A method and device for continuous passive domain adaptation fault diagnosis
By constructing a fault diagnosis model based on the average teacher architecture, and utilizing confidence-weighted knowledge consistency regularization, Mixup data augmentation, and mutual information maximization strategies, the persistent domain drift and catastrophic forgetting problems of rotating machinery equipment are solved, achieving efficient fault diagnosis under passive domain adaptation.
Patent Information
- Application Number
- CN202610279397.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-09
- Publication Date
- 2026-06-26
AI Technical Summary
Existing fault diagnosis models, when aided by passive domain data of rotating machinery, struggle to address the issues of persistent distribution drift and catastrophic forgetting, leading to decreased diagnostic accuracy and false alarms/missed alarms. In particular, they are unable to maintain robustness under complex operating conditions.
We employ confidence-weighted knowledge consistency regularization, Mixup data augmentation, mutual information maximization, and stochastic recovery strategies, combined with a progressive hyperparameter annealing strategy, to construct a fault diagnosis model based on an average teacher architecture, achieving continuous passive domain adaptation.
Without accessing source domain data, the model can quickly adapt to new operating conditions, maintain its ability to recognize old operating conditions, and possess anti-forgetting and robustness, thus achieving long-term accurate fault diagnosis.
Smart Images
Figure CN122286471A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of industrial equipment condition monitoring, fault diagnosis, and deep learning, and particularly to a continuous passive domain adaptive fault diagnosis method and apparatus. Background Technology
[0002] Intelligent fault diagnosis technology is a key core technology for achieving predictive maintenance and safe and reliable operation of modern industrial systems. In recent years, with the rapid development of deep learning technology, deep learning-based intelligent diagnostic methods have demonstrated superior performance in fault mode recognition compared to traditional signal processing methods, particularly in the condition monitoring and fault diagnosis of typical rotating machinery such as bearings, gearboxes, motors, pumps, and fans. However, these deep learning-based fault diagnosis models are usually based on the ideal assumption of independent and identically distributed data, which requires that the historical data (source domain) used to train the model and the test data (target domain) used in actual deployment follow the same probability distribution.
[0003] In real-world, complex industrial applications, this ideal assumption often proves untenable. Rotating machinery typically operates under complex conditions such as variable speed, variable load, frequent start-stop cycles, and coupled transmission chains. Monitoring signals, including vibration and current, exhibit strong non-stationarity. These dynamic conditions cause the statistical distribution of the collected monitoring data to drift continuously over time, a phenomenon known as "persistent domain drift." When a model trained under a specific operating condition (source domain) is directly applied to a new, unseen operating condition (target domain), the model's diagnostic accuracy drops sharply due to feature distribution mismatch, leading to severe false alarms and missed alarms. This poses significant safety hazards and economic losses to industrial production.
[0004] To address the issue of inconsistent data distribution, existing technical solutions include: The first category is unsupervised domain adaptation. This type of method aligns the feature distributions of the source and target domains. However, a hard constraint of unsupervised domain adaptation methods is that both labeled source domain data and unlabeled target domain data must be accessed simultaneously during the adaptation phase. In practical industrial deployments, source domain data often involves trade secrets or sensitive information (such as in the nuclear power and military industries), and is subject to strict privacy protection restrictions.
[0005] The second type is passive domain adaptation. To avoid dependence on source domain data, passive domain adaptation methods adapt using only pre-trained model parameters and the current target domain data. While this solves the data privacy problem, existing passive domain adaptation methods are one-off adaptation schemes designed for a single, fixed target domain. However, the operating conditions of industrial equipment often change in a continuous sequence. For rotating machinery, changes in speed / load between different operating conditions can significantly alter the energy distribution and spectral structure of fault characteristics. Simply reusing passive domain adaptation methods will cause the model to rapidly destroy its original parameter structure when adapting to the new operating conditions, resulting in the model completely losing its ability to identify faults under old operating conditions. This phenomenon is known as catastrophic forgetting.
[0006] Therefore, there is an urgent need for a fault diagnosis method that can, under the premise of strictly satisfying the participation of passive domain data (protecting privacy and reducing storage), quickly and accurately adapt to new operating conditions through efficient feature alignment strategies, effectively overcome catastrophic forgetting through explicit memory retention mechanisms, and maintain long-term robustness for the characteristics of rotating machinery such as non-stationarity and multi-condition switching. Summary of the Invention
[0007] To address the shortcomings of existing fault diagnosis methods in resolving persistent distribution drift and catastrophic forgetting with the aid of passive domain data for rotating machinery, this invention proposes a continuous passive domain adaptive fault diagnosis method and apparatus for rotating machinery. The specific technical solution is as follows: A continuous passive domain adaptive fault diagnosis method includes the following steps: S1: Obtain the labeled source domain datasets of rotating machinery under different operating conditions, and a series of unlabeled target domain datasets arriving in chronological order; S2: Construct a fault diagnosis model and train it on the labeled source domain dataset by minimizing the supervised cross-entropy loss function to obtain initial model parameters with initial feature extraction capability. S3: Construct an adaptive network based on an average teacher architecture, comprising student and teacher models with identical structures, and initialize the student and teacher models with the initial model parameters; employ a confidence-weighted knowledge consistency regularization strategy, utilizing the soft probability distribution generated by the teacher model as knowledge, combined with confidence filtering and Mixup data augmentation, to guide student model updates; adopt a mutual information maximization strategy, enhancing the plasticity of the student model by minimizing prediction entropy and maximizing prediction diversity; simultaneously employ a stochastic recovery strategy, explicitly retaining old knowledge by randomly rolling back some teacher model parameters to the previous stage; and dynamically adjust key hyperparameters in the adaptation process using a progressive hyperparameter annealing strategy, thereby enabling the fault diagnosis model to learn the target domain; S4: Apply the fault diagnosis model that has completed target domain learning in S3 to actual fault diagnosis.
[0008] Furthermore, the formula for the supervised cross-entropy loss function in S2 is as follows: in, As an indicator function, when the sample y i The real label is c The value is 1 if it is true, and 0 otherwise. p i,c For the model on sample X i Category c The predicted probability.
[0009] Furthermore, in S3, a confidence-weighted knowledge consistency regularization strategy is adopted, utilizing the soft probability distribution generated by the teacher model as knowledge, combined with confidence screening and Mixup data augmentation, to guide student model updates. Specifically, this includes: (1) Using the teacher model to process the input sample x i Perform prediction and output the logits vector z. i The soft probability distribution k is obtained by temperature scaling. i k i The c The calculation for each element is as follows: in, For temperature parameters; (2) Calculate the maximum probability And calculate the confidence weights accordingly. : in, c The confidence threshold; (3) For the input sample x i ,x j And the soft probability distribution (knowledge) of teachers k i ,k j Perform Mixup enhancement to generate virtual samples and mixed knowledge : in The mixing coefficient; (4) Calculate the consistency regularization loss of the student model. : in, For student models to virtual samples The predicted output probability.
[0010] Furthermore, the loss function corresponding to the mutual information maximization strategy Minimize loss by entropy and diversity maximization loss Composition, expressed as follows: in, The student model is applied to sample x. i Category c The predicted probability, It is a category c Edge distribution on the current target domain.
[0011] Furthermore, the first t The overall optimization objective function of the student model during the adaptation process in each target domain is... For consistency regularization loss The loss function corresponding to the mutual information maximization strategy sum.
[0012] Furthermore, in S3, the parameter update method for the teacher model is exponential moving average update, with the specific formula as follows: in, For the first t Teacher model parameters for each stage For the first t Student model parameters at each stage m To update the momentum coefficient.
[0013] Furthermore, the specific process of the random recovery strategy in S3 is as follows: Generate a random mask matrix M that follows a Bernoulli distribution, where each element has a recovery probability of b The value is 1: After updating the teacher model using the exponential moving average, the teacher model parameters are adjusted according to the following formula. i t : in, i t-1These are the teacher model parameters saved at the end of the previous adaptation phase, where I is a matrix of all 1s. This indicates an element-wise multiplication operation.
[0014] Furthermore, in S3, the key hyperparameters in the dynamic adjustment and adaptation process of the progressive hyperparameter annealing strategy specifically include: Each hyperparameter varies with training epochs. e According to the cosine annealing strategy, from the initial value Change to target value The calculation formula is: in, Momentum representing the teacher model m Recovery probability b And the confidence threshold of the student model c Any one of these three hyperparameters, E This refers to the total number of training rounds.
[0015] A continuous passive domain adaptation fault diagnosis device is used to implement a continuous passive domain adaptation fault diagnosis method, comprising: a fault database acquisition module, a source domain pre-training module, a target domain adaptation module, and a fault diagnosis module; The fault database acquisition module is used to acquire vibration signals of rotating machinery and divide them into a labeled source domain dataset and multiple unlabeled target domain datasets arranged in chronological order. The source domain pre-training module is used to construct a fault diagnosis model and uses supervised cross-entropy loss to train the fault diagnosis model on the labeled source domain dataset to obtain initial model parameters with initial feature extraction capability. The target domain adaptation module is used to construct an adaptation network based on an average teacher architecture, containing student and teacher models with identical structures, and initializes the student and teacher models with the initial model parameters. It employs a confidence-weighted knowledge consistency regularization strategy, utilizing the soft probability distribution generated by the teacher model as knowledge, combined with confidence filtering and Mixup data augmentation, to guide student model updates. A mutual information maximization strategy is used to enhance the plasticity of the student model by minimizing prediction entropy and maximizing prediction diversity. Simultaneously, a stochastic recovery strategy is employed, explicitly retaining old knowledge by randomly rolling back some teacher model parameters to the previous stage. Furthermore, a progressive hyperparameter annealing strategy is used to dynamically adjust key hyperparameters during the adaptation process, thereby enabling the fault diagnosis model to learn the target domain. The fault diagnosis module is used to apply the fault diagnosis model that has completed target domain learning to actual fault diagnosis.
[0016] The beneficial effects of this invention are as follows: The fault diagnosis method provided by this invention fully considers the problem of continuous data distribution drift caused by dynamic operating conditions in actual industrial scenarios. It adopts a continuous passive domain adaptation paradigm, which can continuously adapt to continuously changing unlabeled target domain data based solely on pre-trained model parameters. This invention does not require access to source domain or historical target domain data during the adaptation phase, and under the premise of strictly meeting the data privacy protection and storage space limitations of industrial scenarios, it does not require the memorization and replay of historical data.
[0017] This invention addresses the challenge of balancing stability and plasticity by constructing a progressively coordinated average teacher model. Specifically, the confidence-weighted knowledge consistency regularization strategy effectively suppresses pseudo-label noise in unsupervised environments; the mutual information maximization strategy endows the model with a keen ability to capture new situations; the stochastic recovery strategy locks in historical memory through an explicit parameter rollback mechanism; and the progressive hyperparameter annealing strategy achieves dynamic optimization throughout the entire process. The synergistic effect of these strategies enables the model to accurately diagnose current faults and firmly grasp historical experience during long-term dynamic operation, exhibiting strong anti-forgetting capabilities and robustness. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the continuous passive domain adaptive fault diagnosis method in an embodiment of the present invention.
[0019] Figure 2 This is a schematic diagram of a rotating mechanical device as an example. Detailed Implementation
[0020] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.
[0021] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be subject to the following interpretations.
[0022] (1) Source Domain: This refers to the dataset or domain in which the model is pre-trained. The source domain typically contains a large number of data samples with explicit labels or supervision information. These samples are used to train and build the initial model, enabling the model to learn effective features and patterns. The source domain can be a task, a domain, or a dataset, depending on the application scenario.
[0023] The target domain refers to the new dataset or domain to which the model will be applied. Typically, the target domain contains only a small number of labeled samples or none at all (i.e., in unsupervised learning). In this invention, the target domain contains no labeled samples whatsoever. The target domain can be a task, domain, or dataset that differs from the source domain.
[0024] In transfer learning and domain adaptation, the source and target domains may differ in terms of data distribution, feature distribution, and category distribution. This invention addresses the difference in data distribution. The source domain is the starting point for knowledge transfer; the knowledge and features learned by the model in the source domain are transferred to the target domain to improve the model's performance in the target domain.
[0025] (2) Source-free domain adaptation refers to the process where the model is trained on a pre-trained source domain model using only unlabeled target domain data to adapt to the new target domain data distribution. This process does not allow access to any source domain data.
[0026] (3) Catastrophic Forgetting: When a neural network learns a new task or adapts to a new distribution, the parameters change, resulting in a significant decrease in its performance on the old task or distribution.
[0027] (4) Continual Learning: When building a unified network model or gradually adding new features to a network model, the usual assumption is that training data for all tasks is always available; however, as the number of tasks increases, storing and retraining such data becomes impractical. To address the problem that training data for existing features is unavailable when adding new features to a network model, the continuous learning method is proposed. This method trains the network using only new task data while retaining the original features.
[0028] This invention proposes a continuous passive domain adaptive fault diagnosis method for rotating machinery under dynamic operating conditions, such as... Figure 1 As shown, the specific steps include steps one through four.
[0029] Step 1: Collect vibration signals of rotating machinery under different operating conditions and create a labeled source domain dataset of the rotating machinery; obtain a series of unlabeled target domain datasets that arrive in chronological order.
[0030] The labeled source domain dataset is represented as follows: In the formula, x i This represents the i-th fault sample.y i for x i The corresponding tags This represents the number of samples in the source domain. This represents the distribution followed by the faulty sample X in the source domain S. This represents the distribution that the label Y of the source domain S follows.
[0031] Unlabeled T The dataset of each target domain is represented as follows: , of which t Data from each target domain Represented as: In the formula, express The number of unlabeled samples included. This represents the fault sample set of the t-th target domain. X The distribution it follows.
[0032] In this embodiment, the data distributions of the source domain and each target domain are different, and the model adapts to the first... When targeting a single target domain, data from the source domain and previous target domains cannot be accessed.
[0033] Step 2: Source Domain Pre-training. Train the fault diagnosis model on the source domain dataset. Use supervised cross-entropy loss to train the fault diagnosis model on the labeled source domain dataset to obtain pre-trained model parameters with initial feature extraction capabilities.
[0034] Fault diagnosis model , C For the number of categories, These are model parameters. Diagnostic model. First in the source domain Pre-training is performed on the model to learn the parameters of the source domain model. Subsequently, in the serialized unlabeled target domain Adaptive training will be conducted on them.
[0035] Fault diagnosis model in the source domain During pre-training, the supervised cross-entropy loss function is minimized. To optimize the model parameters, the formula is as follows: in, As an indicator function, when the sample y i The real label is c The value is 1 if it is true, and 0 otherwise. p i,cFor the model on sample X i Category c The predicted probability.
[0036] Step 3: Sequence Target Domain Adaptation. An adaptation network based on an average teacher architecture is constructed, containing student and teacher models with identical structures. A confidence-weighted knowledge consistency regularization strategy is employed, utilizing the soft probability distribution generated by the teacher model as knowledge. This is combined with confidence filtering and Mixup data augmentation to guide student model updates. A mutual information maximization strategy is used to enhance the plasticity of the student model by minimizing prediction entropy and maximizing prediction diversity. Simultaneously, a stochastic recovery strategy is employed, explicitly retaining old knowledge by randomly rolling back some teacher model parameters to the previous stage. Furthermore, a progressive hyperparameter annealing strategy is used to dynamically adjust key hyperparameters during the adaptation process, thereby enabling the fault diagnosis model to learn the target domain.
[0037] Adaptive networks based on average teacher architecture include student models with identical structures. Teacher Model The parameters all use the final model parameters of the fault diagnosis model learned in the previous stage. Perform initialization , The student model is responsible for rapid learning (plasticity), while the teacher model is responsible for slow integration (stability).
[0038] (3.1) Training and updating of the student model The confidence-weighted knowledge consistency regularization strategy specifically includes: Using the teacher model to analyze the input sample x i Perform prediction and output the logits vector z. i The soft probability distribution k is obtained by temperature scaling. i k i The c The calculation for each element is as follows: in, For temperature parameters; Calculate the maximum probability And calculate the confidence weights accordingly. : in, c This is the confidence threshold.
[0039] To enhance robustness, for input sample x i ,x j And the soft probability distribution (knowledge) of teachers k i ,k jPerform Mixup enhancement to generate virtual samples and mixed knowledge : in This is the mixing coefficient.
[0040] Calculate the consistency regularization loss of the student model. : in, For student models to virtual samples The predicted output probability.
[0041] The loss function of the mutual information maximization strategy is the entropy minimization loss. and diversity maximization loss Composition, expressed as follows: in, The student model is applied to sample x. i Category c The predicted probability, It is a category c Edge distribution on the current target domain.
[0042] The total mutual information loss is: Maximizing mutual information enhances the plasticity of student models, enabling them to quickly adapt to the current domain distribution.
[0043] No. t The overall objective function for the student model during the adaptation process of each target domain is: The student model parameters are updated using stochastic gradient descent.
[0044] (3.2) Update and revision of the teacher model The teacher model's parameters are updated using an exponential moving average, with the specific formula as follows: in, For the first t Teacher model parameters for each stage For the first t Student model parameters at each stage m To update the momentum coefficient.
[0045] The specific process of the random recovery strategy is as follows: Generate a random mask matrix M that follows a Bernoulli distribution, where each element has a recovery probability of b The value is 1: After updating the teacher model using the exponential moving average, the teacher model parameters are adjusted according to the following formula. i t : in, i t-1 These are the teacher model parameters saved at the end of the previous adaptation phase, where I is a matrix of all 1s. This indicates an element-wise multiplication operation.
[0046] (3.3) Hyperparameter update of student and teacher models A progressive hyperparameter annealing strategy is used to update the hyperparameter momentum of the teacher model using exponential moving average. m Recovery probability b And the hyperparameter confidence threshold of the student model. c .
[0047] The above hyperparameters vary with training epochs. e According to the cosine annealing strategy, from the initial value Change to target value The calculation formula is: in, represent m , c or b Any one of the hyperparameters in, E This represents the total number of training epochs. This approach means the model is more adaptable in the early stages of adaptation, while focusing more on stability in later stages.
[0048] Step 4: Apply the fault diagnosis model learned in Step 3 to actual fault diagnosis.
[0049] To verify the fault diagnosis model's ability to continuously adapt to new domains and overcome catastrophic forgetting, the performance of the fault diagnosis model obtained in step three can be tested on both the old and new domains using data from all the learned domains. The details are as follows: The fault diagnosis model completes the first t After adaptation to the target domain, it needs to perform well on all learned domains; therefore, the test data should include... The data above was used to verify the fault diagnosis model's ability to continuously adapt to different domains and overcome catastrophic forgetting. It was assumed that each domain (including the source domain and the target domain, where the source domain is denoted as domain 0 and the target domains are arranged sequentially) corresponds to a training phase. After training in each phase, the fault diagnosis model was tested on all domains. In this case, an accuracy matrix can be obtained to evaluate the performance of the fault diagnosis model across all domains, and the accuracy matrix is represented as follows: , T Indicates the total number of target domains.
[0050] Specific evaluation metrics include final performance. Average performance In the formula, R T,t Indicates in T Training on the domain and t The accuracy of the fault diagnosis model tested on the domain. R t,s Indicates in t Training on the domain and s The accuracy of the fault diagnosis model tested on the domain.
[0051] The following combination Figure 2 The superiority of the method of the present invention is illustrated by obtaining a bearing fault diagnosis dataset of the rotating mechanical equipment shown and simulating a scenario of continuous data distribution drift.
[0052] This bearing fault diagnosis dataset contains 10 fault types and 24 different operating conditions (load and speed variations). As shown in Table 1, 6 of these operating conditions were used in this study, and a sequence scenario containing 6 different operating conditions was set up: 1. 2. 3. 4. 5. 6. .
[0053] Table 1 Six operating conditions used As shown in Table 2, the experiment was repeated ten times under six different operating conditions to calculate the final and average performance of the method of this invention and five commonly used continuous passive domain adaptation methods (i.e., TENT, CoTTA, GSFDA, UCSN, and CoSDA).
[0054] As shown in Table 2, the method of the present invention achieves continuous passive domain adaptive fault diagnosis by effectively adapting to the new target domain while maintaining the performance in the historical domain, without source domain data.
[0055] Table 2 Diagnostic results under persistent data distribution drift It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.
Claims
1. A continuous passive domain adaptive fault diagnosis method, characterized in that, Includes the following steps: S1: Obtain the labeled source domain dataset under different operating conditions of rotating machinery, and a series of unlabeled target domain datasets arriving in chronological order; S2: Construct a fault diagnosis model and train it on the labeled source domain dataset by minimizing the supervised cross-entropy loss function to obtain initial model parameters with initial feature extraction capability. S3: Construct an adaptive network based on an average teacher architecture, comprising student and teacher models with identical structures, and initialize the student and teacher models with the initial model parameters; employ a confidence-weighted knowledge consistency regularization strategy, utilizing the soft probability distribution generated by the teacher model as knowledge, combined with confidence filtering and Mixup data augmentation, to guide student model updates; adopt a mutual information maximization strategy, enhancing the plasticity of the student model by minimizing prediction entropy and maximizing prediction diversity; simultaneously employ a stochastic recovery strategy, explicitly retaining old knowledge by randomly rolling back some teacher model parameters to the previous stage; and dynamically adjust key hyperparameters in the adaptation process using a progressive hyperparameter annealing strategy, thereby enabling the fault diagnosis model to learn the target domain; S4: Apply the fault diagnosis model that has completed target domain learning in S3 to actual fault diagnosis.
2. The continuous passive domain adaptive fault diagnosis method according to claim 1, characterized in that, The formula for the supervised cross-entropy loss function in S2 is as follows: ； in, As an indicator function, when the sample y i The real label is c The value is 1 if the condition is met, and 0 otherwise. p i,c For the model on sample X i Category c The predicted probability.
3. The continuous passive domain adaptive fault diagnosis method according to claim 1, characterized in that, The confidence-weighted knowledge consistency regularization strategy in S3 utilizes the soft probability distribution generated by the teacher's model as knowledge, combined with confidence screening and Mixup data augmentation, to guide student model updates. Specifically, it includes: (1) Using the teacher model to process the input sample x i Perform prediction and output the logits vector z. i The soft probability distribution k is obtained by temperature scaling. i k i The c The calculation for each element is as follows: ； in, For temperature parameters; (2) Calculate the maximum probability And calculate the confidence weights accordingly. : ； in, γ The confidence threshold; (3) For the input sample x i ,x j And the soft probability distribution (knowledge) of teachers k i ,k j Perform Mixup enhancement to generate virtual samples and mixed knowledge : ； ； in The mixing coefficient; (4) Calculate the consistency regularization loss of the student model. : ； in, For student models to virtual samples The predicted output probability.
4. The continuous passive domain adaptive fault diagnosis method according to claim 3, characterized in that, The loss function corresponding to the mutual information maximization strategy Minimize loss by entropy and diversity maximization loss Composition, expressed as follows: ； ； ； in, The student model is applied to sample x i Category c The predicted probability, It is a category c Edge distribution on the current target domain.
5. The continuous passive domain adaptive fault diagnosis method according to claim 4, characterized in that, The first t The overall optimization objective function of the student model during the adaptation process in each target domain is... Consistency regularization loss The loss function corresponding to the mutual information maximization strategy sum.
6. The continuous passive domain adaptive fault diagnosis method according to claim 1, characterized in that, In S3, the parameter update method for the teacher model is exponential moving average update, and the specific formula is as follows: in, For the first t Teacher model parameters for each stage For the first t Student model parameters at each stage m To update the momentum coefficient.
7. The continuous passive domain adaptive fault diagnosis method according to claim 1, characterized in that, In S3, the specific process of the random recovery strategy is as follows: Generate a random mask matrix M that follows a Bernoulli distribution, where each element has a recovery probability of b The value is 1: After updating the teacher model using the exponential moving average, the teacher model parameters are adjusted according to the following formula. θ t : in, θ t-1 These are the teacher model parameters saved at the end of the previous adaptation phase, where I is a matrix of all 1s. This indicates an element-wise multiplication operation.
8. The continuous passive domain adaptive fault diagnosis method according to claim 1, characterized in that, In S3, the key hyperparameters in the dynamic adjustment and adaptation process of the progressive hyperparameter annealing strategy specifically include: Each hyperparameter varies with training epochs. e According to the cosine annealing strategy, from the initial value Change to target value The calculation formula is: in, Momentum representing the teacher model m Recovery probability b And the confidence threshold of the student model γ Any one of these three hyperparameters, E This refers to the total number of training rounds.
9. A continuous passive domain adaptive fault diagnosis device, used to implement the continuous passive domain adaptive fault diagnosis method according to any one of claims 1 to 9, characterized in that, include: The system includes a fault database acquisition module, a source domain pre-training module, a target domain adaptation module, and a fault diagnosis module. The fault database acquisition module is used to acquire vibration signals of rotating machinery and divide them into a labeled source domain dataset and multiple unlabeled target domain datasets arranged in chronological order. The source domain pre-training module is used to construct a fault diagnosis model and uses supervised cross-entropy loss to train the fault diagnosis model on the labeled source domain dataset to obtain initial model parameters with initial feature extraction capability. The target domain adaptation module is used to construct an adaptation network based on an average teacher architecture, containing student and teacher models with identical structures, and initializes the student and teacher models with the initial model parameters. It employs a confidence-weighted knowledge consistency regularization strategy, utilizing the soft probability distribution generated by the teacher model as knowledge, combined with confidence filtering and Mixup data augmentation, to guide student model updates. A mutual information maximization strategy is used to enhance the plasticity of the student model by minimizing prediction entropy and maximizing prediction diversity. Simultaneously, a stochastic recovery strategy is employed, explicitly retaining old knowledge by randomly rolling back some teacher model parameters to the previous stage. Furthermore, a progressive hyperparameter annealing strategy is used to dynamically adjust key hyperparameters during the adaptation process, thereby enabling the fault diagnosis model to learn the target domain. The fault diagnosis module is used to apply the fault diagnosis model that has completed target domain learning to actual fault diagnosis.