A study on an optimized method for wind turbine fault diagnosis based on deep learning algorithms

Abstract

This study proposes an optimized multi-source fault diagnosis method for wind turbines operating under variable speeds, loads, and noise disturbances. The method addresses weak fault characteristics, class imbalance, and unstable recognition across different operating conditions. Synchronous vibration, current, temperature, rotational-speed, and power signals are modeled jointly to construct multi-domain representations. The diagnostic framework further integrates temporal representation, texture coding, class reweighting, operating-condition alignment, and decision calibration. Experimental results demonstrate stable recognition across different fault types and complex operating conditions. The proposed method also improves rare-fault recall and cross-condition diagnostic performance, supporting intelligent fault identification and wind-turbine operation and maintenance.