Fan equipment fault prediction method based on meta-learning and multi-expert model

Abstract

The application discloses a fan equipment fault prediction method based on meta-learning and multi-expert models and belongs to the technical field of equipment fault prediction; the method comprises the following steps: S1, acquiring standardized state monitoring and equipment working condition data of power and transmission equipment from the beginning of the life cycle to the recording moment of the sensor, and constructing a source domain data set; acquiring a state time sequence data set of fan equipment, and constructing a target domain data set; S2, constructing a source domain task training set based on the source domain data set and through a graph regularization method; S3, performing meta-learning training based on the source domain task training set and obtaining a first convolutional neural network fault prediction model suitable for the source domain, migrating the convolutional neural network fault prediction model to the target domain, and obtaining a second convolutional neural network fault prediction model through training; and S4, performing fault prediction through the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and performing multi-expert scoring model weight fitting.

Description

Technical Field

[0001] This invention belongs to the field of equipment failure prediction technology, specifically involving a method for predicting wind turbine equipment failures based on meta-learning and multi-expert models. Background Technology

[0002] Fault prediction methods for wind turbine equipment can be mainly divided into three categories: methods based on physical models, methods based on traditional machine learning, and methods based on deep learning. With the improvement of computer performance and the advancement of artificial intelligence technology, deep learning-based fault prediction technology has developed rapidly and achieved significant results. However, these methods require a large number of fault samples to train an effective deep learning model.

[0003] In real-world operating environments, the maintenance cycle of wind turbine equipment is not static; it can be influenced by factors such as usage frequency, working environment, and turbine type. Maintenance is typically performed after a period of operation, rather than waiting for a failure to occur. Furthermore, some wind turbines have very long lifespans, leading to a scarcity of failure samples. To address this insufficient sample size, several strategies have been adopted: first, accelerating the degradation process of equipment in a laboratory environment by applying extreme loads or using vulnerable materials; second, using exponential failure models to simulate the normal degradation of equipment. While these methods can alleviate the challenge of insufficient real-world failure samples to some extent, they cannot fully simulate real-world operating conditions and failure modes. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for predicting wind turbine equipment faults based on meta-learning and multi-expert models, comprising the following steps: S1. Acquire standardized condition monitoring and equipment operating data of power and transmission equipment from the start of their life cycle to the moment of sensor recording, and construct a source domain dataset; acquire a state time series dataset of wind turbine equipment, and construct a target domain dataset; S2. Construct a source domain task training set based on the source domain dataset and using graph regularization methods; S3. Perform meta-learning training based on the source domain task training set to obtain a first convolutional neural network fault prediction model applicable to the source domain, transfer the convolutional neural network fault prediction model to the target domain and obtain a second convolutional neural network fault prediction model through training. S4. Fault prediction is performed using the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and the weights of the multi-expert scoring model are fitted.

[0005] Preferably, step S1 includes: acquiring relevant data of the electric drive equipment from the start of its life cycle to the moment the sensor records the data, performing data normalization and enhancement processing, constructing a source domain dataset, and using a hash-based sharding method to divide the dataset into a test set and a training set; wherein the training set and the test set are electric drive equipment status monitoring data.

[0006] Preferably, S2 specifically includes: S21. Using the source domain dataset as training samples, each training sample is treated as a node in the graph. The distance between nodes is calculated using Mahalanobis distance, and the Gaussian kernel function is used to convert it into a similarity score, thereby constructing a similarity graph. S22. Similarity is processed through regularization terms and used as a graph regularization constraint; S23. Based on the similarity graph and graph regularization constraints, use the TopK principle or threshold filtering principle to select nodes that meet the predetermined requirements to construct the source domain task training set.

[0007] Preferably, the regularization term is specifically expressed as follows: ; in, and These are nodes and Representation in feature space; Represents a node and nodes Similarity score between them.

[0008] Preferably, S3 specifically includes: based on the source domain task training set, performing source domain inference training and target domain source tracing training using a convolutional neural network fault prediction model; in the source domain inference training stage, the source domain dataset is divided into a training set and a validation set, and the convolutional neural network fault prediction model is trained using the source domain dataset. The training set is used as input, and the first convolutional neural network fault prediction model is optimized using the Adam optimization algorithm. The training steps are repeated iteratively to output the trained first convolutional neural network fault prediction model. The validation set is used as input until the loss function of the first convolutional neural network fault prediction model no longer decreases or reaches the predetermined number of iterations, then the trained first convolutional neural network fault prediction model is output.

[0009] As a preferred approach, the gradient update method in the source domain inference training phase is consistent with the traditional deep learning training method, and multiple gradient descents are used, multiple tasks are randomly selected to calculate the loss separately, and then the meta-learner is updated.

[0010] Preferably, the meta-learner update is specifically represented as follows: ; in, Represents the parameters of the model; Indicates the learning rate; This represents the loss function on the nth test sample.

[0011] Preferably, during the target domain source tracing training phase, a back-inference module is added before the first convolutional neural network fault prediction model to obtain a second convolutional neural network fault prediction model. The basic parameters of the first convolutional neural network fault prediction model are used as pre-training parameters. The second convolutional neural network fault prediction model is trained using the target domain dataset, which is divided into a training set and a validation set. The second convolutional neural network fault prediction model is then optimized using the Adam optimization algorithm. The training steps are repeated iteratively to output the trained second convolutional neural network fault prediction model. The validation set is used as input until the loss function of the second convolutional neural network fault prediction model no longer decreases or reaches the predetermined number of iterations. Then, the trained second convolutional neural network fault prediction model is output.

[0012] Preferably, the reverse inference module uses two temporal convolutional networks: one temporal convolutional network is responsible for encoding future covariates, while the other temporal convolutional network is responsible for encoding past covariates and historical values ​​of the sequence.

[0013] Preferably, step S4 specifically includes the following steps: Sa, simultaneously using the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model to predict the faults of the wind turbine equipment, receiving the wind turbine fault prediction result values ​​of the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and performing multi-expert scoring model weight fitting. Sb, update the parameters of the weighted neural network; Sc, calculate the loss between the weighted output and the true label; Sd, repeat the training steps until the performance of the multi-expert model no longer improves or the predetermined number of iterations is reached; output this multi-expert scoring model; Se, using the multi-expert scoring model for fault prediction, if the fault prediction result exceeds the fault threshold, it is judged that the wind turbine equipment has failed.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This solution utilizes multiple fault datasets for power and transmission equipment. By introducing fault data from other equipment into the source domain meta-learning model, this application achieves cross-domain data fusion, improving the model's generalization ability and prediction accuracy. Furthermore, this method helps optimize the fault feature extraction process and enhances the model's robustness.

[0015] 2. Improve the accuracy of wind turbine equipment fault prediction. This application simultaneously inputs wind turbine fault data into the source domain inference model and the target domain tracing model for prediction, training the fault prediction model from both causal inference and cause-effect tracing perspectives, thus realizing a dual-thread thinking function.

[0016] 3. By combining multi-expert models for weight allocation, the weights of the fault prediction model are adjusted based on real data and prediction results to optimize prediction performance, improve the model's ability to identify complex fault modes, and greatly improve the accuracy of the fault prediction model. Attached Figure Description

[0017] Figure 1 The flowchart shows the wind turbine equipment fault prediction method based on meta-learning and multi-expert models of the present invention. Figure 2 This is a flowchart illustrating the training and use of the wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to embodiments of this application; Figure 3 This is a schematic diagram of the structure of a wind turbine equipment fault prediction system based on meta-learning and multi-expert models according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0018] Example 1: To achieve the objective of this invention, the invention is implemented through the following technical solution.

[0019] S1. Acquire standardized condition monitoring and equipment operating data of power and transmission equipment from the start of their life cycle to the moment of sensor recording, and construct the source domain dataset; acquire the state time series dataset of wind turbine equipment, and construct the target domain dataset.

[0020] S1 includes: acquiring relevant data of the electric drive equipment from the start of its life cycle to the moment of sensor recording, performing data normalization and enhancement processing, constructing a source domain dataset, and using a hash-based sharding method to divide the dataset into a test set and a training set; wherein, the training set and the test set are the status monitoring data of the electric drive equipment.

[0021] First, we construct the source domain dataset, namely the power and transmission equipment state time series dataset, and the target domain dataset, namely the wind turbine equipment state time series dataset. The dataset contains all the data of the power transmission equipment from the beginning of its life cycle to the moment of sensor recording, such as the standardized condition monitoring and equipment operating data of the equipment, including the current, voltage, and speed.

[0022] The preprocessing steps include data normalization and data augmentation. The data normalization uses logarithmic normalization.

[0023] The calculation formula is: ; in, It is the raw data. and These are the minimum and maximum values ​​after taking the logarithm of the data. This is the normalized data. In addition, for the collected time series data, operations can be performed by converting the time series to the frequency domain, such as filtering or gain adjustment, and then converting it back to the time domain to improve the generalization ability of the fault prediction model.

[0024] S2. Construct a source domain task training set based on the source domain dataset and using graph regularization methods.

[0025] S2 specifically includes: S21. Using the source domain dataset as training samples, each training sample is treated as a node in the graph. The distance between nodes is calculated using Mahalanobis distance, and the Gaussian kernel function is used to convert it into a similarity score, thereby constructing a similarity graph. S22. Similarity is processed through regularization terms and used as a graph regularization constraint; S23. Based on the similarity graph and graph regularization constraints, use the TopK principle or threshold filtering principle to select nodes that meet the predetermined requirements to construct the source domain task training set.

[0026] Preferably, the regularization term is specifically expressed as follows: ; in, and These are nodes and Representation in feature space; Represents a node and nodes Similarity score between them.

[0027] Randomly select one fault data point from the source domain with abundant fault samples and use it as the test set for task n. Based on the constructed similarity graph and graph regularization constraints, the K nodes with the highest similarity to the test set nodes are selected as the training set for task n. Therefore, task n can be represented as and The combination of , we denote as .

[0028] The purpose of this step is mainly to preserve the local structure of the original data and to find training samples that are similar to the target task in terms of structure. This allows us to use the similarity information of the data to ensure data consistency and eliminate outliers and noise.

[0029] Graph regularization is used to construct the source domain task set. Specifically, a graph is built on the dataset, where each node represents a training data point, and each edge represents the similarity between node pairs. Let... For the training dataset, where It is the first 1 training sample, construct graph , where V is the set of nodes and E is the set of edges.

[0030] The weights of the edges are calculated using Mahalanobis distance to determine the similarity between nodes. The similarity is calculated as follows: ; Where S is the covariance matrix of the data. It is its inverse matrix. It is the first training samples, It is the first There are training samples, and T is their transpose matrix.

[0031] After calculating the node distances, a Gaussian kernel function is used to convert them into similarity scores. The Gaussian kernel function is calculated as follows: ; in, It's a bandwidth parameter. It is Mahalanobis distance. This indicates that the data is taken as an exponential function.

[0032] The purpose of introducing a graph regularization term is to make similar nodes (time series samples) as close as possible in the feature space, in order to preserve the local geometric structure of the data. The graph regularization term can be expressed as: ; in, and These are nodes and Representation in feature space. node and nodes Similarity score between them.

[0033] S3. Perform meta-learning training based on the source domain task training set to obtain a first convolutional neural network fault prediction model applicable to the source domain. Transfer the convolutional neural network fault prediction model to the target domain and obtain a second convolutional neural network fault prediction model through training.

[0034] S3 specifically includes: based on the source domain task training set, performing source domain inference training and target domain source tracing training through a convolutional neural network fault prediction model; in the source domain inference training stage, the source domain dataset is divided into a training set and a validation set, and the convolutional neural network fault prediction model is trained using the source domain dataset. The training set is used as input, and the first convolutional neural network fault prediction model is optimized using the Adam optimization algorithm. The training steps are repeated iteratively, and the trained first convolutional neural network fault prediction model is output. The validation set is used as input until the loss function of the first convolutional neural network fault prediction model no longer decreases or reaches the predetermined number of iterations, and then the trained first convolutional neural network fault prediction model is output.

[0035] The gradient update method during the source domain inference training phase is consistent with traditional deep learning training methods, employing multiple gradient descent iterations, randomly selecting multiple tasks to calculate the loss separately, and then performing meta-learner updates. The specific representation of meta-learner updates is as follows: ; in, Represents the parameters of the model; Indicates the learning rate; This represents the loss function on the nth test sample.

[0036] In the target domain source tracing training phase, a back reasoning module is added before the first convolutional neural network fault prediction model to obtain a second convolutional neural network fault prediction model. The basic parameters of the first convolutional neural network fault prediction model are used as pre-training parameters. The second convolutional neural network fault prediction model is trained using the target domain dataset, which is divided into a training set and a validation set. The second convolutional neural network fault prediction model is then optimized using the Adam optimization algorithm. The training steps are repeated iteratively to output the trained second convolutional neural network fault prediction model. The validation set is used as input until the loss function of the second convolutional neural network fault prediction model no longer decreases or reaches the predetermined number of iterations. Then, the trained second convolutional neural network fault prediction model is output.

[0037] The reverse inference module uses two temporal convolutional networks: one to encode future covariates and the other to encode past covariates and historical values ​​of the sequence.

[0038] Based on the constructed source domain task set and convolutional neural network fault prediction model, the model is trained and then used for fault prediction of wind turbine equipment.

[0039] The training of the fault prediction model is divided into two stages: source domain inference training and target domain source tracing training.

[0040] In the source domain inference training phase, source domain data is used to train the source domain inference model.

[0041] The Bi-LSTM neural network model consists of three layers: the first layer is the input layer, the second layer is the Bi-LSTM layer, and the third layer is the output layer. The output layer is a Dense layer with the softmax activation function and the cross-entropy loss function. The second layer is composed of a forward LSTM layer and a backward LSTM layer.

[0042] Each LSTM unit consists of the following parts: a unit state, used to store long-term information; a hidden state, used to pass the unit state to the next time step, and also serving as the network's final output; an input gate, controlling the extent to which the current input information enters the unit state; a forget gate, determining which information in the unit state needs to be forgotten; and an output gate, controlling which information in the unit state is output to the hidden state. The source domain inference model is optimized using the Adam optimization algorithm during training. The training steps are repeated iteratively until the source domain inference model's validation set loss no longer decreases or the predetermined number of iterations is reached.

[0043] The gradient update method during the source domain inference training phase is consistent with traditional deep learning training methods. For example, randomly selecting a task... The model is then trained using this task, and the model parameters are updated as follows: ; For simplicity, only one gradient descent step is used here; in practice, multiple gradient descent steps are used. Then we use the updated model on the task... test set Tests were conducted on the above to obtain the losses in this task. .

[0044] The above descriptions all refer to the case of a single task. However, in practice, we will randomly select multiple tasks for updates, and then update the meta-learner after obtaining the losses from these tasks. The calculation formula is shown below: ; In the target domain source tracing training phase, a back-inference module is added before the source domain inference model to obtain the target domain source tracing model. The basic parameters of the source domain inference model are used as pre-training parameters, and the target domain dataset is used to train the target domain source tracing model. The back-inference module uses two temporal convolutional networks: one encodes future covariates, and the other encodes past covariates and historical values ​​of the sequence. The architecture of the temporal convolutional network consists of many temporal blocks, each consisting of: a dilated convolution, a GELU activation function, a dropout layer, and finally a fully connected layer. Each temporal block produces one output value, and the prediction result is obtained by superimposing all the outputs of each block into N layers. The target domain source tracing model is optimized using the Adam optimization algorithm during training. The training steps are repeated iteratively until the validation set loss of the target domain source tracing model no longer decreases or the predetermined number of iterations is reached.

[0045] S4. Fault prediction is performed using the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and the weights of the multi-expert scoring model are fitted.

[0046] S4 specifically includes the following steps: Sa, simultaneously using the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model to predict the faults of the wind turbine equipment, receiving the wind turbine fault prediction result values ​​of the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and performing multi-expert scoring model weight fitting. Sb, update the parameters of the weighted neural network; Sc, calculate the loss between the weighted output and the true label; Sd, repeat the training steps until the performance of the multi-expert model no longer improves or the predetermined number of iterations is reached; output this multi-expert scoring model; Se, using the multi-expert scoring model for fault prediction, if the fault prediction result exceeds the fault threshold, it is judged that the wind turbine equipment has failed.

[0047] Simultaneously, inference models and source tracing models are used to predict faults in wind turbine equipment, and a multi-expert scoring model is fitted with weights based on the prediction results of the two models.

[0048] Receive the wind turbine fault prediction results from the inference model and the source tracing model. Update the parameters of the weighted neural network. Calculate the loss between the weight output and the true label. Repeat the iterative training steps until the loss on the multi-expert model validation set no longer decreases or the predetermined number of iterations is reached. Use the model to predict faults; if the fault prediction result exceeds the fault threshold, the wind turbine equipment is judged to have failed.

[0049] This application improves the accuracy of wind turbine equipment fault prediction. It simultaneously inputs wind turbine fault data into both a source domain inference model and a target domain tracing model for prediction, training the fault prediction model from both causal reasoning and cause-effect tracing perspectives, thus achieving a dual-threaded thinking function.

[0050] Overall, such as Figure 2 As shown, this application process comprises four stages, progressively advancing to achieve fault detection. First, in the source domain inference training stage, the model is trained using a set of source domain tasks. This process involves randomly selecting tasks, calculating the loss function, updating parameters, and repeating this process until the gradient stabilizes or the required number of iterations is reached. Finally, the trained source domain inference model is saved. Next, in the target domain source tracing training stage, after updating the model structure, a similar training process is performed using a set of target domain tasks, and the target domain source tracing model is saved. The third stage is the multi-expert scoring training stage. After initializing the scoring model parameters, the target domain dataset is input, and the scoring model is iteratively optimized and saved. Finally, in the fault detection stage, the models saved in the previous stages (source domain inference model, target domain source tracing model, and multi-expert scoring model) are used to analyze the detection data and output fault prediction results. The entire process, through staged model training and optimization, gradually improves model performance and ultimately achieves accurate fault detection.

[0051] Specifically, this application process consists of four stages. Phase 1: Source Domain Reasoning Training. Specific steps are as follows: Randomly select a training task: randomly select a task from the source domain task set.

[0052] Calculate the loss function: For the selected task, calculate the model's loss function to measure the difference between the model's predictions and the actual values.

[0053] Multi-task average loss and gradient calculation: Calculate the average loss function for multiple tasks and calculate the gradient based on it to provide direction for model parameter updates.

[0054] Update meta-learner parameters: Based on the calculated gradient, update the parameters of the meta-learner to optimize model performance.

[0055] Iterative optimization: Repeat steps 1-4 to continuously update the model parameters until the gradient no longer decreases significantly or the preset number of iterations is reached, ensuring that the model achieves good performance on the source domain task.

[0056] Save the source domain inference model: After training is complete, save the trained source domain inference model for use in subsequent stages.

[0057] Phase Two: Target Domain Source Tracing Training. The specific steps are as follows: Update the model structure: Adjust the model structure according to the characteristics of the target domain task.

[0058] Using the target domain task set: Switch to the target domain task set for training.

[0059] Randomly select a training task: randomly select a task from the target domain task set.

[0060] Calculate the loss function: For the selected target domain task, calculate the loss function of the model.

[0061] Update model parameters: Update model parameters based on the gradients calculated from the loss function to optimize model performance on the target domain task.

[0062] Iterative optimization: Repeat steps 3-5 to continuously update the model parameters until the gradient no longer decreases significantly or the preset number of iterations is reached.

[0063] Save the target domain source tracing model: After training is complete, save the trained target domain source tracing model.

[0064] Phase Three: Multi-Expert Scoring Training. Specific steps are as follows: Initialize multi-expert rating model parameters: Set the initial parameters of the rating model.

[0065] Input target domain dataset: Input the dataset of the target domain into the scoring model.

[0066] Update scoring model parameters: Based on the target domain dataset and preset scoring criteria, update the parameters of the scoring model to enable accurate evaluation of model performance.

[0067] Calculate the loss function: Calculate the loss function of the scoring model to measure the gap between the scoring results and the true performance.

[0068] Iterative optimization: Repeat steps 2-4 to continuously update the scoring model parameters until the loss function no longer decreases significantly or the preset number of iterations is reached.

[0069] Save the scoring model: After training is complete, save the trained multi-expert scoring model.

[0070] Phase Four: Fault Detection. The specific steps are as follows: Using test data: Prepare the data to be tested.

[0071] Load multiple models: Load source domain inference model, source model, target domain tracing model and multi-expert scoring model.

[0072] Output fault prediction results: Analyze the detection data using these models and output comprehensive fault prediction results.

[0073] This application combines multi-expert models for weight allocation, adjusting the weights of the fault prediction model based on real data and prediction results to optimize prediction performance, improve the model's ability to identify complex fault modes, and significantly enhance the accuracy of the fault prediction model. This method introduces fault data from other power transmission equipment into the source domain meta-learning model, achieving cross-domain data fusion and enhancing the model's robustness. This method simultaneously inputs wind turbine fault data into both the source and target domain models for prediction, achieving causal interconnection and dual-threaded thinking, thus improving the accuracy of wind turbine fault prediction. This method combines multi-expert models for weight allocation, adjusting the weights of each model based on prediction results and actual data to improve the model's ability to identify complex fault modes.

Claims

1. A method for predicting wind turbine equipment faults based on meta-learning and multi-expert models, characterized in that, Includes the following steps: S1. Acquire standardized condition monitoring and equipment operating data of power and transmission equipment from the start of their life cycle to the moment of sensor recording, and construct a source domain dataset; acquire a state time series dataset of wind turbine equipment, and construct a target domain dataset; S2. Construct a source domain task training set based on the source domain dataset and using graph regularization methods; S3. Based on the source domain task training set, perform meta-learning training to obtain a first convolutional neural network fault prediction model applicable to the source domain. Transfer the first convolutional neural network fault prediction model to the target domain and obtain a second convolutional neural network fault prediction model through training. S3 specifically includes: based on the source domain task training set, performing source domain inference training and target domain source tracing training through a convolutional neural network fault prediction model; in the source domain inference training stage, the source domain dataset is divided into a training set and a validation set, and the convolutional neural network fault prediction model is trained using the source domain dataset. The training set is used as input, and the first convolutional neural network fault prediction model is optimized using the Adam optimization algorithm. The training steps are repeated iteratively to output the trained first convolutional neural network fault prediction model. The validation set is used as input until the loss function of the first convolutional neural network fault prediction model no longer decreases or reaches the predetermined number of iterations, then the trained first convolutional neural network fault prediction model is output. In the target domain source tracing training phase, a back inference module is added before the first convolutional neural network (CNN) fault prediction model to obtain a second CNN fault prediction model. The basic parameters of the first CNN fault prediction model are used as pre-training parameters. The second CNN fault prediction model is trained using the target domain dataset, which is divided into a training set and a validation set. The second CNN fault prediction model is then optimized using the Adam optimization algorithm. The training steps are repeated iteratively to output the trained second CNN fault prediction model. The validation set is used as input until the loss function of the second CNN fault prediction model no longer decreases or reaches the predetermined number of iterations. The back inference module includes two temporal convolutional networks: one temporal convolutional network is responsible for encoding future covariates, and the other temporal convolutional network is responsible for encoding past covariates and historical values ​​of the state time series dataset. S4. Fault prediction is performed using the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and the weights of the multi-expert scoring model are fitted.

2. The wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to claim 1, characterized in that, S1 includes: acquiring relevant data of the electric drive equipment from the start of its life cycle to the time of sensor recording, performing data normalization and enhancement processing, constructing a source domain dataset, and using a hash-based sharding method to divide the dataset into a test set and a training set; wherein the training set and the test set are the status monitoring data of the electric drive equipment.

3. The wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to claim 1, characterized in that, S2 specifically includes: S21. Using the source domain dataset as training samples, each training sample is treated as a node in the graph. The distance between nodes is calculated using Mahalanobis distance, and the Gaussian kernel function is used to convert it into a similarity score, thereby constructing a similarity graph. S22. Similarity is processed through regularization terms and used as a graph regularization constraint; S23. Based on the similarity graph and graph regularization constraints, use the TopK principle or threshold filtering principle to select nodes that meet the predetermined requirements to construct the source domain task training set.

4. The wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to claim 3, characterized in that, The regularization term is specifically represented as follows: ； in, and These are nodes and Representation in feature space; Represents a node and nodes Similarity score between them.

5. The wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to claim 1, characterized in that, The gradient update method in the source domain inference training stage is consistent with the traditional deep learning training method, and multiple gradient descents are used. Multiple tasks are randomly selected to calculate the loss separately, and then the meta-learner is updated.

6. The wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to claim 5, characterized in that, The specific meaning of updating the meta-learner is as follows: ； in, Represents the parameters of the model; Indicates the learning rate; This represents the loss function on the nth test sample.

7. The wind turbine equipment fault prediction method based on meta-learning and multi-expert models according to claim 1, characterized in that, S4 specifically includes the following steps: Sa, simultaneously using the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model to predict the faults of the wind turbine equipment, receiving the wind turbine fault prediction result values ​​of the first convolutional neural network fault prediction model and the second convolutional neural network fault prediction model, and performing multi-expert scoring model weight fitting. Sb, update the parameters of the weighted neural network; Sc, calculate the loss between the weighted output and the true label; Sd, repeat the training steps until the performance of the multi-expert model no longer improves or the predetermined number of iterations is reached; Output this multi-expert rating model; Se, using the multi-expert scoring model for fault prediction, if the fault prediction result exceeds the fault threshold, it is judged that the wind turbine equipment has failed.

Images