Wind power plant regular check fault identification method based on cloud model

Abstract

The invention provides a wind power plant regular check fault identification method based on a cloud model, and the method comprises the steps: extracting fan operation state features, environment factor features and other key features from processed wind power plant regular check data through feature engineering, and constructing a training data set containing an equipment efficiency classification label; a network architecture of a deep neural network model is designed according to the hierarchical structure of the wind field equipment, and the deep neural network model comprises a convolutional neural network, a long and short term memory network and an attention mechanism; training the deep neural network model by using a focal loss function, and further improving the fault recognition accuracy of the deep neural network model by increasing the attention on the difficult-to-distinguish samples; and performing performance evaluation on the trained deep neural network model by using an independent test set, and evaluating the classification performance of the deep neural network model on each efficiency level by using an accuracy rate, a recall rate and an F1 value index.

Description

Technical Field

[0001] The present invention relates to the field of wind power generation, and in particular to a wind farm scheduled inspection fault identification method based on a cloud model. Background Art

[0002] Safe production and efficient operation and maintenance of wind farms are crucial for the sustainable and healthy development of the wind power industry. Currently, domestic wind farms generally have wind turbine monitoring systems in place, enabling the collection and storage of wind turbine operating data. However, effectively utilizing this massive amount of wind turbine operating data, particularly integrating it with safety production standards and specifications for wind turbine equipment to intelligently diagnose and provide early warnings of wind turbine failures, remains a technical challenge facing wind farm operations and maintenance. Wind turbines are diverse, with different components exhibiting varying failure modes. Furthermore, wind turbine failures are influenced by random factors such as wind speed and direction, resulting in uncertainty and ambiguity. Traditional wind turbine fault diagnosis methods rely primarily on manual experience, often subject to false alarms and missed detections, making them difficult to adapt to the demands of intelligent wind farm operations and maintenance. Therefore, there is an urgent need to explore new wind turbine fault diagnosis methods that can fully exploit the inherent patterns of wind turbine operating big data, adaptively classify and grade wind turbine faults, and improve the level of intelligent wind turbine fault diagnosis. Summary of the Invention

[0003] The present invention provides a wind farm fault identification method based on a cloud model with intelligent diagnosis of wind farm power generation, comprising the following steps:

[0004] A wind farm fault identification method based on a cloud model, characterized by comprising the following steps:

[0005] Obtain wind farm regular inspection data and convert the wind farm regular inspection data into a data format suitable for deep neural network model training; the wind farm regular inspection data includes wind turbine operating status data and environmental factor data;

[0006] Extract key features from wind farm inspection data to construct a training data set; the key features include wind turbine operating status features and environmental factor features; the training data set includes equipment performance classification labels;

[0007] Establishing a network architecture of a deep neural network model based on the hierarchical structure of wind farm equipment, wherein the deep neural network model includes a convolutional neural network, a long short-term memory network, and an attention mechanism;

[0008] The convolutional neural network is used to extract local features of the equipment operation data in the training dataset, the long short-term memory network is used to capture the long-term dependencies of the time series data in the training dataset, and the attention mechanism is used to focus on key node information to build an end-to-end network model that can effectively learn the inherent laws of wind farm regular inspection data;

[0009] Using an independent test set to evaluate the performance of the trained deep neural network model, and evaluating the classification performance of the deep neural network model at each performance level;

[0010] The trained deep neural network model is integrated into the cloud model, integrated or remotely connected with the control terminal and / or wind farm regular inspection mobile application. By associating with production safety standards and equipment data models, real-time monitoring of wind turbine operating status and fault warnings can be achieved. When the equipment efficiency level is abnormal, maintenance suggestions are automatically generated to assist operation and maintenance personnel in making maintenance decisions.

[0011] Furthermore, establishing the network architecture of the deep neural network model includes the following steps:

[0012] Constructing a hybrid neural network including a convolutional layer and a long short-term memory layer, wherein the convolutional layer is used to extract spatial features of the operating data, and the long short-term memory layer is used to capture temporal dependencies in the operating data;

[0013] Introducing an attention mechanism into the hybrid neural network to optimize the direction and strength of information flow by calculating the correlation strength between different devices;

[0014] Integrating the spatial features extracted by the convolutional layer, the temporal features captured by the long short-term memory layer, and the hierarchical correlation information obtained by the attention mechanism to obtain a comprehensive feature representation;

[0015] The comprehensive feature representation is input into the fully connected layer and mapped to the output layer; the operating data of each level of wind farm equipment is obtained, the operating data is used as the input of the neural network model, and the prediction result of the future operating status of the wind farm equipment is obtained through the output of the deep neural network model.

[0016] Furthermore, training the deep neural network model includes the following steps:

[0017] Cleaning and normalizing the wind farm inspection data to obtain a standardized time series data set;

[0018] Using a time series feature extraction tool to extract features from the standardized time series data set to obtain a time series feature set including time domain features and frequency domain features;

[0019] Inputting the temporal feature set into a long short-term memory network, and capturing long-term dependencies in the temporal feature set through a gating mechanism and memory units in the long short-term memory network;

[0020] Applying an attention mechanism to the hidden layer output of the long short-term memory network, calculating the attention weight of each time step through a softmax function, performing weighted summation on the hidden states of the hidden layer output, and obtaining an attention output that integrates key moment information;

[0021] The attention output is sent to the fully connected layer, and the attention output is nonlinearly transformed by the fully connected layer to obtain the health status prediction result of the wind farm regular inspection equipment

[0022] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0023] The present invention discloses a deep learning-based method for classifying wind farm equipment efficiency. This method, targeting the characteristics of wind farm inspection data, uses data preprocessing and feature engineering techniques to extract key features, constructing a deep neural network model that includes a convolutional neural network, a long-short-term memory network, and an attention mechanism. Ultimately, the trained model is integrated into a cloud model and integrated with a control terminal and / or a wind farm inspection mobile application, or remotely connected to the wind farm inspection mobile application. This enables real-time monitoring of wind turbine operating status and fault warnings, and automatically generates maintenance recommendations, effectively improving the accuracy and practicality of wind farm equipment efficiency classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a flow chart of a wind farm scheduled inspection fault identification method based on a cloud model of the present invention.

[0025] Figure 2 This is a schematic diagram of a wind farm scheduled inspection fault identification method based on a cloud model according to the present invention.

[0026] Figure 3 This is another schematic diagram of a wind farm fault identification method based on a cloud model according to the present invention. DETAILED DESCRIPTION

[0027] The following will describe the technical solutions in the embodiments of the present invention in detail with reference to the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention.

[0028] like Figure 1-3 This embodiment provides a wind farm fault identification method based on a cloud model, comprising the following steps:

[0029] S1. Obtain wind farm inspection data and convert it into a data format suitable for deep neural network model training; the wind farm inspection data includes wind turbine operating status data and environmental factor data; after obtaining the wind farm inspection data, use data preprocessing technology to clean and normalize the wind farm inspection data, remove noise and outliers in the wind farm inspection data, and then convert the processed wind farm inspection data into a data format suitable for deep neural network model training. Use principal component analysis to extract features from the wind farm inspection data to obtain wind turbine operating status characteristics and environmental factor characteristics;

[0030] S2. Extract key features from the wind farm inspection data and construct a training data set; the key features include wind turbine operating status features and environmental factor features; the training data set includes equipment performance classification labels; wherein, exploratory analysis is performed on the original inspection data to determine the quality and feature distribution of the data, and to determine the specific operations that require data cleaning and normalization. Using data cleaning technology, reasonable thresholds are set for judging missing values, outliers and other noise data in the wind farm inspection data, and data that exceeds the threshold range is corrected or deleted to improve the quality and reliability of the data. Through normalization processing, the cleaned wind farm inspection data is scaled, and data of different dimensions are mapped to the same scale, eliminating the impact of dimensional differences on model training, improving data comparability and model generalization capabilities. According to the input requirements of the deep neural network model, the normalized wind farm inspection data is converted into a data format suitable for model training, such as converting time series data into a supervised learning problem, and constructing input features and target labels. Feature selection techniques are used to analyze the correlation between each feature in the wind farm inspection data and the target variable. Features with high correlation with the target variable are selected as model input, reducing data dimensionality and reducing the complexity of model training. Data augmentation techniques are used to address the small sample size of wind farm inspection data. Interpolation, rotation, and translation are used to expand the original data, increasing the number and diversity of training samples and improving the generalization and robustness of the model. The processed wind farm inspection data is divided into a training set, a validation set, and a test set. The training set is used to train the deep neural network model, the validation set is used to adjust model hyperparameters and avoid overfitting, and the test set is used to evaluate the model's performance and generalization. In addition, it is also possible to obtain equipment efficiency classification labels corresponding to the wind farm inspection data, and construct a training data set with the wind turbine operating status characteristics, the environmental factor characteristics and the corresponding equipment efficiency classification labels; use the CART decision tree algorithm to perform model training on the training data set to obtain an equipment efficiency evaluation model; obtain the operating status characteristics and environmental factor characteristics of the target wind turbine, input them into the equipment efficiency evaluation model, and obtain the equipment efficiency evaluation result of the target wind turbine; obtain a preset equipment efficiency evaluation threshold, and compare the equipment efficiency evaluation result of the target wind turbine with the equipment efficiency evaluation threshold; if the equipment efficiency evaluation result is lower than the equipment efficiency evaluation threshold, it is judged that the equipment efficiency of the target wind turbine is lower than the preset threshold and needs to be maintained or optimized.

[0031] S3. Establishing a network architecture of a deep neural network model according to the hierarchical structure of the wind farm equipment, wherein the deep neural network model includes a convolutional neural network, a long short-term memory network, and an attention mechanism; establishing the network architecture of the deep neural network model includes the following steps:

[0032] Constructing a hybrid neural network including a convolutional layer and a long short-term memory layer, wherein the convolutional layer is used to extract spatial features of the operating data, and the long short-term memory layer is used to capture temporal dependencies in the operating data;

[0033] Introducing an attention mechanism into the hybrid neural network to optimize the direction and strength of information flow by calculating the correlation strength between different devices;

[0034] Integrating the spatial features extracted by the convolutional layer, the temporal features captured by the long short-term memory layer, and the hierarchical correlation information obtained by the attention mechanism to obtain a comprehensive feature representation;

[0035] The comprehensive feature representation is input into the fully connected layer and mapped to the output layer; the operating data of each level of wind farm equipment is obtained, the operating data is used as the input of the neural network model, and the prediction result of the future operating status of the wind farm equipment is obtained through the output of the deep neural network model.

[0036] Training the model may also include the following steps:

[0037] Acquire a historical data set of equipment performance, wherein the historical data set includes a plurality of equipment performance data samples;

[0038] Establishing a fuzzy concept membership function for equipment effectiveness classification based on the historical data set;

[0039] Generate the probability distribution of device performance level based on expectation, entropy and hyperentropy through the forward cloud generator;

[0040] For each device performance data sample in the historical data set, calculating its membership to each fuzzy concept using the membership function to obtain a probability distribution vector of the device performance level;

[0041] Using the performance level corresponding to the maximum degree of membership in the probability distribution vector as a label for the device performance data sample to form a training data set;

[0042] Selecting a long short-term memory network as a deep learning model, taking the device performance data samples as input, and taking the performance level labels in the training data set as training targets, to train the long short-term memory network model;

[0043] Obtain new equipment performance data and predict the probability distribution of its performance level using the long short-term memory network model;

[0044] Taking the efficiency level with the highest probability in the probability distribution as the final classification result;

[0045] If the maximum probability exceeds a preset probability threshold, the prediction result is considered credible;

[0046] Otherwise, the device performance data is added to the training data set, and the long short-term memory network model is retrained.

[0047] Specifically, an original training dataset containing 1,000 device performance samples was obtained from a historical database, including 200 samples for performance level A, 300 samples for performance level B, and 500 samples for performance level C. For the smaller number of performance level A samples, data augmentation techniques were used to generate new training samples similar to the original samples by randomly rotating each sample from 0 to 360 degrees, translating from -10% to 10%, and scaling it by a factor of 8 to 2. Four new samples were generated for each original sample. These 800 new samples were then added to the original training dataset, resulting in an augmented training dataset containing 2,000 samples. The augmented training dataset was found to have an imbalanced distribution of 1,000 samples for performance level A, 300 for performance level B, and 500 for performance level C. Therefore, data augmentation was performed on the performance level B and C samples, generating 700 and 500 new samples, respectively, ultimately resulting in a balanced training dataset with 1,000 samples for each performance level. Using a 5-fold cross-validation method, the balanced training dataset was randomly divided into five subsets. Four of these subsets were selected as training sets each time, and the remaining subset was used as the validation set. Cross-validation was performed five times. Using the training set, a performance rating classification model was trained using the gradient boosting decision tree algorithm. The number of trees was set to 100, the maximum depth was set to 6, and the learning rate was set to 1. After 50 rounds of iterative optimization, the trained classification model was obtained. The trained classification model was used to predict performance ratings on the validation set, and the prediction accuracy was calculated. The average accuracy of the five cross-validations was 85%, which did not reach the preset threshold of 90%. The model parameters were adjusted by increasing the number of trees to 150, the maximum depth to 8, and the learning rate to 0.5. Training and optimization continued. After these parameter adjustments, the average accuracy of the model on the validation set increased to 92%, reaching the preset threshold. The trained performance rating classification model was saved for subsequent performance rating prediction tasks.

[0048] Training the model may also include the following steps:

[0049] Obtaining an original training data set, wherein the original training data set includes samples of multiple different performance levels;

[0050] For the minority class samples in the original training data set, data augmentation technology is used to generate new training samples similar to the minority class samples by rotating, translating and scaling the minority class samples;

[0051] Adding the new training samples to the original training data set to obtain an expanded training data set;

[0052] Determining whether the number of samples of each performance level in the expanded training data set is distributed evenly;

[0053] If the number of samples of each performance level in the expanded training data set is unbalanced, returning to perform the data augmentation operation until the number of samples of each performance level in the expanded training data set is balanced, thereby obtaining a balanced training data set;

[0054] Using a cross-validation method to divide the balanced training data set into a training set and a validation set;

[0055] Using the training set, a gradient boosting decision tree algorithm is used to train the effectiveness level classification model, and through multiple rounds of iterative optimization, a trained effectiveness level classification model is obtained;

[0056] Using the trained efficacy grade classification model, predicting the efficacy grade of the validation set, and calculating the prediction accuracy;

[0057] If the prediction accuracy does not reach the preset threshold, return to adjust the model parameters of the gradient boosting decision tree algorithm and continue to perform the model training operation;

[0058] If the prediction accuracy reaches the preset threshold, the trained performance level classification model is saved for subsequent performance level prediction.

[0059] Training the model may also include the following steps:

[0060] Obtain a pre-labeled fault sample data set, preprocess the sample data, and extract the fault feature vector;

[0061] According to the extracted fault feature vector, the structure and parameters of the input layer, hidden layer and output layer of the deep neural network model are set with respect to the fault feature;

[0062] Focal loss is used as the loss function. Focal loss introduces a modulation factor to reduce the weight of easy-to-classify samples and increase the weight of difficult-to-classify samples, making the model pay more attention to difficult-to-classify samples.

[0063] The preprocessed fault sample dataset is divided into a training set and a validation set, which are then input into the deep neural network model. The model is trained using the training set data, and the model performance is evaluated using the validation set data.

[0064] During the model training process, the focal loss hyperparameters are dynamically adjusted, and the model weights and biases are continuously updated through the gradient descent algorithm, so that the loss function value of the model on the training set is continuously reduced and the recognition accuracy is continuously improved;

[0065] When the recognition accuracy of the model on the validation set reaches the preset threshold or the number of training rounds reaches the set value, the training is stopped and the trained fault recognition model is obtained;

[0066] The trained fault identification model is applied to actual fault diagnosis scenarios to predict and classify the input fault data, output the fault type and cause, and assist maintenance personnel in fault location and troubleshooting.

[0067] During training, a deep learning model suitable for time series forecasting, such as a long short-term memory (LSTM) or a gated recurrent unit (GRU), is selected. The model architecture is constructed and trained. During training, early stopping and learning rate adjustment strategies are used to monitor performance on the validation set to avoid overfitting. Finally, the model's performance is evaluated on the test set, using metrics such as root mean square error (RMSE) and mean absolute error (MAE) to comprehensively measure the model's predictive ability. Furthermore, the model's prediction results are visualized and compared with the true values ​​to intuitively assess the model's predictive effectiveness, providing a reference for subsequent model optimization and application.

[0068] S4. Using the convolutional neural network to extract local features of the equipment operation data in the training data set, using the long short-term memory network to capture the long-term dependencies of the time series data in the training data set, focusing on key node information through the attention mechanism, and constructing an end-to-end network model that can effectively learn the inherent laws of wind farm inspection data; training the deep neural network model includes the following steps:

[0069] Cleaning and normalizing the wind farm inspection data to obtain a standardized time series data set;

[0070] Using a time series feature extraction tool to extract features from the standardized time series data set to obtain a time series feature set including time domain features and frequency domain features;

[0071] Inputting the temporal feature set into a long short-term memory network, and capturing long-term dependencies in the temporal feature set through a gating mechanism and memory units in the long short-term memory network;

[0072] Applying an attention mechanism to the hidden layer output of the long short-term memory network, calculating the attention weight of each time step through a softmax function, performing weighted summation on the hidden states of the hidden layer output, and obtaining an attention output that integrates key moment information;

[0073] The attention output is sent to a fully connected layer, and a nonlinear transformation is performed on the attention output through the fully connected layer to obtain a health status prediction result of the wind farm regular inspection equipment.

[0074] S5. Using an independent test set, the trained deep neural network model is subjected to performance evaluation to evaluate the classification performance of the deep neural network model at each performance level. The performance evaluation of the trained deep neural network model comprises the following steps:

[0075] Obtaining a pre-prepared independent test set, wherein the test set includes sample data of multiple different performance levels;

[0076] Use PyTorch to load the trained convolutional neural network model into memory and initialize the relevant parameters;

[0077] Preprocessing the sample data in the test set, wherein the preprocessing includes data normalization and feature extraction, and converting the preprocessed data into an input format of the model;

[0078] Inputting the preprocessed test sample data into the convolutional neural network model, and obtaining the performance level prediction result of the model for each sample through forward propagation;

[0079] For each test sample, compare the performance level label predicted by the model with the actual performance level label of the sample, and count the number of samples with correct and incorrect predictions for each performance level;

[0080] According to the number of correct and incorrect prediction samples for each performance level, sklearn is used;

[0081] The precision_score and recall_score functions in metrics calculate the precision and recall rate indicators for each performance level;

[0082] After obtaining the precision and recall indicators of each performance level, sklearn was further used;

[0083] The f1_score function in metrics calculates the F1 value of each performance level, which includes macro-average F1 and micro-average F1;

[0084] The evaluation results of the model on the test set are output and saved in a file. The evaluation results include the precision, recall and F1 value of each performance level, which are used for subsequent model analysis and optimization.

[0085] As an example, specifically, a pre-prepared independent test set is obtained, which contains sample data of multiple different performance levels; the trained convolutional neural network model is loaded into the memory using PyTorch, and the relevant parameters are initialized; the sample data in the test set are preprocessed, and the preprocessing includes data normalization and feature extraction, and the preprocessed data is converted into the input format of the model; the preprocessed test sample data is input into the convolutional neural network model, and the performance level prediction result of the model for each sample is obtained through forward propagation; for each test sample, the performance level label predicted by the model is compared with the true performance level label of the sample, and the number of samples with correct and incorrect predictions for each performance level is counted; based on the number of samples with correct and incorrect predictions for each performance level, the precision_score and recall_score functions in sklearn metrics are used to calculate the precision and recall rate indicators of each performance level; after obtaining the precision and recall rate indicators of each performance level, sklearn is further used. The f1_score function in metrics calculates the F1 value of each performance level, which includes the macro-average F1 and micro-average F1; the evaluation results of the model on the test set are output and saved to a file, and the evaluation results include the precision, recall rate and F1 value of each performance level for subsequent model analysis and optimization; based on the model evaluation results, the performance of the convolutional neural network on samples of different performance levels is analyzed, and the performance levels with more classification errors than the preset threshold are identified. The reasons are further analyzed, and the model is optimized and improved in a targeted manner.

[0086] Taking a data example as a reference, we first randomly selected 1,000 samples from a pre-prepared independent test set, including 300 high-performance samples, 400 medium-performance samples, and 300 low-performance samples. The trained convolutional neural network model was loaded into memory using PyTorch, with the learning rate set to 0.01 and the momentum factor set to 9. The test set data was preprocessed, scaling the data to the range [0, 1] using min-max normalization, and key sample features, such as average response time and resource utilization, were extracted. The preprocessed data was converted to the tensor format required by the convolutional neural network model. The preprocessed test sample data was input into the convolutional neural network model, and forward propagation was used to obtain the model's performance level prediction for each sample. For each test sample, the model's predicted performance level label was compared with the actual performance level label for that sample, and the number of correct and incorrect predictions for high-performance, medium-performance, and low-performance samples was counted. The number of correct and incorrect predictions for each performance level was calculated using sklearn. The precision_score and recall_score functions in metrics calculate precision and recall metrics for high, medium, and low performance levels. The results show that the precision for the high performance level is 92 and the recall is 88; the precision for the medium performance level is 85 and the recall is 90; and the precision for the low performance level is 89 and the recall is 86. After obtaining the precision and recall metrics for each performance level, we further used sklearn. The f1_score function in metrics calculates the F1 score for each performance level, resulting in an F1 score of 90 for the high performance level, 87 for the medium performance level, and 88 for the low performance level. The macro-average F1 score is 88, and the micro-average F1 score is 89. The model evaluation results on the test set are output and saved to a file for subsequent model analysis and optimization. The model evaluation results show that the convolutional neural network's classification errors on the medium performance level samples exceed the preset threshold of 20. Further analysis reveals that the feature distribution of these samples differs from that of the training set. In order to improve the performance of the model on medium-performance samples, you can consider increasing the number of medium-performance samples in the training set and adjusting the model architecture and hyperparameters, such as increasing the number of convolutional layers and adjusting the learning rate, to improve the generalization ability and robustness of the model.

[0087] S6. Integrate the trained deep neural network model into the cloud model, integrate or remotely connect it with the control terminal and / or wind farm regular inspection mobile application, and achieve real-time monitoring of the wind turbine operating status and fault warning by associating it with production safety standards and equipment data models. When the equipment efficiency level is abnormal, maintenance suggestions are automatically generated to assist operation and maintenance personnel in making maintenance decisions.

[0088] When training or learning the model, operational data from wind farm equipment at all levels, including parameters such as turbine blade speed, generator speed, wind direction, and wind speed, is collected at a sampling frequency of once per minute and used as input for the neural network model. A hybrid neural network is then constructed, consisting of three convolutional layers and two long-short-term memory (LSTM) layers. The convolutional kernels are 3x3, 5x5, and 7x7, respectively, to extract spatial features from the operational data. The LSTM layer has 128 hidden units to capture temporal dependencies within the data. An attention mechanism is introduced into the hybrid neural network. By calculating the Pearson correlation coefficient between different devices, a device correlation matrix is ​​generated to optimize the direction and strength of information flow. The spatial features extracted by the convolutional layers, the temporal features captured by the LSTM layers, and the hierarchical correlation information obtained by the attention mechanism are concatenated to produce a 512-dimensional comprehensive feature representation. Finally, this comprehensive feature representation is input into a fully connected layer with 256 neurons and mapped to the output layer via a softmax function. This produces operational status predictions for wind farm equipment for the next 1, 3, and 24 hours, with prediction accuracy exceeding 95%. By analyzing the historical operating data of wind farm equipment and setting early warning thresholds for equipment operating status, relevant equipment is promptly inspected and maintained when the predicted results exceed the threshold range, thereby minimizing wind farm operating failures and improving power generation efficiency.

[0089] During actual operation, sensors installed on the fan collect real-time data on operating parameters such as speed, oil temperature, and vibration at a sampling frequency of 10 times per second. This data is transmitted to a cloud server via Industrial Ethernet at a transmission rate of 100 Mbps. A pre-trained LSTM neural network model is deployed on the cloud server. This model uses a three-layer LSTM architecture with 128 neurons per layer, employing the Adam optimizer with a learning rate set to 0.01. The model analyzes the received fan operating data in real time and, through forward propagation, generates real-time operating status monitoring results for the fan, including speed deviation, abnormal oil temperature, and abnormal vibration. These monitoring results are compared with preset safety thresholds. For example, if the speed deviation exceeds 5% of the rated speed, the oil temperature exceeds 80°C, or the vibration amplitude exceeds 5g, the fan is deemed to be operating abnormally, triggering a corresponding fault warning. Based on the wind turbine's criticality level (e.g., high, medium, and low) and fault type (e.g., bearing wear, gearbox failure), a case-based reasoning approach is used to match and recommend appropriate maintenance measures from a pre-established maintenance knowledge base containing 500 cases. For example, if a fault warning indicating excessive bearing temperature is detected for a high-criticality wind turbine, maintenance measures such as bearing replacement and lubrication adjustment are matched from the knowledge base and distributed to on-site maintenance personnel. Based on the recommended measures, maintenance personnel develop a detailed maintenance plan, such as completing bearing replacement within the next three days. During the maintenance process, wind turbine operating data is continuously collected and uploaded to the cloud for analysis using an LSTM model. Maintenance effectiveness is evaluated by comparing changes in parameters such as vibration levels and temperature before and after the maintenance. If the vibration level drops below 1g and the temperature returns to the normal operating range after the maintenance, the wind turbine is considered to have resumed normal operation. The fault warning and operating data before and after the repair were recorded and used to continuously optimize the LSTM model parameters, such as adjusting the number of network layers and neurons, to continuously improve the model's fault diagnosis and prediction capabilities. At the same time, this case was added to the repair knowledge base to enrich the knowledge base and provide a reference for the diagnosis and repair of similar faults in the future.

[0090] The above embodiments are intended to illustrate the technical solutions of the present invention and are not intended to limit the present invention. The present invention is described in detail with reference to the preferred embodiments only. It should be understood by those skilled in the art that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or equivalents should be included within the scope of the claims of the present invention.

Claims

1. A wind farm fault identification method based on cloud model, characterized in that: The following steps are involved: Obtain wind farm inspection data, and convert the wind farm inspection data into a data format suitable for deep neural network model training; the wind farm inspection data includes wind turbine operating status data and environmental factor data; Extract key features from wind farm inspection data to construct a training data set; the key features include wind turbine operating status features and environmental factor features; The training data set includes equipment performance classification labels; According to the hierarchical structure of the wind farm equipment, a network architecture of a deep neural network model is established, wherein the deep neural network model includes a convolutional neural network, a long short-term memory network and an attention mechanism; The convolutional neural network is used to extract local features of the equipment operation data in the training data set, the long short-term memory network is used to capture the long-term dependencies of the time series data in the training data set, and the attention mechanism is used to focus on key node information to build an end-to-end network model that can effectively learn the internal laws of wind farm regular inspection data; Using an independent test set to perform performance evaluation on the trained deep neural network model, and evaluating the classification performance of the deep neural network model at each performance level; The trained deep neural network model is integrated into the cloud model, and merged or remotely connected with the control terminal and / or wind farm regular inspection mobile application. By associating with production safety standards and equipment data models, real-time monitoring of the wind turbine operating status and fault warning can be achieved. When the equipment efficiency level is abnormal, maintenance suggestions are automatically generated to assist operation and maintenance personnel in making maintenance decisions.

2. A wind farm scheduled inspection fault identification method based on a cloud model according to claim 1, characterized in that: Building the network architecture of a deep neural network model involves the following steps: Constructing a hybrid neural network including a convolutional layer and a long short-term memory layer, wherein the convolutional layer is used to extract the spatial features of the operation data, and the long short-term memory layer is used to capture the temporal dependency in the operation data; Introducing an attention mechanism into the hybrid neural network to optimize the direction and strength of information flow by calculating the correlation strength between different devices; Integrate the spatial features extracted by the convolutional layer, the temporal features captured by the long short-term memory layer, and the hierarchical association information obtained by the attention mechanism to obtain a comprehensive feature representation; The comprehensive feature representation is input into the fully connected layer and mapped to the output layer; the operating data of each level of the wind farm equipment is obtained, the operating data is used as the input of the neural network model, and the prediction result of the future operating status of the wind farm equipment is obtained through the output of the deep neural network model.

3. The method for identifying wind farm faults based on a cloud model according to claim 1, characterized in that: Training the deep neural network model includes the following steps: Cleaning and normalizing the wind farm inspection data to obtain a standardized time series data set; Using a time series feature extraction tool to extract features from the standardized time series data set to obtain a time series feature set including time domain features and frequency domain features; Inputting the temporal feature set into a long short-term memory network, and capturing the long-term dependency in the temporal feature set through a gating mechanism and a memory unit in the long short-term memory network; Applying an attention mechanism to the hidden layer output of the long short-term memory network, calculating the attention weight of each time step through a Softmax function, performing weighted summation on the hidden states of the hidden layer output, and obtaining an attention output that integrates key moment information; The attention output is sent to a fully connected layer, and a nonlinear transformation is performed on the attention output through the fully connected layer to obtain a health status prediction result of the wind farm regular inspection equipment.

4. The method for identifying wind farm faults based on a cloud model according to claim 3 is characterized in that: Training the deep neural network model also includes the following steps: Acquire a historical data set of equipment performance, wherein the historical data set includes a plurality of equipment performance data samples; Establishing a fuzzy concept membership function for equipment effectiveness classification based on the historical data set; Generate the probability distribution of equipment performance level based on expectation, entropy and hyperentropy through the forward cloud generator; For each equipment performance data sample in the historical data set, the membership function is used to calculate its membership to each fuzzy concept, and a probability distribution vector of the equipment performance level is obtained; Using the performance level corresponding to the maximum membership degree in the probability distribution vector as a label of the equipment performance data sample to form a training data set; Selecting a long short-term memory network as a deep learning model, taking the equipment performance data sample as input, taking the performance level label in the training data set as a training target, and training the long short-term memory network model; Acquire new equipment performance data, and predict the probability distribution of its performance level through the long short-term memory network model; Taking the efficiency level with the highest probability in the probability distribution as the final classification result; If the maximum probability exceeds the preset probability threshold, the prediction result is considered to be credible; Otherwise, the device performance data is added to the training data set, and the long short-term memory network model is retrained.

5. The method according to claim 3 or 4, characterized in that: Training the deep neural network model also includes the following steps: Acquire an original training data set, wherein the original training data set includes samples of multiple different performance levels; For the minority class samples in the original training data set, data enhancement technology is used to generate new training samples similar to the minority class samples by rotating, translating and scaling the minority class samples; Adding the new training sample to the original training data set to obtain an expanded training data set; Determining whether the distribution of the number of samples of each performance level in the expanded training data set is balanced; If the number of samples of each performance level in the expanded training data set is unbalanced, return to perform the data enhancement operation until the number of samples of each performance level in the expanded training data set is balanced, thereby obtaining a balanced training data set; Using a cross-validation method to divide the balanced training data set into a training set and a validation set; Using the training set, a gradient boosting decision tree algorithm is used to train the performance level classification model, and a trained performance level classification model is obtained through multiple rounds of iterative optimization; Using the trained performance grade classification model, predicting the performance grade of the validation set, and calculating the prediction accuracy; If the prediction accuracy does not reach the preset threshold, return to adjust the model parameters of the gradient boosting decision tree algorithm and continue to perform the model training operation; If the prediction accuracy reaches the preset threshold, the trained performance level classification model is saved for subsequent performance level prediction.

6. The method according to claim 3, characterized in that Training the deep neural network model also includes the following steps: Obtain a pre-labeled fault sample data set, pre-process the sample data, and extract the fault feature vector; According to the extracted fault feature vector, the structure and parameters of the input layer, hidden layer and output layer of the deep neural network model regarding the fault feature are set; Focalloss is used as the loss function. Focalloss introduces a modulation factor to reduce the weight of easy-to-classify samples and increase the weight of difficult-to-classify samples, so that the model pays more attention to difficult-to-classify samples. The preprocessed fault sample data set is divided into a training set and a validation set, and input into the deep neural network model. The model is trained using the training set data, and the performance of the model is evaluated using the validation set data. During the model training process, the focalloss hyperparameters are dynamically adjusted, and the model weights and biases are continuously updated through the gradient descent algorithm, so that the loss function value of the model on the training set is continuously reduced and the recognition accuracy is continuously improved; When the recognition accuracy of the model on the validation set reaches a preset threshold or the number of training rounds reaches a set value, the training is stopped to obtain a trained fault recognition model; The trained fault identification model is applied to actual fault diagnosis scenarios to predict and classify the input fault data, output the fault type and cause, and assist maintenance personnel in locating and troubleshooting faults.

7. The method according to claim 1, characterized in that The performance evaluation of the trained deep neural network model includes the following steps: Obtaining a pre-prepared independent test set, wherein the test set includes sample data of multiple different performance levels; Use PyTorch to load the trained convolutional neural network model into memory and initialize relevant parameters; Preprocessing the sample data in the test set, wherein the preprocessing includes data normalization and feature extraction, and converting the preprocessed data into an input format of the model; Input the preprocessed test sample data into the convolutional neural network model, and obtain the performance level prediction result of the model for each sample through forward propagation; For each test sample, compare the performance level label predicted by the model with the actual performance level label of the sample, and count the number of samples with correct and incorrect predictions for each performance level; According to the number of correct and incorrect samples predicted for each performance level, sklearn is used; The precision_score and recall_score functions in metrics calculate the precision and recall indicators for each performance level; After obtaining the precision and recall indicators of each performance level, sklearn was further used; The f1_score function in metrics calculates the F1 value of each performance level, which includes macro-average F1 and micro-average F1; The evaluation results of the model on the test set are output and saved in a file. The evaluation results include the precision, recall and F1 value of each performance level for subsequent model analysis and optimization.

Images