Fan blade icing detection method based on lightweight gated spatiotemporal transformer

Abstract

This invention discloses a wind turbine blade icing detection method based on a lightweight gated spatiotemporal Transformer, belonging to the field of wind turbine blade icing detection technology. The method first collects multivariate data from the wind farm's SCADA system and preprocesses it, then inputs it into a core model for identification. This model mainly consists of two core components: a spatiotemporal gated network and a lightweight Transformer encoder. The former utilizes asymmetric convolution and bidirectional GRU to extract spatial correlations and bidirectional temporal dependencies, and adaptively enhances key spatiotemporal features and suppresses noise through parallel feature and temporal attention and gated fusion mechanisms. The latter perfectly balances global dependency modeling and parameter compression through gated multi-head attention and gated linear units. This invention maintains high detection accuracy while reducing the number of parameters, possessing practical engineering value and real-time deployment capability at the edge.

Description

Technical Field

[0001] This invention relates to the field of wind turbine blade icing detection technology, and more specifically, to a wind turbine blade icing detection method based on a lightweight gated spatiotemporal Transformer. Background Technology

[0002] With the ongoing transformation of the global energy structure, wind power has become one of the most important new energy sources. However, wind turbines operating in cold, high-humidity, and high-altitude regions are highly susceptible to icing on their blade surfaces. Icing disrupts the original aerodynamic shape, leading to decreased lift and increased drag, resulting in a significant reduction in the turbine's output power. Furthermore, uneven icing on the blades disrupts the rotor's mass balance, significantly exacerbating vibrations and increasing structural fatigue loads on the drivetrain, yaw system, and tower. In extreme cases, icing can not only pose a safety hazard due to flying ice but also force emergency shutdowns, causing substantial economic losses and maintenance costs for wind farms. Therefore, achieving early and accurate detection of wind turbine blade icing is of crucial engineering significance for ensuring safe operation and improving power generation efficiency. Traditional icing detection relies heavily on manual inspections or the installation of physical sensors on the blades. These methods are not only costly but also prone to damage and failure in harsh freezing weather conditions. In recent years, with the widespread adoption of the Industrial Internet of Things (IIoT), wind turbine SCADA data acquisition and monitoring control systems have accumulated massive amounts of operational data. Currently, data-driven methods based on deep learning are gradually replacing traditional methods and becoming the mainstream technical approach for wind turbine blade anti-icing and condition monitoring research because they do not require additional hardware deployment and can mine multi-dimensional operational status characteristics.

[0003] Despite this, existing deep learning-based icing detection methods still face two significant challenges in practical industrial applications: First, insufficient mining of complex spatiotemporal features and poor cross-scenario generalization ability. Wind turbines are highly complex multivariate coupled systems, and their SCADA data includes sensor signals in dozens of dimensions such as wind speed, power, temperature, and pitch angle. Most existing models suffer from incomplete characterization of cross-variable global dependencies when processing this high-dimensional data, making it difficult to effectively capture the deep nonlinear coupling relationships between different physical quantities. Simultaneously, icing is a slow, cumulative, dynamic process, requiring models to be capable of processing long-term time-series data. However, existing methods often face problems such as poor parallel computing efficiency, long-range memory decay, or GPU memory overflow when processing long sequences. Second, the contradiction between model performance and computational cost is becoming increasingly prominent. This involves blindly widening the network structure, stacking complex attention modules, or using large pre-trained models to improve recognition performance. While this approach improves accuracy on specific test sets, it significantly increases the size of model parameters and computational cost. Based on this, a high-precision, robust, and lightweight wind turbine blade icing detection model was developed. Summary of the Invention

[0004] In view of the above situation and to overcome the shortcomings of the existing technology, the purpose of this invention is to propose a wind turbine blade icing detection method based on a lightweight gated spatiotemporal Transformer, which can effectively solve the above problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for detecting icing on wind turbine blades based on a lightweight gated spatiotemporal Transformer includes the following steps:

[0007] S1. Collect raw data from multiple sensors of the wind farm's SCADA system, and filter out 27 feature fields highly correlated with icing detection, including: time t, wind speed, generator speed, active power, wind direction, 25-second average wind direction, yaw position, yaw speed, pitch angle (pitch1angle, pitch2angle, pitch3angle), pitch speed (pitch1speed, pitch2speed, pitch3speed), pitch motor temperature (pitch1mototmp, pitch2mototmp, pitch3mototmp), and acceleration (acceleration). x acc yThese are characteristic parameters: ambient temperature, cabin interior temperature, ng5 module temperature (pitch1ng5tmp, pitch2ng5tmp, pitch3ng5tmp), and ng5 module charger DC current (pitch1ng5DC, pitch2ng5DC, pitch3ng5DC).

[0008] S2. Raw Data Reading and Feature Filtering: In sampling mode, quantitative sampling is used to balance the categories; the default process uses the full dataset and prints the label distribution for easy observation of the degree of data imbalance. This results in a clean DataFrame.

[0009] S3. Stratified sampling: To prevent class bias, the data is grouped by the frozen label, and within each group, the training and testing are divided in an 8:2 ratio.

[0010] S4. Normalization: The Min-Max normalization method is used to map all feature parameters to a specific interval [0,1].

[0011] S5. Class Imbalance Handling: An adaptive synthetic oversampling algorithm is used to augment the minority class samples. The calculation is as follows:

[0012] S6. Sequence Construction and Data Loading: This module divides the two-dimensional feature matrix into sequence samples of length seq_len using a sliding window. Each sequence contains features from consecutive time steps, and the labels also retain the corresponding time periods, forming a sequence-level supervision signal.

[0013] S7. Construction based on a lightweight gated spatiotemporal Transformer model: The LiST-Former model includes a sequentially connected spatiotemporal gated network and a lightweight Transformer encoder; wherein, the spatiotemporal gated network consists of a spatiotemporal feature extraction module and a gated multi-attention mechanism.

[0014] S8. Input the time series samples into the LiST-Former model for end-to-end training, and obtain the optimal detection model by minimizing the cross-entropy loss function through the Adam optimizer.

[0015] S9. Use the optimal detection model to detect the real-time SCADA data of the wind turbine under test, and output the binary classification probability result of the icing state through the Softmax function.

[0016] The specific method for sequence construction and data loading in step S6 is as follows: After sequence construction, it is encapsulated into a custom Dataset, and then handed over to the Data Loader to handle batching, shuffling, and discarding residual samples. Finally, three iterators—training, validation, and testing—are output, which are the data sources for the model training and evaluation loop.

[0017] The specific method for constructing the lightweight gated spatiotemporal Transformer model in step S7 is as follows: The lightweight gated spatiotemporal Transformer model consists of a sequentially connected spatiotemporal gated network and a lightweight Transformer encoder; wherein, the spatiotemporal gated network includes a spatiotemporal feature extraction module and a multi-attention fusion module; the spatiotemporal feature extraction module consists of a spatial modeling stage and a temporal modeling stage; the multi-attention fusion module consists of a gated fusion mechanism combining parallel feature dimension attention branches and temporal dimension attention branches. The lightweight Transformer encoder consists of a gated fusion attention mechanism and a gated linear unit structure, and finally outputs the prediction result through a feedforward classifier.

[0018] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0019] This invention aims to provide a wind turbine blade icing detection method that is highly accurate, robust, and easy to deploy at the edge. Through multi-dimensional technological innovation, it solves the contradiction between incomplete spatiotemporal feature characterization and excessive model complexity in existing technologies.

[0020] To address the issues of insufficient global dependency characterization and poor generalization, this invention utilizes asymmetric convolutional branches to model local details and cross-sensor correlations in the spatial dimension; employs residual bidirectional gated recurrent units to capture bidirectional long-range dependencies in the temporal dimension; and designs a dual-dimensional attention and gating fusion mechanism based on features and time to achieve key feature focusing and noise suppression. To address the difficulty of edge deployment due to the large number of parameters, this invention constructs a lightweight Transformer encoder. This encoder uses a "trust gate" signal to adjust the gating multi-head attention to reduce redundancy, and employs gated linear units to finely control the information flow, combined with the decomposition characteristics of asymmetric convolution to achieve significant parameter compression. Attached Figure Description

[0021] Figure 1 This is a schematic flowchart of the method of the present invention;

[0022] Figure 2 This is a schematic diagram of the lightweight gated spatiotemporal Transformer model structure of the present invention;

[0023] Figure 3 This is a schematic diagram of the spatiotemporal feature extraction module of the present invention.

[0024] Figure 4 This is a schematic diagram of the multi-attention mechanism structure of the present invention;

[0025] Figure 5 This is a schematic diagram of the lightweight Transformer encoder structure of the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0027] See Figure 1 This embodiment provides a method for detecting the icing state of wind turbine blades based on a lightweight gated spatiotemporal Transformer, including the following steps:

[0028] S1. Collect raw data from multiple sensors of the wind farm's SCADA system. This data consists of datasets from two wind turbines in a specific wind farm, collected over two months. From this dataset, 27 feature fields highly correlated with icing detection were selected, including: time t, wind speed, generator speed, active power, wind direction, 25-second average wind direction, yaw position, yaw speed, pitch angles (pitch1angle, pitch2angle, pitch3angle), pitch speeds (pitch1speed, pitch2speed, pitch3speed), pitch motor temperature (pitch1mototmp, pitch2mototmp, pitch3mototmp), and acceleration (acceleration). x acc y These are characteristic parameters: ambient temperature, cabin interior temperature, ng5 module temperature (pitch1ng5tmp, pitch2ng5tmp, pitch3ng5tmp), and ng5 module charger DC current (pitch1ng5DC, pitch2ng5DC, pitch3ng5DC).

[0029] S2. Raw Data Reading and Feature Filtering: In optional sampling mode, quantitative sampling is used to balance the categories; the default process uses the full dataset and prints the label distribution for easy observation of the degree of data imbalance. This results in a clean DataFrame.

[0030] S3. Stratified sampling: To prevent class bias, the data is grouped by the frozen label, and within each group, the training and testing are divided in an 8:2 ratio.

[0031] S4. Normalization: First, fit based on all samples, then transform the training, validation, and test sets respectively, and use the Min-Max normalization method to map all feature parameters to a specific interval [0,1].

[0032] S5. Class Imbalance Handling: An adaptive synthetic oversampling algorithm is used to augment the minority class samples. The number of classes after oversampling is printed out to confirm the effect. The calculation is as follows:

[0033] S6, Sequence Construction and Data Loading: This module divides the two-dimensional feature matrix into sequence samples of length seq_len using a sliding window. Each sequence contains features from consecutive time steps, and the labels also retain the corresponding time periods, forming sequence-level supervision signals. After sequence construction, the sequences are encapsulated into a custom dataset, and then the Data Loader is responsible for batching, shuffling, and discarding residual samples. Finally, it outputs three iterators: training, validation, and testing, which are the data sources for the model training and evaluation loops.

[0034] S7. Construction based on a lightweight gated spatiotemporal Transformer model: The structure diagram of the LiST-Former model is shown below. Figure 2 The model comprises a sequentially connected spatiotemporal gating network and a lightweight Transformer encoder; wherein the spatiotemporal gating network consists of a spatiotemporal feature extraction module and a gating multi-attention mechanism, and the flowchart of the spatiotemporal feature extraction module model is shown below. Figure 3 See the flowchart for the gated multi-attention mechanism. Figure 4 See the flowchart for the lightweight Transformer encoder. Figure 5 The implementation steps are as follows:

[0035] (1) Spatial feature extraction stage: Asymmetric convolution is used to separate and extract local and global spatial features. Vertical convolution (kernel size=3) captures column dimension patterns, and horizontal convolution (kernel size=1) captures row dimension patterns. After feature fusion, batch normalization and GELU activation function are applied to transform the input dimension from [batch, seq len, input dim] to [batch, seq len, hidden dim].

[0036] (2) Time-gated processing stage: The sequential dependencies in the time series are captured by a bidirectional GRU network. The bidirectional GRU can consider both past and future time information at the same time, and the output dimension is 2×hidden dim. Then, it is compressed back to hidden dim through a linear projection layer. At the same time, a residual connection mechanism is introduced to enhance the gradient propagation effect and prevent the gradient vanishing problem in deep networks.

[0037] (3) Dual Attention Mechanism Stage: This stage includes two parallel branches: feature attention and temporal attention. Feature attention calculates the self-attention weights between feature dimensions, dynamically weighting the most important feature dimensions. Temporal attention injects temporal information through learnable positional encoding and uses convolutional operations in the temporal dimension to enhance local temporal correlation. The outputs of the two branches are adaptively fused through a gating fusion mechanism, where the gating signal dynamically adjusts the fusion ratio of the two attention methods based on the input data.

[0038] (4) Lightweight Transformer Encoder Stage: The fused features are processed in depth through a multi-layer stacked Transformer encoder. Each layer contains a gated multi-head self-attention mechanism and a feedforward network. The gated mechanism can effectively suppress noisy features, the feedforward network uses the GLU activation function to enhance nonlinear expression, and the Pre-Norm structure is adopted to improve training stability. The multi-layer stacking further deepens the feature representation.

[0039] S8. Input the time-series samples into the LiST-Former model for end-to-end training, and obtain the optimal detection model by minimizing the cross-entropy loss function using the Adam optimizer.

[0040] S9. Using the optimal detection model, detect the real-time SCADA data of the wind turbine under test, and output the binary classification probability result of the icing state through the Softmax function. The evaluation index is:

[0041] (1) Accuracy:

[0042] (2) Accuracy:

[0043] (3) Recall rate:

[0044] (4) F1 score:

[0045] The beneficial effects of this invention are: it provides a high-precision, lightweight method for detecting icing on wind turbine blades, achieving a balance between in-depth spatiotemporal feature characterization and model complexity by constructing a LiST-Former model. In the spatial dimension, this method utilizes asymmetric convolutional branches to model local details and cross-sensor correlations; in the temporal dimension, it employs residual bidirectional gated recurrent units to capture bidirectional long-range dependencies, and combines this with a two-dimensional gated attention mechanism to achieve key feature focusing and noise suppression. To adapt to edge deployment, this invention constructs a lightweight Transformer encoder, which uses a "trust gate" signal and gated linear units to finely control information flow and reduce redundancy. Combined with the decomposition characteristics of asymmetric convolution, it reduces the number of model parameters while maintaining high accuracy.

Claims

1. A method for detecting icing on wind turbine blades based on a lightweight gated spatiotemporal Transformer, characterized in that, Includes the following steps: S1. Acquire SCADA data from the main wind turbine unit: Collect raw data from multiple sensors and filter out 27 feature fields highly relevant to icing detection, including: time, wind speed, generator speed, active power, wind direction, 25-second average wind direction, yaw position, yaw speed, and pitch angle. 1_ angle, pitch 2_ angle, pitch 3_ angle, pitch 1_ speed, pitch 2_ speed, pitch 3_ speed), pitch motor temperature 1_ moto _ tmp, pitch 2_ moto _ tmp, pitch 3_ moto _ tmp), acceleration (acc_x, acc_y), ambient temperature, cabin interior temperature, ng5 module temperature (pitch) 1_ ng5 _ tmp, pitch 2_ ng5 _ tmp, pitch 3_ ng5 _ tmp), ng5 module charger DC current (pitch) 1_ ng5 _ DC, pitch 2_ ng5 _ DC, pitch 3_ ng5 _ DC) These characteristic parameters; S2. Raw Data Reading and Feature Filtering: In sampling mode, quantitative sampling is used to balance the categories; the default process uses the full dataset and prints the label distribution for easy observation of the degree of data imbalance. This results in a clean DataFrame. S3. Stratified sampling: To prevent class bias, the data is grouped by the frozen label, and within each group, the training and testing are divided in an 8:2 ratio. S4. Normalization: The Min-Max normalization method is used to map all feature parameters to a specific interval [0,1]. S5. Class imbalance handling: Adaptive synthetic oversampling algorithm is used to expand minority class samples; S6. Sequence Construction and Data Loading: This module divides the two-dimensional feature matrix into sequence samples of length seq len using a sliding window. Each sequence contains features from consecutive time steps, and the labels also retain the corresponding time periods, forming a sequence-level supervision signal. S7. Construction based on a lightweight gated spatiotemporal Transformer model: The LiST-Former model includes a sequentially connected spatiotemporal gated network and a lightweight Transformer encoder; wherein, the spatiotemporal gated network consists of a spatiotemporal feature extraction module and a gated multi-attention mechanism; S8. Input the time series samples into the LiST-Former model for end-to-end training, and obtain the optimal detection model by minimizing the cross-entropy loss function through the Adam optimizer. S9. The optimal detection model is used to detect the real-time SCADA data of the wind turbine under test, and the binary classification probability result of the icing state is output through the Softmax function. According to step S6 of claim 1, sequence construction and data loading, after sequence construction, it is encapsulated into a custom Dataset, and then handed over to the Data Loader to handle batching, shuffling, and discarding of residual samples, etc. Finally, three iterators—training, validation, and testing—are output, which are the data sources for the model training and evaluation loop. According to claim 1, step S7 is based on a lightweight gated spatiotemporal Transformer model, wherein the spatiotemporal gated network includes a spatiotemporal feature extraction module and a multi-attention fusion module; the spatiotemporal feature extraction module consists of a spatial modeling stage and a temporal modeling stage; the multi-attention fusion module is composed of a gating fusion mechanism that combines parallel feature dimension attention branches and temporal dimension attention branches.

2. The method described in claim 1, step S7, is based on a lightweight gated spatiotemporal Transformer model, characterized in that: In the lightweight gated spatiotemporal Transformer model, the lightweight Transformer encoder consists of a gated fusion attention mechanism and a gated linear unit structure, and finally outputs the prediction result through a feedforward classifier.

3. The wind turbine blade icing detection method based on a lightweight gated spatiotemporal Transformer according to claim 3, characterized in that: The spatiotemporal feature extraction module comprises a spatial modeling stage and a temporal modeling stage. In the spatial modeling stage, asymmetric convolutional blocks are used to extract local short-term dynamic features from each sensor through 1×3 vertical convolutional branches, and a 1×1 horizontal convolutional branch is used to achieve a linear combination of cross-sensor features to model global spatial correlation patterns. In the temporal modeling stage, a bidirectional gated recurrent unit (Bi-GRU) with residual connections is used to capture bidirectional long-range temporal dependencies by processing sequences in both forward and backward directions, and then the residuals are added to the input.

4. The wind turbine blade icing detection method based on a lightweight gated spatiotemporal Transformer according to claim 3, characterized in that: The multi-attention fusion module consists of a gating fusion mechanism that combines feature dimension attention branches and time dimension attention branches. The feature dimension attention branches achieve global alignment and focus of information. The temporal attention branch introduces learnable temporal location encoding and extracts local change patterns through one-dimensional convolution to generate attention coefficients; the gating fusion mechanism generates a gating coefficient matrix G through the Sigmoid activation function, which adaptively adjusts the fusion ratio of the two attention outputs.

5. The wind turbine blade icing detection method based on a lightweight gated spatiotemporal Transformer according to claim 4, characterized in that: The lightweight Transformer encoder incorporates a gated fusion attention mechanism and a gated linear unit (GLU) structure. The gated fusion attention mechanism introduces a gating signal on top of standard multi-head self-attention, enabling dynamic confidence control for each attention head at each time step. It divides the attention head into multiple attention heads, and then concatenates the outputs of these heads. The GLU structure uses a gated linear unit (GLU) instead of the ReLU activation function, dividing the information channel and the gated channel through linear projection, enhancing nonlinear feature interaction while achieving parameter compression.

Images