Decoupling operational variabilities from gear fault signatures: an adaptive wavelet packet and multi-scale attention deep learning framework

Reliable gearbox fault diagnosis under variable operating conditions remains a critical challenge in non-destructive condition monitoring of industrial rotating machinery, where speed- and load-induced signal variations systematically mask fault-induced signatures and degrade the generalisation capability of conventional diagnostic models. To overcome this drawback, in this paper, we propose a novel Adaptive Wavelet-Attention Deep Learning (AWADL) framework which combines entropy-guided Adaptive Wavelet Packet Decomposition (AWPD) with a Multi-Scale Attention Convolutional Neural Network (MSA-CNN) for robust and interpretable multi-class gearbox fault classification. The AWPD module uses a Shannon entropy best-basis criterion to adaptively structure the optimum wavelet decomposition tree for each fault condition to obtain physically meaningful sub-band representations which preserve fault-sensitive spectral content while reducing the effect of operational variability. To learn condition-invariant fault representation, channel and spatial attention module (CBAM) has been introduced along with parallel multi-scale convolutional branches (kernel sizes: 3, 5, 7) in MSA-CNN architecture. The proposed framework shows condition-invariant classification, which effectively reduces the variability of the input signal caused by speed and load, while using adaptive attention weighting for interpretability analysis, and provides physically meaningful justification for the decisions made by the model.