Applicability of different neural network architectures in UWB signal processing for different object classification

Abstract

In the paper, the performance of different Artificial neural network (ANN) architectures - CNN, RNN, Transformers, CNN/LSTM, ResNet, and MLP is discussed in UWB impulse radio radar signal classification. The signals are obtained by reflecting and passing UWB pulses through different material objects that give different information for the classification purpose. The ANN architectures are compared on their classification precision, training time, and memory requirements. The training data consists of 144 objects including regular and crashed PET bottles, glass bottles, and metal cans. The results show the accuracy's of classification if mono-static (reflected signals are analyzed), bistatic (propagated trough signals), and multi-static setups of UWB radar are used. For single-channel cases GRU (99.65%), Resnet (99.69%), Transformer (99.66%) architectures are preferable, while the highest multi-static evaluation accuracy reaches 99.90% for the Transformer. As expected the more dense ANN networks perform better classification.