| الملخص: | ObjectiveWind energy plays a crucial role in achieving sustainable energy goals. As a critical structural component, the wind turbine foundation significantly influences the operational stability, safety, and long-term performance of wind turbine systems. However, structural health monitoring (SHM) of wind turbine foundations often faces challenges with data integrity due to environmental factors, sensor malfunctions, or data transmission issues. These missing data can severely impact the accuracy of structural health assessments, thereby affecting maintenance decisions and operational safety. To tackle the persistent data gaps in the monitoring system of adjustable wind turbine foundations, this study introduces a frequency-space-time domain attention network (FST-ATTNet). This model is designed to enhance the modeling capabilities for complex time-series data, improve the accuracy of missing data reconstruction, and ensure the reliability of health monitoring, ultimately guaranteeing wind turbines' safe and efficient operation. Moreover, it presents potential solutions for similar data reconstruction challenges across various engineering disciplines.MethodsThe FST-ATTNet model introduces an innovative data repair framework by integrating features from the frequency, time, and spatial domains. In the frequency domain, the model employs discrete cosine transform (DCT) to extract periodic and global patterns from time series data, effectively mitigating the high-frequency noise caused by the Gibbs phenomenon in traditional discrete Fourier transform (DFT). A frequency-domain attention mechanism is introduced to enhance this process, adaptively assigning weights to frequency components and prioritizing those most relevant for data reconstruction. In the time domain, Bidirectional Gated Recurrent Units (BiGRU) capture both forward and backward dependencies within the time series, ensuring a comprehensive understanding of local sequence patterns. The Kolmogorov-Arnold Network (KAN) incorporates a B-spline activation function, further enhancing the model's ability to capture complex nonlinear temporal changes. In the spatial domain, the Temporal Convolutional Network (TCN) models long-range dependencies by expanding causal convolutions, thereby capturing local and global spatial relationships. The Squeeze-and-Excitation Network (SENet) further boosts spatial feature extraction by dynamically adjusting the importance of different feature channels. By combining these various attention mechanisms, FST-ATTNet successfully integrates frequency, time, and spatial domain features, achieving superior modeling of complex time series patterns and robust reconstruction of missing data. The model is validated using monitoring data of the measured strain on a pile-based adjustable wind turbine foundation, and its performance is evaluated using the coefficient of determination (<italic>R</italic>²) and mean squared error (<italic>MSE</italic>) metrics.Results and Discussions Validation experiments based on measured data show that FST-ATTNet has the following advantages: (1) Superior performance compared to traditional models: FST-ATTNet outperforms traditional sequence models in data reconstruction tasks, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional GRU (BiGRU), Temporal Convolutional Network (TCN), and Transformer, achieving excellent results with <italic>R</italic>² of 0.98 and <italic>MSE</italic> of 0.66. In contrast, the <italic>R</italic>² of LSTM and GRU models is only 0.86, and the <italic>MSE</italic> is 2.35 and 2.34, respectively, which are limited by their unidirectional or local feature extraction capabilities. The Transformer model performs the worst, with <italic>R</italic>² of 0.72 and <italic>MSE</italic> as high as 2.77, likely due to its inability to effectively capture local patterns and frequency domain features in multivariate time series data. FST-ATTNet, through deep integration of frequency, time, and spatial information, can capture complex patterns, including periodicity, local dynamics, and long-term dependencies, significantly improving reconstruction accuracy. (2) Robustness in high missing-rate scenarios: The model excels in handling severe data-missing scenarios. In the continuous data missing test, even with a missing rate as high as 40%, FST-ATTNet maintains high reconstruction accuracy with <italic>R</italic>² of 0.94 and <italic>MSE</italic> of 0.95. At the 45% missing rate, performance slightly declines, with <italic>R</italic>² of 0.87 and <italic>MSE</italic> of 1.31, but still outperforms other models. In the feature missing test, when 25% of the monitoring features are entirely missing, the model achieves <italic>R</italic>² of 0.91 and <italic>MSE</italic> of 0.75, demonstrating its ability to handle complex multi-feature missing scenarios commonly found in actual monitoring systems. (3) Insights from ablation experiments: Ablation experiments provide key insights into the contribution of each component of FST-ATTNet. After removing the frequency-domain enhanced attention mechanism, <italic>R</italic>² decreases to 0.92, and <italic>MSE</italic> increases to 1.13. After removing SENet, <italic>R</italic>² is 0.91, and <italic>MSE</italic> is 1.22, indicating that these attention mechanisms play a crucial role in feature enhancement. Removing all attention mechanisms results in further performance degradation, with <italic>R</italic>² of 0.90 and <italic>MSE</italic> of 1.20, highlighting their importance in the prioritization of selective features. Removing the KAN network results in <italic>R</italic>² of 0.92 and <italic>MSE</italic> of 1.09, indicating its contribution to modeling complex time series patterns. Using only frequency domain, time domain, or spatial information results in significant performance drops, with <italic>R</italic>² of 0.75, 0.68, and 0.86, respectively, and <italic>MSE</italic> of 1.98, 2.44, and 1.33, indicating the necessity of integrating frequency, time, or spatial information. Spatial information is especially critical in high-missing scenarios. (4) Applicability of the model: To evaluate the model's applicability, FST-ATTNet was applied to anchor cable monitoring data with a missing rate of up to 50%, achieving excellent results with <italic>R</italic>² of 0.92 and <italic>MSE</italic> of 0.80. The model achieved near-perfect reconstruction for datasets with strong periodicity at a 25% missing rate (<italic>R</italic>² = 0.97, <italic>MSE</italic> = 0.40). However, performance slightly declined at a 50% missing rate (<italic>R</italic>² = 0.92, <italic>MSE</italic> = 0.80), with deviations at the peaks primarily due to the training data not fully covering the entire cycle. Nonetheless, FST-ATTNet demonstrates adaptability across different monitoring scenarios and a unique ability to handle cyclic patterns in periodic data reconstruction.ConclusionsThe FST-ATTNet model offers a reliable and robust solution to the problem of continuous data loss in the health monitoring of pile-type adjustable wind turbine foundations. By deeply integrating frequency, time, and spatial domain information and incorporating advanced attention mechanisms, the model achieves exceptional reconstruction accuracy in high missing-rate scenarios, significantly outperforming traditional sequence models. Furthermore, the successful application of the model to other monitoring datasets (such as anchor cable data) demonstrates its versatility and broad applicability in structural health monitoring. FST-ATTNet not only enhances the reliability of wind turbine foundation monitoring but also provides innovative solutions to similar data repair challenges in other engineering domains, offering crucial support for the safety and efficiency of wind turbine systems.
|
|---|
