Study on fault sensitivity and feature distribution characteristics of wind turbine condition monitoring data

This study investigates diagnostic index distribution patterns across fault modes in wind turbines (WTs) to establish boundary conditions for condition monitoring models. Using historical operational monitoring data and maintenance logs, we analyze the failure statistics of key WT components, the distribution of critical diagnostic indexes under fault conditions, and the co-occurrence relationships among various indexes across different operational states. We further investigate the power characteristics of generators and turbines under different fault modes, examining the fault sensitivity of diagnostic indexes and their mapping relationships with power output and wind speed. Finally, we propose a novel adaptive deep learning network model to validate the identified patterns, demonstrating its effectiveness in fault prediction under varying operational conditions. The results show more positive correlations between monitoring indexes at partial load during faults compared to rated power. The proposed adaptive deep learning model Adaptive Recurrent Neural Network (AdaRNN) reduces prediction errors to <5% by dynamically aligning temporal distributions through its temporal distribution characterization and matching framework. These findings enable prioritized monitoring of critical index correlations during high-power operation and wind-speed-specific modeling, providing field-ready solutions for predictive maintenance of WTs and similar mechanical systems.