A bearing fault diagnosis method and system with high noise and label scarcity

Abstract

The application provides a bearing fault diagnosis method and system with high noise and label scarcity, and relates to the technical field of bearing fault detection. The method comprises the following steps: constructing a bearing fault diagnosis model, which comprises an adaptive wavelet enhancement module, a double-branch feature extraction network, a feature fusion module and a classifier connected in sequence; inputting a vibration acceleration signal into the adaptive wavelet enhancement module to generate an enhanced view; inputting the enhanced view into the double-branch feature extraction network to extract local time domain features and global frequency domain features respectively; adaptively fusing the local time domain features and the global frequency domain features through the feature fusion module to obtain fused features; inputting the fused features into the classifier, training the bearing fault diagnosis model through a dynamic weighted multi-task loss function, and obtaining a trained final diagnosis model; and inputting a bearing vibration signal into the final diagnosis model to output a bearing fault category.

Description

Technical Field

[0001] This invention relates to the field of bearing fault detection technology, and in particular to a bearing fault diagnosis method and system that addresses high noise levels and scarce tags. Background Technology

[0002] Rotating machinery plays a vital role in industrial production systems, and bearings, as key supporting components, directly affect the stability and reliability of the equipment. Efficient and accurate bearing fault diagnosis provides a reliable basis for equipment maintenance decisions, helps improve the continuity and safety of industrial production, and has significant engineering application value.

[0003] Existing bearing fault diagnosis methods mostly rely on vibration signal analysis and deep learning models. Fault type determination is achieved through feature extraction and pattern recognition of the collected signals. Continuous advancements in related technologies, such as data feature mining and model structure design, have propelled the intelligentization of fault diagnosis, providing diverse technical pathways for condition monitoring in high-noise industrial environments and significantly contributing to improving the operation and maintenance level of mechanical equipment.

[0004] However, in real-world industrial scenarios, the collected bearing vibration signals are susceptible to noise interference, and the number of labeled samples is limited. Existing methods still have room for improvement in noise suppression, few-sample learning, and feature utilization efficiency. Summary of the Invention

[0005] To address the technical challenges of existing methods in terms of noise suppression, few-shot learning, and feature utilization efficiency, which still have room for improvement.

[0006] The technical solution provided by this invention is as follows: A first aspect of this invention provides a bearing fault diagnosis method with high noise and scarce tags, comprising: S1: Acquire the vibration acceleration signal of the target bearing and the vibration signal of the bearing to be diagnosed; S2: Construct a bearing fault diagnosis model based on physical guided contrastive learning. The bearing fault diagnosis model includes an adaptive wavelet enhancement module, a dual-branch feature extraction network, a feature fusion module, and a classifier connected in sequence. S3: Input the vibration acceleration signal to the adaptive wavelet enhancement module to perform noise suppression and fault feature enhancement on the vibration acceleration signal and generate an enhanced view; S4: Input the enhanced view into the dual-branch feature extraction network to extract local temporal features and global frequency domain features respectively; S5: The feature fusion module adaptively fuses local time-domain features and global frequency-domain features to obtain fused features; S6: Input the fused features into the classifier, and train the bearing fault diagnosis model through a dynamically weighted multi-task loss function to obtain the final diagnosis model after training; S7: Input the bearing vibration signal into the final diagnostic model and output the bearing fault category.

[0007] A second aspect of the present invention provides a bearing fault diagnosis system with high noise and scarce tags, comprising: a processor and a memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the high-noise and tag-scarce bearing fault diagnosis method as described in the first aspect.

[0008] The beneficial effects of the technical solution provided by this invention include: In this embodiment of the invention, to address the problems of vibration signals in industrial scenarios being susceptible to noise interference, scarce labeled samples, and low feature utilization efficiency, an adaptive wavelet enhancement module is used to suppress noise and enhance fault features in the vibration acceleration signal, thereby improving signal quality. A dual-branch feature extraction network is used to extract local time-domain features and global frequency-domain features respectively, improving feature utilization efficiency. A dynamically weighted multi-task loss function is used to train the bearing fault diagnosis model, completing model learning under label-scarce conditions to obtain the final diagnostic model, thus achieving accurate and stable bearing fault diagnosis in high-noise and label-scarce scenarios. Attached Figure Description

[0009] Figure 1 A flowchart illustrating a bearing fault diagnosis method with high noise and scarce tags provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a bearing fault diagnosis system with high noise and scarce tags, provided as an embodiment of the present invention. Detailed Implementation

[0010] Reference manual attached Figure 1 The diagram shows a flowchart of a bearing fault diagnosis method with high noise and scarce tags provided by an embodiment of the present invention.

[0011] This invention provides a method for diagnosing bearing faults characterized by high noise and scarce tags, which may include the following steps: S1: Collect the vibration acceleration signal of the target bearing and the vibration signal of the bearing to be diagnosed.

[0012] Among them, the vibration acceleration signal is the core signal reflecting the bearing's operating status and is used for subsequent fault feature extraction and diagnosis. The bearing vibration signal is the vibration data generated during the operation of the bearing, which contains its health status characteristics and is the basis for fault diagnosis.

[0013] Specifically, an acceleration sensor is placed on the target bearing housing to collect vibration signals at a preset sampling frequency.

[0014] It should be noted that those skilled in the art can set the preset sampling frequency according to actual needs, and this invention does not limit that.

[0015] For example, the sampling frequency can be set to 12.8kHz, and the sampling duration can be determined according to actual needs.

[0016] S2: Construct a bearing fault diagnosis model based on physical guided contrastive learning. The bearing fault diagnosis model includes an adaptive wavelet enhancement module, a dual-branch feature extraction network, a feature fusion module, and a classifier connected in sequence.

[0017] Among them, physical-guided contrastive learning is the core mechanism for guiding model training, the bearing fault diagnosis model is the core framework for realizing fault diagnosis, the adaptive wavelet enhancement module is used for signal noise suppression and feature enhancement, the dual-branch feature extraction network is used to extract features of different dimensions, the feature fusion module is used to integrate multi-dimensional features, and the classifier is used to output the fault category.

[0018] It should be noted that the adaptive wavelet enhancement module achieves physically guided data enhancement based on learnable parameters, the dual-branch feature extraction network includes a time-domain branch and a frequency-domain branch, and the feature fusion module adopts a channel attention mechanism.

[0019] S3: Input the vibration acceleration signal into the adaptive wavelet enhancement module to perform noise suppression and fault feature enhancement on the vibration acceleration signal and generate an enhanced view.

[0020] Among them, the enhanced view is a signal view that is easier to extract fault features after noise suppression and feature enhancement.

[0021] Specifically, the module adaptively scales the signal bandwidth using learnable wavelet domain coefficients to generate an enhanced view while maintaining the consistency of the signal's physical semantics.

[0022] In one possible implementation, S3 specifically includes sub-steps S301 to S304: S301: Input the vibration acceleration signal to the adaptive wavelet enhancement module, perform discrete wavelet transform on the vibration acceleration signal, and obtain multiple sets of wavelet coefficients.

[0023] Discrete wavelet transform is a processing method for decomposing vibration acceleration signals, and wavelet coefficients are the signal decomposition results obtained after discrete wavelet transform.

[0024] For example, performing discrete wavelet transform on 1024 collected vibration signal samples can yield wavelet coefficients at multiple scales.

[0025] S302: Multiply each group of wavelet coefficients element-wise with the enhancement coefficients in the adaptive wavelet enhancement module to obtain multiple scaled wavelet coefficients.

[0026] Among them, the enhancement coefficient is a learnable parameter used to adjust the wavelet coefficient in the adaptive wavelet enhancement module, and the scaling wavelet coefficient is the adjustment result after multiplying the wavelet coefficient with the enhancement coefficient.

[0027] It should be noted that the enhancement coefficients are optimized through gradient backpropagation during model training to achieve adaptive enhancement of fault-related frequency bands and suppression of noise frequency bands.

[0028] S303: Perform discrete wavelet inverse transform on each scaled wavelet coefficient to reconstruct the enhanced signal.

[0029] Among them, the discrete wavelet inverse transform is a processing method that restores the scaled wavelet coefficients to the signal, and the enhanced signal is a signal whose noise is suppressed and fault characteristics are enhanced after reconstruction.

[0030] S304: By using residual connection, the vibration acceleration signal is combined with the enhancement signal to generate an enhanced view.

[0031] Among them, residual connection is a feature fusion method used to preserve the original signal features and combine them with enhanced signal features.

[0032] Optionally, the discrete wavelet transform uses the Daubechies-4 wavelet basis functions.

[0033] Among them, the Daubechies-4 wavelet basis functions are the fundamental functions used for signal decomposition in discrete wavelet transform.

[0034] It should be noted that the Daubechies-4 wavelet gene function, with its good orthogonality and tight support characteristics, is particularly suitable for extracting fault features from vibration signals.

[0035] In this embodiment of the invention, by inputting the vibration acceleration signal into the adaptive wavelet enhancement module, and through discrete wavelet transform, learnable enhancement coefficients, and residual connections, noise is effectively suppressed, fault characteristics are enhanced, and signal quality is improved while maintaining the physical semantic consistency of the signal.

[0036] S4: Input the enhanced view into the dual-branch feature extraction network to extract local temporal features and global frequency domain features respectively.

[0037] Among them, local time-domain features reflect the local changes of the signal in the time domain, while global frequency-domain features reflect the overall distribution of the signal in the frequency domain.

[0038] Optionally, the dual-branch feature extraction network includes a time-domain branch and a frequency-domain branch.

[0039] The temporal branch is the branch in the dual-branch feature extraction network used to extract local temporal features, while the frequency branch is the branch used to extract global frequency features.

[0040] In one possible implementation, S4 specifically includes sub-steps S401 to S403: S401: Input the enhanced view into the temporal branch, and extract local temporal features through a one-dimensional convolutional layer and residual blocks.

[0041] One-dimensional convolutional layers are network layers used to extract one-dimensional features of signals, while residual blocks are network structures used to mitigate gradient vanishing and preserve features.

[0042] Specifically, the temporal branch consists of stacked one-dimensional convolutional layers and residual blocks, used to extract multi-scale local temporal shock features from the enhanced view.

[0043] S402: Perform a Fast Fourier Transform on the enhanced view to obtain the corresponding spectral features.

[0044] The Fast Fourier Transform (FFT) is a method of converting time-domain signals into frequency-domain signals, and spectral characteristics are features that reflect the frequency-domain properties of a signal.

[0045] For example, performing a Fast Fourier Transform on a 1024-point enhanced view yields an amplitude spectrum feature of 512 points.

[0046] S403: Input the spectral features into the frequency domain branch and extract global frequency domain features through a multi-head self-attention mechanism.

[0047] Among them, the multi-head self-attention mechanism is a network mechanism that can capture the global correlation of features, and is used to fully mine the global information of spectral features.

[0048] Specifically, the frequency domain branch consists of multiple Transformer encoder layers, which utilize a multi-head self-attention mechanism to capture the long-range dependencies between frequency components in the spectrum.

[0049] In this embodiment of the invention, by inputting an enhanced view into a dual-branch feature extraction network, local impulse features are extracted through the time-domain branch and global frequency-domain correlations are captured through the frequency-domain branch, thereby comprehensively mining signal features, avoiding the limitations of single-dimensional features, and improving feature representation capabilities.

[0050] S5: Through the feature fusion module, local time-domain features and global frequency-domain features are adaptively fused to obtain fused features.

[0051] Among them, the fusion feature is a more representative fault feature that integrates local time domain features and global frequency domain features.

[0052] In one possible implementation, S5 specifically includes sub-steps S501 to S504: S501: The local time domain features and the global frequency domain features are concatenated along the channel dimension to obtain the concatenated features.

[0053] Among them, the channel dimension is one of the important dimensions of the feature, and the concatenated feature is the intermediate feature obtained by merging two features in the channel dimension.

[0054] For example, if the time domain feature is 64 channels and the frequency domain feature is 64 channels, then the spliced ​​feature will be 128 channels.

[0055] S502: Input the concatenated features into the global average pooling layer in the feature fusion module to generate a global channel description.

[0056] Among them, the global average pooling layer is a network layer used to extract global information from the spliced ​​features, and the channel global description is a description vector that reflects the global information of each channel of the spliced ​​features.

[0057] Specifically, global average pooling is performed on the splicing features to obtain the statistics for each channel.

[0058] S503: Input the global description of the channels into the fully connected layer in the feature fusion module to generate channel attention weights.

[0059] Among them, the fully connected layer is a network layer used to transform the global description of the channels, and the channel attention weight is a weight parameter used to distinguish the importance of different channel features.

[0060] It should be noted that fully connected layers typically employ a two-layer structure, with an activation function in between, to enhance non-linear expressive power.

[0061] S504: Based on the channel attention weights, the splicing features are weighted and fused to obtain the fused features.

[0062] Weighted fusion is a method that assigns different weights to different channel features based on channel attention weights, and then integrates them to obtain fused features.

[0063] Specifically, the channel attention weights are multiplied with the splicing features channel by channel to achieve adaptive fusion.

[0064] In this embodiment of the invention, the feature fusion module adaptively fuses two features, combining channel splicing, global pooling, and attention weights to achieve precise weighted integration of features, resulting in more representative fused features that provide reliable support for subsequent classification and diagnosis.

[0065] S6: Input the fused features into the classifier, and train the bearing fault diagnosis model through a dynamically weighted multi-task loss function to obtain the final diagnosis model after training.

[0066] Among them, the dynamically weighted multi-task loss function is used to guide model training and balance different losses, and the final diagnostic model is a mature model that can be used for fault diagnosis after training.

[0067] Optionally, the multi-task loss function specifically includes dynamic weight coefficients, contrastive learning loss, and classification loss.

[0068] Among them, the dynamic weight coefficient is a parameter used to dynamically adjust the proportion of different losses, the contrastive learning loss is used to optimize the feature discrimination, and the classification loss is used to optimize the fault classification accuracy.

[0069] It should be noted that the contrastive learning loss uses InfoNCE loss, which is used to bring the feature representations of the same signal closer together in different augmented views and to push away the feature representations of different samples. The classification loss uses Focal Loss, which is used to focus on hard-to-classify samples and suppress the loss contribution of easy-to-classify samples.

[0070] Specifically, the total loss function is defined as: in, This represents a dynamically weighted multi-task loss function. Indicates the contrast learning loss. Represents classification loss, Indicates the dynamic weighting coefficient. t Indicates the training period.

[0071] For example, initial value It can be set to 0.99, which decays linearly or cosinely to near 0 over the training period, so that the early stage of the training process focuses on unsupervised contrastive representation learning, and the later stage focuses on supervised fault classification.

[0072] In one possible implementation, S6 specifically includes sub-steps S601 to S606: S601: Based on vibration acceleration signals, construct a dataset containing training and validation sets.

[0073] The dataset is a set of signals used for model training and validation. The training set is used for model training, and the validation set is used to validate model performance.

[0074] Specifically, the preprocessed long sequence signal is divided into samples of fixed length, a very small portion of the samples are randomly selected and labeled with faults, and the rest are treated as unlabeled samples.

[0075] For example, the sample length can be set to 1024 points, and the label rate can be set to 0.65% of the total number of samples to construct a label-scarce dataset containing training, validation, and test sets.

[0076] S602: Construct a dynamically weighted multi-task loss function based on contrastive learning loss and classification loss.

[0077] Dynamic weighting adjusts the proportions of the two losses through dynamic weight coefficients, making the multi-task loss function more suitable for model training.

[0078] S603: Based on dynamic weight coefficients, adjust the proportions of contrastive learning loss and classification loss in the multi-task loss function.

[0079] The adjustment of the proportion is to balance the impact of the two types of loss on model training and improve model performance.

[0080] It should be noted that the dynamic weighting coefficient The model dynamically changes with the training cycle, ensuring that in the early stages of training, the model mainly performs unsupervised contrastive learning to build a robust feature space, while in the later stages of training, it focuses on supervised classification to refine the decision boundary.

[0081] S604: Based on the multi-task loss function and training set, end-to-end training is performed on the bearing fault diagnosis model to obtain a preliminary diagnosis model.

[0082] End-to-end training is a continuous training method from signal input to model output, while the preliminary diagnostic model is a model that has undergone basic training but has not been fine-tuned.

[0083] Specifically, all model parameters are optimized simultaneously during training, including the learnable enhancement coefficients in the adaptive wavelet enhancement module.

[0084] S605: Based on the prediction results and true labels of the samples in the validation set by the preliminary diagnostic model, difficult samples are screened out.

[0085] The prediction result is the initial diagnostic model's judgment of the fault category of the validation set samples, while the true label is the actual fault category corresponding to the validation set samples.

[0086] For example, if a sample in the validation set is initially diagnosed as having an inner ring fault, but the true label is an outer ring fault, then the sample is classified as a hard sample.

[0087] It should be noted that the index set for screening difficult samples can be represented as: in, The set of indices representing the hard samples. This represents a single bearing vibration acceleration signal sample in the validation set. Represents the samples in the validation set The corresponding actual fault label, Represents the validation set. Represents the samples in the validation set The corresponding index number, argmax This represents the operation of retrieving the category label corresponding to the maximum value of the function's output. This represents the prediction function of the preliminary diagnostic model. This means for all.

[0088] Furthermore, for samples with ambiguous classification boundaries, a difficult sample set is constructed: in, This represents a difficult sample set.

[0089] It should be noted that a hierarchical learning rate strategy is adopted, which fine-tunes the fully connected classification layer of the model with a lower learning rate, while freezing or updating the feature extraction network parameters with a smaller learning rate, and using a difficult sample set for additional training rounds to optimize the model's decision boundary.

[0090] S606: Based on difficult samples, the preliminary diagnostic model is fine-tuned to obtain the final diagnostic model.

[0091] Fine-tuning is the process of adjusting the parameters of the preliminary diagnostic model for difficult samples in order to improve the model's diagnostic accuracy for these samples.

[0092] Specifically, a hierarchical learning rate strategy is adopted to fine-tune the fully connected classification layer of the model with a lower learning rate, while freezing or updating the feature extraction network parameters with a smaller learning rate.

[0093] Optionally, difficult samples are specifically those in the validation set whose prediction results are inconsistent with the true labels.

[0094] Inconsistent samples are those that the initial diagnostic model predicted incorrectly; fine-tuning these samples can improve the overall diagnostic accuracy of the model.

[0095] In this embodiment of the invention, the model is trained by a dynamically weighted multi-task loss function, combined with a fine-tuning strategy for difficult samples, to adapt to scenarios with scarce labels, balance feature learning and classification accuracy, and optimize model parameters through hierarchical learning rates to improve diagnostic accuracy and generalization ability.

[0096] S7: Input the bearing vibration signal into the final diagnostic model and output the bearing fault category.

[0097] Among them, the bearing vibration signal needs to be used to determine the fault type, and the bearing fault category is the fault determination result output by the final diagnostic model.

[0098] For example, the final diagnostic model can output categories such as inner ring fault, outer ring fault, rolling element fault, or normal.

[0099] Specifically, the signal to be diagnosed is processed sequentially through an adaptive wavelet enhancement module, a dual-branch feature extraction network, and a feature fusion module, and then input into a classifier to obtain the final fault diagnosis result.

[0100] In this embodiment of the invention, the fault category is output after the signal to be diagnosed is input into the final diagnostic model and processed by each module. The diagnostic process is simple and efficient, and can accurately identify a variety of bearing faults, providing a reliable basis for equipment maintenance and is highly practical.

[0101] Reference manual attached Figure 2 The diagram shows a structural schematic of a bearing fault diagnosis system with high noise and scarce tags provided by an embodiment of the present invention.

[0102] This invention provides a bearing fault diagnosis system 20 with high noise and scarce tags, including: a processor 201 and a memory 202; The memory 202 stores programs or instructions that can run on the processor 201. When the program or instructions are executed by the processor 201, they implement the steps of the above-described high-noise and tag-scarce bearing fault diagnosis method and achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0103] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A bearing fault diagnosis method with high noise and scarce tags, characterized in that, include: S1: Acquire the vibration acceleration signal of the target bearing and the vibration signal of the bearing to be diagnosed; S2: Construct a bearing fault diagnosis model based on physical guided contrastive learning. The bearing fault diagnosis model includes an adaptive wavelet enhancement module, a dual-branch feature extraction network, a feature fusion module, and a classifier connected in sequence. S3: Input the vibration acceleration signal into the adaptive wavelet enhancement module to perform noise suppression and fault feature enhancement on the vibration acceleration signal and generate an enhanced view; S4: Input the enhanced view into the dual-branch feature extraction network to extract local temporal features and global frequency domain features respectively; S5: The feature fusion module adaptively fuses the local time-domain features and the global frequency-domain features to obtain fused features; S6: Input the fused features into the classifier, and train the bearing fault diagnosis model through a dynamically weighted multi-task loss function to obtain the final diagnosis model after training; S7: Input the bearing vibration signal into the final diagnostic model and output the bearing fault category.

2. The bearing fault diagnosis method with high noise and scarce tags according to claim 1, characterized in that, S3 specifically includes: S301: Input the vibration acceleration signal to the adaptive wavelet enhancement module, perform discrete wavelet transform on the vibration acceleration signal, and obtain multiple sets of wavelet coefficients; S302: Multiply each group of wavelet coefficients element-wise with the enhancement coefficients in the adaptive wavelet enhancement module to obtain multiple scaled wavelet coefficients; S303: Perform inverse discrete wavelet transform on each of the scaling wavelet coefficients to reconstruct the enhanced signal; S304: The vibration acceleration signal and the enhancement signal are combined through residual connection to generate the enhanced view.

3. The bearing fault diagnosis method with high noise and scarce tags according to claim 2, characterized in that, The discrete wavelet transform uses the Daubechies-4 wavelet basis functions.

4. The bearing fault diagnosis method with high noise and scarce tags according to claim 1, characterized in that, The dual-branch feature extraction network includes a time-domain branch and a frequency-domain branch.

5. The bearing fault diagnosis method with high noise and scarce tags according to claim 4, characterized in that, S4 specifically includes: S401: Input the enhanced view into the temporal branch, and extract the local temporal features through a one-dimensional convolutional layer and residual blocks; S402: Perform a fast Fourier transform on the enhanced view to obtain the corresponding spectral features; S403: Input the spectral features into the frequency domain branch, and extract the global frequency domain features through a multi-head self-attention mechanism.

6. The bearing fault diagnosis method with high noise and scarce tags according to claim 1, characterized in that, S5 specifically includes: S501: The local time-domain features and the global frequency-domain features are concatenated along the channel dimension to obtain the concatenated features; S502: Input the spliced ​​features into the global average pooling layer in the feature fusion module to generate a channel global description; S503: Input the global description of the channel into the fully connected layer in the feature fusion module to generate channel attention weights; S504: Based on the channel attention weights, the splicing features are weighted and fused to obtain the fused features.

7. The bearing fault diagnosis method with high noise and scarce tags according to claim 1, characterized in that, The multi-task loss function specifically includes dynamic weight coefficients, contrastive learning loss, and classification loss.

8. The bearing fault diagnosis method with high noise and scarce tags according to claim 7, characterized in that, S6 specifically includes: S601: Based on the vibration acceleration signal, construct a dataset containing a training set and a validation set; S602: Based on the contrastive learning loss and the classification loss, construct the dynamically weighted multi-task loss function; S603: Based on the dynamic weight coefficients, adjust the proportions of the contrastive learning loss and the classification loss in the multi-task loss function; S604: Based on the multi-task loss function and the training set, perform end-to-end training on the bearing fault diagnosis model to obtain a preliminary diagnosis model; S605: Based on the prediction results and true labels of the samples in the validation set by the preliminary diagnostic model, difficult samples are selected; S606: Based on the difficult samples, the preliminary diagnostic model is fine-tuned to obtain the final diagnostic model.

9. A bearing fault diagnosis method with high noise and scarce tags according to claim 8, characterized in that, The difficult samples are specifically those in the validation set whose prediction results are inconsistent with the true labels.

10. A bearing fault diagnosis system characterized by high noise and scarce tags, wherein... include: Processor and memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the bearing fault diagnosis method with high noise and scarce tags as described in any one of claims 1 to 9.

Images