Applying Machine Learning to Adaptive Array Signal Processing Weight Generation

Array-based sensing with the goal of signal reconstruction to support downstream processing presents challenges in many domains, including radar and sonar. Existing methods of generating filter weights are stochastically optimal in the beam power response, but do not necessarily result in a better solution for signal reconstruction. Coherent signal-to-noise ratio (SNR) correlates the actual signal to the output of the beamformer and better relates the quality of the reconstructed timeseries. However, coherent SNR requires knowledge of the waveform and so is not available for passive applications. From a large corpus of training data, machine learning approaches learn to estimate solutions improving coherent SNR using data available in situ. This article proposes an approach using machine learning to aid in array signal processing weight generation. A cost function is developed to assess the quality of a set of weights based on coherent SNR responses, and this cost function is used to train a neural network. While this approach works when the stationary assumption holds, this article also considers the case of limited duration, loud dynamic interferers. An expanded architecture trains another neural network, inspired by extreme value distributions, which is then used to select the samples most likely to come from the so-called stationary distribution. This work demonstrates that machine learning has leverage to improve on existing weight generation techniques for the signal reconstruction task.