Cross-condition bearing fault diagnosis method and system based on decoupled adversarial distillation

Abstract

This invention discloses a cross-condition bearing fault diagnosis method and system based on decoupled adversarial distillation, belonging to the field of mechanical fault diagnosis technology. It constructs a teacher-student dual-network framework. First, the teacher network employs a dual strategy of adversarial learning and distance covariance minimization to decouple the statistical distribution of fault features and operating condition interference, thus removing operating condition interference information. Second, the student network is designed with a lightweight structure, transferring higher-order geometric features of the teacher network through a manifold topology containing angle and distance constraints, and enhancing cross-domain robustness by combining an adversarial adaptation mechanism. The system integrates an edge computing terminal and an industrial control interface, supporting real-time online monitoring and abnormal operating condition rejection. Ultimately, only the student network is deployed to the edge. This invention effectively solves the feature confusion problem under cross-conditions, improves the accuracy of diagnosing unseen operating conditions, and simultaneously meets the low-power and real-time requirements of edge devices.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis technology, and in particular to a cross-condition bearing fault diagnosis method, system and storage medium based on decoupled adversarial distillation for edge computing scenarios. Background Technology

[0002] With the development of industrial automation and intelligent manufacturing, the health status of rotating machinery (such as bearings) is a critical component, directly impacting production safety. While deep learning-based intelligent diagnostic technologies have achieved high accuracy under single operating conditions, in real-world industrial scenarios, equipment often operates under complex conditions such as varying speeds and loads, leading to significant data distribution shifts. However, existing technologies still have significant shortcomings in addressing this issue regarding feature decoupling and lightweight model edge deployment.

[0003] For example, patent application CN114492533A discloses a method for constructing a bearing fault diagnosis model under varying operating conditions. This method introduces a generative module and utilizes adversarial training to reduce differences in operating conditions. However, this method mainly relies on adversarial learning to obfuscate the edge distributions of the source and target domains, achieving only global alignment at the feature level. Since the essential features of the fault and the information on operating condition interference are highly coupled in the physical signal, simple adversarial obfuscation cannot completely remove the interference from the operating conditions in a statistical sense. As a result, when the model faces unseen operating conditions, operating condition noise still remains in the feature space, affecting generalization performance.

[0004] For example, patent application CN118746439A proposes an integrated domain generalization network based on feature decoupling. While this method attempts feature decoupling, its core architecture relies on a multi-branch structure, requiring the construction of multiple domain expert models for integrated diagnosis. This multi-expert integration strategy leads to model structural redundancy, significantly increasing the number of parameters and computational complexity, making it difficult to meet the stringent requirements of low power consumption and low latency for edge devices in industrial settings. Furthermore, the decoupling in this method primarily relies on implicit adversarial supervision signals, lacking explicit statistical independence constraints on the latent feature space, making it difficult to guarantee the complete orthogonality between fault features and operating condition features.

[0005] Furthermore, existing lightweight methods (such as conventional knowledge distillation) typically focus only on the consistency of the probability distributions of the teacher and student network outputs, neglecting the higher-order geometric relationships (such as manifold topology) between samples in the feature space. In cross-condition scenarios, this feature manifold structure often contains more essential fault category information than a single feature; ignoring this makes it difficult for lightweight student networks to truly inherit the generalization ability of the teacher network. In addition, existing methods mostly adopt the paradigm of training independent models for single devices, resulting in high model update and maintenance costs, which makes it difficult to meet the stringent requirements of low latency and low power consumption for edge intelligence in the Industrial Internet of Things.

[0006] Therefore, there is an urgent need for a method that can explicitly decouple and remove operating condition interference, avoid complex and redundant model structures, and use relational knowledge distillation to transfer high-order structural knowledge, so as to achieve high-precision cross-operating condition fault diagnosis under low computing power constraints. Summary of the Invention

[0007] This invention aims to address the problems of poor robustness, insufficient feature decoupling, and overly complex models in existing bearing fault diagnosis methods across different operating conditions, which are difficult to deploy to edge devices. Specifically, existing methods struggle to simultaneously achieve high robustness across operating conditions and lightweight models on edge devices, and fail to effectively achieve statistical-level decoupling between fault features and operating condition information. To solve these technical problems, this invention provides a fault diagnosis method based on decoupling and adversarial distillation for edge computing. A decoupled teacher network is constructed with distance covariance constraints, and a lightweight student network is trained using knowledge distillation that includes the relationship between angle and distance constraints. The method includes the following steps: Step S01: Acquire and preprocess bearing vibration signals under multiple operating conditions. Acquire bearing vibration signals under different operating conditions such as speed and load to obtain vibration data with fault labels and operating condition labels; perform preprocessing operations such as denoising, segmentation, and normalization on the vibration data to construct multi-condition bearing vibration samples for model training.

[0008] Step S02: Construct and pre-train a decoupled teacher network. A decoupled variational autoencoder structure is used to map the input bearing vibration signal to two latent spaces: a fault latent subspace and a fault latent subspace. and working condition potential subspace .

[0009] Step S03: Multi-objective joint optimization training of the teacher network. The teacher network is trained using labeled bearing samples under various operating conditions, and the parameters are updated by minimizing the composite loss function. The composite loss function is defined as: ; The specific definitions of each sub-loss are as follows: The reconstruction loss is used to ensure that the latent variables retain the effective information of the original signal, and is calculated using the mean squared error (MSE): ; in, It is the original vibration signal. This is the reconstructed signal output by the decoder. The KL divergence loss is used to constrain the distribution of the latent variables in the encoder output to approximate a standard normal distribution, thus achieving regularization. ; in, and These are the mean and variance of the latent variables output by the encoder, respectively. Let be the dimension of the latent space. Supervised classification loss, which assigns semantic meaning to the latent space, uses the cross-entropy loss function: ; in, and These are the actual fault labels and operating condition labels, respectively. Let be the predicted probability of the corresponding classifier. The adversarial classification loss is implemented through a gradient inversion layer, aiming to maximize the prediction entropy of the adversarial classifier and force the latent space to decouple. ; in, It is an adversarial classifier that attempts to predict operating conditions from the fault space. It is an adversarial classifier that attempts to predict faults from the condition space. During backpropagation, the gradient of this loss is inverted and backpropagated to the encoder. The distance covariance decoupling loss is used to measure and minimize the fault sub-potential space. With working condition potential subspace Nonlinear statistical correlation between them: ; in, and They are respectively and The centralized distance matrix. Specifically, when calculating the loss, a masking operation is implemented by resetting the weights of non-cross-space terms to 0, that is, only the correlation between the fault potential vector and the operating condition potential vector is calculated, while the correlation between vectors within the same subspace is ignored, thereby focusing on decoupling the relationship between the two subspaces. ,in , Indicates the row, column, or global mean; Similarly based on calculate.

[0010] Step S04: Construct a lightweight student network using a convolutional neural network (CNN) for online fault diagnosis tasks. The student network's structure is more lightweight than the teacher network, making it suitable for real-time inference on resource-constrained edge devices.

[0011] Step S05: Train the student network based on decoupling adversarial distillation. Freeze the teacher network parameters, input the same preprocessed sample into the student network, and update the student network by minimizing the total student loss. The total student loss includes: relational knowledge distillation loss, which is used to transfer the higher-order geometric structure knowledge of the teacher network in the fault potential subspace to the student network. The relational knowledge distillation loss includes angle loss and distance loss. ; The first term is used to maintain the angular relationship between sample triplets, and the second term is used to maintain the normalized Euclidean distance ratio between sample pairs. L2 distillation loss: ; in, and These represent samples in the teacher and student networks, respectively. Potential representation of .

[0012] The adversarial operating condition loss introduces a gradient inversion layer connected to the operating condition discriminator after the student feature extraction layer. By maximizing the prediction error of the operating condition discriminator, it forces the student network to extract general fault features independent of operating conditions. ; in, The output of the operating condition discriminator indicates whether samples in the fault potential subspace of the student network can be used to distinguish different operating conditions.

[0013] Step S06: Deploy the student network and perform online fault diagnosis. Deploy the trained student network on the edge device for real-time bearing fault diagnosis. When the student network receives a bearing vibration signal under new operating conditions, it can perform accurate diagnosis based on the fault characteristics obtained from the distillation process.

[0014] Compared with the prior art, the technical solution provided by the present invention has the following beneficial effects: (1) Deep statistical decoupling with strong robustness: Based on adversarial training, this invention innovatively introduces distance covariance constraints. Existing adversarial training can only align marginal distributions and is difficult to completely eliminate complex working condition couplings; while this invention filters out explicit working condition features through adversarial mechanisms and further eliminates nonlinear statistical correlations between latent variables in conjunction with distance covariance. The two work together to achieve a deeper level of statistical independence. This dual decoupling mechanism effectively solves the overfitting problem caused by residual working condition noise in traditional methods when no working conditions are observed.

[0015] (2) Manifold structured transfer with high accuracy: Unlike traditional single-point feature distillation, this invention utilizes structured loss that includes angle and distance constraints to completely transfer the geometric topology learned by the teacher model in the feature space to the student model. Since the geometric distribution of fault categories is invariant under varying operating conditions, this topology-based transfer enables the lightweight model to inherit the discriminative ability of the complex teacher model in deep manifold space.

[0016] (3) Highly lightweight and suitable for edge deployment: The final deployed model is only a very lightweight student network, and all auxiliary branches used for decoupling have been removed, which significantly reduces the amount of computation and storage requirements, making it very suitable for real-time diagnosis of resource-constrained industrial edge devices. Attached Figure Description

[0017] Figure 1 This is a block diagram of the overall architecture of a cross-condition bearing fault diagnosis system according to an embodiment of the present invention; Figure 2 This is a flowchart of a bearing fault diagnosis method based on decoupling counter-distillation according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a decoupled teacher network according to an embodiment of the present invention; Figure 4 This is a training flowchart of a decoupled teacher network according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a teacher-student joint training framework based on decoupled adversarial distillation according to an embodiment of the present invention. Figure 6 This is a schematic diagram of a lightweight student network and online fault diagnosis process according to an embodiment of the present invention.

[0018] The diagram shows: bearing and sensor assembly 101, data acquisition module 102, data preprocessing module 103, model training and update module 104, storage module 105, online fault diagnosis module 106, result display and alarm module 107, decoupled teacher network module 200, teacher encoder 210, and subspace. 232, Fault Classifier 251, Operating Condition Classifier 252, Adversarial Fault Classifier 253, Student Network Module 300, Preprocessed Vibration Samples 301, Student Features 321, Confidence Determination / Rejection Module 325, Operating Condition Discriminator 330, Output Fault Diagnosis Results 340, Edge Computing Device 350. Detailed Implementation

[0019] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the technical solutions of this invention will be described in complete and detailed manner below with reference to the accompanying drawings and preferred embodiments. These embodiments are only used to explain this invention and do not constitute any limitation on its scope of protection.

[0020] Reference Figure 1This invention provides a cross-condition bearing fault diagnosis system, which is logically divided into an offline training phase and an online diagnosis phase. The system mainly includes: a bearing and sensor assembly 101: a single-axis or three-axis accelerometer mounted on a rotating machinery bearing housing for capturing vibration signals; a data acquisition module 102: used to convert analog signals into digital signals; in this embodiment, the sampling rate is set to 25.6kHz; a data preprocessing module 103: performs bandpass filtering (0.5kHz~8kHz) on the raw signal to remove high-frequency noise and low-frequency interference; and uses a sliding window method for data segmentation, with a window length of 2048 points and an overlap rate of 50%; finally, Z-score normalization is performed on each sample; and a decoupled teacher network module 200: deployed on a high-performance server or industrial core. The system includes: a workstation for learning decoupled fault characteristics in the offline phase; a lightweight student network module 300 designed as a lightweight one-dimensional convolutional neural network (1D-CNN), which can be trained by distillation on the server side or deployed at the edge; a model training and update module 104 responsible for performing pre-training of the teacher network and distillation training of the student network under the guidance of the teacher; a storage module 105 for storing the trained student network model parameters; and an online fault diagnosis module 106 and a result display and alarm module 107 deployed on an edge computing device 350 (such as an embedded industrial control computer based on the ARM Cortex-A53 architecture) for loading the student model and performing real-time inference.

[0021] like Figure 2 As shown, the method in this embodiment includes the following steps: Step S01: Acquire bearing vibration signals under multiple operating conditions. Bearing vibration signals are acquired under multiple operating conditions using the bearing and vibration sensor 101, and sampled by the data acquisition module 102.

[0022] Step S02: Vibration signal preprocessing and sample construction. The data preprocessing module 103 performs noise reduction, filtering, segmentation and normalization on the acquired signal, constructs a fixed-length preprocessed vibration sample 301, and labels each sample with a fault label and an operating condition label.

[0023] Step S03: Construct a decoupled teacher network. For example... Figure 3 As shown, construct based on The teacher network structure includes a teacher encoder 210 that extracts features using one-dimensional convolution. Preferably, a spatial attention module can be combined to enhance the extraction capability of key features. The latent space maps the encoded features to the mean and log-variance, and latent vectors are obtained through reparameterization techniques. The latent vectors are then partitioned dimensionally, with the first 32 dimensions forming the fault latent subspace. 231, the latter 16 dimensions are the potential subspace of the working condition. 232 (dimensionality ratio 2:1); the decoder structure is symmetrical to the encoder and is used to reconstruct the original signal; the fault classifier 251 and the adversarial condition classifier 254 are connected to... The operating condition classifier 252 and the fault-resistant classifier 253 are connected to .

[0024] Step S04: Train the teacher network. For example... Figure 3 The network structure shown calculates the composite loss function based on the input labeled samples. The parameters are then updated using the Adam optimizer. The specific training process is as follows: Figure 4 As shown, the process includes the following sub-steps: Steps S201-S202: Initialize the teacher network parameters and input labeled multi-condition bearing preprocessed vibration samples; Step S203: The samples are mapped by the encoder to obtain latent vectors, and then divided into fault latent subspaces according to preset dimensions. and working condition potential subspace Steps S204-S207: Calculate the reconstruction loss and KL divergence (S204), the supervised classification loss for faults and operating conditions (S205), the adversarial classification loss (S206), and the distance-covariance decoupling loss (S207), respectively. Steps S208-S209: Calculate the weighted sum of the above losses, backpropagate to update the teacher network parameters, and determine if the model has converged. If it has not converged, return to step S202 to continue the next iteration; if it has converged, end the training and save the teacher encoder parameters. In this embodiment, the weights of the loss function are set as follows: ; Parameter settings To enhance the decoupling constraints of latent features; the adversarial loss is implemented using a gradient inversion layer 331, adversarially opposing the classifier's attempt to predict incorrect labels while the encoder attempts to deceive it. The distance covariance loss is calculated by the distance covariance decoupling module 260. In each mini-batch (batch size...) ) on, calculate respectively and The distance matrix is ​​calculated, and the distance covariance between the two is minimized. The weights are set to... Once the network converges, freeze all parameters of the teacher network

[200] , retaining only the teacher encoder and the fault potential subspace for subsequent steps.

[0025] Step S05: Construct and train a lightweight student network. Construct a three-layer CNN with only 0.8M parameters (less than 1.5M) and 1.2 GFLOPs of floating-point operations as the student network 300. This network removes the complex decoder and attention modules, focusing on fast inference; the student network is trained using adversarial distillation based on decoupling relationships, such as... Figure 5As shown, a frozen teacher network is used to guide a student network. The same batch of preprocessed samples is simultaneously input into both the teacher and student networks. The calculated loss includes relational knowledge distillation, which is essentially a transfer learning method based on feature manifold topology. Here, angle loss is used to constrain the relative orientation of sample triples in the feature space, and distance loss is used to maintain the relative density between samples. This geometric constraint enables the student network to reconstruct the high-dimensional feature distribution of the teacher network with extremely low parameter count. The weights are set to... The distance loss constraint ensures that the Euclidean distance ratio between sample pairs is consistent, and the temperature parameter is set to... The weights are set to The adversarial loss is applied after student feature 321, followed by a gradient reversal layer and a condition discriminator 330. The flip coefficient of the gradient reversal layer... The training epochs linearly increase from 0 to 1 (for 50 epochs), forcing student features to be inseparable from work conditions. The weights are set to... L2 distillation loss ( ): Optional feature alignment loss is added, weights Finally, minimize the above weighted total loss and update the student network parameters.

[0026] Step S06: Edge Deployment and Online Diagnostics. For example... Figure 6 As shown, the trained student network is deployed to the edge computing device 350. The device collects real-time vibration data, which is preprocessed and then input into the student network. The student network outputs the Softmax probability distribution of the fault categories. The confidence / rejection module 325 checks the maximum probability value. ,like (Set threshold) Determine the corresponding fault category, output fault diagnosis result 340, if If the condition is determined to be "rejected" or "unknown new operating condition", the result display and alarm module 107 will issue a yellow warning, prompting manual review or retraining.

[0027] Those skilled in the art will understand that the specific parameters mentioned in the above embodiments (such as...) (Window length 2048) is only a preferred example; in practical applications, it can be within the scope defined in the claims (e.g., Any adjustments made to these provisions fall within the scope of protection of this invention.

Claims

1. A cross-condition bearing fault diagnosis method based on decoupled antagonistic distillation, characterized in that, The method includes the following steps: S01. Multi-condition vibration data acquisition and preprocessing: Under multiple different working conditions, the vibration signal of the bearing is acquired by the vibration sensor. After the data acquisition module acquires the signal, the preprocessing module filters, segments and normalizes the signal to obtain a preprocessed vibration sample with fault label and working condition label. S02. Construct and pre-train a decoupled teacher network: Using a β-VAE-based encoder-decoder structure, preprocessed samples are mapped into independent fault latent subspaces and operating condition latent subspaces, which are then reconstructed by the decoder. A fault classifier is set at the output of the fault latent subspace for supervised training. The teacher network is trained using a dual-constraint strategy to achieve statistical decoupling between fault features and operating condition disturbances. S021. Set up an adversarial classifier for working conditions, configure it to predict working condition labels from the potential fault subspace, and use a gradient inversion layer to invert the adversarial loss gradient and back-pass it to the encoder to suppress the explicit working condition discrimination information encoded in the potential fault subspace. S022. For each mini-batch, calculate the distance covariance between the fault potential subspace and the operating condition potential subspace, add it to the total loss function and minimize it, eliminate the high-order nonlinear statistical correlation between the two subspaces, and force them to satisfy the statistical independence condition. S03. Construct and distill the student network: Construct a lightweight one-dimensional convolutional neural network as the student network, including a feature extractor, a fault classifier, and a condition discriminator; utilize the frozen teacher network, and constrain the topological structure of the student feature space to be consistent with the geometric structure of the teacher's fault potential subspace through a geometric structure loss function containing angle preservation and normalized distance preservation terms, thereby transferring knowledge of higher-order fault manifolds; at the same time, introduce a gradient reversal layer after the student feature extractor to constrain the student features to be inseparable from the condition discriminator, forming a student adversarial loss; S04. Edge Deployment and Online Diagnosis: The trained student network is deployed to the edge computing device. During the online phase, real-time vibration data is processed in a window and then input into the student network, and the probability of the fault category is output. The diagnostic results are converted into alarm signals or control commands through the industrial control interface and sent to the field control system.

2. The method according to claim 1, characterized in that: In step S01, the preprocessing specifically includes: acquiring data at a sampling rate of not less than 25.6kHz, implementing bandpass filtering from 0.5kHz to 8kHz, and segmenting the data using a sliding window with a length of 1024 or 2048 points and an overlap rate of 50% to 75%.

3. The method according to claim 1, characterized in that: In step S02, the β value of β-VAE ranges from 0.5 to 10; the dimensionality ratio of the fault potential subspace to the operating condition potential subspace is 3:1 to 1:1; when calculating the distance covariance, a mask is used to accumulate only the cross-subspace related terms for the zero-mean potential vectors, and its loss weight... It ranges from 0.001 to 0.

1.

4. The method according to claim 1, characterized in that: In step S03, the angle preservation term is calculated based on sample triples, and the normalized distance preservation term is calculated based on sample pairs; the weights of the two terms in the total loss... and The values ​​range from 0.1 to 2.0; the distance holding parameter uses temperature parameters. Scaling normalization, Values ​​range from 0.5 to 5.

0.

5. The method as described in claim 1, characterized in that: In step S03, the gradient inversion layer coefficients of the student adversarial loss are increased from 0 to 1 with each training epoch, in a piecewise linear growth manner, for a duration of 20 to 100 epochs; the weights of the student adversarial loss... Set to 0.1 to 0.

6.

6. The method according to claim 1, characterized in that: In step S03, the total student loss also includes L2 feature alignment loss, which is used to constrain the mean square error between student features and teacher fault features, and its weights... The range is 0 to 0.

1.

7. The method according to claim 1, characterized in that: The student network contains no fewer than three convolutional blocks and adaptive pooling layers, with ≤1.5M model parameters and ≤2GFLOPs for a single forward floating-point operation; the edge computing device is an embedded processor based on the ARM architecture.

8. The method according to claim 1, characterized in that: In step S04, the category probability output by the student network is temperature scaled and a maximum Softmax probability threshold is set. When the maximum probability is lower than the threshold, it is determined to be a rejection state, and the edge computing device is triggered to output at least one of the following through the industrial control interface: fault diagnosis result, alarm signal or control command.

9. A cross-condition bearing fault diagnosis system based on decoupling relationship to counteract distillation, characterized in that, include: The vibration acquisition unit is used to acquire the vibration signal of the bearing in real time under various working conditions, and to perform filtering and normalization preprocessing on the signal. Memory, used to store computer programs; industrial control interface, used to connect to field control systems and output diagnostic results or control commands; The processor is connected to the vibration acquisition unit, the memory, and the industrial control interface, respectively. The processor is used to execute the computer program stored in the memory to implement the cross-condition bearing fault diagnosis method based on decoupling relationship to counteract distillation as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images