A fault diagnosis method, system and storage medium based on double-branch cooperation

Abstract

The application discloses a kind of based on the fault diagnosis method of double-branch cooperation, comprising the following steps: step S1, multi-source industrial data acquisition and double-flow asymmetric data enhancement preprocessing;Step S2, unsupervised and supervised double-branch feature extraction of cooperation;Step S3, cross-branch consistency cooperation based on dynamic threshold and pseudo-label self-adaptive generation;Step S4, combined loss optimization training of self-adaptive asymmetric knowledge distillation;Step S5, fault diagnosis and health assessment output.The application breaks through semantic barrier by constructing the double-parallel architecture of unsupervised self-supervised contrast representation branch and supervised fine-grained fault classification branch, using the cross-branch pseudo-label generation mechanism of dynamic confidence threshold, and combining the dynamic evolution of double-branch constrained by self-adaptive asymmetric knowledge distillation loss function, to realize high-precision intelligent diagnosis of equipment fault in the industrial scene lacking fault label.

Description

Technical Field

[0001] This invention relates to intelligent fault diagnosis technology based on deep learning. Background Technology

[0002] With the development of modern industry towards automation, intelligence, and continuous operation, real-time health monitoring and fault diagnosis of core industrial equipment such as rotating machinery, CNC machine tools, and wind turbines have become key technologies for ensuring production safety and reducing maintenance costs. In recent years, intelligent fault diagnosis methods based on deep learning have been widely applied in the field of industrial predictive maintenance due to their powerful automatic feature extraction capabilities. For example, the patent publication with publication number CN112084974A, publication date December 15, 2020, entitled "A Multi-Label Rolling Bearing Fault Diagnosis Method Based on Meta-Learning," is a relevant example.

[0003] However, in practical industrial applications, existing intelligent fault diagnosis methods still have the following significant shortcomings and technical bottlenecks:

[0004] 1. Reliance on massive and high-quality labeled fault samples: Existing deep learning fault diagnosis models mostly rely on traditional supervised learning architectures. For these methods to achieve high diagnostic accuracy, they must be based on the premise that all types of fault samples have complete and sufficient labels. However, in real-world factory environments, equipment operates normally most of the time, and faults are rare events. This results in a severely asymmetrical distribution of data collected on-site, characterized by "massive amounts of unlabeled normal data and extremely scarce labeled fault data." Traditional supervised models are prone to overfitting when faced with such extremely imbalanced small sample data, leading to a significant deviation in actual diagnostic performance from current state-of-the-art levels.

[0005] 2. Accumulation and Cascade Amplification of False Label Errors in Semi-Supervised Mechanisms: To utilize massive amounts of unlabeled data, existing technologies often introduce semi-supervised learning or self-training mechanisms. A primary classifier predicts "false labels" for the unlabeled data and incorporates them into the training set. However, in the early stages of training or under strong noise interference, the primary classifier has extremely low recognition accuracy for fault features, easily generating incorrect false labels. This misclassification information is continuously confirmed and amplified during network iteration, leading to a severe "confirmation bias" phenomenon, ultimately causing model training collapse and a surge in both false positive and false negative rates.

[0006] 3. Semantic Discontinuity Between Unsupervised Representation and Supervised Classification: Current research attempts to use unsupervised learning (such as traditional autoencoders and large-scale model self-supervised pre-training) to mine global representations of unlabeled data. However, due to the lack of deterministic labels, pure unsupervised learning tends to learn features that are biased towards macroscopic global statistical regularities (such as being sensitive to normal operating condition boundaries), while failing to capture discriminative features targeting microscopic, fine-grained defects (such as the subtle difference between minute spalling on the inner ring and damage on the outer ring of a bearing). Existing technologies lack an effective mechanism for deep collaboration and dynamic complementarity between unsupervised "globally robust representation capabilities" and supervised "fine-grained fault discrimination capabilities," preventing the integration of their advantages. Summary of the Invention

[0007] The technical problem this invention aims to solve in the health status monitoring of modern industrial equipment is the severe asymmetric distribution of "massive unlabeled normal data and extremely scarce labeled fault data", the easy collapse of existing semi-supervised learning mechanisms due to false labeling, and the semantic gap between pure unsupervised representation networks and supervised classification networks, which makes it impossible to balance global robustness and fine-grained discriminability. The invention proposes a fault diagnosis method based on the collaboration of unsupervised and supervised dual branches.

[0008] To achieve the above objectives, the technical solution adopted by this invention is: a fault diagnosis method based on dual-branch collaboration, comprising the following steps:

[0009] Step S1: Multi-source industrial data acquisition and dual-stream asymmetric data enhancement preprocessing;

[0010] Step S2: Unsupervised and supervised collaborative dual-branch feature extraction;

[0011] Step S3: Cross-branch consistency collaboration and pseudo-label adaptive generation based on dynamic threshold;

[0012] Step S4: Optimize training using the joint loss of adaptive asymmetric knowledge distillation;

[0013] Step S5: Fault diagnosis and health assessment output.

[0014] In step S1, data is first collected synchronously, and then a dual-stream asymmetric data enhancement preprocessing mechanism is used.

[0015] The synchronous data acquisition is to synchronously and continuously acquire multi-source one-dimensional time-series signals from the sensor group to construct a heterogeneous mixed training dataset containing a very small number of labeled fault samples and a large number of unlabeled samples.

[0016] The dual-stream asymmetric data augmentation preprocessing mechanism is designed for the unlabeled samples. By constructing an asymmetric data augmentation operator system, it maps the original one-dimensional time series signal into a spatial multi-view representation with different perturbation intensities.

[0017] The spatial multi-view representation includes the construction of weakly enhanced views and the construction of strongly enhanced views;

[0018] The construction of the weakly enhanced view involves applying a slight geometric perturbation operator to the original one-dimensional time-series signal, dynamically adjusting the signal amplitude through random scaling, and fine-tuning the time sampling rate using Gaussian resampling. This process generates weakly enhanced variant samples while maintaining the basic time-frequency topology and essential semantics of the original one-dimensional time-series signal without distortion. ;

[0019] The construction of the enhanced view involves applying a heavily semantically disruptive perturbation operator to the original one-dimensional time-series signal, using a high-order Gaussian white noise injection operator to simulate extreme environmental noise, and using a random frequency mask to nullify specific characteristic frequency bands in the frequency domain. Furthermore, a time-domain slice disordered recombination operator is combined to break the continuity and phase dependency of local time sequences. Under the premise of forcibly stripping local redundant information and improving the model's generalization barrier, a strongly enhanced variant sample is generated. .

[0020] In step S2, a spatially parallel dual-branch deep neural network topology is constructed, including an unsupervised self-supervised contrastive representation branch and a supervised fine-grained fault classification branch.

[0021] The unsupervised self-supervised contrastive representation branch is named Branch A, and it uses a one-dimensional deep residual network based on a momentum contrastive architecture to represent unlabeled weakly enhanced variant samples. With strongly enhanced variant samples The common input branch A extracts the global degradation feature vector by maximizing the mutual information between different enhanced versions of the same original signal. ;

[0022] The supervised fine-grained fault classification branch is named branch B. It adopts a deep residual classification network with the same structure as branch A but with independent weight parameters. A very small number of labeled fault samples in the dataset are input into branch B. Using the deterministic supervision information of the labels, the samples are directly mapped to the fine-grained physical space of the equipment's defects, and fault feature vectors are extracted. .

[0023] Step S3 includes the following steps:

[0024] Step S3.1, Dynamic Confidence Threshold Calculation: In each batch of model training, the mean overall prediction confidence of the current model for global unlabeled data is statistically calculated in real time. ;

[0025] Step S3.2, Soft Label Prediction and Evaluation: This involves identifying weakly enhanced variant samples from unlabeled samples. Input unsupervised branch A, output its conditional probability distribution score in the representation space. ;

[0026] Step S 3.3, High-purity pseudo-label assignment: Set a confidence level gate if and only if At that time, the one-dimensional one-hot vector with the highest predicted probability in branch A is activated and used as the hard pseudo-label for that sample. ;

[0027] Step S3.4: Consistency regularization constraint, which applies to the corresponding strongly enhanced variant samples. Given a supervised branch B, the classification layer of branch B is forced to predict the hard pseudo-labels generated by branch A. By guiding strong reinforcement with weak reinforcement consistency regularization, robust common representations learned in unsupervised branches are transmitted to supervised branches without loss.

[0028] In step S4, the training phase adopts a multi-stream end-to-end fully optimized mode to construct a multi-task joint loss function. And update the network parameters;

[0029]

[0030] in, For balance coefficient, For supervised classification of losses, For unsupervised self-supervised comparison loss, For adaptive asymmetric knowledge distillation loss.

[0031] Supervised classification loss It is for a very small number of labeled samples, and the standard cross-entropy loss is calculated at the output of branch B;

[0032] The unsupervised self-supervised comparison loss This involves applying a loss function to all samples within a batch, within branch A, to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs.

[0033] The adaptive asymmetric knowledge distillation loss It is used to integrate the physical advantages of the two branches and uses asymmetric KL divergence to constrain the consistency of the soft label probability distribution of the two branches on massive unlabeled data:

[0034]

[0035] in, This is the probability score for the current sample to be judged as having a fault;

[0036] When the equipment is operating in a healthy or slightly degraded phase As the value approaches zero, the loss function degenerates into one dominated by the unsupervised branch A, strongly constraining the supervised branch to deviate from the normal boundary;

[0037] When the equipment is operating in the failure evolution phase As the loss function rapidly increases, it adaptively switches to be dominated by the supervised branch B, which in turn guides the unsupervised branch A to capture micro-distortions.

[0038] In step S5, the unknown operating signal collected by the sensor is directly input into the supervised fine-grained fault classification branch from step S1 to step S4 in real time. The classification layer of branch B directly outputs whether the current device has a fault, the degree of fault deterioration, and the specific fine-grained fault type.

[0039] A fault diagnosis system based on dual-branch collaboration is disclosed. The system includes a sensor group installed on the target industrial equipment. The sensor group is connected to and outputs sensing signals to a processing device. The unlabeled data output by the sensor group is sent to a sample enhancement module of the processing device. The sample enhancement module sends the processed signal to an unsupervised self-supervised comparative characterization branch module. A small amount of labeled data and labeled fault samples output by the sensor group are sent to a supervised fine-grained fault classification branch module of the processing device. Both the unsupervised self-supervised comparative characterization branch module and the supervised fine-grained fault classification branch module are connected to and output the processed data to a fault diagnosis and health assessment output module. The fault diagnosis and health assessment output module outputs the analysis structure to a display device. The fault diagnosis system based on dual-branch collaboration executes the fault diagnosis method based on dual-branch collaboration.

[0040] A storage medium, the storage medium being a computer-readable storage medium for storing software program code for executing the fault diagnosis method based on dual-branch collaboration.

[0041] This invention constructs a dual-parallel architecture of an unsupervised self-supervised comparative representation branch and a supervised fine-grained fault classification branch. It breaks through semantic barriers by using a cross-branch pseudo-label generation mechanism with dynamic confidence thresholds and combines an adaptive asymmetric knowledge distillation loss function to constrain the dynamic evolution of the two branches, thereby achieving highly accurate intelligent diagnosis of equipment faults in industrial scenarios lacking fault labels. Attached Figure Description

[0042] The following is a brief explanation of the content and markings in each of the accompanying drawings in this specification:

[0043] Figure 1 This is a flowchart of the fault diagnosis method based on unsupervised and supervised dual-branch collaboration of the present invention;

[0044] Figure 2 This is a schematic diagram of the dual-branch network structure for unsupervised and supervised collaboration of the present invention. Detailed Implementation

[0045] The following description, with reference to the accompanying drawings, details the specific implementation of the present invention, including the shape and structure of each component, the relative positions and connections between the parts, the function and working principle of each part, the manufacturing process, and the operation and use methods, to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention.

[0046] like Figure 2 As shown, a fault diagnosis system based on dual-branch collaboration can be applied to target industrial equipment as needed. It requires the installation of a sensor group on the target industrial equipment. The types and data of the sensor group can be set as needed. The sensor group needs to synchronously and continuously collect multi-source one-dimensional time-series signals during the operation of the equipment, including but not limited to vibration signals, current signals and audio signals. These signals will be sent to the processing equipment, which is a computer device capable of processing data. The processing equipment includes a sample enhancement module, an unsupervised and self-supervised comparative characterization branch module, a supervised fine-grained fault classification branch module, and a fault diagnosis and health assessment output module.

[0047] Unlabeled data output from the sensor array is sent to the sample augmentation module of the processing device. This module maps the original one-dimensional time-series signal into a spatial multi-view representation with varying perturbation intensities. The sample augmentation module then sets up an unsupervised self-supervised comparative representation branch module for the processed data. A small amount of labeled data and labeled fault samples from the sensor array are sent to the supervised fine-grained fault classification branch module of the processing device. Finally, the unsupervised self-supervised comparative representation branch module and the supervised fine-grained fault classification branch module output the processed data to the fault diagnosis and health assessment output module. This module determines whether the target industrial equipment has a fault, the degree of fault deterioration, and the specific fine-grained fault type. The fault diagnosis and health assessment output module outputs the analysis structure to a display device, which may include, but is not limited to, a display screen, indicator lights, and alarm devices. Simultaneously, the processing device stores all historical data of the target industrial equipment (all data information obtained from the determination of whether a fault exists, the degree of fault deterioration, and the specific fine-grained fault type).

[0048] The processing device contains a storage medium, which is a computer-readable storage medium for storing software program code used to perform fault diagnosis based on dual-branch collaboration.

[0049] The following is a detailed explanation of the fault diagnosis method based on unsupervised and supervised dual-branch collaboration, such as... Figure 1 As shown, the method includes the following steps:

[0050] Step S1: Multi-source industrial data acquisition and dual-stream asymmetric data augmentation and preprocessing

[0051] 1.1 Synchronous Data Acquisition: By deploying sensor arrays on industrial equipment, multi-source one-dimensional time-series signals (including vibration signals, current signals, and audio signals) are collected synchronously and continuously during equipment operation to construct a heterogeneous hybrid training dataset containing a very small number of labeled fault samples and a massive number of unlabeled samples.

[0052] 1.2 Dual-stream Asymmetric Data Augmentation Preprocessing Mechanism: For the unlabeled samples, this invention constructs an asymmetric data augmentation operator system to map the original one-dimensional time-series signal into a spatial multi-view representation with different perturbation intensities. The specific process is as follows:

[0053] Construction of the weakly enhanced view: A slight geometric perturbation operator is applied to the original one-dimensional time-series signal. The signal amplitude is dynamically adjusted through random scaling, and the time sampling rate is fine-tuned by combining Gaussian resampling. While maintaining the basic time-frequency topology and essential semantics of the original one-dimensional time-series signal without distortion, weakly enhanced variant samples are generated. .

[0054] Construction of the enhanced view: A heavily semantically disruptive perturbation operator is applied to the original one-dimensional time-series signal. A high-order Gaussian white noise injection operator simulates extreme environmental noise, and a random frequency mask is used to nullify specific characteristic frequency bands in the frequency domain. Furthermore, a time-domain slice disordered recombination operator is combined to break the continuity and phase dependence of local time sequences. Under the premise of forcibly stripping local redundant information and improving the model's generalization barrier, strongly enhanced variant samples are generated. .

[0055] Step S2: Unsupervised and supervised collaborative bi-branch feature extraction

[0056] Constructing a spatially parallel two-branch deep neural network topology:

[0057] 2.1 Unsupervised Self-Supervised Contrast Representation Branch: This branch, named Branch A, employs a one-dimensional deep residual network based on a momentum contrast architecture. It represents unlabeled, weakly enhanced variant samples... With strongly enhanced variant samples Common input branch A. This branch completely deviates from the dependence on manually labeled data. By maximizing the mutual information between different enhanced versions of the same original signal, it widens the spatial distance between different samples, thereby learning and extracting a global degradation feature vector that has a high generalization and robustness to complex background noise in factories. .

[0058] 2.2 Supervised Fine-Grained Fault Classification Branch: This branch, named Branch B, employs a deep residual classification network with the same structure as Branch A but independent weight parameters. A very small number of labeled fault samples from the dataset are input into Branch B. Utilizing the deterministic supervision information of the labels, the data is directly mapped to the fine-grained physical space of the equipment's defects, extracting fine-grained fault feature vectors that are highly discriminative of the microscopic evolution of defects. .

[0059] Step S3: Cross-branch consistency collaboration based on dynamic threshold and adaptive pseudo-label generation

[0060] To bridge the semantic gap between unsupervised and supervised branches, this invention designs a cross-branch consistency collaborative flow, the implementation of which is as follows:

[0061] Step S3.1, Dynamic Confidence Threshold Calculation: In each batch of model training, the mean overall prediction confidence of the current model for the global unlabeled data is statistically calculated in real time. This threshold rises non-linearly and adaptively as training progresses.

[0062] Step S3.2, Soft Label Prediction and Evaluation: This involves predicting and evaluating weakly enhanced variants of unlabeled samples. Input unsupervised branch A, output its conditional probability distribution score in the representation space. .

[0063] Step S3.3, High-purity pseudo-label assignment: Establish confidence gating. If and only if At that time, the one-dimensional one-hot vector with the highest predicted probability in branch A is activated and used as the hard pseudo-label for that sample. .

[0064] Step S3.4, Consistency Regularization Constraint: For the corresponding strongly enhanced variant samples... Given a supervised branch B, the classification layer of branch B is forced to predict the hard pseudo-labels generated by branch A. By guiding the consistency regularization of strong reinforcement with weak reinforcement, the robust common representations learned in the unsupervised branch are transmitted to the supervised branch without loss.

[0065] Step S4: Optimization Training of Joint Loss Based on Adaptive Asymmetric Knowledge Distillation

[0066] The training phase of this invention employs a multi-stream, end-to-end fully optimized mode to construct a multi-task joint loss function. Update network parameters:

[0067]

[0068] in, This is the balance coefficient.

[0069] 4.1 Supervised Classification Loss For a very small number of labeled samples, the standard cross-entropy loss is calculated at the output of branch B to ensure accurate classification of specific fault types.

[0070] 4.2 Unsupervised vs. Self-Supervised Comparison Loss For all samples within a batch, a loss function is used within branch A to constrain the distance between positive sample pairs and maximize the distance between negative sample pairs, thereby stabilizing the topological boundary of the feature space.

[0071] 4.3 Adaptive Asymmetric Knowledge Distillation Loss To integrate the physical advantages of the two branches, this invention utilizes asymmetric KL divergence to constrain the consistency of the soft label probability distributions of the two branches on massive unlabeled data.

[0072]

[0073] in, This is the probability score for the current sample to be judged as faulty. When the equipment is operating in a healthy or slightly degraded phase, When the value approaches 0, the loss function degenerates into one dominated by the unsupervised branch A, with strong constraints on the supervised branch deviating from the normal boundary; when the equipment is operating in the fault evolution stage... As the loss function rapidly increases, it adaptively switches to be dominated by the supervised branch B, which in turn guides the unsupervised branch A to capture micro-distortions, thus achieving co-evolution of the two branches.

[0074] Step S5: Fault Diagnosis and Health Assessment Output

[0075] The real-time unknown operating signals collected by sensors deployed on the target industrial equipment are directly input into the supervised fine-grained fault classification branch, which has been jointly trained and optimized through steps S1 to S4. After the signals propagate forward through the network, the classification layer of branch B directly outputs whether the equipment has a fault, the degree of fault deterioration, and the specific fine-grained fault type.

[0076] Compared with the prior art, the present invention has the following significant beneficial effects and technical advantages:

[0077] 1. This invention reduces the reliance on manual labeling in industrial fault diagnosis. Traditional supervised learning cannot be trained when fault labels are lacking. By constructing a dual-branch architecture of unsupervised and supervised learning, this invention can transform massive amounts of easily accessible, label-free normal operation data in factories into powerful low-level feature representations. Even in extreme "small sample" scenarios with only a very small number of labeled fault samples, it can still maintain high accuracy in fault diagnosis performance.

[0078] 2. Effectively prevents the accumulation and cascading amplification of pseudo-label errors in semi-supervised learning. Traditional semi-supervised methods often introduce "incorrect labels" when using pseudo-labels, leading to model collapse due to inaccurate initial classifiers. This invention significantly improves the purity of pseudo-labels through a dynamic confidence threshold and a mechanism of "weak enhancement generating pseudo-labels and strong enhancement forcing predictions." Combined with a dual-branch decoupling parameter design, the unsupervised branch acts as a "stabilizer," effectively preventing the cascading amplification of misclassified information in the network.

[0079] 3. By employing asymmetric knowledge distillation, this invention achieves complementary advantages and co-evolution between two branches. It innovatively designs an adaptive asymmetric knowledge distillation loss. This design fully leverages the respective physical advantages of "unsupervised learning being sensitive to normal operating conditions" and "supervised learning being sensitive to fault characteristics." Through decoupled KL divergence dynamic weights, the two branches can act as "mentors" to each other at different degradation stages, achieving true co-evolution between the two branches and enabling the model to capture unseen faults under varying operating conditions.

[0080] The present invention has been described above with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A fault diagnosis method based on dual-branch collaboration, characterized in that, Includes the following steps: Step S1: Multi-source industrial data acquisition and dual-stream asymmetric data enhancement preprocessing; Step S2: Unsupervised and supervised collaborative dual-branch feature extraction; Step S3: Cross-branch consistency collaboration and pseudo-label adaptive generation based on dynamic threshold; Step S4: Optimize training using the joint loss of adaptive asymmetric knowledge distillation; Step S5: Fault diagnosis and health assessment output.

2. The fault diagnosis method based on dual-branch collaboration according to claim 1, characterized in that, In step S1, data is first collected synchronously, and then a dual-stream asymmetric data enhancement preprocessing mechanism is used. The synchronous data acquisition is to synchronously and continuously acquire multi-source one-dimensional time-series signals from the sensor group to construct a heterogeneous mixed training dataset containing a very small number of labeled fault samples and a large number of unlabeled samples. The dual-stream asymmetric data augmentation preprocessing mechanism is designed for the unlabeled samples. By constructing an asymmetric data augmentation operator system, it maps the original one-dimensional time series signal into a spatial multi-view representation with different perturbation intensities.

3. The fault diagnosis method based on dual-branch collaboration according to claim 2, characterized in that, The spatial multi-view representation includes the construction of weakly enhanced views and the construction of strongly enhanced views; The construction of the weakly enhanced view involves applying a slight geometric perturbation operator to the original one-dimensional time-series signal, dynamically adjusting the signal amplitude through random scaling, and fine-tuning the time sampling rate using Gaussian resampling. This process generates weakly enhanced variant samples while maintaining the basic time-frequency topology and essential semantics of the original one-dimensional time-series signal without distortion. ; The construction of the enhanced view involves applying a heavily semantically disruptive perturbation operator to the original one-dimensional time-series signal, using a high-order Gaussian white noise injection operator to simulate extreme environmental noise, and using a random frequency mask to nullify specific characteristic frequency bands in the frequency domain. Furthermore, a time-domain slice disordered recombination operator is combined to break the continuity and phase dependency of local time sequences. Under the premise of forcibly stripping local redundant information and improving the model's generalization barrier, a strongly enhanced variant sample is generated. .

4. The fault diagnosis method based on dual-branch collaboration according to claim 3, characterized in that, In step S2, a spatially parallel dual-branch deep neural network topology is constructed, including an unsupervised self-supervised contrastive representation branch and a supervised fine-grained fault classification branch. The unsupervised self-supervised contrastive representation branch is named Branch A, and it uses a one-dimensional deep residual network based on a momentum contrastive architecture to represent unlabeled weakly enhanced variant samples. With strongly enhanced variant samples The common input branch A extracts the global degradation feature vector by maximizing the mutual information between different enhanced versions of the same original signal. ; The supervised fine-grained fault classification branch is named branch B. It adopts a deep residual classification network with the same structure as branch A but with independent weight parameters. A very small number of labeled fault samples in the dataset are input into branch B. Using the deterministic supervision information of the labels, the samples are directly mapped to the fine-grained physical space of the equipment's defects, and fault feature vectors are extracted. .

5. The fault diagnosis method based on dual-branch collaboration according to claim 4, characterized in that, Step S3 includes the following steps: Step S3.1, Dynamic Confidence Threshold Calculation: In each batch of model training, the mean overall prediction confidence of the current model for global unlabeled data is statistically calculated in real time. ; Step S3.2, Soft Label Prediction and Evaluation: This involves identifying weakly enhanced variant samples from unlabeled samples. Input unsupervised branch A, output its conditional probability distribution score in the representation space. ; Step S 3.3, High-purity pseudo-label assignment: Set a confidence level gate if and only if At that time, the one-dimensional one-hot vector with the highest predicted probability in branch A is activated and used as the hard pseudo-label for that sample. ; Step S3.4: Consistency regularization constraint, which applies to the corresponding strongly enhanced variant samples. Given a supervised branch B, the classification layer of branch B is forced to predict the hard pseudo-labels generated by branch A. By guiding strong reinforcement with weak reinforcement consistency regularization, robust common representations learned in unsupervised branches are transmitted to supervised branches without loss.

6. The fault diagnosis method based on dual-branch collaboration according to claim 5, characterized in that, In step S4, the training phase adopts a multi-stream end-to-end fully optimized mode to construct a multi-task joint loss function. And update the network parameters; in, For balance coefficient, For supervised classification of losses, For unsupervised self-supervised comparison loss, For adaptive asymmetric knowledge distillation loss.

7. The fault diagnosis method based on dual-branch collaboration according to claim 6, characterized in that: Supervised classification loss It is for a very small number of labeled samples, and the standard cross-entropy loss is calculated at the output of branch B; The unsupervised self-supervised comparison loss This involves applying a loss function to all samples within a batch, within branch A, to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. The adaptive asymmetric knowledge distillation loss It is used to integrate the physical advantages of the two branches and uses asymmetric KL divergence to constrain the consistency of the soft label probability distribution of the two branches on massive unlabeled data: in, This is the probability score for the current sample to be judged as having a fault; When the equipment is operating in a healthy or slightly degraded phase As the value approaches zero, the loss function degenerates into one dominated by the unsupervised branch A, strongly constraining the supervised branch to deviate from the normal boundary; When the equipment is operating in the failure evolution phase As the loss function rapidly increases, it adaptively switches to be dominated by the supervised branch B, which in turn guides the unsupervised branch A to capture micro-distortions.

8. The fault diagnosis method based on dual-branch collaboration according to any one of claims 1-7, characterized in that: In step S5, the unknown operating signal collected by the sensor is directly input into the supervised fine-grained fault classification branch from step S1 to step S4 in real time. The classification layer of branch B directly outputs whether the current device has a fault, the degree of fault deterioration, and the specific fine-grained fault type.

9. A fault diagnosis system based on dual-branch collaboration, characterized in that, The system includes a sensor array installed on the target industrial equipment. The sensor array is connected to and outputs sensing signals to a processing device. The unlabeled data output by the sensor array is sent to a sample enhancement module of the processing device. The sample enhancement module sends the processed signal to an unsupervised self-supervised comparative characterization branch module. A small amount of labeled data and labeled fault samples output by the sensor array are sent to a supervised fine-grained fault classification branch module of the processing device. Both the unsupervised self-supervised comparative characterization branch module and the supervised fine-grained fault classification branch module are connected to and output the processed data to a fault diagnosis and health assessment output module. The fault diagnosis and health assessment output module outputs the analysis structure to a display device. The fault diagnosis system based on dual-branch collaboration executes the fault diagnosis method based on dual-branch collaboration as described in any one of claims 1-8.

10. A storage medium, said storage medium being a computer-readable storage medium for storing software program code, characterized in that: The software program code is used to execute the fault diagnosis method based on dual-branch collaboration as described in any one of claims 1-8.

Images