Torque fault signal extraction in hybrid electric powertrains through a wavelet-supported processing of residuals
Francesco Melluso, Mario Spirto, Armando Nicolella, Pierangelo Malfi, Ciro Tordela, Chiara Cosenza, Sergio Savino, Vincenzo Niola
Abstract
• Development of an innovative Torque FD Technique based on shaft angular speed monitoring only, suitable for HEPS. • The ODE-RLS-DWT framework overcomes the fault masking effect and the limitations of data-driven methods. • Experimental Validation across various hybrid operating modes on the HEPS Test Bench. • Consistency of the technique even under non-ideal conditions, including measurement noise and model parameter variations. • Performance metrics show an average ACC of 98.88 %, a FAR of 0.63 % and an average processing time of 0.7 ms across 97 tests. The growing implementation of Hybrid Electric Propulsion Systems (HEPS) in the aeronautic field introduces new challenges for health monitoring and fault detection (FD). This is due to different motor dynamics and limited sensor accessibility. Internal Combustion Engine (ICE) torque represents one of the most critical quantities for monitoring HEPS reliability. This work proposes a non-invasive torque FD framework that couples physical modelling and signal processing approaches. It addresses the previously mentioned issues, supervising the angular speed measurement only. The residual, obtained from the difference between measured and model-predicted shaft speeds, is refined through a Recursive Least Squares estimator to enhance the adaptability of the proposed modelling approach. A Discrete Wavelet Transform isolates transient features associated with ICE characteristics. This enables the extraction of torque-fault signatures from non-ideal conditions and overcomes the fault-masking effect of adaptive estimators. Experimental validation on a dedicated HEPS test bench confirms consistent performance of the method under various hybrid operating conditions, parameter deviations, and measurement noise. The proposed method achieves an average accuracy of 98.88%, a false alarm rate of 0.63%, and an average processing time of 0.7 ms.