Machine Learning Methods for Super-Resolution in Sparse Sensor Arrays

Abstract

Due to their robustness to weather and environmental conditions, radars are an important sensor in automotive and industrial applications. However, their utility in more advanced applications is increasingly limited by their resolution, which increases linearly with the number of radar elements in a uniform linear array (ULA). In this work, neural networks are trained to approximate the signals from a large aperture array from the signals obtained from smaller aperture sub-arrays. The training set consists of simulated radar responses to ideal point reflectors in noiseless vacuum, and the network is trained to minimize the squared error between the network output and the normalized log-magnitude of the large aperture array signal in its Fourier Transform domain. In general, given signals from two sets of 12-element sub-arrays, the neural network can reproduce results more than 10 times closer to the signals of the 1024-element array than signals from either of the input sub-array in terms of squared error in the Transform domain. The outputs of the neural network have over 91.97% probabilityof detection, 𝑃𝐷, with 1.02% probability of false alarm, 𝑃𝐹𝐴, compared to 67.53% 𝑃𝐷 and 5.57% 𝑃𝐹𝐴 with data from either sub-array. The results show the possibility for extracting more information by exploiting structure in real-world data with inexpensive, small sensors, and have major implications for the use of radar sensors in the automotive and industrial applications.