Learning-Based Scattering Transform for Explainable Classification

Vessel noise classification is generally considered as a challenging task due to its need for robustness and reliability. Thus, classification in this domain mainly relied on expert feature. Raw waveform architectures have been historically avoided, despite their performances in other domains. This paper proposes a Learning-based Scattering Transform (LST) that efficiently learns temporal dependencies within cyclostationary signals, such as vessel noises. The LST is implememented as a Convolutional Neural Network (CNN) with short filters whose structure mimics a multiscale signal decomposition. By this way, the architecture of our neural network is intrinsically explainable. Numerical simulations compare our method to an other explainable model and classic convolutional neural networks.