Explainable Deep Learning Detection of Gaussian Propeller Noise with Unknown Signal-to-Noise Ratio

Due to its need for robustness and reliability, underwater target detection is a challenging task for deep learning applications. Though many attempts were made to deal with this problem using expert features, few works assessed the benefit of designing deep raw waveform architecture despite its performance in other domains. This paper is focused on explainable raw waveform based neural network for underwater propeller detection. To this purpose, we design a class of Bayes explainable deep neural networks that contains neural networks whose architecture matches the structure of the optimal Bayes detector. This class is derived from a realistic acoustic model of underwater propeller noise. It is established that the approximation error of our class is as small as desired. We also show that this class can be efficiently implemented as a convolutional neural network. Numerical simulations study the risk and explainability of our class compared to a usual convolutional neural network.