minority class sample, the algorithm randomly selects samples from its k-nearest neighbors (default k=5) and synthesizes new, diverse artificial samples through linear interpolation on the feature vector connections. This process is performed in the feature space rather than the original data space, effectively avoiding the risk of overfitting from simple replication. After oversampling, the number of samples in all classes reaches a basic balance, ensuring that subsequent model training will not be biased towards the majority class due to data distribution skew.

[0042] Specifically, the cross-modal fusion unit includes: an image feature extraction branch, which uses a VGG-style convolutional network to extract spatial hierarchical features of the two-dimensional temporal waveform image; a handcrafted feature extraction branch, which uses a multilayer perceptron to perform nonlinear transformation on the handcrafted feature vectors; a feature adaptation layer, which aligns the output features of the two branches to the same semantic space; and a multi-head attention module, which performs adaptive interactive fusion on the aligned dual-branch features; the VGG-style convolutional network includes 3 to 4 convolutional blocks, each convolutional block containing 2 to 3 convolutional layers and 1 max-pooling layer; the multilayer perceptron includes 2 to 3 hidden layers; and the multi-head attention module has 4 to 6 heads.

[0043] Specifically, the feature construction and enhancement unit includes: a signal decomposition module, which uses variational mode decomposition (VMD) to adaptively decompose the one-dimensional vibration signal into multiple intrinsic mode function (IMF) components; the adaptive feature fusion unit is a component fusion unit, used to fuse the features extracted from each IMF component; the parameters of the variational mode decomposition are set as follows: number of modes K=3-8, penalty factor α=2000-4000, noise tolerance τ=0-0.05, and the iteration stopping condition is that the modal energy difference between two adjacent iterations is less than 1e-5.

[0044] Specifically, it also includes a feature extraction module, which is a WDCNN-SENet composite network used to extract fault-sensitive features of each IMF component; wherein, the WDCNN includes 5 convolutional layers, the first layer is a wide convolution, and the rest are narrow convolutions; the compression ratio of the SENet module is 16:32.

[0045] Specifically, the component fusion unit employs at least one of the attention score-based weighted fusion, feature splicing fusion, max pooling fusion, or adaptive fusion strategies to dynamically aggregate the specific features of each IMF component.

[0046] Specifically, the small sample classification unit is a classifier based on a prototype network. It forms a category prototype by calculating the feature mean of each category of samples in the support set, and classifies the query sample into the nearest prototype category by using a distance metric.

[0047] Specifically, the small sample classification unit is trained under the meta-learning paradigm. The training process combines contrastive learning loss and prototype diversity regularization loss, and integrates data augmentation strategies such as signal noise addition, scaling, or segmented permutation.

[0048] Specifically, the training process of the cross-modal fusion unit employs progressive unfreezing of branch weights, gradient pruning, dynamic adjustment of learning rate, and early stopping mechanism to stabilize the optimization process and prevent overfitting.

[0049] This invention provides a fault diagnosis system that integrates multimodal and variational mode decomposition for small sample sizes. This system aims to address the technical problems of existing bearing fault diagnosis technologies, such as reliance on large numbers of labeled samples, poor generalization ability in small sample scenarios, insufficient feature mining, and weak robustness under complex operating conditions.

[0050] The embodiments of the present invention have the following beneficial effects: 1. This invention innovatively proposes two feature construction schemes: dual-modal cross-modal fusion and VMD decomposition-component fusion. These schemes fully extract fault information from vibration signals from the perspectives of morphological statistics and frequency domain decoupling, respectively. The dual-modal scheme effectively compensates for the information limitations of single-modal representation through complementary representation of 18-dimensional handcrafted feature vectors and two-dimensional time-domain images, significantly improving diagnostic accuracy. The VMD decomposition scheme decouples high-frequency fault features from low-frequency noise, laying the foundation for subsequent accurate feature extraction and avoiding noise interference with fault features, effectively solving the problem of insufficient feature extraction in traditional diagnostic methods.

[0051] 2. This invention, through the collaborative design of a small-sample balancing module, multi-strategy component fusion, and a meta-learning training paradigm, achieves high-precision diagnosis in extremely small-sample scenarios where only a few samples are available for each type of fault. The episodic training paradigm of meta-learning, combined with contrastive learning and prototype diversity regularization loss, significantly improves the model's feature generalization ability. The targeted design of SMOTE oversampling and VMD decomposition alleviates the class imbalance problem from both the data and feature levels, breaking through the dependence of traditional diagnostic models on a large number of labeled samples and addressing the core pain point of scarce fault samples in real-world industrial scenarios.

[0052] 3. The cross-modal fusion unit in this embodiment of the invention achieves adaptive interactive fusion of bimodal features through the collaborative design of VGG-style convolutional networks, multilayer perceptrons, and multi-head attention. Compared with traditional fusion strategies such as simple concatenation, it can more fully explore complementary information between modalities and improve feature discriminativeness. The WDCNN-SENet composite network enhances the ability to capture fault-sensitive features by combining wide convolutions and channel attention, outperforming traditional deep convolutional models without attention mechanisms. At the same time, training optimization mechanisms such as progressive unfreezing of branch weights, gradient pruning, and early stopping effectively avoid gradient explosion and overfitting problems during training, balancing training stability and diagnostic efficiency.

[0053] 4. The technical solutions adopted in the embodiments of the present invention are all based on commonly used industrial vibration signal acquisition equipment and general deep learning frameworks, with low hardware deployment costs and strong portability; the data augmentation strategy can adapt to complex working conditions such as noise interference and signal amplitude fluctuations in industrial scenarios, and improve the robustness of the model in harsh environments; at the same time, the two technical routes can be flexibly selected according to the scarcity of samples, which significantly improves the system's scenario adaptability and facilitates its promotion and application in different industrial scenarios.

[0054] In one specific embodiment of the present invention, a small-sample bearing fault diagnosis system based on dual-modal cross-modal fusion is described.

[0055] 1.1 System Hardware Environment

[0056] The fault diagnosis system in this embodiment runs on the following hardware platform: CPU is Intel Core i7-12700H; GPU is NVIDIA RTX 3090; memory is 64GB DDR5 4800MHz; storage is 2TB NVMe solid-state drive; operating system is Ubuntu 22.04 LTS; deep learning framework is PyTorch 2.0; programming language is Python 3.9.

[0057] 1.2 Data Sources and Preprocessing

[0058] The experimental data used the bearing failure dataset from Case Western Reserve University (CWRU), selecting bearing model 6202-2RSJEM. It includes four states: normal, inner ring failure (IR), outer ring failure (OR), and rolling element failure (BR). Each failure state includes three damage levels: 0.007 in, 0.014 in, and 0.021 in, for a total of 13 subcategories.

[0059] Data acquisition parameters: sampling frequency 12kHz, motor load 1hp, speed 1797r / min; 10 sets of continuous vibration signals were acquired for each state, with each set containing 1024 sampling points. Small sample scenario setting: 5 samples were selected as the support set (training prototype) and 15 samples as the query set (test classification) for each type of fault. The small sample equalization module of claim 3 was used, and the SMOTE algorithm was used to oversample the support set in the feature space (sampling rate 2 times) to alleviate class imbalance.

[0060] 1.3 Feature Construction and Enhancement Unit Implementation

[0061] The feature construction and enhancement unit in this embodiment is a multimodal data representation module, and the specific implementation steps are as follows: (1) Two-dimensional time-domain waveform image generation: The one-dimensional vibration signal (length 1024) is directly mapped to a 28×28 grayscale image. The mapping method is min-max normalization (normalizing the signal amplitude to [0,255]). Then, it is converted into a two-dimensional matrix through the reshape operation to preserve the temporal structure and morphological changes of the signal. (2) 18-dimensional manual feature vector extraction: Calculated according to the definition, the specific calculation method is as follows: Time-domain statistics (6 dimensions): peak value = max(|x|), peak-to-peak value = max(x) - min(x), mean = mean(x), variance = var(x), kurtosis = skew(x), skewness = kurt(x), where x is the vibration signal sequence; Frequency domain metrics (3D): The frequency domain sequence X is obtained by performing a Fast Fourier Transform (FFT) on x, with center frequency = sum(f·|X|²) / sum(|X|²), mean square frequency = sum(f²·|X|²) / sum(|X|²), and frequency variance = sum((f-center frequency)²·|X|²) / sum(|X|²), where f is the frequency sequence; Envelope features (2D): The analytic signal is obtained by performing a Hilbert transform on x, and the envelope signal e is obtained by taking the modulus. The envelope peak value = max(e) and the envelope kurtosis = kurt(e). Wavelet energy features (7 dimensions): The db4 wavelet is used to decompose x into 5 levels, resulting in 5 detail components (d1-d5) and 1 approximate component (a5). The energy percentage of each component is calculated as component energy / total energy, resulting in 6 percentages. The 5-level decomposition yields 6 components (d1-d5+a5), with 6 energy percentages. One wavelet entropy (the entropy value of the energy of each component after wavelet decomposition) is added to complete the 7 dimensions. The total handcrafted feature vector is 6+3+2+7=18 dimensions.

[0062] 1.4 Implementation of Cross-Modal Fusion Unit

[0063] The cross-modal fusion unit includes an image feature extraction branch, a manual feature extraction branch, a feature adaptation layer, and a multi-head attention module. The specific structure and parameters are as follows: Image feature extraction branch (VGG-style convolutional network): 3 convolutional blocks, each containing 2 convolutional layers + 1 max pooling layer; convolutional layer parameters: conv1(3×3,64), conv2(3×3,64), pooling layer (2×2, stride 2); conv3(3×3,128), conv4(3×3,128), pooling layer (2×2, stride 2); conv5(3×3,256), conv6(3×3,256), pooling layer (2×2, stride 2); output feature dimension is 256×3×3, which is then processed by global average pooling to obtain a 256-dimensional feature vector; Manual feature extraction branch (multilayer perceptron): 2 hidden layers, input dimension 18, first hidden layer (18→128, activation function ReLU), second hidden layer (128→256, activation function ReLU), output 256-dimensional feature vector; Feature adaptation layer: A fully connected layer (256→256) is used to linearly transform the dual-branch output features and align them to the same semantic space. The adaptation layer parameters are initialized using a He normal distribution. Multi-head attention module: Number of heads = 4, the 256-dimensional features are split according to the number of heads (64 dimensions per head), cross-modal attention weights are calculated in parallel, the attention function adopts Scaled Dot-Product Attention, and the output is concatenated and passed through a fully connected layer (256→256) to obtain fused features.

[0064] 1.5 Training Optimization Mechanism for Cross-Modal Fusion Units

[0065] To stabilize the optimization process and prevent overfitting, the following measures are taken: gradual unfreezing of branch weights, gradient pruning, dynamic adjustment of learning rate, and early stopping mechanism. The specific implementation steps are as follows: Branch weights are gradually unfrozen: Training is divided into 3 phases, with a total of 100 epochs; Phase 1 (epochs 1-30): Freeze convolutional blocks 1-2 of the image feature extraction branch, and train only the hand-crafted feature extraction branch, feature adaptation layer, and multi-head attention module; the optimizer is Adam, the initial learning rate is 1e-4, and the batch size is 32. Phase 2 (epochs 31-60): Unfreeze convolutional block 2 of the image feature extraction branch, keep convolutional block 1 frozen, and continue training the remaining modules; adjust the learning rate to 5e-5; Phase 3 (epochs 61-100): All modules are fully unfrozen and trained jointly; the learning rate is adjusted to 1e-5. Gradient clipping: The norm clipping method is adopted, with a clipping threshold of 1.0. When the gradient norm exceeds the threshold, the gradient is scaled according to the threshold to avoid gradient explosion. Dynamic learning rate adjustment: A cosine annealing strategy is adopted, and the learning rate decays to 0.8 times the current value every 20 epochs. At the same time, combined with the validation set accuracy, if the accuracy does not improve for 3 consecutive epochs, the learning rate decays by an additional 50%. Early stopping mechanism: Set the early stopping patience value to 5. If the classification accuracy of the validation set does not exceed the current optimal value for 5 consecutive epochs, then stop training and save the optimal model parameters.

[0066] 1.6 Implementation of Small Sample Classification Units

[0067] The few-shot classification unit is a classifier based on a prototype network. Training adopts a meta-learning paradigm, with the following specific steps: (1) Prototype construction: In the support set, the mean of the fusion features of each type of fault is calculated to obtain the prototype vector of the four types of faults (dimension 256). (2) Distance metric: Euclidean distance is used to calculate the distance between the fusion features of the query set samples and each prototype vector. The category corresponding to the prototype with the smallest distance is the diagnostic result. Meta-learning training: The "episodic training" paradigm is adopted. Each episode contains one support set (5 samples / class) and one query set (15 samples / class). The loss function is contrastive learning loss + prototype diversity regularization loss, with a weight ratio of 1:0.5. (3) Meta-learning training: The “episodic training” paradigm is adopted. Each episode contains 1 support set (5 samples / class) and 1 query set (15 samples / class); the loss function is contrastive learning loss + prototype diversity regularization loss, with a weight ratio of 1:0.5. (4) Data augmentation: Integrate three strategies: signal noise addition (adding Gaussian noise, signal-to-noise ratio SNR=20dB), scaling (signal amplitude scaling by 0.8-1.2 times), and segment replacement (dividing the signal into 4 segments and randomly replacing the segment order). One augmentation method is randomly selected for each episode.

[0068] 1.7 Experimental Results

[0069] In this embodiment, the diagnostic accuracy reaches 98.7% in a small sample scenario (5 samples / class), which is 6-9 percentage points higher than the single-modal model (image branch accuracy 92.3%, manual feature branch accuracy 89.5%). Through the early stopping mechanism, the training epochs are reduced from 100 to 78, and the training efficiency is improved by 22%. Gradient clipping reduces the fluctuation of model training loss by 40%, and the optimization process is more stable.

[0070] In another specific embodiment of the present invention, a small-sample bearing fault diagnosis system based on VMD-WDCNN-SENet is described.

[0071] 2.1 System Hardware Environment

[0072] As with the previous embodiment, ensure the comparability of experiments.

[0073] 2.2 Data Sources and Preprocessing

[0074] The data source is the same as in the previous embodiment (CWRU dataset). Small sample scenario setting: 3 samples are selected as the support set for each type of fault, and 12 samples are selected as the query set. No additional oversampling is required (feature diversity is improved after VMD decomposition, and the impact of class imbalance is reduced).

[0075] 2.3 Feature Construction and Enhancement Unit Implementation

[0076] The feature construction and enhancement unit is a signal decomposition module (VMD), and the specific parameters and steps are as follows: VMD parameter settings: number of modes K=5, penalty factor α=3000, noise tolerance τ=0.02, iteration stopping condition is that the modal energy difference between two adjacent iterations is less than 1e-5; Signal decomposition process: The one-dimensional vibration signal (length 1024) is input into the VMD module and adaptively decomposed into 5 intrinsic mode functions (IMF1-IMF5). IMF1-IMF3 correspond to high-frequency fault features, and IMF4-IMF5 correspond to low-frequency background noise and trend terms. IMF1-IMF3 are retained for subsequent feature extraction.

[0077] 2.4 Feature Extraction Module Implementation

[0078] The feature extraction module is a WDCNN-SENet composite network, with the following structure: WDCNN network: 5 convolutional layers. The first layer is a wide convolution (1×64, stride 1, padding=32) with 64 output channels; the second to fifth layers are narrow convolutions (1×3, stride 1, padding=1) with 128, 256, 512, and 1024 output channels respectively; each convolutional layer is followed by a batch normalization (BN) layer and a ReLU activation function, and the third and fifth convolutional layers are followed by a max pooling layer (1×2, stride 2). SENet module: After inserting the 5th convolutional layer of WDCNN, the compression ratio is 24; Specific process: Global average pooling (1024→1) → Fully connected layer 1 (1024→42, ReLU) → Fully connected layer 2 (42→1024, Sigmoid) → Multiply the channel weights to calibrate the importance of fault-sensitive feature channels. Feature extraction process: The three IMF components are input into three independent WDCNN-SENet networks, and the outputs three 1024-dimensional component-specific feature vectors.

[0079] 2.5 Component Fusion Unit Implementation

[0080] The component fusion unit adopts a combined strategy of "attention-weighted fusion + feature concatenation fusion": Attention-weighted fusion: For the three 1024-dimensional component features, the attention score (weight) of each component is calculated through a fully connected layer (1024→1). After the attention score is normalized by Softmax, it is weighted and summed with the corresponding component features to obtain the 1024-dimensional weighted features. Feature splicing and fusion: The three component features are directly spliced ​​together to obtain 3072-dimensional spliced ​​features; Final fusion: The weighted features (1024 dimensions) and the spliced ​​features (3072 dimensions) are reduced in dimensionality through a fully connected layer (4096→256) to obtain a unified feature representation of 256 dimensions.

[0081] 2.6 Implementation of Small Sample Classification Units

[0082] The prototype network classifier structure is consistent with that of the previous embodiment, except that: the episode for meta-learning training is set to "3 samples / class support set + 12 samples / class query set"; the weight ratio of contrastive learning loss to prototype diversity regularization loss in the loss function is 1:0.3; and data augmentation only uses signal noise addition (SNR=15dB) and segmented permutation (3-segment permutation).

[0083] 2.7 Experimental Results

[0084] In this embodiment, the diagnostic accuracy reaches 97.5% in a small sample scenario (3 samples / class), which is 5.2 percentage points higher than the WDCNN model without the SENet module and 8.1 percentage points higher than the single component feature model. The component fusion strategy improves the intra-class compactness of the feature space by 35% and the inter-class separability by 28%, verifying the effectiveness of the fusion strategy.

[0085] Conclusion: The two embodiments of this invention respectively verify the feasibility of the two technical routes, both of which can achieve high-precision bearing fault diagnosis in small sample scenarios: The dual-modal cross-modal fusion route of the first embodiment mines comprehensive signal features through multimodal representation and combines training strategies such as progressive unfreezing to balance diagnostic accuracy and training stability; The VMD-WDCNN-SENet route of the second embodiment decouples fault features through signal decomposition and enhances feature robustness by combining attention fusion, adapting to the scarcer small sample scenarios and having significant industrial application value.

[0086] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0087] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A bearing fault diagnosis method for small sample sizes, characterized in that, include: Acquire bearing vibration signals; The bearing vibration signal is decomposed in multiple dimensions to obtain enhanced fault characteristics; The enhanced fault features are adaptively fused to obtain a unified feature representation after fusion. Based on meta-learning or few-shot learning paradigms, fault classification is performed using the unified feature representation to obtain fault classification results.

2. The bearing fault diagnosis method for small sample sizes according to claim 1, characterized in that, The bearing vibration signal is decomposed into multiple dimensions, including: The bearing vibration signal is synchronously converted into a two-dimensional time-domain waveform image and a manually generated feature vector; Alternatively, variational mode decomposition can be used to adaptively decompose the bearing vibration signal into multiple intrinsic mode function components.

3. The bearing fault diagnosis method for small sample sizes according to claim 2, characterized in that, Also includes: During the training phase, a fixed number of samples are drawn evenly from each type of fault state, and the SMOTE algorithm is used to oversample the feature space.

4. The bearing fault diagnosis method for small sample sizes according to claim 2, characterized in that, The enhanced fault features are adaptively fused to obtain a unified feature representation, including: using an attention mechanism to fuse the features extracted from the two-dimensional time-domain waveform image and the hand-crafted feature vectors respectively; Alternatively, features extracted from each intrinsic mode function component can be fused.

5. The bearing fault diagnosis method for small sample sizes according to claim 4, characterized in that, An attention mechanism is used to fuse the features extracted from the two-dimensional time-domain waveform image and the hand-crafted feature vectors, including: an image feature extraction branch, which uses a VGG-style convolutional network to extract the spatial hierarchical features of the two-dimensional time-domain waveform image; The manual feature extraction branch uses a multilayer perceptron to perform a nonlinear transformation on the manual feature vector; Align the output features of the two branches to the same semantic space; Adaptive interactive fusion is performed on the aligned bi-branch features.

6. The bearing fault diagnosis method for small sample sizes according to claim 4, characterized in that, A WDCNN-SENet composite network is used to extract fault-sensitive features of each intrinsic mode function component.

7. The bearing fault diagnosis method for small sample sizes according to claim 4, characterized in that, The features extracted from each intrinsic mode function component are fused, including: Employing at least one of attention-score-based weighted fusion, feature splicing fusion, max pooling fusion, or adaptive fusion strategies, specific features of each intrinsic mode function component are dynamically aggregated.

8. The bearing fault diagnosis method for small sample sizes according to claim 1, characterized in that, Based on meta-learning or few-shot learning paradigms, fault classification is performed using the unified feature representation to obtain fault classification results, including: A prototype-based classifier forms a category prototype by calculating the feature mean of each category sample in the support set, and classifies the query sample to the nearest prototype category by using a distance metric.

9. A bearing fault diagnosis system for small sample sizes, employing the bearing fault diagnosis method for small sample sizes as described in any one of claims 1-8, comprising: Signal acquisition unit, used to acquire bearing vibration signals; The feature construction and enhancement unit is used to perform multi-dimensional decomposition on the bearing vibration signal to obtain enhanced fault features. An adaptive feature fusion unit is used to adaptively fuse the enhanced fault features to obtain a unified feature representation after fusion. The few-sample classification unit, based on meta-learning or few-sample learning paradigm, uses the unified feature representation to classify faults and obtain fault classification results.

Images