Wind turbine blade icing detection method based on adaptive weight loss function

Abstract

The application discloses a wind turbine blade icing detection method based on an adaptive weight loss function, and belongs to the field of wind turbine state detection. First, blade icing data is acquired from a data acquisition and monitoring control system, invalid data is deleted, and the blade icing data is divided into a training set, a verification set and a test set. Then, a blade icing state detection model based on an attention network is constructed, and an adaptive weight loss function is designed to train the detection model. In the model training stage, the data set is divided into multiple batches, the icing data in different batches is classified according to the number of the icing data, and the icing data is adaptively assigned weights based on different categories of the batches, so that the recognition rate of the icing state detection model for the icing data is improved, and the overall classification performance of the detection model is enhanced. Compared with a traditional loss function, the adaptive weight loss function designed in the application does not need to set hyperparameters, has better adaptability, and also improves the accuracy of the icing detection.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine condition monitoring, specifically relating to a method for detecting icing on wind turbine blades based on an adaptive weighted loss function. Background Technology

[0002] Wind turbine blades are key components for energy conversion in wind turbine generators. In actual wind farms, wind turbines are often located in relatively cold climates such as high altitudes and high latitudes to capture more wind energy. Therefore, wind turbine blades are susceptible to icing due to low temperatures and snowfall. Blade icing alters the airfoil, reducing wind energy capture capacity and increasing the energy required for blade rotation, resulting in power loss. Furthermore, the ice layer formed by icing changes the blade's modal parameters; if de-icing is not carried out promptly, it can induce blade breakage. Therefore, timely and accurate detection of icing conditions on wind turbine blades is crucial for the healthy operation of wind turbine generators.

[0003] Due to the importance of blade icing detection, researchers have proposed various condition detection schemes, including acoustic emission monitoring, ultrasonic detection, and Supervisory Control and Data Acquisition (SCADA) system monitoring. Among these, SCADA systems utilize sensors deployed throughout the wind turbine assembly process to collect multi-scale operational status information of the wind turbine in real time. This allows for systematic and comprehensive condition monitoring of key components in each subsystem, including the blades. Furthermore, since SCADA systems are installed alongside the wind turbine assembly, they offer the advantage of low operating costs. However, using SCADA systems to detect blade icing in wind turbines still presents some challenges. One such challenge is that for an operational wind turbine, the normal operating period of its components, including the blades, is significantly longer than the icing period. Therefore, the SCADA system collects a relatively small proportion of icing data and a relatively large proportion of normal data in the wind turbine blade operation data, exhibiting typical class imbalance. This increases the difficulty for relevant condition detection models to accurately classify the icing data from the dataset. To reduce the difficulty of detecting blade icing status using SCADA systems, appropriate strategies need to be adopted to address the imbalanced datasets in SCADA systems and enhance the model's classification performance. Currently, strategies for addressing imbalanced datasets mainly include data-level resampling and model-level reweighting. Data-level resampling modifies the distribution of the dataset itself to reduce imbalance. Specific methods include undersampling and oversampling. Undersampling involves taking only a portion of the SCADA dataset while discarding another, which can easily lead to the loss of key information. Oversampling, on the other hand, resamples data from a minority of classes to reduce imbalance, but this can cause information redundancy and increase the difficulty of model training. Model-level reweighting refers to designing a loss function during the model training phase to adjust the weights assigned to different classes of data in the dataset, thereby enhancing the ability to identify a particular class. Traditional loss functions include standard cross-entropy, balanced cross-entropy, Focal_loss, and cost-sensitive loss functions.

[0004] Because loss functions can enhance the identification of minority class samples by the state detection model without oversampling or undersampling the original operating data of the SCADA system, they are increasingly being used in the field of wind turbine blade icing state detection. However, traditional loss functions such as balanced cross-entropy and Focal_loss require manual setting of hyperparameters to assign weights to each class of data. Furthermore, for different icing state detection models, the hyperparameters need to be continuously changed to achieve better classification results, which increases the difficulty of model optimization. In addition, standard cross-entropy and cost-sensitive loss functions often do not provide ideal optimization results for blade icing state detection. To reduce the difficulty of model optimization and improve the classification performance of the state detection model on imbalanced datasets of the SCADA system, a loss function should be designed that does not require manual setting of hyperparameters but adaptively assigns weights to each class of data and achieves better classification results. Summary of the Invention

[0005] To address the shortcomings of traditional loss functions and the requirements for new loss functions, this invention aims to design a loss function that adaptively assigns weights to corresponding data based on the different proportions of icing data and normal data in the SCADA system's blade icing dataset. This improves the classification performance of the blade icing detection model on datasets with significant imbalances, thereby evaluating the icing status of the blades.

[0006] To achieve the above objectives, this invention is implemented through the following scheme: a wind turbine blade icing detection method based on an adaptive weight loss function. This scheme includes the following:

[0007] Step 1: The SCADA system acquires wind turbine blade icing data and labels it according to different data categories: normal data is labeled with label 0, icing data with label 1, and invalid data with label -1. The invalid data refers to data recorded by the SCADA system during turbine shutdown maintenance; this data is directly removed from the blade icing dataset. The icing dataset after removing invalid data is called the icing detection dataset. Further, the icing detection dataset is divided into training, validation, and test sets.

[0008] Step 2: Establish a blade icing state detection model based on a spatiotemporal attention network. The spatiotemporal attention network refers to a model based on a parallel structure of a fully convolutional network (FCNN) and a long short-term memory network (LSTM).

[0009] Step 3: Construct an adaptive weight loss function to train the blade icing state detection model obtained in Step 2. The constructed loss function is used to adaptively assign weights to icing data and normal data in the training set, enhancing the model's classification ability for icing data and thus improving its overall classification performance on the test set.

[0010] Step 4: Use the test set of the icing detection dataset to evaluate the classification performance of the model trained in Step 3.

[0011] Step two includes the following sub-steps.

[0012] (2.1) Construct a time attention module, which includes a global average pooling layer, a gated recurrent neural network (GRU) network, and a fully connected layer to calculate the weights of the data corresponding to each sampling time in the icing detection dataset.

[0013] (2.2) Construct a spatial attention module, which consists of a global average pooling layer, multiple convolutional and max pooling layers, and a fully connected layer to calculate the weights of each spatial variable in the icing detection dataset.

[0014] (2.3) Construct a parallel network of LSTM and FCNN, input the operation results of the spatiotemporal attention mechanism into the parallel network, extract the temporal correlation and spatial correlation of the icing detection dataset respectively, and then concatenate the operation results of the parallel network and use the Softmax classifier for classification output.

[0015] Furthermore, the above steps are divided into the following sub-steps.

[0016] (3.1) Divide the training set of the icing detection dataset into multiple batches of data, and the resulting batches are called sample groups. Import the training set into the model according to the sample groups for training.

[0017] (3.2) Analyze the distribution of icing data in each sample group. For a sample group containing icing data, calculate the average number of icing data points in a single sample group, specifically:

[0018]

[0019] In the formula, X is the total number of data of this type in all sample groups containing icing data, and N is the number of sample groups containing icing data.

[0020] (3.3) The sample groups are classified according to the number of icing data they contain, into three categories: no icing data sample group, sparse sample group, and dense sample group. The no icing data sample group contains only normal data, the sparse sample group contains fewer icing data than the average value A calculated in step (3.2), and the dense sample group contains more icing data than the average value A.

[0021] (3.4) Design a weight allocation formula for the non-icing data sample group, specifically as follows:

[0022]

[0023] In the formula, m represents the number of normal data in a single sample group.

[0024] Substituting β into the formula for the loss function, the specific result is:

[0025] loss = -βlog(p) t (3)

[0026] p t The expression is:

[0027]

[0028] In the formula, p represents the probability that the data is predicted to be in an icy state.

[0029] (3.5) Design a weight allocation formula for the sparse sample group, specifically as follows:

[0030]

[0031] In the formula, m and n represent the number of normal data and icing data in a single sample group, respectively, and the formula for the loss function is the same as (3).

[0032] (3.6) Design a weight allocation formula for dense sample groups.

[0033] Introducing adaptive adjustment coefficients λ and γ, the specific expressions are as follows:

[0034]

[0035]

[0036] m and n represent the number of normal data and icing data in a single sample group, respectively.

[0037] The weighting formula for dense sample groups is:

[0038]

[0039] In the formula, λ and γ are adaptive adjustment coefficients, and the formula for the loss function is the same as that in (3).

[0040] Due to the design scheme described above, this invention has the following advantages over traditional loss functions:

[0041] This loss function adaptively assigns weights to different categories of data based on the sample groups divided in the training set during the training of the icing state detection model, eliminating the need for manual hyperparameter setting and reducing the optimization difficulty of the model. Because it adaptively assigns higher weights to icing category data, this loss function outperforms traditional loss functions in identifying icing data. The adaptive weighting of different data classes means that this loss function does not require hyperparameter readjustment for different state detection models, demonstrating good adaptability. Attached Figure Description

[0042] Figure 1 A schematic diagram of the parallel network structure of FCNN and LSTM to introduce the attention mechanism;

[0043] Figure 2 A flowchart for constructing an adaptive weight loss function;

[0044] Figure 3 The distribution of icing data for each sample group in the training set of the leaf icing dataset during model training;

[0045] Figure 4 This diagram illustrates the comparison between the optimization effect of the adaptive weight loss function involved in this invention on the blade icing state detection model on the test set and other traditional loss functions. Detailed Implementation

[0046] To enable those skilled in the art to better understand the invention and its advantages, the implementation of the invention is further described below with reference to the accompanying drawings and specific blade icing datasets. The specific datasets described herein are merely illustrative and not intended to limit the scope of the invention.

[0047] The specific implementation steps of this invention are as follows:

[0048] Step 1: The SCADA system acquires wind turbine blade icing data and labels it according to different data categories: normal data is labeled with label 0, icing data with label 1, and invalid data with label -1. The invalid data refers to data recorded by the SCADA system during turbine shutdown maintenance; this data is directly removed from the blade icing dataset. The icing dataset after removing invalid data is called the icing detection dataset. Further, the icing detection dataset is divided into training, validation, and test sets.

[0049] Step 2: Establish a blade icing state detection model based on a spatiotemporal attention network. The spatiotemporal attention network refers to a model based on a parallel structure of a fully convolutional network (FCNN) and a long short-term memory network (LSTM). This step includes the following sub-steps:

[0050] (2.1) Construct a time attention module, which includes a global average pooling layer, a gated recurrent neural network (GRU) network, and a fully connected layer to calculate the weights of the data corresponding to each sampling time in the icing detection dataset.

[0051] (2.2) Construct a spatial attention module, which consists of a global average pooling layer, multiple convolutional and max pooling layers, and a fully connected layer to calculate the weights of each spatial variable in the icing detection dataset.

[0052] (2.3) Construct a parallel network of LSTM and FCNN, input the operation results of the spatiotemporal attention mechanism into the parallel network, extract the temporal correlation and spatial correlation of the icing detection dataset respectively, and then concatenate the operation results of the parallel network and use the Softmax classifier to classify and output the results.

[0053] Step 3: Construct an adaptive weight loss function to train the blade icing state detection model obtained in Step 2. The constructed loss function is used to adaptively assign weights to icing data and normal data in the training set, enhancing the model's classification ability for icing data and thus improving its overall classification performance on the test set.

[0054] Step three is specifically divided into the following sub-steps:

[0055] (3.1) Divide the training set of the icing detection dataset into multiple batches of data, and the resulting batches are called sample groups. Import the training set into the model according to the sample groups for training.

[0056] (3.2) Analyze the distribution of icing data in each sample group. For a sample group containing icing data, calculate the average number of icing data points in a single sample group, specifically:

[0057]

[0058] In the formula, X is the total number of data of this type in all sample groups containing icing data, and N is the number of all sample groups containing icing data.

[0059] (3.3) The sample groups are classified according to the number of icing data they contain, into three categories: the non-icing data group, the sparse sample group, and the dense sample group. The non-icing data group contains only normal data, the sparse sample group contains fewer icing data than the average value A calculated in step (3.2), and the dense sample group contains more icing data than the average value A.

[0060] (3.4) is the weighting formula designed for the non-icing sample data, specifically:

[0061]

[0062] In the formula, m represents the number of normal data in a single sample group.

[0063] Substituting β into the formula for the loss function, the specific result is:

[0064] loss = -βlog(p) t (3)

[0065] Where p t The expression is:

[0066]

[0067] In the formula, p represents the probability that the data is predicted to be in an icy state.

[0068] (3.5) Design a weight allocation formula for the sparse sample group, specifically as follows:

[0069]

[0070] In the formula, m and n represent the number of normal data and icing data in a single sample group, respectively. The formula for the loss function is the same as (3).

[0071] (3.6) Design a weight allocation formula for dense sample groups.

[0072] Introducing adaptive adjustment coefficients λ and γ, the specific expressions are as follows:

[0073]

[0074]

[0075] m and n represent the number of normal data and icing data in a single sample group, respectively.

[0076] The formula for weighting dense sample groups is:

[0077]

[0078] In the formula, λ and γ are adaptive adjustment coefficients, and the formula for the loss function is the same as that in (3).

[0079] Step 4: Use the test set of the icing detection dataset to evaluate the classification performance of the model trained in Step 3.

[0080] Example 1

[0081] The advantages of the loss function involved in this invention are verified using a specific wind turbine blade icing dataset. This dataset consists of blade icing data from a wind turbine in a wind farm. This dataset has been widely used in the field of blade icing state detection to verify the classification performance of various state detection models.

[0082] The aforementioned leaf icing dataset is preprocessed according to step one above, including labeling and deleting invalid data. This preprocessed dataset is called the icing detection dataset. Further, the icing detection dataset is divided into a training set, a validation set, and a test set.

[0083] The structure of the parallel network of FCNN and LSTM with attention mechanism is as follows: Figure 1 As shown, it includes a temporal attention mechanism module, a spatial attention mechanism module, and a main network module.

[0084] According to such Figure 2 The flowchart shown constructs an adaptive weight loss function. This function is used to enhance the model's ability to classify icing data during model training, thereby improving the overall classification performance.

[0085] During model training, the training set of the icing detection dataset is divided into sample groups, each containing 128 data points. The distribution of icing data in each sample group is as follows: Figure 3 As shown in the figure, the average value refers to the average number of data points of this type in the sample group containing icing data.

[0086] Figure 3 In the sample group, except for groups 1, 7, 12, 16, 18, 19, and 22 which contain icing data, the remaining sample groups only contain normal data. According to the specific implementation method, sample groups without icing data are classified as the icing-free data sample groups discussed in the specific implementation method; groups 1, 16, and 22, which contain significantly fewer icing data than the average, are classified as sparse sample groups; and groups 7, 12, and 19, which contain more icing data than the average, are classified as dense sample groups. Furthermore, for sample group 18, although the number of icing data is slightly less than the average, the difference is only 1 / 6, which is much smaller than the difference between the number of icing data and the average in sample groups 1, 6, and 22. Therefore, this scheme also classifies sample group 18 as a dense sample group.

[0087] To verify the advantages of the loss function involved in this invention, the optimization effect of this loss function on the icing state detection model was compared with several other traditional loss functions on the test set. These traditional loss functions include Focal_loss, balanced cross-entropy loss function, cost-sensitive loss function, and standard cross-entropy loss function.

[0088] For the loss function Focal_loss, the hyperparameters α and β need to be manually set. The hyperparameter β is set to 2, 3, or 4, and the range of α for each β value is (0.05, 0.95), with a step size of 0.1. The best results achieved by Focal_loss under the above different hyperparameter settings are compared with the loss function of this invention. After verification, the optimization effect is the best among the hyperparameters of this loss function when β=2, α=0.65, β=3, α=0.05, β=4, α=0.15, and β=4, α=0.55.

[0089] For the balanced cross-entropy loss function, the hyperparameter α needs to be manually set. The value of α is in the range of (0.05 to 0.99), with a step size of 0.05. The optimal optimization effect of the loss function is selected under each hyperparameter setting. After verification, the optimization effect is best when the hyperparameter α is set to 0.25 and 0.6.

[0090] Cost-sensitive and balanced cross-entropy loss functions do not contain hyperparameters, so their optimization effects are directly compared.

[0091] Figure 4 The figure compares the optimal optimization performance of various traditional loss functions on the test set with the adaptive weight loss function involved in this invention. The figure includes three evaluation metrics: icing data recognition rate reflects the model's accuracy in recognizing icing data on blades in the dataset; the comprehensive recognition rate is the average of the model's recognition accuracies for normal data and icing data; and F1_Score is a key metric used in wind turbine condition detection to measure model classification performance. As shown in the figure, the wind turbine blade icing detection method based on the adaptive weight loss function involved in this invention outperforms the optimal metrics achieved by other loss functions under different hyperparameters in terms of icing data recognition rate, comprehensive recognition rate, and F1_score. The icing data recognition rate is 0.76, the comprehensive recognition rate is 0.86, and the F1_score is 0.66.

[0092] In summary, the adaptive weight loss function proposed in this invention can significantly enhance the model's ability to identify icing category data in wind turbine blade icing datasets with typical imbalances compared to other traditional loss functions, providing a new approach for processing imbalanced data.

[0093] Those skilled in the art will understand that the above descriptions are merely preferred examples in conjunction with specific datasets and are not intended to limit the invention. Although the present invention has been described in detail with reference to the aforementioned specific datasets, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for detecting icing on wind turbine blades based on an adaptive weighted loss function, characterized in that... The method includes the following steps: Step 1: Obtain wind turbine blade icing data from the SCADA system and label the data according to different categories, with normal data labeled as label 0, icing data labeled as label 1, and invalid data labeled as label-1; divide the icing detection dataset into training set, validation set, and test set; Step 2: Establish a wind turbine blade icing state detection model based on a spatiotemporal attention network. The spatiotemporal attention network refers to a model based on the parallel structure of a fully convolutional network FCNN and a long short-term memory network LSTM. Step 3: Construct an adaptive weight loss function to train the blade icing state detection model obtained in Step 2; utilize the constructed loss function to adaptively assign weights to icing data and normal data in the training set, thereby enhancing the model's classification ability for icing data and improving the overall classification performance of the model on the test set; Step 3 includes the following sub-steps: 3.1 Divide the training set of the icing detection dataset into multiple batches of data, which are called sample groups. Import the training set into the model according to the sample groups for training. 3.2 Analyze the distribution of icing data in each sample group. For a sample group containing icing data, calculate the average number of icing data points in a single sample group, specifically: ； In the formula, X is the total number of data of this type in all sample groups containing icing data, and N is the number of sample groups containing icing data. 3.3 After analyzing the distribution of icing data in different sample groups, the sample groups are classified according to the number of icing data they contain, into three categories: no icing data sample group, sparse sample group, and dense sample group. Among them, the no icing data sample group contains only normal data, the sparse sample group contains fewer icing data than the average value A calculated in step 3.2, and the dense sample group contains more icing data than the average value A. 3.4 Design a weighting formula for the non-icing data sample group, specifically as follows: ； In the formula, m represents the number of normal data in a single sample group; Substituting β into the formula for the loss function, the specific result is: ； In the formula The expression is: ； In the formula, p represents the probability that the data is predicted to be in an icy state; 3.5 Designs a weight allocation formula for sparse sample groups, specifically as follows: ； In the formula, m and n represent the number of normal data and icing data in a single sample group, respectively; Substituting β into the formula for the loss function, the specific result is: ； in The expression is: ； In the formula, p represents the probability that the data is predicted to be in an icy state; 3.6 Design a weighting formula for dense sample groups; introduce two adaptive adjustment coefficients λ and γ, where λ is the ratio of the number of icing data points to the number of normal data points in the sample group, and γ is the ratio of the number of normal data points to the number of icing data points in the sample group; the specific expression is: ； ； m and n represent the number of normal data and icing data in a single sample group, respectively; The weighting formula for dense sample groups is: ； Substituting β into the formula for the loss function, the specific result is: ； In the formula The expression is: ； In the formula, p represents the probability that the data is predicted to be in an icy state; Step 4: Use the test set to evaluate the classification performance of the model trained in Step 3.

2. The wind turbine blade icing detection method based on adaptive weight loss function according to claim 1, characterized in that step two specifically includes the following sub-steps: 2.1 Construct a time attention module, which includes a global average pooling layer, a gated recurrent neural network (GRU), and a fully connected layer to calculate the weights of the data corresponding to each sampling time in the icing detection dataset. 2.2 Construct a spatial attention module, which includes a global average pooling layer, multiple convolutional and max pooling layers, and a fully connected layer. The purpose is to calculate the weights of each spatial variable in the icing detection dataset. 2.3 Construct a parallel network of LSTM and FCNN, input the computation results of the spatiotemporal attention mechanism into the network, extract the temporal and spatial correlations of the icing detection dataset, and then concatenate the computation results of the parallel network and use the Softmax classifier for classification output.

3. The method for detecting icing on wind turbine blades based on an adaptive weighted loss function according to claim 1, characterized in that: In step one, the invalid data refers to the data recorded by the SCADA system during the unit's shutdown maintenance. This invalid data is directly removed from the blade icing dataset. The icing dataset after removing the invalid data is called the icing detection dataset.

Images