A wind turbine icing fault diagnosis method, system, device and storage medium
By using LeNet5like networks and transfer learning algorithms, the problems of high sensor cost, slow calculation speed and insufficient model generalization ability in wind turbine icing fault diagnosis are solved, and fast and accurate wind turbine icing diagnosis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTH CHINA ELECTRIC POWER UNIV
- Filing Date
- 2023-06-27
- Publication Date
- 2026-07-31
AI Technical Summary
Existing methods for diagnosing icing faults in wind turbines suffer from problems such as high sensor costs, high maintenance costs, slow calculation speed, insufficient model generalization ability, and long calculation time, making it difficult to obtain reliable diagnostic results on different wind turbines.
By employing a LeNet5like network combined with a transfer learning algorithm, monitoring data and icing conditions of wind turbine units are obtained. The LeNet5like network is trained using the first training set to obtain a pre-trained network, and the network parameters are fine-tuned using the second training set to establish an icing fault diagnosis model.
This improved the network model's ability to extract fault features, enhanced its generalization and accuracy, enabled rapid and accurate diagnosis of wind turbine icing, and shortened the model building time.
Smart Images
Figure CN116771615B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wind turbine icing fault diagnosis, and in particular to a method, system, device and storage medium for wind turbine icing fault diagnosis. Background Technology
[0002] New energy sources, primarily photovoltaic and wind power, are continuously developing, with installed capacity increasing year by year. However, large wind turbine units built in harsh environments, such as offshore and high-altitude areas, are prone to equipment failures. Among these failures, icing of wind turbine blades is a major issue, which can lead to blade breakage, ice spillage injuries, and other serious hazards.
[0003] Currently, diagnostic research methods for wind turbine icing faults are mainly divided into two categories: icing mechanism analysis methods and data-driven methods. Traditional icing mechanism analysis methods mainly use sensors and internal icing mechanisms to construct icing mathematical models for icing diagnosis. However, this method has low diagnostic accuracy, poor computational power, and slow calculation speed; moreover, the cost of sensors is high and maintenance costs are high, and it relies too much on experience to build models, resulting in poor universality.
[0004] With the widespread application of Supervisory Control and Data Acquisition (SCADA) systems in wind farms, a large amount of data containing abnormal operating conditions has become readily available. Many diagnostic methods that use historical SCADA data to build wind turbine models through neural networks have been proposed, achieving economic advantages in detecting icing faults in wind turbines. However, these methods ignore the differences in the characteristic distribution of different wind turbines, resulting in insufficient generalization ability on different wind turbine test sets, excessively long retraining time for the model, and difficulty in building multi-wind turbine detection models.
[0005] As can be seen from the above research, the traditional wind turbine icing fault diagnosis method has two main problems: (1) To establish an accurate icing mathematical model, multiple sensors need to be installed, which will add extra production and maintenance costs; (2) In actual scenarios, the monitoring data of different wind turbines are quite different, which makes it difficult for even the model trained with data from the same type of wind turbine to obtain reliable diagnostic results on data from other wind turbines. The established model is difficult to extend to multi-wind turbine scenarios and lacks robustness.
[0006] In recent years, deep learning has developed rapidly. Among them, transfer learning has shown remarkable effectiveness in handling heterogeneous data and improving the generalization ability of models, as it can transfer the parameters of a trained model to a new model to help train the new model. However, the transfer learning models established in existing research have slow network computation speeds, long model building times, and significant difficulties in acquiring and processing image data, making them unsuitable for establishing multi-wind turbine icing diagnosis models.
[0007] Traditional LeNet-5 neural networks have relatively shallow layers, resulting in low fitting rates when trained on SCADA data. While the aforementioned transfer learning method achieves better generalization than traditional wind turbine icing fault diagnosis methods, there is still room for further improvement. Summary of the Invention
[0008] The purpose of this invention is to provide a method, system, device, and storage medium for diagnosing icing faults in wind turbines, thereby improving the network model's ability to extract fault features, enhancing the model's generalization ability, and improving the model's accuracy while ensuring speed, thus effectively achieving rapid and accurate diagnosis of wind turbine icing.
[0009] To achieve the above objectives, the present invention provides the following solution:
[0010] A method for diagnosing icing faults in wind turbine units includes:
[0011] Acquire monitoring data and icing status of wind turbine units;
[0012] Based on the monitoring data and icing conditions, a training set is determined and the training set is divided into a first training set and a second training set.
[0013] The LeNet5like network is trained using the first training set to obtain a pre-trained network;
[0014] The transfer learning algorithm is used to fine-tune the network parameters of the pre-trained network using the second training set to obtain an icing fault diagnosis model.
[0015] Based on the monitoring data of the wind turbine unit to be tested, the icing fault diagnosis model is used to perform fault diagnosis and obtain the icing fault diagnosis result.
[0016] Optionally, after acquiring the monitoring data and icing status of the wind turbine, the method further includes: preprocessing the monitoring data and icing status, wherein the preprocessing includes normalization and data deletion.
[0017] Optionally, the LeNet5like network includes a first convolutional layer, a first batch of normalized layers, a first ReLU layer, a first pooling layer, a first deactivation layer, a second convolutional layer, a second batch of normalized layers, a second ReLU layer, a second pooling layer, a third convolutional layer, a third batch of normalized layers, a third ReLU layer, a third pooling layer, and two linear layers connected in sequence.
[0018] Optionally, the objective function expression for fine-tuning the network parameters is:
[0019]
[0020] Among them, y tTo digitally represent real-world fault conditions, y p C provides a digital representation of network diagnostic fault conditions. k For the feature parameters extracted from layer k, C k-1 Here are the extracted feature parameters for layer k-1, and x is the order of the training set data.
[0021] The present invention also provides a wind turbine icing fault diagnosis system, comprising:
[0022] The acquisition module is used to acquire monitoring data and icing conditions of wind turbine units;
[0023] A partitioning module is used to determine a training set based on the monitoring data and icing conditions, and to divide the training set into a first training set and a second training set.
[0024] The training module is used to train the LeNet5like network using the first training set to obtain a pre-trained network.
[0025] The network parameter fine-tuning module is used to fine-tune the network parameters of the pre-trained network using the second training set with a transfer learning algorithm to obtain an icing fault diagnosis model.
[0026] The fault diagnosis module is used to perform fault diagnosis based on the monitoring data of the wind turbine to be tested using the icing fault diagnosis model, and obtain the icing fault diagnosis result.
[0027] Optionally, the LeNet5like network includes a first convolutional layer, a first batch of normalized layers, a first ReLU layer, a first pooling layer, a first deactivation layer, a second convolutional layer, a second batch of normalized layers, a second ReLU layer, a second pooling layer, a third convolutional layer, a third batch of normalized layers, a third ReLU layer, a third pooling layer, and two linear layers connected in sequence.
[0028] The present invention also provides an electronic device, comprising:
[0029] One or more processors;
[0030] A storage device on which one or more programs are stored;
[0031] When the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method as described.
[0032] The present invention also provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described above.
[0033] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0034] This invention acquires monitoring data and icing conditions of wind turbine generators; determines a training set based on the monitoring data and icing conditions, and divides the training set into a first training set and a second training set; trains a LeNet5like network using the first training set to obtain a pre-trained network; uses a transfer learning algorithm to fine-tune the network parameters of the pre-trained network using the second training set to obtain an icing fault diagnosis model; and performs fault diagnosis using the icing fault diagnosis model based on the monitoring data of the wind turbine generator to be tested, obtaining the icing fault diagnosis result. This invention utilizes a LeNet5like network and a transfer learning algorithm to improve the network model's ability to extract fault features, enhance the model's generalization ability, and improve the model's accuracy while ensuring speed, effectively achieving rapid and accurate diagnosis of wind turbine icing. Attached Figure Description
[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 This is a diagram of the LeNet5like network structure.
[0037] Figure 2 A schematic diagram illustrating a method for diagnosing icing faults in wind turbine units;
[0038] Figure 3 A graph showing the accuracy of training tests for W1 wind turbine data;
[0039] Figure 4 Loss plot for training and testing data of W1 wind turbine;
[0040] Figure 5 A graph showing the accuracy of training tests for W2 wind turbine data;
[0041] Figure 6 Loss plot for training and testing data of W2 wind turbine;
[0042] Figure 7 A graph showing the accuracy of data transfer learning training tests for the W2 wind turbine.
[0043] Figure 8 Loss graph for W2 wind turbine data transfer learning training test;
[0044] Figure 9 The flowchart of the wind turbine icing fault diagnosis method provided by the present invention. Detailed Implementation
[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0046] The purpose of this invention is to provide a method, system, device, and storage medium for diagnosing icing faults in wind turbines, thereby improving the network model's ability to extract fault features, enhancing the model's generalization ability, and improving the model's accuracy while ensuring speed, thus effectively achieving rapid and accurate diagnosis of wind turbine icing.
[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0048] like Figure 2 and Figure 9 As shown, the present invention provides a method for diagnosing icing faults in wind turbine units, comprising:
[0049] Step 101: Obtain monitoring data and icing status of the wind turbine.
[0050] After acquiring the monitoring data and icing status of the wind turbine, the process also includes: preprocessing the monitoring data and icing status, which includes normalization and data deletion.
[0051] Step 102: Determine the training set based on the monitoring data and icing conditions, and divide the training set into a first training set and a second training set.
[0052] The monitoring data in this invention is SCADA data. By integrating wind turbine SCADA data with icing conditions and performing data preprocessing, a training dataset is obtained. The sample data in the training dataset is obtained through preprocessing of data collected by the acquisition system installed on each wind turbine, specifically including:
[0053] The dataset used in this invention was collected from the SCADA system data of two three-bladed wind turbines (W1 and W2) at a wind farm, from 20:20 on November 1, 2020 to 21:38 on January 1, 2021. The SCADA system sampled every 7 seconds, collecting a total of 584,380 data samples and 80 data variables. 26 continuous numerical characteristic variables, including wind turbine state parameters, environmental parameters, and operating condition parameters, were preserved through wind turbine mechanism models and human experience. Specific descriptions of the detected variable data are shown in Table 1. To facilitate data input and computation, all data has been normalized and integrated with the wind turbine icing status annotated by engineers to form a training dataset. The first training set in this invention is the processed W1 wind turbine dataset, and the second training set is the processed W2 wind turbine dataset.
[0054] Table 1 Explanation of SCADA Fan Detection Variables
[0055] Serial Number Detection variables Serial Number Detection variables 1 wind speed 14 Pitch motor 2 temperature 2 generator speed 15 Pitch motor 3 temperature 3 Grid-side active power 16 X-direction acceleration 4 Windward angle 17 Y-direction acceleration 5 25s average wind angle 18 Ambient temperature 6 Yaw position 19 Cabin temperature 7 Blade 1 pitch angle 20 Leaf 1ng5 temperature 8 Blade 2 pitch angle 21 Leaf 2ng5 temperature 9 Blade 3 pitch angle 22 Leaf 3ng5 temperature 10 Blade speed 1 23 Blade 1ng5 DC current 11 Blade 2 speed 24 Blade 2ng5 DC current 12 Blade 3 speed 25 Blade 3ng5 DC current 13 Pitch motor 1 temperature 26 Data group identifier
[0056] Missing and erroneous data were removed, and normal data was downsampled using faulty data as the standard. The resulting W1 and W2 datasets were then divided into training and test sets proportionally. Data partitioning types and status information are shown in Table 2.
[0057] Table 2. Statistics on Dataset Types and Status
[0058]
[0059] Step 103: Train the LeNet5like network using the first training set to obtain a pre-trained network. The LeNet5like network includes, in sequence, a first convolutional layer, a first batch of normalized layers, a first ReLU layer, a first pooling layer, a first deactivation layer, a second convolutional layer, a second batch of normalized layers, a second ReLU layer, a second pooling layer, a third convolutional layer, a third batch of normalized layers, a third ReLU layer, a third pooling layer, and two linear layers. Both linear layers are fully connected.
[0060] The LeNet5like network model was trained using data from wind turbine W1 in the collected sample data to obtain the icing fault diagnosis model for wind turbine W1. During model training, the input consisted of 26 parameters listed in Table 1, and the output was the result of whether icing occurred. The processed W1 wind turbine dataset was then input into the LeNet5like network for feature extraction. The network learned the icing characteristics of wind turbine W1, and the icing fault diagnosis model for wind turbine W1 was established. The specific structure of the LeNet5like network is as follows: Figure 1 As shown.
[0061] Table 3 Parameters of each layer of the LeNet5like network
[0062]
[0063] Table 3 shows the input and output parameters of each layer during the training of the LeNet5like network. One column lists parameter information such as stride, kernel size, and padding pixels. The LeNet5like network proposed in this invention is an adjustment and improvement upon the traditional LeNet-5 network. The types and number of network layers have been adjusted, with the addition of three batch normalization layers, one pooling layer, and one deactivation layer. This reduces the parameters required for computation in each neural network layer. The ReLU function, with its stronger extraction capabilities, has been replaced, reducing the data size required for computation and enhancing robustness. The Adam (Adaptive Moment Estimation) algorithm is used to continuously adapt the learning rate and optimize the parameters, bringing the model to its optimal solution.
[0064] Step 104: Use the transfer learning algorithm to fine-tune the network parameters of the pre-trained network using the second training set to obtain the icing fault diagnosis model.
[0065] The formula for calculating the parameters of the LeNet5like transfer learning network is as follows:
[0066]
[0067] min∑||C k -C k-1 || 2 (2)
[0068]
[0069] Equation (1) is the first optimization objective, y t To digitally represent real-world fault conditions, y p The network diagnostic fault situation is digitally represented to minimize the target fault diagnosis error; Equation (2) is the second optimization objective, C k For the feature parameters extracted from layer k, C k-1 For the extracted feature parameters of layer k-1, the divergence between the representations of W1 and W2 should be minimized, while keeping most of the parameters in the convolutional pooling layer unchanged, in order to minimize the parameter changes of the extracted features in the pre-trained network structure; the two optimization objectives are combined to obtain the final optimization objective function (3) used to fine-tune the network structure parameters.
[0070] The objective function expression for fine-tuning network parameters is:
[0071]
[0072] Among them, y t To digitally represent real-world fault conditions, yp C provides a digital representation of network diagnostic fault conditions. k For the feature parameters extracted from layer k, C k-1 Here are the extracted feature parameters for layer k-1, and x is the order of the training set data.
[0073] Using data from wind turbine W2 in the dataset, the network parameters of the icing fault diagnosis model corresponding to wind turbine W1 were fine-tuned to obtain the icing fault diagnosis model corresponding to W2. Based on the pre-trained feature extraction LeNet5like network, the network parameters of the established W1 diagnosis network were fine-tuned using the target W2 dataset to complete the transfer diagnosis model of wind turbine W2.
[0074] The transfer learning training process is conducted in the following manner:
[0075] The obtained W1 wind turbine icing fault diagnosis model was used as a pre-trained network. The final optimization objective function for fine-tuning the network structure parameters was constructed using Equation (3) to perform network parameter fine-tuning transfer learning on the W2 wind turbine data, and a fault diagnosis model was established. After training, an icing fault diagnosis model suitable for the W2 wind turbine was obtained.
[0076] Step 105: Based on the monitoring data of the wind turbine unit to be tested, perform fault diagnosis using the icing fault diagnosis model to obtain the icing fault diagnosis result.
[0077] The SCADA data of the W2 wind turbine is input into the established migration diagnostic network to obtain the icing status and complete the diagnosis of icing faults.
[0078] The diagnostic criteria formula for icing results is as follows:
[0079]
[0080]
[0081]
[0082] In the formula, TP (True Positive) represents a wind turbine whose icing condition is determined to be normal and predicted as normal by the classification model; FN (False Negative) represents a wind turbine whose icing condition is determined to be normal but predicted as icing by the classification model; FP (False Positive) represents a wind turbine whose icing condition is determined to be icing but predicted as normal by the classification model; and TN (True Negative) represents a wind turbine whose icing condition is determined to be icing and predicted as icing by the classification model. Precision is the accuracy rate, Recall is the recall rate, and F1 is the overall metric.
[0083] The values for the three evaluation criteria above are all in the range of [0,1]. The closer the value is to 1, the better the model performs in this aspect.
[0084] This invention addresses the need to establish icing fault diagnosis models for different wind turbine units. It uses a pre-trained icing fault early warning model for the same type of wind turbine as a pre-trained network, and then employs a transfer learning module to train fault diagnosis models for different wind turbines. The training iterations are set to 30 times. After each training iteration, the training accuracy and loss, as well as the test accuracy and loss, are printed, and a graph showing the change in training, testing, loss, and accuracy is finally plotted. The LeNet5like network is used to train on the W1 dataset, and the results are as follows: Figure 3 and Figure 4 As shown.
[0085] like Figure 3 and Figure 4 As shown, after 15 iterations, the fault classification accuracy converged rapidly and stabilized at over 98.4%, with the highest accuracy rate being 99.26%. The loss function value decreased rapidly in the first 10 iterations, then decreased slowly and approached 0, reaching convergence after 20 training iterations. The test set loss was basically maintained below 0.05, with the minimum error loss being 0.0226.
[0086] Table 4 Comparison of Experimental Results for W1 Wind Turbine
[0087] method Recall rate (%) Precision rate (%) F1(％) KNN 94.74 99.41 97.02 Decision Tree 94.04 91.91 92.64 Random Forest 97.54 92.60 95.00 LeNet-5 97.40 97.24 97.27 LeNet5like 99.20 99.61 99.41
[0088] As shown in Table 4 of the experimental results, based on formulas (4), (5), and (6), compared with traditional machine learning algorithms and the traditional LeNet-5 network, the present invention achieves the highest recall, precision, and F1 score, with average improvements of approximately 3.27%, 4.32%, and 3.93%, respectively. Significant improvements are achieved when the accuracy exceeds 90%. This demonstrates that the LeNet5like network better extracts the correlation features between variables, significantly improving accuracy.
[0089] like Figure 5 and Figure 6 As shown, the model training results are close to the model testing results, and there is no underfitting or overfitting, proving the effectiveness of the proposed fault diagnosis model. The training time was 1 minute and 51 seconds. If the model trained on W1 is used to test the W2 wind turbine data, the highest accuracy and minimum loss are 93.38% and 0.5987, respectively. The data shows that the diagnostic results are severely inaccurate and not suitable for diagnosis. Using a LeNet5like network to train on the W2 data, the training time was 2 minutes and 56 seconds, the minimum loss was 0.0092, and the highest accuracy was 99.77%.
[0090] Based on the W1 wind turbine icing diagnosis model as a pre-trained network, and after training using a transfer learning module, the accuracy of the W2 wind turbine data transfer learning training test is as follows: Figure 7 and Figure 8 As shown. The training time was 2 minutes and 28 seconds, the minimum loss was 0.0689, and the highest accuracy was 98.90%. Compared to training on the W2 data without using the transfer learning module, the training time of the pre-trained network using transfer learning was shortened by 28 seconds, an improvement of approximately 15.91%, while the accuracy and loss were not significantly different from the original result.
[0091] Table 5 Comparison of Experimental Results for W2 Fan Data
[0092] method Recall rate (%) Precision rate (%) F1(％) KNN 95.09 98.85 96.93 Decision Tree 93.88 95.93 94.89 Random Forest 98.25 91.52 94.76 LeNet-5 94.97 97.70 96.32 LeNet5like 99.22 99.86 99.54
[0093] As shown in Table 5 of the experimental results, when the parameter metrics of the LeNet5like network with the transfer learning module are compared with those of traditional machine learning methods and traditional LeNet-5 networks, the LeNet5like network shows an average improvement of 3.67%, 3.86%, and 3.82% in the three parameter metrics compared with other methods.
[0094] It can be seen that the LeNet5like-based transfer learning method for icing fault diagnosis of wind turbines provided by this invention has the advantages of good robustness, high accuracy and fast modeling speed when using SCADA data to build a wind turbine icing fault diagnosis model.
[0095] This invention provides a LeNet5like-based transfer learning method for wind turbine icing fault diagnosis. It uses easily processed and readily available SCADA datasets as the training input for the neural network. Addressing the characteristics of low-frequency sampling and small sample size in the dataset, the traditional shallow LeNet-5 neural network is improved to accelerate network computation, enhance the accuracy of target fault diagnosis, and increase the model's generalizability and robustness. Through fine-tuning of network parameters using transfer learning, a multi-wind turbine diagnostic model can be established more quickly based on existing wind turbine models, improving the accuracy of feature transfer for target wind turbines. This method is suitable for multi-wind turbine icing fault diagnosis.
[0096] This invention improves upon the traditional LeNet-5 network structure by increasing the number and types of network layers, changing the activation function, and fine-tuning the parameters of each layer to create a fault diagnosis model framework. A transfer learning module is also incorporated to form a LeNet5like network. The network model is pre-trained using fault features learned from the W1 wind turbine, and a fault diagnosis model for the W2 wind turbine is quickly established using network parameter fine-tuning. Analysis of actual wind power SCADA datasets shows that, compared to traditional machine learning methods, the proposed method achieves an accuracy of 98.90%, improves training accuracy by approximately 2% compared to the traditional LeNet-5 network, and reduces training time by 28 seconds compared to an untrained network, representing an improvement of approximately 15.91%. This method enhances the network model's ability to extract fault features, strengthens its generalization ability, and improves accuracy while maintaining speed, effectively enabling rapid and accurate diagnosis of wind turbine icing.
[0097] The present invention also provides a wind turbine icing fault diagnosis system, comprising:
[0098] The acquisition module is used to acquire monitoring data and icing conditions of wind turbine units.
[0099] The partitioning module is used to determine the training set based on the monitoring data and icing conditions, and to partition the training set into a first training set and a second training set.
[0100] The training module is used to train the LeNet5like network using the first training set to obtain a pre-trained network.
[0101] The network parameter fine-tuning module is used to fine-tune the network parameters of the pre-trained network using the second training set through a transfer learning algorithm to obtain an icing fault diagnosis model.
[0102] The fault diagnosis module is used to perform fault diagnosis based on the monitoring data of the wind turbine to be tested using the icing fault diagnosis model, and obtain the icing fault diagnosis result.
[0103] As an optional implementation, the LeNet5like network includes a first convolutional layer, a first batch of normalized layers, a first ReLU layer, a first pooling layer, a first deactivation layer, a second convolutional layer, a second batch of normalized layers, a second ReLU layer, a second pooling layer, a third convolutional layer, a third batch of normalized layers, a third ReLU layer, a third pooling layer, and two linear layers connected in sequence.
[0104] The present invention also provides an electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon; wherein, when the one or more programs are executed by the one or more processors, the one or more processors perform the method as described.
[0105] The present invention also provides a computer storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method as described above.
[0106] This invention is based on the traditional LeNet-5 network, with adjustments and improvements made to the types and number of network layers, and changes to the network function and optimizer function; a network parameter fine-tuning transfer learning module is added to enable the rapid establishment of a diagnostic model for icing faults in wind turbine units.
[0107] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.
[0108] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A wind turbine icing fault diagnosis method, characterized in that, include: Acquire monitoring data and icing status of wind turbine units; After acquiring the monitoring data and icing status of the wind turbine, the process also includes: preprocessing the monitoring data and icing status, the preprocessing including normalization and data deletion; Based on the monitoring data and icing conditions, a training set is determined and the training set is divided into a first training set and a second training set. The LeNet5like network is trained using the first training set to obtain a pre-trained network; the LeNet5like network includes a first convolutional layer, a first batch of normalization layers, a first ReLU layer, a first pooling layer, a first deactivation layer, a second convolutional layer, a second batch of normalization layers, a second ReLU layer, a second pooling layer, a third convolutional layer, a third batch of normalization layers, a third ReLU layer, a third pooling layer, and two linear layers connected in sequence. The transfer learning algorithm is used to fine-tune the network parameters of the pre-trained network using the second training set to obtain the icing fault diagnosis model; the optimization objective function expression for the network parameter fine-tuning is: in, To provide a digital representation of the actual fault situation. Digital representation of network diagnostic faults. For the feature parameters extracted from layer k, Here are the feature parameters extracted from layer k-1, and x is the order of the training set data. Based on the monitoring data of the wind turbine unit to be tested, the icing fault diagnosis model is used to perform fault diagnosis and obtain the icing fault diagnosis result.
2. A wind turbine icing fault diagnosis system, characterized in that, The wind turbine icing fault diagnosis system uses the wind turbine icing fault diagnosis method of claim 1, and the wind turbine icing fault diagnosis system includes: The acquisition module is used to acquire monitoring data and icing conditions of wind turbine units; A partitioning module is used to determine a training set based on the monitoring data and icing conditions, and to divide the training set into a first training set and a second training set. The training module is used to train the LeNet5like network using the first training set to obtain a pre-trained network. The network parameter fine-tuning module is used to fine-tune the network parameters of the pre-trained network using the second training set with a transfer learning algorithm to obtain an icing fault diagnosis model. The fault diagnosis module is used to perform fault diagnosis based on the monitoring data of the wind turbine to be tested using the icing fault diagnosis model, and obtain the icing fault diagnosis result.
3. An electronic device, comprising: include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the wind turbine icing fault diagnosis method as described in claim 1.
4. A computer storage medium, characterized in that, It stores a computer program, wherein the computer program, when executed by a processor, implements the wind turbine icing fault diagnosis method as described in claim 1.