Comparative learning method and system for wind turbine blade icing detection

Abstract

The present invention belongs to the technical field of wind power generation, and relates to a comparative learning method and system for detecting ice on wind turbine blades. The method comprises: collecting data and performing data preprocessing; adaptive spectrum-time enhancement processing; optimizing time masks and frequency masks based on a combination of sparsity loss, pattern preservation loss, and diversity loss; obtaining enhanced vibration signals as positive sample pairs; extracting negative sample pairs; identifying abnormal samples through anomaly scores; fine-tuning an unsupervised comparative learning model using labeled samples, and outputting icing anomaly detection results through a lightweight classification head. The present invention proposes a method combining adaptive spectrum-time enhancement and negative sample pool comparative learning to generate reliable positive and negative comparative sample pairs, which can accurately detect icing anomalies under conditions of scarce labels and data imbalance; can effectively alleviate the problem of false negative sampling, solve the problem of class imbalance, and achieve high-quality feature learning.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power generation, and in particular relates to a comparative learning method and system for detecting icing on wind turbine blades. Background Art

[0002] In cold environments, blade icing seriously affects the aerodynamic performance, structural integrity and operational safety of wind turbines. Even slight ice accumulation can cause severe aerodynamic imbalance, trigger blade structural vibration, reduce power output, and pose safety risks due to ice falling. Therefore, blade icing detection is crucial for reliable wind power generation.

[0003] Existing blade icing detection methods include physical methods and data-driven methods. Physical methods, such as electrothermal heating and anti-icing coatings, aim to directly prevent or remove icing, but often involve high operating costs and limited adaptability. Data-driven methods attempt to simulate the physical process of icing using thermodynamic principles, but rely heavily on expert knowledge and assumptions, limiting scalability across different turbine types and weather conditions. Driven by advances in sensor technology and the deployment of Supervisory Control and Data Acquisition (SCADA) systems, traditional machine learning models such as Principal Component Analysis (PCA) and eXtreme Gradient Boosting (XGBoost) rely on handcrafted features to detect anomalies. However, these features are often insufficient to capture the nonlinear and multivariate nature of real-world icing events.

[0004] Existing models typically rely on large labeled datasets for supervised training. However, due to the rarity of icing events, the difficulty of accurate labeling, and the significant cost associated with expert labeling, high-quality labeled icing data is extremely scarce. Furthermore, the severe imbalance between normal and icing samples complicates training, often leading to poor performance of traditional supervised learning methods in detection tasks. Therefore, how to effectively leverage limited labeled data and abundant unlabeled data to accurately detect rare icing events is an urgent problem to be solved. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a comparative learning method and system for wind turbine blade icing detection.

[0006] In a first aspect, the present invention provides a comparative learning method for detecting icing on wind turbine blades, comprising:

[0007] Collect operational and environmental data of wind turbines under icing and unicing conditions and perform data preprocessing to obtain input data; data preprocessing includes cleaning, normalizing, and partitioning the data;

[0008] Adaptive spectrum-time enhancement processing is performed on the input data, time enhancement is performed on the input data using a time mask to obtain a time-enhanced signal, and spectrum enhancement is performed on the input data using a frequency mask to obtain a spectrum-enhanced signal;

[0009] Based on the combination of sparsity loss, mode preservation loss and diversity loss, the loss function of adaptive spectrum-time enhancement is obtained and the time mask and frequency mask are optimized according to the loss function;

[0010] According to the optimized time mask and frequency mask, an enhanced vibration signal including a time-enhanced signal and a spectrum-enhanced signal is obtained as a positive sample pair;

[0011] Extract negative sample pairs through the negative sample pool contrast learning method, build a dynamic negative sample pool, calculate the anomaly score based on the similarity of samples within the batch, and filter and update the negative samples in the negative sample pool;

[0012] An unsupervised contrastive learning model is used to determine the anomaly score of a sample by maximizing the similarity of positive sample pairs in the feature space and minimizing the similarity of negative sample pairs. The anomaly score is then used to identify abnormal samples.

[0013] The unsupervised contrastive learning model is fine-tuned using labeled samples, and the icing anomaly detection results are output through a lightweight classification head.

[0014] In a second aspect, the present invention provides a comparative learning system for wind turbine blade icing detection, comprising a comparative learning system for wind turbine blade icing detection, comprising a collection and preprocessing unit, an enhancement processing unit, a combination optimization unit, a positive sample pair output unit, an extraction and negative sample output unit, a model building and recognition unit, and a fine-tuning and output unit;

[0015] A collection and preprocessing unit is used to collect the operating data and environmental data of wind turbines under icing and non-icing conditions and perform data preprocessing to obtain input data; data preprocessing includes cleaning, normalizing and dividing the data;

[0016] an enhancement processing unit, configured to perform adaptive spectrum-time enhancement processing on the input data, perform time enhancement on the input data using a time mask to obtain a time-enhanced signal, and perform spectrum enhancement on the input data using a frequency mask to obtain a spectrum-enhanced signal;

[0017] a combination optimization unit, configured to obtain a loss function for adaptive spectrum-time enhancement based on a combination of sparsity loss, mode preservation loss, and diversity loss, and optimize a time mask and a frequency mask according to the loss function;

[0018] A positive sample pair output unit is used to obtain an enhanced vibration signal including a time-enhanced signal and a spectrum-enhanced signal according to the optimized time mask and frequency mask as a positive sample pair;

[0019] The extraction and negative sample output unit is used to extract negative sample pairs through the negative sample pool comparative learning method, build a dynamic negative sample pool, calculate the anomaly score based on the similarity of samples within the batch, and filter and update the negative samples in the negative sample pool;

[0020] The model building and identification unit is used to use an unsupervised contrastive learning model to determine the anomaly score of the sample by maximizing the similarity of positive sample pairs in the feature space and minimizing the similarity of negative sample pairs, and identify abnormal samples based on the anomaly score;

[0021] The fine-tuning and output unit is used to fine-tune the unsupervised contrastive learning model using labeled samples and output the icing anomaly detection results through a lightweight classification head.

[0022] On the basis of the above technical solution, the present invention can also be improved as follows.

[0023] Furthermore, when temporally enhancing the input data using a temporal mask to obtain a temporally enhanced signal, a binary mask is constructed and applied to the original time series to generate two complementary views; each view is enhanced using a parameterized transformation function; the parameterized transformation function applies two transformation heads, including a transformation head for retaining information segments and a transformation head for masking irrelevant segments, each of which consists of a linear layer and a sigmoid activation function to perform different parameterized mappings.

[0024] Furthermore, when frequency masking is used to perform spectral enhancement on the input data to obtain a spectral enhanced signal, the original input is converted to the frequency domain using fast Fourier transform, and an adaptive masking mechanism is adopted to selectively highlight the information area in the amplitude spectrum. The obtained binary mask is used to decompose it into two complementary views, and a parameterized transformation function is applied to enhance each component. Two transformation heads are used to adaptively modify the retained information amplitude and mask amplitude. Each transformation head consists of a linear layer and a Sigmoid activation function to perform different parameterized mappings.

[0025] Furthermore, let the sparsity loss be , the mode preservation loss is , the diversity loss is , the loss function is , the hyperparameters contributing to the mode preservation loss are , the hyperparameter of the diversity loss contribution is , then the loss function of adaptive spectrum-time enhancement is expressed as:

[0026] .

[0027] Furthermore, negative sample pairs are extracted through the negative sample pool contrast learning method, including:

[0028] The encoder is used to extract negative sample pairs; for time domain input, the feature extractor in the encoder uses stacked hole convolution layers to extract time domain features; for frequency domain input, the feature extractor in the encoder uses a multi-band convolution module to extract frequency domain features; parallel one-dimensional convolution filters with different kernel sizes are applied to the time domain features and frequency domain features for filtering; and they are averaged through splicing and pooling to generate the final representation vector.

[0029] Furthermore, a dynamic negative sample pool is constructed, anomaly scores are calculated based on the similarity of samples within a batch, and negative samples in the negative sample pool are screened and updated, including: using a machine learning model to track samples with the smallest similarity across batches as potential anomaly samples to construct a negative sample pool, updating the negative sample pool during the training process of the machine learning model and combining the negative sample pool samples for comparative learning.

[0030] Furthermore, anomaly scores are calculated based on the similarity of samples within a batch, including:

[0031] The encoder maps each sample in a batch to an embedding to obtain an embedding set; the anomaly score is calculated based on the cumulative similarity between each embedding and all other embeddings in the batch; let the anomaly score be , embedded as , Represents similarity calculation, Indicates the embeddings, and the batch size is , then the expression of the anomaly score is:

[0032] .

[0033] Furthermore, for abnormal samples whose anomaly scores are greater than the set threshold, the coefficient of variation values ​​of all feature dimensions are calculated, and the coefficient of variation values ​​of all feature dimensions are sorted from large to small. The abnormal samples of the first several feature dimensions are retained. For each feature dimension in each batch, the calculation process of the coefficient of variation is:

[0034] set up For the Features, Indicates the standard deviation. Indicates the mean value, The coefficient of variation of a feature is , is a positive constant, then the expression of the coefficient of variation is:

[0035] .

[0036] Furthermore, after filtering and updating the negative samples in the negative sample pool, the negative samples in the negative sample pool are combined with the original negative sample set to form a new negative sample set, and the unsupervised contrastive learning model is trained using the new negative sample set. The loss function of the unsupervised contrastive learning model is calculated, including: assuming is the positive sample in the positive sample pair, is the negative sample extracted from the original negative sample set, is the temperature scaling parameter, For a new negative sample set, the loss function of the unsupervised contrastive learning model is ,but:

[0037] .

[0038] The beneficial effects of the present invention are as follows: the present invention proposes a method combining adaptive spectrum-time enhancement and negative sample pool contrastive learning to generate reliable positive and negative contrast sample pairs. Experimental results show that this method can accurately detect icing anomalies under conditions of scarce labels and data imbalance, and is superior to existing methods that rely on large amounts of labeled data; in addition, experiments show that negative sample pool contrastive learning can effectively alleviate the false negative sampling problem caused by category imbalance in the unsupervised contrastive learning framework. The present invention not only solves the category imbalance problem, but also achieves high-quality feature learning, which is very suitable for actual wind turbine anomaly detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of a comparative learning method for wind turbine blade icing detection provided in Example 1 of the present invention;

[0040] Figure 2 This is a schematic diagram of a comparative learning system for wind turbine blade icing detection provided in Example 2 of the present invention. DETAILED DESCRIPTION

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0042] Example 1

[0043] As an example, Figure 1 As shown, to solve the above technical problems, this embodiment provides a comparative learning method for wind turbine blade icing detection, including:

[0044] Collect operational and environmental data of wind turbines under icing and unicing conditions and perform data preprocessing to obtain input data; data preprocessing includes cleaning, normalizing, and partitioning the data;

[0045] Adaptive spectrum-time enhancement processing is performed on the input data, time enhancement is performed on the input data using a time mask to obtain a time-enhanced signal, and spectrum enhancement is performed on the input data using a frequency mask to obtain a spectrum-enhanced signal;

[0046] Based on the combination of sparsity loss, mode preservation loss and diversity loss, the loss function of adaptive spectrum-time enhancement is obtained and the time mask and frequency mask are optimized according to the loss function;

[0047] According to the optimized time mask and frequency mask, an enhanced vibration signal including a time-enhanced signal and a spectrum-enhanced signal is obtained as a positive sample pair;

[0048] Extract negative sample pairs through the negative sample pool contrast learning method, build a dynamic negative sample pool, calculate the anomaly score based on the similarity of samples within the batch, and filter and update the negative samples in the negative sample pool;

[0049] An unsupervised contrastive learning model is used to determine the anomaly score of a sample by maximizing the similarity of positive sample pairs in the feature space and minimizing the similarity of negative sample pairs. The anomaly score is then used to identify abnormal samples.

[0050] The unsupervised contrastive learning model is fine-tuned using labeled samples, and the icing anomaly detection results are output through a lightweight classification head.

[0051] This paper proposes a wind turbine icing anomaly detection method based on contrastive learning to address the problems of class imbalance and scarce labeled data. It includes three main stages: data preprocessing, contrastive learning and fine-tuning.

[0052] The data preprocessing stage includes collecting operating data of wind turbines under different conditions, including operating data and environmental data of wind turbines in normal scenarios and icing scenarios, operating data of wind turbines such as rotor speed and power output, and environmental conditions such as temperature, wind speed and wind direction; then the data is cleaned, normalized and divided to ensure the consistency and reliability of the data, reduce noise and remove irrelevant features, and improve the stability of the subsequent learning process.

[0053] The contrastive learning phase includes Adaptive Spectral-Temporal Augmentation (ASTA) and Negative Pool Contrastive Learning (NPCL). ASTA generates enhanced views of the original time series data in both the temporal and spectral domains, preserving key signal structures while introducing controllable perturbations. NPCL alleviates class imbalance by maintaining a dynamic repository of potentially anomalous samples, ensuring that the contrastive learning process can effectively distinguish between normal and anomalous cases.

[0054] During the fine-tuning phase, a small portion of labeled data is used to optimize the learned encoder representation, and a lightweight classification head is applied to enhance the model's ability to detect icing anomalies, improving detection accuracy while maintaining computational efficiency.

[0055] After data preprocessing, we construct positive pairs for contrastive learning through adaptive spectral-temporal augmentation. ASTA generates two enhanced views: a temporal view and a spectral view, which dynamically preserves key information of the original sequence while introducing the necessary diversity required for contrastive learning.

[0056] The temporal enhancement process starts with the input data. In order to obtain the feature embedding that guides the enhancement process, an embedding function is applied to map the time domain sequence to a latent space while retaining the basic structural characteristics of the original sequence. Let the time domain sequence be , the embedding function is , is time, then:

[0057] .

[0058] Optionally, when temporally enhancing the input data using a temporal mask to obtain a temporally enhanced signal, a binary mask is constructed and applied to the original time series to generate two complementary views; each view is enhanced using a parameterized transformation function; the parameterized transformation function applies two transformation heads, including a transformation head for retaining information segments and a transformation head for masking irrelevant segments, each transformation head consisting of a linear layer and a Sigmoid activation function, performing different parameterized mappings.

[0059] The adaptive spectral-temporal enhancement method uses an adaptive masking mechanism, called the masking factor, to selectively highlight the informative regions in the temporal dimension. The goal is to construct a binary mask that distinguishes the informative regions from the less relevant parts. Let the binary mask be , and is a learnable parameter, represents the Sigmoid activation function, Represents a binarizer that applies a threshold operation to discretize the output into , represent the retained and masked regions respectively. Once the binary mask is obtained, it is applied to the original time domain sequence To generate two complementary views, the binary mask is calculated as follows:

[0060] .

[0061] set up Indicates that the most informative regions in the sequence are retained. These two components ensure that ASTA can distinguish important and less relevant parts of the original sequence for targeted processing.

[0062] ;

[0063] .

[0064] ASTA then enhances each component using a parameterized transformation function, applying two transformation heads: one for the information fragment to be retained , another unrelated fragment used for masking .set up and is a learnable weight matrix, and is the bias term, Represents the Sigmoid activation function. Each transformation head consists of a linear layer and a Sigmoid activation function, performing different parameterized mappings:

[0065] ;

[0066] .

[0067] Assume that the enhanced time view is reconstructed as ,but:

[0068] .

[0069] By learning different transformation paths, ASTA introduces controllable changes to irrelevant regions while preserving meaningful patterns, ensuring that the enhanced views incorporate sufficient diversity for contrastive learning while maintaining the essential temporal structure.

[0070] The spectrum enhancement process first converts the input data into the frequency domain using fast Fourier transform. is the frequency domain expression of the input data, Denotes the fast Fourier transform operation, which focuses on the amplitude spectrum, while the adaptive spectrum time domain enhancement preserves the phase information to maintain time coherence and prevent distortion. is the embedding function, For a fully connected operation, in order to obtain feature embeddings that guide the enhancement process, an embedding function is applied:

[0071] .

[0072] Similar to the time enhancement process, adaptive spectrum time domain enhancement uses an adaptive mask mechanism to selectively highlight the information area in the amplitude spectrum. The resulting binary mask is used to Decomposed into two complementary views, is a binary mask, is a binarization operation, is the information amplitude to be retained, represents the mask amplitude, is a learnable bias parameter, and is a learnable parameter, then:

[0073] ;

[0074] ;

[0075] .

[0076] Optionally, let the sparsity loss be , the mode preservation loss is , the diversity loss is , the loss function is , the hyperparameters contributing to the mode preservation loss are , the hyperparameter of the diversity loss contribution is , then the loss function of adaptive spectrum-time enhancement is expressed as:

[0077] .

[0078] The learning process of adaptive spectral temporal enhancement is guided by three key objectives, each targeting a key aspect of enhancement quality, namely extracting the most informative components, preserving inherent patterns, and promoting diversity.

[0079] Let the binary mask be , is the input dataset, express The 0 norm of , the sparsity loss is expressed as:

[0080] .

[0081] set up express The enhanced view set, is the feature mapping function used in contrastive learning, is the original input sample, To retain The key structural pattern of , the pattern preservation loss is expressed as:

[0082] .

[0083] set up express The original irrelevant fragment, express The enhanced set of irrelevant fragments, To ensure that the enhanced views are sufficiently deviated from the uninformative regions of the original sequence and enhance the diversity of contrastive learning, the diversity loss is expressed as:

[0084] .

[0085] Optionally, negative sample pairs are extracted through contrastive learning using a negative sample pool, including:

[0086] The encoder is used to extract negative sample pairs; for time domain input, the feature extractor in the encoder uses stacked hole convolution layers to extract time domain features; for frequency domain input, the feature extractor in the encoder uses a multi-band convolution module to extract frequency domain features; parallel one-dimensional convolution filters with different kernel sizes are applied to the time domain features and frequency domain features for filtering; and they are averaged through splicing and pooling to generate the final representation vector.

[0087] Extractor in contrastive learning with negative sample pooling in time domain: For time domain input, the feature extractor adopts a series of stacked dilated convolution layers. Specifically, 6 layers of one-dimensional dilated convolution are used with a kernel size of 3 and a dilation ratio ranging from 1 to 64. This design enables the encoder to capture the multi-scale time domain patterns inherent in complex wind turbine operating conditions. Extractor in contrastive learning with negative sample pooling in frequency domain: For frequency domain input, the feature extractor adopts a multi-band convolution module, applying one-dimensional convolution with a kernel size of 1 to each of the four frequency subbands. This enables the encoder to extract spectral patterns in specific frequency bands and distinguish abnormal features (such as icing features) from normal features.

[0088] After the feature extractor, parallel one-dimensional convolutional filters with different kernel sizes are applied, and the outputs are averaged through concatenation and pooling to generate the final representation vector. This multi-kernel strategy captures diverse time or frequency domain patterns while generating compact feature embeddings.

[0089] Optionally, a dynamic negative sample pool is constructed, anomaly scores are calculated based on the similarity of samples within a batch, and negative samples in the negative sample pool are screened and updated, including: using a machine learning model to track samples with the smallest similarity across batches as potential anomaly samples to construct a negative sample pool, updating the negative sample pool during the training process of the machine learning model and combining the negative sample pool samples for comparative learning.

[0090] and constitutes a positive sample pair, and and Construct a negative sample pair for contrastive learning. In order to alleviate the class imbalance problem, a dynamic negative sample pool mechanism is introduced. Negative sample pool contrastive learning first constructs a negative sample pool by tracking the most dissimilar samples across batches and treating them as potential abnormal samples. By continuously updating the negative sample pool during training It is combined with negative sample pool samples for contrastive learning. The contrast goal explicitly forces the representation of normal samples to be separated from the representation of these potential abnormal samples, thereby achieving robust anomaly detection without labeled data.

[0091] Optionally, calculate anomaly scores based on the similarity of samples within a batch, including:

[0092] The encoder maps each sample in a batch to an embedding to obtain an embedding set; the anomaly score is calculated based on the cumulative similarity between each embedding and all other embeddings in the batch; let the anomaly score be , embedded as , Represents similarity calculation, Indicates the embeddings, and the batch size is , then the expression of the anomaly score is:

[0093] .

[0094] The anomaly score quantifies The degree of similarity to the rest of the samples in the batch. Lower values ​​indicate that the sample is less similar to the majority distribution. right Sort all samples in ascending order and select the top samples with the lowest scores. These selected embeddings form the set , represents the most dissimilar sample in the batch These samples are identified as potential abnormal samples and added to the negative sample pool middle.

[0095] if If the similarity with most samples in the batch is lower than the set threshold, then it belongs to the minority group and is a strong candidate for inclusion in the negative sample pool. Initially, there are fewer samples in the negative sample pool. As training progresses, more abnormal candidate samples will be identified and included. Once the negative sample pool reaches its capacity , the newly identified potential anomaly samples must dynamically replace the existing samples. To achieve this, the anomaly scores of the new candidate samples are Compare with the anomaly score of the current negative sample pool member. Formally, in Identify the most likely normal samples , defined as:

[0096] .

[0097] examine If All samples in have anomaly scores higher than , no replacement is performed, otherwise, Iteratively select the most dissimilar sample (the one with the lowest anomaly score) in Time Replacement By dynamically updating the most dissimilar embedding in each training batch , the negative sample pool will retain as many potential abnormal samples as possible. This ensures that negative sample pairs are constructed more effectively in the case of class imbalance and enhances the robustness of contrastive learning.

[0098] Optionally, for abnormal samples with anomaly scores greater than a set threshold, calculate the coefficient of variation values ​​of all feature dimensions, sort the coefficient of variation values ​​of all feature dimensions from large to small, and retain the abnormal samples of the first several feature dimensions. For each feature dimension in each batch, the coefficient of variation calculation process is:

[0099] set up For the Features, Indicates the standard deviation. Indicates the mean value, The coefficient of variation of a feature is , is a positive constant, then the expression of the coefficient of variation is:

[0100] .

[0101] Optionally, after filtering and updating the negative samples in the negative sample pool, the negative samples in the negative sample pool are combined with the original negative sample set to form a new negative sample set, and the unsupervised contrastive learning model is trained using the new negative sample set to calculate the loss function of the unsupervised contrastive learning model, including: assuming is the positive sample in the positive sample pair, is the negative sample extracted from the original negative sample set, is the temperature scaling parameter, For a new negative sample set, the loss function of the unsupervised contrastive learning model is ,but:

[0102] .

[0103] This loss function measures the ability to distinguish between positive and negative sample pairs in unsupervised contrastive learning. Specifically, after constructing a negative sample pool, the negative sample pool is combined with the original negative sample set to form a new negative sample set, which serves as a reference for contrastive learning. For each positive sample pair, the loss function for unsupervised contrastive learning is constructed by calculating its similarity ratio within the combined set and adopting a normalized exponential function with a temperature scaling parameter. This method can effectively enhance the model's ability to aggregate positive sample pairs and distinguish negative samples, thereby improving the representation of icing state characteristics.

[0104] After the unsupervised contrastive learning phase, the temporal and spectral encoders trained by T-NPCL and S-NPCL have learned reliable feature representations. Then, a lightweight fine-tuning step is performed using only a limited number of labeled samples for anomaly detection. First, the temporal encoder is trained with and spectrum encoder The learned representations are concatenated to form a joint embedding, and the joint embedding is set to ,but:

[0105] .

[0106] The classification head is trained using binary cross entropy loss via a lightweight classifier consisting of two fully connected layers with ReLU activation function, followed by a final Sigmoid layer that outputs anomaly probabilities.

[0107] This paper proposes adaptive spectral-temporal enhancement (ASTA) and negative pooling contrastive learning (NPCL) to generate reliable positive and negative contrast sample pairs. Experimental results demonstrate that this method can accurately detect icing anomalies under conditions of scarce labels and data imbalance, outperforming existing methods that rely on large amounts of labeled data. Furthermore, experiments show that negative pooling contrastive learning can effectively alleviate the false negative sampling problem caused by class imbalance in unsupervised contrastive learning frameworks. These results confirm that this paper not only solves the class imbalance problem but also achieves high-quality feature learning, making it well-suited for the practical task of wind turbine anomaly detection.

[0108] Example 2

[0109] Based on the same principle as the method shown in Example 1 of the present invention, as shown in the attached Figure 2 As shown, an embodiment of the present invention further provides a comparative learning system for wind turbine blade icing detection, comprising a collection and preprocessing unit, an enhancement processing unit, a combination optimization unit, a positive sample pair output unit, an extraction and negative sample output unit, a model building and recognition unit, and a fine-tuning and output unit;

[0110] A collection and preprocessing unit is used to collect the operating data and environmental data of wind turbines under icing and non-icing conditions and perform data preprocessing to obtain input data; data preprocessing includes cleaning, normalizing and dividing the data;

[0111] an enhancement processing unit, configured to perform adaptive spectrum-time enhancement processing on the input data, perform time enhancement on the input data using a time mask to obtain a time-enhanced signal, and perform spectrum enhancement on the input data using a frequency mask to obtain a spectrum-enhanced signal;

[0112] a combination optimization unit, configured to obtain a loss function for adaptive spectrum-time enhancement based on a combination of sparsity loss, mode preservation loss, and diversity loss, and optimize a time mask and a frequency mask according to the loss function;

[0113] A positive sample pair output unit is used to obtain an enhanced vibration signal including a time-enhanced signal and a spectrum-enhanced signal according to the optimized time mask and frequency mask as a positive sample pair;

[0114] The extraction and negative sample output unit is used to extract negative sample pairs through the negative sample pool comparative learning method, build a dynamic negative sample pool, calculate the anomaly score based on the similarity of samples within the batch, and filter and update the negative samples in the negative sample pool;

[0115] The model building and identification unit is used to use an unsupervised contrastive learning model to determine the anomaly score of the sample by maximizing the similarity of positive sample pairs in the feature space and minimizing the similarity of negative sample pairs, and identify abnormal samples based on the anomaly score;

[0116] The fine-tuning and output unit is used to fine-tune the unsupervised contrastive learning model using labeled samples and output the icing anomaly detection results through a lightweight classification head.

[0117] Optionally, when temporally enhancing the input data using a temporal mask to obtain a temporally enhanced signal, a binary mask is constructed and applied to the original time series to generate two complementary views; each view is enhanced using a parameterized transformation function; the parameterized transformation function applies two transformation heads, including a transformation head for retaining information segments and a transformation head for masking irrelevant segments, each transformation head consisting of a linear layer and a Sigmoid activation function, performing different parameterized mappings.

[0118] Optionally, when frequency masking is used to perform spectral enhancement on the input data to obtain a spectral enhanced signal, the original input is converted to the frequency domain using a fast Fourier transform, and an adaptive masking mechanism is used to selectively highlight the information area in the amplitude spectrum. The obtained binary mask is used to decompose the image into two complementary views, and a parameterized transformation function is applied to enhance each component. Two transformation heads are used to adaptively modify the retained information amplitude and mask amplitude. Each transformation head consists of a linear layer and a Sigmoid activation function to perform different parameterized mappings.

[0119] Optionally, let the sparsity loss be , the mode preservation loss is , the diversity loss is , the loss function is , the hyperparameters contributing to the mode preservation loss are , the hyperparameter of the diversity loss contribution is , then the loss function of adaptive spectrum-time enhancement is expressed as:

[0120] .

[0121] Optionally, negative sample pairs are extracted through contrastive learning using a negative sample pool, including:

[0122] The encoder is used to extract negative sample pairs; for time domain input, the feature extractor in the encoder uses stacked hole convolution layers to extract time domain features; for frequency domain input, the feature extractor in the encoder uses a multi-band convolution module to extract frequency domain features; parallel one-dimensional convolution filters with different kernel sizes are applied to the time domain features and frequency domain features for filtering; and they are averaged through splicing and pooling to generate the final representation vector.

[0123] Optionally, a dynamic negative sample pool is constructed, anomaly scores are calculated based on the similarity of samples within a batch, and negative samples in the negative sample pool are screened and updated, including: using a machine learning model to track samples with the smallest similarity across batches as potential anomaly samples to construct a negative sample pool, updating the negative sample pool during the training process of the machine learning model and combining the negative sample pool samples for comparative learning.

[0124] Optionally, calculate anomaly scores based on the similarity of samples within a batch, including:

[0125] The encoder maps each sample in a batch to an embedding to obtain an embedding set; the anomaly score is calculated based on the cumulative similarity between each embedding and all other embeddings in the batch; let the anomaly score be , embedded as , Represents similarity calculation, Indicates the embeddings, and the batch size is , then the expression of the anomaly score is:

[0126] .

[0127] Optionally, for abnormal samples with anomaly scores greater than a set threshold, calculate the coefficient of variation values ​​of all feature dimensions, sort the coefficient of variation values ​​of all feature dimensions from large to small, and retain the abnormal samples of the first several feature dimensions. For each feature dimension in each batch, the coefficient of variation calculation process is as follows:

[0128] set up For the Features, Indicates the standard deviation. Indicates the mean value, The coefficient of variation of a feature is , is a positive constant, then the expression of the coefficient of variation is:

[0129] .

[0130] Optionally, after filtering and updating the negative samples in the negative sample pool, the negative samples in the negative sample pool are combined with the original negative sample set to form a new negative sample set, and the unsupervised contrastive learning model is trained using the new negative sample set to calculate the loss function of the unsupervised contrastive learning model, including: assuming is the positive sample in the positive sample pair, is the negative sample extracted from the original negative sample set, is the temperature scaling parameter, For a new negative sample set, the loss function of the unsupervised contrastive learning model is ,but:

[0131] .

[0132] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A comparative learning method for wind turbine blade icing detection, characterized in that: include: Collect the operating data and environmental data of wind turbines under icing and non-icing conditions and perform data preprocessing to obtain input data; Data preprocessing includes cleaning, normalizing and partitioning the data; Adaptive spectrum-time enhancement processing is performed on the input data, time enhancement is performed on the input data using a time mask to obtain a time-enhanced signal, and spectrum enhancement is performed on the input data using a frequency mask to obtain a spectrum-enhanced signal; Based on the combination of sparsity loss, mode preservation loss and diversity loss, the loss function of adaptive spectrum-time enhancement is obtained and the time mask and frequency mask are optimized according to the loss function; According to the optimized time mask and frequency mask, an enhanced vibration signal including a time-enhanced signal and a spectrum-enhanced signal is obtained as a positive sample pair; Extract negative sample pairs through the negative sample pool contrast learning method, build a dynamic negative sample pool, calculate the anomaly score based on the similarity of samples within the batch, and filter and update the negative samples in the negative sample pool; An unsupervised contrastive learning model is used to determine the anomaly score of a sample by maximizing the similarity of positive sample pairs in the feature space and minimizing the similarity of negative sample pairs. The anomaly score is then used to identify abnormal samples. The unsupervised contrastive learning model is fine-tuned using labeled samples, and the icing anomaly detection results are output through a lightweight classification head.

2. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that: When temporally enhancing the input data using a temporal mask to obtain a temporally enhanced signal, a binary mask is constructed and applied to the original time series to generate two complementary views. Each view is enhanced using a parameterized transformation function. The parameterized transformation function applies two transformation heads: one for preserving information segments and one for masking irrelevant segments. Each transformation head consists of a linear layer and a sigmoid activation function, performing different parameterized mappings.

3. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that: When using frequency mask to perform spectral enhancement on input data to obtain a spectral enhanced signal, the original input is converted to the frequency domain using fast Fourier transform, and an adaptive masking mechanism is adopted to selectively highlight the information area in the amplitude spectrum. The obtained binary mask is used to decompose it into two complementary views, and a parameterized transformation function is applied to enhance each component. Two transformation heads are used to adaptively modify the retained information amplitude and mask amplitude. Each transformation head consists of a linear layer and a Sigmoid activation function to perform different parameterized mappings.

4. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that The sparsity loss is , the mode preservation loss is , the diversity loss is , the loss function is , the hyperparameters contributing to the mode preservation loss are , the hyperparameter of the diversity loss contribution is , then the loss function of adaptive spectrum-time enhancement is expressed as: 。 5. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that: Negative sample pairs are extracted through the negative sample pool contrast learning method, including: The encoder is used to extract negative sample pairs; for time domain input, the feature extractor in the encoder uses stacked hole convolution layers to extract time domain features; for frequency domain input, the feature extractor in the encoder uses a multi-band convolution module to extract frequency domain features; parallel one-dimensional convolution filters with different kernel sizes are applied to the time domain features and frequency domain features for filtering; and they are averaged through splicing and pooling to generate the final representation vector.

6. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that: Build a dynamic negative sample pool, calculate anomaly scores based on the similarity of samples within the batch, and screen and update negative samples in the negative sample pool, including: using the machine learning model to track the samples with the smallest similarity across batches as potential anomaly samples to build a negative sample pool, updating the negative sample pool during the training process of the machine learning model and combining the negative sample pool samples for comparative learning.

7. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that: Calculate anomaly scores based on the similarity of samples within a batch, including: The encoder maps each sample in a batch to an embedding to obtain an embedding set; the anomaly score is calculated based on the cumulative similarity between each embedding and all other embeddings in the batch; let the anomaly score be , embedded as , Represents similarity calculation, Indicates the embeddings, and the batch size is , then the expression of the anomaly score is: 。 8. The comparative learning method for wind turbine blade icing detection according to claim 7, characterized in that: For abnormal samples whose anomaly scores are greater than the set threshold, the coefficient of variation values ​​of all feature dimensions are calculated, and the coefficient of variation values ​​of all feature dimensions are sorted from large to small. The abnormal samples of the first several feature dimensions are retained. For each feature dimension in each batch, the calculation process of the coefficient of variation is as follows: set up For the Features, Indicates the standard deviation. Indicates the mean value, The coefficient of variation of a feature is , is a positive constant, then the expression of the coefficient of variation is: 。 9. The comparative learning method for wind turbine blade icing detection according to claim 1, characterized in that: After screening and updating the negative samples in the negative sample pool, the negative samples in the negative sample pool are combined with the original negative sample set to form a new negative sample set. The new negative sample set is used to train the unsupervised contrastive learning model and calculate the loss function of the unsupervised contrastive learning model, including: setting is the positive sample in the positive sample pair, is the negative sample extracted from the original negative sample set, is the temperature scaling parameter, For a new negative sample set, the loss function of the unsupervised contrastive learning model is ,but: 。 10. A comparative learning system for wind turbine blade icing detection, characterized in that: It includes collection and preprocessing unit, enhancement processing unit, combination optimization unit, positive sample pair output unit, extraction and negative sample output unit, model building and recognition unit, and fine-tuning and output unit; A collection and preprocessing unit, used for collecting the operation data and environmental data of the wind turbine under icing and non-icing conditions and performing data preprocessing to obtain input data; Data preprocessing includes cleaning, normalizing and partitioning the data; an enhancement processing unit, configured to perform adaptive spectrum-time enhancement processing on the input data, perform time enhancement on the input data using a time mask to obtain a time-enhanced signal, and perform spectrum enhancement on the input data using a frequency mask to obtain a spectrum-enhanced signal; a combination optimization unit, configured to obtain a loss function for adaptive spectrum-time enhancement based on a combination of sparsity loss, mode preservation loss, and diversity loss, and optimize a time mask and a frequency mask according to the loss function; A positive sample pair output unit is used to obtain an enhanced vibration signal including a time-enhanced signal and a spectrum-enhanced signal according to the optimized time mask and frequency mask as a positive sample pair; The extraction and negative sample output unit is used to extract negative sample pairs through the negative sample pool comparative learning method, build a dynamic negative sample pool, calculate the anomaly score based on the similarity of samples within the batch, and filter and update the negative samples in the negative sample pool; The model building and identification unit is used to use an unsupervised contrastive learning model to determine the anomaly score of the sample by maximizing the similarity of positive sample pairs in the feature space and minimizing the similarity of negative sample pairs, and identify abnormal samples based on the anomaly score; The fine-tuning and output unit is used to fine-tune the unsupervised contrastive learning model using labeled samples and output the icing anomaly detection results through a lightweight classification head.

Images