Reliable anomaly intelligent detection for electromechanical system of wind turbines based on probabilistic physics-informed autoencoders

The reliability and safety of electromechanical systems in wind turbines are essential for sustaining operational stability and energy efficiency. Anomaly detection, as a fundamental element of prognostics and health management, facilitates the assessment of system health status. With the proliferation of industrial big data, data-driven anomaly detection methodologies have gained substantial attention. Nevertheless, the inherent uncertainty and pronounced stochasticity of such data present considerable challenges for purely data-driven models. Even minimal false alarms can trigger unwarranted maintenance activities, potentially resulting in significant economic losses or critical safety concerns. To address these challenges, this study introduces a Physics-Informed Gated Recurrent Unit network (PyGRU) to construct a normal behavioural model for prognostic purposes. PyGRU is formulated based on a general ordinary differential equation framework, capturing the relationship between latent features and model predictions. The effectiveness and monitoring capabilities of the PyGRU-based anomaly detection approach are validated through application to a real-world wind turbine operational dataset. This integrated method reduces epistemic uncertainty for in-distribution data while accurately representing increased uncertainty for out-of-distribution inputs, thereby improving the robustness and reliability of AI-driven anomaly detection systems.