A charging facility semi-supervised fault diagnosis method combining ensemble learning and self-training
By employing data equalization and self-training to generate pseudo-labeled samples, combined with a Stacking ensemble learning model, the problem of insufficient and unevenly distributed labeled samples in charging facility fault diagnosis was solved, thereby improving the model's generalization ability and diagnostic accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIANGTAN UNIV
- Filing Date
- 2025-09-05
- Publication Date
- 2026-07-31
AI Technical Summary
Existing charging facility fault diagnosis models rely on a large amount of labeled data, and when the fault data is unevenly distributed, the diagnosis results tend to favor the majority class, resulting in poor model generalization and difficulty in accurately diagnosing with a small number of labeled samples.
The Borderline-SMOTE algorithm is used to generate minority class samples for data equalization. Combined with the Stacking ensemble learning model and self-training method, GBDT, CatBoost and convolutional neural network are used as base classifiers, and SVM is used as meta-learner. Pseudo-label samples are generated through K-fold cross-validation and self-training to expand the training set.
It improves the generalization ability and robustness of the charging facility fault diagnosis model, enabling high accuracy and high recall in fault diagnosis with a small number of labeled samples, and adapts to the fault diagnosis needs in multiple scenarios.
Smart Images
Figure SMS_1 
Figure SMS_2 
Figure HDA0005583222710000011
Abstract
Description
Technical Field
[0001] This invention belongs to the field of charging facility fault diagnosis technology, and in particular relates to a semi-supervised fault diagnosis method for charging facilities that combines integrated learning and self-training. Background Technology
[0002] In recent years, electric vehicles (EVs), as an important form of green transportation, have shown a rapid growth trend. The surge in the number of EVs has directly driven a large-scale demand for charging infrastructure. Charging infrastructure is a collective term for a series of key basic equipment that provide power to EVs. Failures in charging infrastructure can significantly impact charging users and even normal traffic operations. However, charging infrastructure mostly operates outdoors, where factors such as temperature and humidity are uncontrollable. It inevitably faces problems such as equipment aging, component wear and tear, and electrical faults. These, combined with factors like grid fluctuations and human error, contribute to the failure of charging infrastructure. Therefore, achieving online fault diagnosis for charging infrastructure and promptly identifying faulty facilities to ensure their reliable operation is of great significance.
[0003] With the development of artificial intelligence technology, existing research has focused on fault diagnosis models for charging facilities based on machine learning and deep learning, such as Support Vector Machines, Extreme Learning Machines, Random Forests, Gradient Boosting, and Deep Neural Networks, most of which employ supervised learning methods. However, supervised learning training requires a large amount of labeled data, which is time-consuming, labor-intensive, and requires extensive expert experience. Furthermore, in the actual operation of charging facilities, the probability of different faults occurring varies, leading to an imbalanced distribution of the collected fault data categories. Training the model on an imbalanced dataset can cause the model's diagnostic results to be biased towards the majority class, resulting in bias and affecting the model's generalization ability. Therefore, addressing the problems of insufficient labeled samples and imbalanced fault data distribution in charging facility fault diagnosis is crucial for improving the model's generalization ability and diagnostic accuracy.
[0004] The Stacking ensemble learning framework combines base learners with significant differences to form a learning model, fully leveraging the advantages of different base learners to improve prediction accuracy and robustness, making it suitable for charging facility fault diagnosis in various scenarios. Semi-supervised learning is a learning method combining supervised and unsupervised learning, enabling training using unlabeled samples even with limited labeled data. Compared to supervised and unsupervised learning, it requires fewer labeled training samples. Self-training is a semi-supervised learning method that aims to improve model performance using a small amount of labeled data and a large amount of unlabeled data. Therefore, a semi-supervised fault diagnosis method for charging facilities based on an ensemble learning model and self-training is proposed. Summary of the Invention
[0005] This invention aims to solve one of the technical problems existing in the prior art. To this end, this invention proposes a semi-supervised fault diagnosis method for charging facilities that combines ensemble learning and self-training, comprising the following steps:
[0006] 1) Collect charging facility fault characteristic data. For small amounts of labeled data with a large difference in the number of samples for each fault category, use the Borderline-SMOTE algorithm to generate minority class samples and perform equalization processing.
[0007] 2) Construct a Stacking ensemble learning model. The first layer uses GBDT, CatBoost and convolutional neural network as base classifiers, the second layer uses SVM as meta-learner, and then combines K-fold cross-validation technology and self-training to form a semi-supervised fault diagnosis model for charging facilities.
[0008] 3) Online fault diagnosis: The pre-processed charging facility fault characteristics are measured in real time and input into the trained semi-supervised fault diagnosis model of the charging facility to obtain the fault type result under the working condition online.
[0009] Specifically, for the collection of charging facility fault feature data, for small amounts of labeled data where the number of samples for each fault category differs significantly, the Borderline-SMOTE algorithm is used to generate minority class samples for equalization processing, as detailed below:
[0010] 1) Collect fault characteristic data of charging facilities through sensor acquisition, preprocess the data to form a small amount of labeled dataset and a large amount of unlabeled dataset;
[0011] 2) For small amounts of labeled data where the number of samples in each fault category varies significantly, the Borderline-SMOTE algorithm is used. All minority class samples are partitioned based on their K-nearest neighbor distances. A boundary sample is selected, and then, based on the Euclidean distance between samples, the K-nearest neighbors of that sample are chosen. Finally, linear interpolation is performed between that sample and its K-nearest neighbors to generate new minority class samples, forming a balanced labeled dataset. The generation of new minority class samples can be expressed as follows:
[0012] x new =x m +ε(x m,j -x m (1) Where, x new For the generated minority class samples, x m For the m-th minority class boundary sample selected, x m，j For x new The j-th nearest neighbor sample of the same class, where ε is a random number between (0, 1);
[0013] The aforementioned construction of the Stacking ensemble learning model employs GBDT, CatBoost, and convolutional neural networks as base classifiers in the first layer, and SVM as a meta-learner in the second layer. Combined with K-fold cross-validation and self-training, this forms a semi-supervised fault diagnosis model for charging facilities, as detailed below:
[0014] 1) Constructing a Stacking ensemble learning model: including base classifier algorithms and meta-learner algorithms. The first layer base classifier is constructed using GBDT, CatBoost, and convolutional neural networks, which have strong fault diagnosis performance and large algorithm differences. The second layer meta-learner is SVM.
[0015] 2) Initial training of the Stacking ensemble learning model: The equalized labeled dataset is divided into training set D in an 8:2 ratio. train and test set D test , training set D train The training set M for base classifiers is divided according to 4-fold cross-validation. b1 M b2 M b3 Base classifier test set M test Using the base classifier training set, train three base classifiers: GBDT, CatBoost, and a convolutional neural network. Then, use the trained base classifiers on the base classifier test set M. test Make predictions to obtain the prediction results of the base classifier; training set D train Each sample in the dataset is predicted once. The prediction results of the base classifiers are combined in the second dimension to obtain the training data for the meta-learner, and the SVM is trained.
[0016] 3) Employ a self-training method to expand the training set and iteratively train the Stacking ensemble learning model until all unlabeled datasets are pseudo-labeled or the number of iterations reaches a set value. The specific steps are as follows:
[0017] ①The three base classifiers, GBDT, CatBoost and convolutional neural network, are referred to as S1, S2 and S3 respectively. The total number of iterations is set to T. A certain number of unlabeled data are selected from the collected unlabeled dataset as the unlabeled data training set for the first iteration.
[0018] ② In the t-th iteration, the samples x in the unlabeled training set are... t The data are input into three base classifiers—GBDT, CatBoost, and a convolutional neural network—which have completed initial training, respectively, to obtain their respective prediction results. The prediction results are then evaluated, and samples x whose predictions from classifiers S2 and S3 are consistent and whose confidence levels are both greater than a preset threshold γ are considered. cre1t Generating pseudo-labels and forming an extended training set M for S1.ex1t The prediction results for S1 and S3 are consistent and the confidence level of both is greater than the preset threshold γ. cre2t Pseudo-labels are generated and an extended training set M for S2 is formed. ex2t The prediction results for S1 and S2 are consistent and the confidence level of both is greater than the preset threshold γ. cre3t Pseudo-labels are generated and an extended training set M for S3 is formed. ex3t ;
[0019] ③Use the expanded training set M respectively ex1t ∪M ex1t-1 ∪...∪M b1 M ex2t ∪M ex2t-1 ∪...∪M b2 M ex3t ∪M ex3t-1 ∪...∪M b3 Retrain the three base classifiers and the meta-learner SVM, and update the model parameters;
[0020] ④ On the unlabeled data selected in the previous iteration, select additional unlabeled data to construct a new unlabeled data training set, and repeat the labeling of the unlabeled data training set samples from the previous iteration to continue iterating and expanding the training set; during the iteration process, dynamically adjust the confidence threshold γ according to the performance changes of the base classifier, and appropriately reduce the threshold when the performance improves in order to include more samples.
[0021] ⑤ When all the unlabeled datasets collected have been pseudo-labeled or the number of iterations has reached the set value T, stop training and output the three base classifiers and the meta-learner SVM that have been trained.
[0022] 4) Model Evaluation: Using the trained base classifier, evaluate the model on the test set D. test Make predictions and obtain prediction results. Then, merge the results in the second dimension to obtain the meta-learner test data. Use the trained SVM to perform performance testing on the meta-learner test data to obtain fault diagnosis test results. Evaluate the model's performance based on evaluation metrics.
[0023] The online fault diagnosis involves inputting preprocessed real-time measurement data of charging facility fault characteristics into a trained semi-supervised fault diagnosis model for charging facilities to obtain the fault type result under the specified operating condition online, as detailed below:
[0024] The offline-trained semi-supervised fault diagnosis model for charging facilities is deployed in the charging control unit of the charging facility. The system collects fault characteristic measurement data of the charging facility in real time and transmits it to the charger control unit. After preprocessing, the system obtains the fault type result under the working condition online through the trained semi-supervised fault diagnosis model for the charging facility. If a fault exists, the system immediately triggers an alarm mechanism and automatically records key information such as the fault occurrence time, characteristic parameters, and diagnosis results. Operators can then handle the fault in a timely manner according to the fault type.
[0025] The embodiments of the present invention have at least the following beneficial technical effects:
[0026] 1) The semi-supervised fault diagnosis method for charging facilities based on the Stacking ensemble learning model proposed in this invention uses GBDT, CatBoost and convolutional neural network, which have strong fault diagnosis performance and large algorithm differences, to jointly construct the first base classifier. This can overcome the problem of insufficient adaptability of single base classifier models and enhance the generalization ability and robustness of fault diagnosis models.
[0027] 2) In the semi-supervised fault diagnosis model for charging facilities proposed in this invention, non-uniform data sample processing and self-training are adopted to equalize the labeled data and generate pseudo-labeled samples using unlabeled samples, so that these samples can participate in training. This can overcome the problem caused by the unbalanced distribution of labeled sample data. Under the condition of a small number of labeled samples, the fault diagnosis model has high accuracy and generalization ability. Attached Figure Description
[0028] Figure 1 This is a flowchart of a semi-supervised fault diagnosis method for charging facilities that combines integrated learning and self-training, provided by an embodiment of the present invention. Detailed Implementation
[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the following embodiments are by no means intended to limit the present invention.
[0030] A semi-supervised fault diagnosis method for charging facilities that combines ensemble learning and self-training, such as Figure 1 As shown, it includes the following steps:
[0031] 1) Collect charging facility fault characteristic data. For small amounts of labeled data with a large difference in the number of samples for each fault category, use the Borderline-SMOTE algorithm to generate minority class samples and perform equalization processing.
[0032] 2) Construct a Stacking ensemble learning model. The first layer uses GBDT, CatBoost and convolutional neural network as base classifiers, the second layer uses SVM as meta-learner, and then combines K-fold cross-validation technology and self-training to form a semi-supervised fault diagnosis model for charging facilities.
[0033] 3) Online fault diagnosis: The pre-processed charging facility fault characteristics are measured in real time and input into the trained semi-supervised fault diagnosis model of the charging facility to obtain the fault type result under the working condition online.
[0034] Specifically, for the collection of charging facility fault feature data, for small amounts of labeled data where the number of samples for each fault category differs significantly, the Borderline-SMOTE algorithm is used to generate minority class samples for equalization processing, as detailed below:
[0035] 1) Collect fault characteristic data of charging facilities through sensor acquisition, preprocess the data to form a small amount of labeled dataset and a large amount of unlabeled dataset;
[0036] 2) For small amounts of labeled data where the number of samples in each fault category varies significantly, the Borderline-SMOTE algorithm is used. All minority class samples are partitioned based on their K-nearest neighbor distances. A boundary sample is selected, and then, based on the Euclidean distance between samples, the K-nearest neighbors of that sample are chosen. Finally, linear interpolation is performed between that sample and its K-nearest neighbors to generate new minority class samples, forming a balanced labeled dataset. The generation of new minority class samples can be expressed as follows:
[0037] x new =x m +ε(x m,j -x m (1)
[0038] Where, x new For the generated minority class samples, x m For the m-th minority class boundary sample selected, x m，j For x new The j-th nearest neighbor sample of the same class, where ε is a random number between (0, 1);
[0039] The aforementioned construction of the Stacking ensemble learning model employs GBDT, CatBoost, and convolutional neural networks as base classifiers in the first layer, and SVM as a meta-learner in the second layer. Combined with K-fold cross-validation and self-training, this forms a semi-supervised fault diagnosis model for charging facilities, as detailed below:
[0040] 1) Constructing a Stacking ensemble learning model: including base classifier algorithms and meta-learner algorithms. The first layer base classifier is constructed using GBDT, CatBoost, and convolutional neural networks, which have strong fault diagnosis performance and large algorithm differences. The second layer meta-learner is SVM.
[0041] 2) Initial training of the Stacking ensemble learning model: The equalized labeled dataset is divided into training set D in an 8:2 ratio. train and test set D test , training set D train The training set M for base classifiers is divided according to 4-fold cross-validation. b1 M b2 M b3 Base classifier test set M test Using the base classifier training set, train three base classifiers: GBDT, CatBoost, and a convolutional neural network. Then, use the trained base classifiers on the base classifier test set M. test Make predictions to obtain the prediction results of the base classifier; training set D train Each sample in the dataset is predicted once. The prediction results of the base classifiers are combined in the second dimension to obtain the training data for the meta-learner, and the SVM is trained.
[0042] 3) Employ a self-training method to expand the training set and iteratively train the Stacking ensemble learning model until all unlabeled datasets are pseudo-labeled or the number of iterations reaches a set value. The specific steps are as follows:
[0043] ①The three base classifiers, GBDT, CatBoost and convolutional neural network, are referred to as S1, S2 and S3 respectively. The total number of iterations is set to T. A certain number of unlabeled data are selected from the collected unlabeled dataset as the unlabeled data training set for the first iteration.
[0044] ② In the t-th iteration, the samples x in the unlabeled training set are... t The data are input into three base classifiers—GBDT, CatBoost, and a convolutional neural network—which have completed initial training, respectively, to obtain their respective prediction results. The prediction results are then evaluated, and samples x whose predictions from classifiers S2 and S3 are consistent and whose confidence levels are both greater than a preset threshold γ are considered. cre1t Generating pseudo-labels and forming an extended training set M for S1. ex1t The prediction results for S1 and S3 are consistent and the confidence level of both is greater than the preset threshold γ. cre2t Pseudo-labels are generated and an extended training set M for S2 is formed. ex2t The prediction results for S1 and S2 are consistent and the confidence level of both is greater than the preset threshold γ. cre3tPseudo-labels are generated and an extended training set M for S3 is formed. ex3t ;
[0045] ③Use the expanded training set M respectively ex1t ∪M ex1t-1 ∪...∪M b1 M ex2t ∪M ex2t-1 ∪...∪M b2 M ex3t ∪M ex3t-1 ∪...∪M b3 Retrain the three base classifiers and the meta-learner SVM, and update the model parameters;
[0046] ④ On the unlabeled data selected in the previous iteration, select additional unlabeled data to construct a new unlabeled data training set, and repeat the labeling of the unlabeled data training set samples from the previous iteration to continue iterating and expanding the training set; during the iteration process, dynamically adjust the confidence threshold γ according to the performance changes of the base classifier, and appropriately reduce the threshold when the performance improves in order to include more samples.
[0047] ⑤ When all the unlabeled datasets collected have been pseudo-labeled or the number of iterations has reached the set value T, stop training and output the three base classifiers and the meta-learner SVM that have been trained.
[0048] 4) Model Evaluation: Using the trained base classifier, evaluate the model on the test set D. test Make predictions and obtain prediction results. Then, merge the results in the second dimension to obtain the meta-learner test data. Use the trained SVM to perform performance testing on the meta-learner test data to obtain fault diagnosis test results. Evaluate the model's performance based on evaluation metrics.
[0049] The online fault diagnosis involves inputting preprocessed real-time measurement data of charging facility fault characteristics into a trained semi-supervised fault diagnosis model for charging facilities to obtain the fault type result under the specified operating condition online, as detailed below:
[0050] The offline-trained semi-supervised fault diagnosis model for charging facilities is deployed in the charging control unit of the charging facility. The system collects fault characteristic measurement data of the charging facility in real time and transmits it to the charger control unit. After preprocessing, the system obtains the fault type result under the working condition online through the trained semi-supervised fault diagnosis model for the charging facility. If a fault exists, the system immediately triggers an alarm mechanism and automatically records key information such as the fault occurrence time, characteristic parameters, and diagnosis results. Operators can then handle the fault in a timely manner according to the fault type.
[0051] To verify the performance of this invention, the proposed method was implemented using Python software on a computer configured with a 12th Gen Intel(R) Core(TM) i5-12500 3.00GHz CPU and 32GB of memory. The method was trained and tested on the publicly available Baidu DianShi charging facility dataset, and simulation experiments were conducted for comparison. Each sample in the Baidu DianShi charging facility dataset contains a feature vector and a fault label for one charging facility. The feature vector includes the charging facility's K1K2 drive signal, electronic lock drive signal, emergency stop signal, access control signal, voltage total harmonic distortion (THD), and current total harmonic distortion (THD). Fault types are categorized as faulty and fault-free.
[0052] 1) 500 sets of labeled data and 5000 sets of unlabeled data were generated from the publicly available Baidu Dianshi charging facility dataset. After preprocessing, the 500 sets of labeled data included 400 sets of fault-related data and 100 sets of fault-free data.
[0053] 2) The Borderline-SMOTE algorithm is used to generate new fault-free class samples. The fault-free class data after equalization is 400 sets. The 800 sets of labeled data after equalization are divided into labeled training set and labeled test set in an 8:2 ratio.
[0054] 3) Using 640 sets of labeled training samples and 5000 sets of unlabeled samples, the semi-supervised fault diagnosis model for charging facilities proposed in this invention was trained by combining K-fold cross-validation technology and self-training. Then, the trained semi-supervised fault diagnosis model for charging facilities was tested with 160 sets of labeled test data, and the accuracy and recall were 94.33% and 95.65%, respectively.
[0055] 4) Simulation Experiment Comparison
[0056] With 640 sets of labeled training samples, the semi-supervised fault diagnosis model of this invention was self-trained using 5000 sets of unlabeled samples. The performance comparison on the test set with GBDT, CatBoost, convolutional neural networks, and the unself-trained Stacking ensemble learning model is shown in Table 1. As shown in Table 1, the model of this invention outperforms the unself-trained model in both accuracy and recall. This demonstrates that the proposed semi-supervised fault diagnosis model for charging facilities based on ensemble learning and self-training can fully utilize unlabeled data to generate pseudo-labeled samples under conditions of a small number of labeled training samples, allowing these samples to participate in training and exhibiting superior classification performance.
[0057] Table 1. Performance comparison of different models with a small number of labeled training samples.
[0058]
[0059] Using 640 sets of labeled training samples and 160 sets of labeled test data, the performance comparison of the model of this invention on the test set was obtained by varying the number of unlabeled samples. Table 2 shows that, with fewer labeled training samples, as the number of unlabeled samples participating in the self-training of the Stacking ensemble learning model increases, a large number of pseudo-labeled samples are added to the training, resulting in a significant improvement in the evaluation performance of the semi-supervised fault diagnosis model for charging facilities, and the model exhibits better generalization ability.
[0060] Table 2 Performance comparison under different numbers of unlabeled samples
[0061]
Claims
1. A semi-supervised fault diagnosis method for charging facilities combining ensemble learning and self-training, characterized in that, The semi-supervised fault diagnosis method for charging facilities that combines ensemble learning and self-training includes the following steps: 1) Collect charging facility fault characteristic data. For a small amount of labeled data with a large difference in the number of samples for each fault category, use the Borderline-SMOTE algorithm to generate minority class samples and perform equalization processing. 2) Construct a Stacking ensemble learning model. The first layer uses GBDT, CatBoost and convolutional neural network as base classifiers, the second layer uses SVM as meta-learner, and then combines K-fold cross-validation technology and self-training to form a semi-supervised fault diagnosis model for charging facilities. 3) Online fault diagnosis: The pre-processed charging facility fault characteristics are measured in real time and input into the trained semi-supervised fault diagnosis model of the charging facility to obtain the fault type result under the working condition online.
2. The semi-supervised fault diagnosis method for charging facilities combining ensemble learning and self-training as described in claim 1, characterized in that, The aforementioned Stacking ensemble learning model employs GBDT, CatBoost, and convolutional neural networks as base classifiers in the first layer, and SVM as a meta-learner in the second layer. Combined with K-fold cross-validation and self-training, it forms a semi-supervised fault diagnosis model for charging facilities, as detailed below: 1) Constructing a Stacking ensemble learning model: including base classifier algorithms and meta-learner algorithms. The first layer base classifier is constructed using GBDT, CatBoost, and convolutional neural networks, which have strong fault diagnosis performance and large algorithm differences. The second layer meta-learner is SVM. 2) Initial training of the Stacking ensemble learning model: The equalized labeled dataset is divided into training set D in an 8:2 ratio. train and test set D test , training set D train The training set M for base classifiers is divided according to 4-fold cross-validation. b1 M b2 M b3 Base classifier test set M test Using the base classifier training set, train three base classifiers: GBDT, CatBoost, and a convolutional neural network. Then, use the trained base classifiers on the base classifier test set M. test Make predictions to obtain the prediction results of the base classifier; training set D train Each sample in the dataset is predicted once. The prediction results of the base classifiers are combined in the second dimension to obtain the training data for the meta-learner, and the SVM is trained. 3) Employ a self-training method to expand the training set and iteratively train the Stacking ensemble learning model until all unlabeled datasets are pseudo-labeled or the number of iterations reaches a set value. The specific steps are as follows: ①The three base classifiers, GBDT, CatBoost and convolutional neural network, are referred to as S1, S2 and S3 respectively. The total number of iterations is set to T. A certain number of unlabeled data are selected from the collected unlabeled dataset as the unlabeled data training set for the first iteration. ② In the t-th iteration, the samples x in the unlabeled training set are... t The data are input into three base classifiers—GBDT, CatBoost, and a convolutional neural network—which have completed initial training, respectively, to obtain their respective prediction results. The prediction results are then evaluated, and samples X whose predictions from classifiers S2 and S3 are consistent and whose confidence levels are both greater than a preset threshold γ are considered. cre1t Generating pseudo-labels and forming an extended training set M for S1. ex1t The prediction results for S1 and S3 are consistent and the confidence level of both is greater than the preset threshold γ. cre2t Generating pseudo-labels and forming an extended training set M for S2. ex2t The prediction results for S1 and S2 are consistent and the confidence level of both is greater than the preset threshold γ. cre3t Generating pseudo-labels and forming an extended training set M for S3. ex3t ; ③Use the expanded training set M respectively ex1t UM ex1t-1 U…U Mb1 M ex2t UM ex2t-1 U…UM b2 M ex3t UM ex3t-1 U…UM b3 Retrain the three base classifiers and the meta-learner SVM, and update the model parameters; ④ On the unlabeled data selected in the previous iteration, select additional unlabeled data to construct a new unlabeled data training set, and repeat the labeling of the unlabeled data training set samples from the previous iteration to continue iterating and expanding the training set; during the iteration process, dynamically adjust the confidence threshold γ according to the performance changes of the base classifier, and appropriately reduce the threshold when the performance improves in order to include more samples. ⑤ When all the unlabeled datasets collected have been pseudo-labeled or the number of iterations has reached the set value T, stop training and output the three base classifiers and the meta-learner SVM that have been trained. 4) Model Evaluation: Using the trained base classifier, evaluate the model on the test set D. test Make predictions, obtain the prediction results, and then merge them in the second dimension to obtain the meta-learner test data; The trained SVM is used to perform performance tests on the meta-learner test data to obtain fault diagnosis test results, and the model performance is evaluated based on the evaluation metrics.