Wavelet convolutional bidirectional temporal network and modified informer for RUL prediction of wind turbines under non-stationary conditions

Abstract

Remaining useful life prediction for rotating machinery under non-stationary operating conditions demands effective extraction of degradation features from multi-sensor signals with varying spectral characteristics. This study proposes a WKCL-MBiTCN-Modified Informer hybrid framework that integrates a wavelet kernel convolutional layer (WKCL) with trainable Morlet wavelet basis functions as a front-end signal enhancement module. WKCL performs simultaneous time–frequency decomposition and feature extraction, improving spectral SNR by 2.7 dB over a fixed wavelet decomposition and 4.5 dB over the raw signal at degradation-characteristic frequencies. The parallel architecture combines multi-scale bidirectional temporal convolution with pyramidal self-attention-based spatial feature extraction. An asymmetric weighted Huber loss function penalizes late predictions more heavily to align training with engineering safety requirements. Experiments on the Commercial Modular Aero-Propulsion System Simulation (CMAPSS) and N-CMAPSS turbofan benchmark datasets demonstrate that the proposed method achieves average root mean square error of 13.06 and 3.89 respectively, outperforming recent baselines including Mamba-based models. Attention-based interpretability analysis confirms fault-mode-dependent sensor importance consistent with domain knowledge.