Robust wind turbine monitoring for digital twin integration: A physics-informed covariance-preserving deep learning approach

Accurate state estimation in wind turbines is crucial for integrating real-time monitoring into Digital Twins (DT). We present Eigen Decomposition-KalmanNet (ED-KN), a novel physics-based machine learning approach designed for real-time state estimation of wind turbines within DT frameworks. ED-KN combines the physics-informed and transparent structure of Kalman filter (KF) with the adaptability of deep learning to address critical challenges such as sensor noise, system nonlinearities, and complex operational conditions of wind turbines. A key innovation is the development of eigen decomposition-based Positive Definite Enforcing Layer, which ensures stable and reliable error covariance estimation throughout the process. Another key contribution is the integration of the estimated error covariance directly into the training process to enhance the accuracy of state estimation. Additionally, a Kalman gain masking technique is proposed that addresses scale discrepancies between state and measurement variables that cannot be resolved through normalization. We apply the ED-KN to National Renewable Energy Laboratory’s 5-MW wind turbine model, demonstrating its superior performance compared to KF under various noisy conditions. Using the estimated states by ED-KN, cumulative fatigue damage was estimated. Results indicate that ED-KN provides a solid foundation for DT development in wind turbines, optimizing their performance and extending the system’s lifespan.