A Spatiotemporal Dynamic Wavelet Network for Infrared Thermography-Based Machine Prognostics

IRT is increasingly exploited to track mechanical degradation in a noncontact manner, readily available for further prognostics. Recently, wavelet networks have coalesced DL and wavelet transform (WT), expected to achieve data-driven and interpretable prognostics. However, traditional wavelet networks neither possess enough adaptability to extract degradation-related features nor sufficiently fuse learned wavelet coefficients. Thus, this article presents a spatiotemporal DyConv-based wavelet network to handle the above difficulties in industry. First, a spatiotemporal dynamic convolution layer is presented to flexibly modulate kernels according to input samples and the multidimensional kernel space. Second, a learnable LS structure is constructed to perform signal-adapted WT while incorporating crucial properties to link the optimization of lifting filters and degradation-related feature learning. Finally, a bilinear feature fusion is implemented to jointly represent extracted wavelet energy across decomposition levels, facilitating synergistic optimization. The superiority of the proposed method is illustrated through infrared degradation image datasets.