Rispecnet: a Geometry Independent Deep Learning Oriented Method for Coherent and Non-Coherent Signal Number Estimation

Abstract

In this study, “the total number of non-coherent signals including only direct-path signals” and “the total number of coherent signals including direct-path and multi-path signals” estimation algorithm tailored for sensor array signal processing systems is proposed. The applications of these algorithms are numerous. Existing Direction Finding (DF) techniques often face challenges related to the rank loss of the noise-free covariance matrix due to uncertainties regarding the number of direct-paths and multi-paths. Traditional and modified spatial smoothing algorithms have opened up possibilities for estimating coherent signals’ Direction of Arrival (DoA). However, all these algorithms need to know the exact number of coherent and non-coherent signal numbers in the presence of interference. Besides, they can work properly only if a very specific antenna array geometry is provided. Many existing signal enumeration methods fail when distinguishing coherent and non-coherent signal numbers in an environment where a multi-path effect may exist. In this paper, a novel deep learning-based method which is called Received Individual Sensor Powers and Eigenvalues Combination Network (RISPECNet) is proposed to make this distinction, with the addition of existing information that is already calculated by the covariance matrix in prevailing methods. Compared with conventional approaches for distinguishing coherent signals, the necessity for Forward Backward Spatial Smoothing (FBSS) is eliminated, resulting in lower mathematical complexity that can be utilized in real-time applications. Additionally, the proposed algorithm works on various geometries by using networks specifically trained for the relevant geometry to make it possible to realize it using sensors without being bound to a specific sensor geometry. It is tested on simulated and also real-world data in non-coherent only scenarios and coherent and non-coherent scenarios together. Simulation results show that the proposed method has an enhanced estimation capability in lower and higher SNR Levels. For real world data sets which includes only non-coherent signals, as the SNR rises from 3 dB to 18 dB, RISPECNet's accuracy improves significantly, ranging from 95.1% to 99.7%, surpassing the performance of nearly all other methods. On the other hand, for simulated dataset which consists of coherent and non-coherent signals, the proposed architecture's accuracy values increases from 93.7% to 98.1% when the SNR increases from 0 dB to 40 dB. So, RISPECNet outperforms the existing methods in terms of accuracy without the need for FBSS. © 2025 The Franklin Institute