Double-step collaborative semi-supervised fault diagnosis method and system

Abstract

The invention provides a double-step collaborative semi-supervised fault diagnosis method and system, and relates to the technical field of fault diagnosis, and the method comprises the steps: constructing a collaborative feature guided learning architecture; the collaborative feature guide type learning architecture comprises a feature extractor module, a discriminator module and a historical steady state guide module; jointly training a collaborative feature guided learning architecture through the data with the labels and the data without the labels to obtain an initial diagnosis model; in the fine tuning stage, the parameters of the feature extractor module are fixed, only the parameters of the discriminator module are optimized to adapt to the target fault category, and a fine-tuned fault diagnosis model is obtained; and performing fault diagnosis on the equipment monitoring data based on the trained model, and outputting a fault type classification result. According to the method, the problems that an existing semi-supervised fault diagnosis method is unstable in training, large in parameter fluctuation and fuzzy in judgment boundary under the condition of few samples are effectively solved, and the training stability and generalization ability of the model under the condition of extremely uneven data categories are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a two-step collaborative semi-supervised fault diagnosis method and system. Background Art

[0002] Fault diagnosis in industrial equipment has always been a key issue in ensuring reliable system operation. Condition monitoring of large electrical equipment, such as power transformers, often involves analyzing sensor data collected during operation to determine the type of internal fault. For example, oil-gas analysis technology is widely used in transformers, inferring the type of fault based on the composition and concentration of dissolved gases in transformer oil. This type of diagnostic approach, based on monitoring indicators, has achieved early success. Different faults generate distinct characteristic signals, and analyzing these signal characteristics can effectively identify the fault type.

[0003] Early fault diagnosis mainly relied on manually summarized rule-based methods, such as gas ratio judgment and triangle diagram method. They classified monitoring indicators according to preset rule thresholds or graphic areas to infer the fault type. However, such empirical methods have obvious limitations: for example, the ratio method may produce combinations that cannot give a clear diagnosis in some cases, and the triangle diagram method has the problem of overlapping different fault areas, such as Figure 1 As shown in the figure, some fault modes are difficult to distinguish. This shows that the diagnosis results based solely on fixed thresholds or empirical charts are uncertain.

[0004] With the development of intelligent algorithms, researchers have begun applying machine learning and artificial intelligence (AI) methods to equipment fault diagnosis. For example, models such as artificial neural networks (ANNs), support vector machines (SVMs), fuzzy logic, and Bayesian networks are being used to directly learn fault patterns from raw monitoring data. These AI methods can perform fault classification without the need for manual feature extraction, but they require sufficient and balanced labeled data to train stable and reliable models. In reality, the frequency of different fault types varies significantly, making obtaining large amounts of labeled data difficult and costly. Consequently, the amount of data used for training is often limited and the category distribution is uneven. Furthermore, in modern online monitoring environments, the amount of unlabeled operational data is vast, while manually labeled fault data is very scarce and the categories are uneven. This lack of labeling and the long-tailed category distribution expose the limitations of traditional purely supervised learning methods.

[0005] For small amounts of unbalanced fault data, existing studies have proposed a variety of improvement strategies to improve the classification performance of the model, including methods at the data level, loss level, and model level:

[0006] 1) Data resampling methods: Oversampling or undersampling the dataset is used to balance the number of samples in each class. For example, random oversampling (ROS) increases the number of minority class samples, while random undersampling (RUS) reduces the number of majority class samples, physically rebalancing the dataset. Resampling methods alleviate the imbalance problem to a certain extent, but oversampling does not increase the actual information of the minority class. Instead, the repetition may cause the model to overfit the minority class, reducing generalization performance. Undersampling discards a large number of majority class samples, potentially losing key information and causing the model to underfit.

[0007] 2) Loss Weighting: This approach increases the relative importance of minority class samples in training by assigning different weights to different classes in the loss function. The difficulty of this approach lies in setting the weight coefficients. Improper weight selection can cause model training to deviate from the optimal direction, potentially leading to a decrease in overall performance.

[0008] 3) Model structure improvement methods: Improve performance under imbalanced data by changing the model structure or training strategy. For example, methods such as transfer learning, meta-learning, and two-stage learning can be introduced to enable the model to obtain more effective feature representations for small sample distributions. However, in scenarios with extremely imbalanced categories, these model-level improvements are still limited.

[0009] In summary, when faced with a small number of fault data with an inherent long-tail distribution, traditional methods often fail to achieve ideal results. New diagnostic ideas are urgently needed to improve the robustness and recognition accuracy of the model. Summary of the Invention

[0010] In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a two-step collaborative semi-supervised fault diagnosis method and system, which can effectively improve the accuracy of fault diagnosis.

[0011] To achieve the above object, the present invention provides the following solutions:

[0012] A two-step collaborative semi-supervised fault diagnosis method comprising:

[0013] Constructing a collaborative feature-guided learning architecture; the collaborative feature-guided learning architecture includes a feature extractor module, a discriminator module, and a historical steady-state guidance module;

[0014] An initial diagnostic model is obtained by jointly training the collaborative feature-guided learning architecture with labeled data and unlabeled data;

[0015] In the fine-tuning stage, the parameters of the feature extractor module are fixed, and only the parameters of the discriminator module are optimized to adapt to the target fault category, thereby obtaining a fine-tuned fault diagnosis model;

[0016] Perform fault diagnosis on equipment monitoring data based on the trained model and output fault type classification results.

[0017] Preferably, the feature extractor module adopts a three-layer fully connected neural network structure; the input layer of the feature extractor module receives the monitoring data after normalization preprocessing, and then passes through three hidden layers in sequence; the number of neurons in the hidden layers is set to 256, 128 and 64 respectively; each hidden layer adopts the ReLU activation function to enhance the nonlinear expression ability; the output layer of the feature extractor module outputs a 64-dimensional feature representation vector for the input of the subsequent discriminator module.

[0018] Preferably, the discriminator module consists of a two-layer fully connected network; the input layer of the discriminator module receives the 64-dimensional feature representation vector output by the feature extractor; the input layer of the discriminator module is provided with 16 neurons and is activated by ReLU; the number of nodes in the output layer of the discriminator module is the number of fault categories, and the Softmax activation function is used to output the predicted probability distribution of each category.

[0019] Preferably, the parameter update formula of the historical steady-state guidance module is:

[0020] θ t '=αθ t-1 '+(1-α)θ t

[0021] φ t '=αφ t-1 '+(1-α)φ t

[0022] Among them, θ t With φ t are the parameters of the feature extractor module and the discriminator module at the current time t, θ t ' and φ t ' are the parameters corresponding to the historical steady-state guidance module at the current time t, and α is the smoothing coefficient.

[0023] Preferably, the smoothing coefficient α is dynamically adjusted according to the sample category distribution during the training process, and the steps of dynamic adjustment are specifically as follows:

[0024] For minority class samples whose number of class samples is less than the preset threshold, reduce the α value to enhance the impact of its parameter update;

[0025] For majority class samples whose number of class samples is greater than or equal to the preset threshold, the α value is increased to maintain the overall training stability.

[0026] Preferably, the collaborative feature-guided learning architecture is trained by jointly training the labeled data and the unlabeled data to obtain an initial diagnostic model, including:

[0027] Taking the feature extractor module and the discriminator module as the main model;

[0028] Perform supervised learning on labeled data and optimize the parameters of the main model using cross-entropy loss;

[0029] Performing semi-supervised learning on unlabeled data, generating a reference output through the historical steady-state guidance module, and calculating an unsupervised loss based on a consistency constraint between the outputs of the main model and the historical steady-state guidance module;

[0030] The parameters of the historical steady-state guidance module are dynamically updated through an exponential sliding average mechanism, and the historical state of the main model is synchronously smoothed to obtain the initial diagnostic model.

[0031] Preferably, the calculation formula of the cross entropy loss is:

[0032]

[0033] Among them, C is the total number of categories, y ic For sample x i The true category indicator variable, For the sample x of category c i The predicted probability distribution, N L is the total amount of labeled data, L sup is the cross entropy loss.

[0034] Preferably, the calculation formula of the unsupervised loss is:

[0035]

[0036] Among them, p ref (x j ) is the output of the historical steady-state guidance module, N U is the total amount of unlabeled data, L unsup is the unsupervised loss.

[0037] Preferably, the total loss function of the joint training is calculated as:

[0038] L total =L sup +λL unsup +βL reg +γL ent

[0039] Among them, L ent is the entropy regularization term, Lreg is the weight decay regularization term, L reg =β||θ|| 2 , λ, β, γ are hyperparameters, and θ is the parameter of the feature extraction module.

[0040] A two-step collaborative semi-supervised fault diagnosis system comprising:

[0041] An architecture construction unit, configured to construct a collaborative feature-guided learning architecture comprising a feature extractor module, a discriminator module, and a historical steady-state guidance module;

[0042] a joint training unit, configured to jointly train the collaborative feature-guided learning architecture using labeled data and unlabeled data to obtain an initial diagnostic model;

[0043] A model fine-tuning unit, configured to fix the parameters of the feature extractor module and only optimize the parameters of the discriminator module during the fine-tuning phase to adapt to the target fault category and obtain a fine-tuned fault diagnosis model;

[0044] The fault diagnosis unit is used to perform fault diagnosis on the equipment monitoring data based on the trained model and output the fault type classification results.

[0045] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0046] The present invention provides a two-step collaborative semi-supervised fault diagnosis method and system, comprising: constructing a collaborative feature-guided learning architecture; the collaborative feature-guided learning architecture includes a feature extractor module, a discriminator module, and a historical steady-state guidance module; jointly training the collaborative feature-guided learning architecture with labeled data and unlabeled data to obtain an initial diagnosis model; in the fine-tuning stage, fixing the parameters of the feature extractor module and only optimizing the parameters of the discriminator module to adapt to the target fault category to obtain a fine-tuned fault diagnosis model; performing fault diagnosis on the equipment monitoring data based on the trained model and outputting a fault type classification result. The present invention effectively overcomes the problems of existing semi-supervised fault diagnosis methods such as unstable training under few-sample conditions, large parameter fluctuations, and fuzzy discrimination boundaries, improves the training stability and generalization ability of the model under extremely uneven data categories, and can effectively improve the accuracy of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 A schematic diagram of a triangle diagram method provided in the prior art;

[0049] Figure 2 A flow chart of a method provided by an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of the process flow of the collaborative feature-guided learning architecture provided by an embodiment of the present invention;

[0051] Figure 4 A schematic diagram of the system structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0053] The purpose of the present invention is to provide a two-step collaborative semi-supervised fault diagnosis method and system, which effectively overcomes the problems of unstable training, large parameter fluctuations, and fuzzy judgment boundaries of existing semi-supervised fault diagnosis methods under conditions of few samples, improves the training stability and generalization ability of the model under extremely uneven data categories, and can effectively improve the accuracy of fault diagnosis.

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0055] Figure 2 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 2 As shown, the present invention provides a two-step collaborative semi-supervised fault diagnosis method, comprising:

[0056] Step 100: Constructing a collaborative feature guided learning architecture; the collaborative feature guided learning architecture includes a feature extractor module, a discriminator module, and a historical steady-state guidance module;

[0057] Step 200: jointly train a collaborative feature-guided learning architecture with labeled data and unlabeled data to obtain an initial diagnostic model;

[0058] Step 300: In the fine-tuning stage, the parameters of the feature extractor module are fixed, and only the parameters of the discriminator module are optimized to adapt to the target fault category, thereby obtaining a fine-tuned fault diagnosis model;

[0059] Step 400: Perform fault diagnosis on the equipment monitoring data based on the trained model and output a fault type classification result.

[0060] Specifically, the collaborative feature-guided learning architecture proposed in this paper mainly consists of three key components: a feature extractor module, a discriminator module, and a historical steady-state guidance module. The feature extractor module is responsible for extracting high-quality feature representations from raw monitoring data, the discriminator module is used to classify fault types based on feature representations, and the historical steady-state guidance module serves as an auxiliary unit, providing stable historical reference information during training to guide the main model in learning robust features.

[0061] The feature extractor module employs a three-layer, fully connected neural network architecture. The input layer receives normalized, preprocessed monitoring data with an input dimension of d. This data then passes through three hidden layers, with 256, 128, and 64 neurons, respectively. Each layer uses a ReLU activation function to enhance nonlinear representation. Finally, it outputs a 64-dimensional feature representation vector h, which serves as input to the subsequent discriminator module.

[0062] The discriminator module has a relatively simple structure, consisting of two fully connected layers. The input layer receives the 64-dimensional feature representation output by the feature extractor. The first layer has 16 neurons and uses ReLU activation. The second layer is the output layer, with the number of nodes equal to the number of fault categories. It uses the Softmax activation function to output the predicted probability distribution for each category.

[0063] The overall structure of the historical steady-state guidance module is identical to that of the feature extractor and discriminator modules. Its parameters do not directly participate in backpropagation during training. Instead, they are dynamically updated using the historical parameters of the main model through a specific exponential moving average (EMA) method. As a smoothed collection of the model's past states, this module provides a stable teacher signal during training, effectively alleviating training instability caused by minimally labeled data and uneven classification.

[0064] The present invention designs a complete data processing flow in both training and reasoning processes, and establishes a rigorous mathematical modeling formula system.

[0065] In the training phase, we first obtain raw data from the monitoring system, including limited labeled samples. and a large number of unlabeled samples All input data must be normalized so that the eigenvalues of each dimension are mapped to the interval [0, 1] to ensure the numerical stability of training.

[0066] The normalized data is fed into the feature extractor module in turn, and the low-dimensional feature representation is extracted through the mapping function. The mapping process can be expressed as:

[0067] h=f(x;θ)

[0068] Among them, x is the input sample, h is the output feature representation, and θ is the parameter of feature extraction.

[0069] The feature representation of the labeled data is input into the discriminator module, and the Softmax function outputs the predicted probability distribution p main (x). The supervised loss adopts the standard cross entropy loss function form:

[0070]

[0071] Among them, C is the total number of categories, y ic For sample x i The true category indicator variable.

[0072] For unlabeled samples, the system also calls the historical steady-state guidance module to generate reference output. The parameter update of this module adopts the EMA method, and the specific formula is:

[0073] θ t '=αθ t-1 '+(1-α)θ t

[0074] Among them, θ t is the current main model parameter, θ t ' is the parameter of the steady-state guidance module, and α is the smoothing coefficient (usually set to 0.99). Subsequently, the consistency loss is calculated by the output results of the main model and the steady-state reference module for the same unlabeled sample, and the formula is:

[0075]

[0076] where p ref (x j ) is the output of the historical steady-state guidance module.

[0077] In order to further enhance the generalization ability of the model, the present invention introduces a weight decay regularization term L in the training process. reg =β||θ|| 2 and entropy regularization term Finally, the total loss function is:

[0078] L total =L sup +λL unsup +βL reg +γL ent

[0079] Among them, λ, β, and γ are hyperparameters.

[0080] In the fine-tuning stage, only labeled samples are used, the feature extractor parameters are fixed, only the discriminator parameters are updated, and cross-entropy loss optimization is adopted to finally obtain a complete fault diagnosis model.

[0081] The data processing flow in the inference phase is as follows: Figure 3 As shown in the figure, the specific steps include data collection and normalization preprocessing, feature extractor inference process, discriminator output prediction label, and post-processing logic based on set rules, and finally output fault discrimination results.

[0082] The proposed historical steady-state guidance module innovatively introduces an improved mechanism for EMA parameter updates, effectively overcoming the problems of existing semi-supervised fault diagnosis methods with unstable training, large parameter fluctuations, and blurred discrimination boundaries under low-sample conditions. Unlike traditional EMA, which only performs a single-dimensional sliding average on the feature extractor parameters, the present invention improves the EMA update mechanism by targeting the synergistic characteristics of the feature extractor and discriminator modules, jointly updating the parameters of the two modules as a whole through EMA. This ensures that the dynamic smoothness of the historical steady-state guidance module in both the feature space and the discriminant space is maintained simultaneously, thereby improving the training stability and generalization ability of the model under extremely uneven data categories.

[0083] Specifically, the improved EMA parameter update method of the present invention not only uses sliding average for the feature extractor parameter θ, but also synchronously performs sliding average for the discriminator parameter φ. The parameter update formula of the historical steady-state guidance module is:

[0084] θ t '=αθ t-1 '+(1-α)θ t

[0085] φ t '=αφ t-1 '+(1-α)φ t

[0086] Among them, θ t With φ t are the current feature extractor and discriminator parameters of the main model, θ t ' and φ t ' is the corresponding parameter of the steady-state reference module, and α is the smoothing coefficient.

[0087] The main improvements of this parameter update method compared to the general EMA mechanism are:

[0088] First, a joint smoothing mechanism of feature and classification modules is introduced to effectively avoid the feature-discriminant inconsistency problem caused by smoothing only the feature space;

[0089] Second, under the conditions of few labels and uneven categories, the EMA update frequency is dynamically adjusted when each category sample participates in training. That is, when the number of category samples is extremely small, the α value is appropriately reduced to enhance the influence of minority class samples on the historical steady-state reference output, thereby alleviating the model bias problem caused by uneven categories; when training the majority class, the α value is restored to a high value to ensure the overall stability of the training process.

[0090] Through the above improvements, the EMA update mechanism adopted in the present invention better adapts to the design requirements of the historical steady-state guidance module, and can dynamically adjust the degree of smoothing when the distribution of samples of different categories is extremely uneven, thereby achieving fine-grained control of the training process and effectively alleviating the training instability caused by uneven categories and insufficient samples.

[0091] Furthermore, by introducing an output consistency constraint, this invention effectively incorporates unlabeled data into the fault discrimination model training process for the first time. This constraint minimizes the difference in output between the main model and the historical steady-state guidance module on unlabeled data, ensuring the model achieves stable output on unlabeled data. This enhances the ability to identify minority class fault samples and mitigates the decision bias caused by class imbalance in traditional methods.

[0092] At the same time, the present invention adopts a design in which the feature extractor and the discriminator structure are separated. During the fine-tuning stage, only the discriminator parameters need to be adjusted, and there is no need to retrain the feature extractor. This significantly reduces the retraining cost when the model adapts to different data sets or new fault categories in actual applications, and improves the model's portability and practicality.

[0093] In summary, the personalized improvement of the EMA parameter update mechanism in this invention, combined with the output consistency constraint term and structural separation design, has achieved a technological breakthrough in traditional semi-supervised learning methods, significantly improving the model's training stability, discrimination accuracy, and engineering adaptability under conditions of very few labeled samples and severe category imbalance.

[0094] To verify the applicability and technical advantages of the collaborative feature-guided learning architecture in actual large-scale electrical equipment fault diagnosis scenarios, we conducted experimental verification based on an internationally available dataset of typical transformer faults. The experimental data was derived from the IEC TC 10DGA dataset, published by the International Electrotechnical Commission. This dataset contains a large amount of transformer oil dissolved gas analysis (DGA) monitoring data, which is widely used in power equipment fault diagnosis algorithm research and performance comparison and is currently one of the more authoritative fault identification benchmark datasets in the engineering field.

[0095] This experiment used all 117 labeled fault data from the dataset, covering five typical fault conditions: partial discharge (PD), low-energy discharge (D1), high-energy discharge (D2), low-temperature overheating (T12), and high-temperature overheating (T3). Notably, the number of fault data in each category varied significantly in the dataset, with the minority category samples far fewer than the majority category samples, exhibiting a typical long-tail imbalanced distribution. This is highly consistent with the distribution of fault data from actual engineering equipment, effectively validating the performance of the proposed method in low-sample, imbalanced scenarios.

[0096] In the experimental design, the present invention divided 117 labeled data points into a training set and a test set in a 7:3 ratio. The training set contained a small amount of labeled data, while additional unlabeled data was introduced during the training process to simulate the large amount of unlabeled data in the real-world scenario of online equipment monitoring. The number of samples in the test set for each fault category was relatively balanced, allowing for a more objective evaluation of the model's ability to distinguish between different categories.

[0097] During the training phase, the labeled and unlabeled data in the training set are first normalized and preprocessed to normalize the gas concentration values to the interval [0, 1], eliminating the interference of dimensional differences on the training process. Subsequently, the collaborative feature-guided learning architecture of the present invention is used for model training. During the semi-supervised pre-training phase, the model simultaneously utilizes the supervisory signals of a small number of labeled samples and the consistency constraint signals of unlabeled samples to fully explore the potential information in the unlabeled data, thereby improving the robustness of the feature extractor and the accuracy of the discriminator. During the fine-tuning phase, the pre-trained feature extractor parameters are fixed, and the discriminator parameters are adjusted only for the labeled data to further improve the model classification performance.

[0098] During the testing phase, the trained model is verified using the divided test set, and the classification accuracy (ACC) and F1 score are used as evaluation indicators to comprehensively reflect the overall diagnostic capability of the model and its recognition effect on minority faults.

[0099] Experimental results show that the method of the present invention achieved an overall classification accuracy of 94.57% on the test set, with an average F1 score of 94.55%. The recognition of various types of faults reached a high level, especially in the low-energy discharge and high-temperature overheating categories with a small number of samples, showing excellent recognition capabilities.

[0100] The method of the present invention demonstrates significant performance improvements and practical advantages over existing technologies. On a publicly available standard fault dataset, this method achieved an overall classification accuracy of approximately 94.57%, significantly higher than the accuracy of traditional imbalanced processing methods (such as random oversampling (ROS), random undersampling (RUS), and loss weighting). Furthermore, the method's average F1 score also improved significantly, demonstrating its enhanced ability to identify fault categories with fewer samples. In contrast, traditional resampling methods struggle to achieve a balanced performance across all categories due to issues such as overfitting or information loss.

[0101] On an actual data set collected in this embodiment with an extremely disparate distribution of fault categories (the ratio of majority class to minority class samples is as high as 25:1), this method still achieved a high accuracy of about 91.67%. Even in the case of extreme class imbalance, the model still maintains excellent fault detection performance: its accuracy only drops by about 2.9% compared to when the categories are balanced. This result demonstrates the robustness of this method in long-tail data scenarios - while some traditional methods will show a significant drop in accuracy when the imbalance is extremely high, this method only experiences a slight performance degradation. In summary, the technology of the present invention demonstrates significant superiority in fault diagnosis tasks with low labeling rates and long-tail distributions, and is expected to reduce the need for manual labeling and improve the reliability of anomaly detection.

[0102] Corresponding to the above method, such as Figure 4 As shown, this embodiment also provides a two-step collaborative semi-supervised fault diagnosis system, including:

[0103] An architecture construction unit, configured to construct a collaborative feature-guided learning architecture comprising a feature extractor module, a discriminator module, and a historical steady-state guidance module;

[0104] a joint training unit, configured to jointly train the collaborative feature-guided learning architecture using labeled data and unlabeled data to obtain an initial diagnostic model;

[0105] A model fine-tuning unit, configured to fix the parameters of the feature extractor module and only optimize the parameters of the discriminator module during the fine-tuning phase to adapt to the target fault category and obtain a fine-tuned fault diagnosis model;

[0106] The fault diagnosis unit is used to perform fault diagnosis on the equipment monitoring data based on the trained model and output the fault type classification results.

[0107] The beneficial effects of the present invention are as follows:

[0108] (1) This paper proposes to introduce semi-supervised learning into fault diagnosis and designs a two-stage collaborative training framework of "pre-training + fine-tuning". By integrating unsupervised training signals with supervised learning, the fault classification performance under imbalanced data conditions is significantly improved. Compared with traditional methods that rely solely on labeled training or simple resampling, this collaborative architecture fully explores the value of unlabeled data and shows stronger adaptability to fault data with a small number of samples and a long tail distribution.

[0109] (2) Output stability constraint factors mitigate the impact of imbalance. During the semi-supervised pre-training phase, this paper introduces an output stability constraint as an unsupervised regularization term, effectively reducing the bias caused by imbalanced labeled data on model training. By imposing output consistency constraints on unlabeled samples, the model is guided to focus on the common features of samples from each category, thereby balancing the impact of majority and minority class samples on model parameter updates and alleviating the model bias problem caused by the scarcity of minority class samples.

[0110] (3) The historical steady-state guidance module of the present invention improves training stability and adopts a historical steady-state guidance mechanism to utilize unlabeled data. Compared with traditional pseudo-labeling, self-training and other methods, this mechanism can provide a smoother and more stable learning target. The historical steady-state guidance module, as an exponential average of the historical states of the feature extractor and discriminator, can significantly enhance the stability of model training, smooth the classification decision boundary, and reduce the prediction fluctuation caused by too few minority class samples. This strategy ensures that the model can still achieve reliable fault recognition performance even when labels are extremely scarce and the category distribution is extremely unbalanced.

[0111] (4) The model architecture of the present invention, which separates feature extraction and discrimination, innovatively divides the model into two parts, a feature extractor and a discriminator, for separate training. Pre-training focuses on the extraction and learning of general features, while fine-tuning focuses on the optimization of classification decisions. This separate architecture allows the feature extraction module obtained through pre-training to be easily transferred to new tasks or data. During fine-tuning, only the discriminator needs to be adjusted to complete the adaptation. The advantage is that when faced with different but related fault data, there is no need to train the entire model from scratch, which greatly improves the reusability and scalability of the model.

[0112] (5) The present invention is applicable to extremely unbalanced long-tail data. The present invention is still applicable to extremely unbalanced long-tail data, significantly improving the reliability and robustness of fault diagnosis. In experiments, it can be observed that as the degree of data imbalance increases, the accuracy of the present method only decreases slightly, demonstrating strong adaptability to long-tail distributions. Therefore, the method of the present invention covers a wide range of applications from mild to highly unbalanced data distributions and can provide effective fault diagnosis support for complex data distributions in various real-world scenarios.

[0113] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0114] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A two-step collaborative semi-supervised fault diagnosis method, characterized in that: include: Constructing a collaborative feature-guided learning architecture; The collaborative feature-guided learning architecture includes a feature extractor module, a discriminator module, and a historical steady-state guidance module; An initial diagnostic model is obtained by jointly training the collaborative feature-guided learning architecture with labeled data and unlabeled data; In the fine-tuning stage, the parameters of the feature extractor module are fixed, and only the parameters of the discriminator module are optimized to adapt to the target fault category, thereby obtaining a fine-tuned fault diagnosis model; Perform fault diagnosis on equipment monitoring data based on the trained model and output fault type classification results.

2. The two-step collaborative semi-supervised fault diagnosis method according to claim 1, characterized in that: The feature extractor module adopts a three-layer fully connected neural network structure; the input layer of the feature extractor module receives the monitoring data after normalization preprocessing, and then passes it through three hidden layers in sequence; the number of neurons in the hidden layers is set to 256, 128 and 64 respectively; each hidden layer adopts the ReLU activation function to enhance the nonlinear expression ability; the output layer of the feature extractor module outputs a 64-dimensional feature representation vector for the input of the subsequent discriminator module.

3. The two-step collaborative semi-supervised fault diagnosis method according to claim 2, characterized in that: The discriminator module is composed of a two-layer fully connected network; the input layer of the discriminator module receives the 64-dimensional feature representation vector output by the feature extractor; The input layer of the discriminator module is provided with 16 neurons and is activated by ReLU; the number of nodes in the output layer of the discriminator module is the number of fault categories, and the Softmax activation function is used to output the predicted probability distribution of each category.

4. The two-step collaborative semi-supervised fault diagnosis method according to claim 3, characterized in that: The parameter update formula of the historical steady-state guidance module is: i t '=αθ t-1 '+(1-α)θ t f t '=after t-1 '+(1-a)φ t Among them, θ t With φ t are the parameters of the feature extractor module and the discriminator module at the current time t, θ t ' and φ t ' are the parameters corresponding to the historical steady-state guidance module at the current time t, and α is the smoothing coefficient.

5. The two-step collaborative semi-supervised fault diagnosis method according to claim 4, characterized in that: The smoothing coefficient α is dynamically adjusted according to the sample category distribution during the training process, and the steps of dynamic adjustment are specifically as follows: For minority class samples whose number of class samples is less than the preset threshold, reduce the α value to enhance the impact of its parameter update; For majority class samples whose number of class samples is greater than or equal to the preset threshold, the α value is increased to maintain the overall training stability.

6. The two-step collaborative semi-supervised fault diagnosis method according to claim 1, characterized in that: The collaborative feature-guided learning architecture is trained by jointly training the labeled and unlabeled data to obtain an initial diagnostic model, including: Taking the feature extractor module and the discriminator module as the main model; Perform supervised learning on labeled data and optimize the parameters of the main model using cross-entropy loss; Performing semi-supervised learning on unlabeled data, generating a reference output through the historical steady-state guidance module, and calculating an unsupervised loss based on a consistency constraint between the outputs of the main model and the historical steady-state guidance module; The parameters of the historical steady-state guidance module are dynamically updated through an exponential sliding average mechanism, and the historical state of the main model is synchronously smoothed to obtain the initial diagnostic model.

7. The two-step collaborative semi-supervised fault diagnosis method according to claim 6, characterized in that: The calculation formula of the cross entropy loss is: Among them, C is the total number of categories, y ic For sample x i The true category indicator variable, For the sample x of category c i The predicted probability distribution, N L is the total amount of labeled data, L sup is the cross entropy loss.

8. The two-step collaborative semi-supervised fault diagnosis method according to claim 7, characterized in that: The calculation formula of the unsupervised loss is: Among them, p ref (x j ) is the output of the historical steady-state guidance module, N U is the total amount of unlabeled data, L unsup is the unsupervised loss.

9. The two-step collaborative semi-supervised fault diagnosis method according to claim 8, characterized in that: The total loss function for joint training is calculated as: L total =L sup +λL unsup +βL reg +γL ent Among them, L ent is the entropy regularization term, L reg is the weight decay regularization term, L reg =β||θ|| 2 , λ, β, γ are hyperparameters, and θ is the parameter of the feature extraction module.

10. A two-step collaborative semi-supervised fault diagnosis system, characterized in that: include: Architecture building unit, used to build collaborative feature-guided learning architecture; The collaborative feature-guided learning architecture includes a feature extractor module, a discriminator module, and a historical steady-state guidance module; a joint training unit, configured to jointly train the collaborative feature-guided learning architecture using labeled data and unlabeled data to obtain an initial diagnostic model; A model fine-tuning unit, configured to fix the parameters of the feature extractor module and only optimize the parameters of the discriminator module during the fine-tuning phase to adapt to the target fault category and obtain a fine-tuned fault diagnosis model; The fault diagnosis unit is used to perform fault diagnosis on the equipment monitoring data based on the trained model and output the fault type classification results.

Images