A Computationally Lightest and Robust Neural Network Receiver for Ultra Wideband Time Hopping Communication Systems

This work systematically develops and proposes the simplest form of Neural Network (NN) architecture for efficient detection of signals in an impulse radio (IR) ultra wideband (UWB) time hopping (TH) communication system. The simplicity of the NN receiver ascribes to the subtle change in the training method, which contrary to the conventional ones, requires the network to be trained with only 10000 symbols in an ideal condition without any channel impairments to allow the receiver to learn only the desired signal patterns. The detection efficiency of the NN receiver is attributed to a key parameter entitled as an optimal average cost function which is formulated from the soft decision statistics that is reached by the NN receiver in 3000 training epochs owing to the prudent selection of all hyper-parameters. Both the simplicity and detection efficiency of the NN receiver are quantified by i) a comparison as well as a comprehensive performance evaluation conducted with the classic Rake receiver structures in a broad range of chaotic communication scenarios involving mobility and ii) derivation of simple expressions to compute the computational complexities. The performance results demonstrate that the proposed NN receiver cuts-down the computational complexity by 76% while attaining and maintaining a very high performance gain of 30 dB in signal-to-noise ratio on average for all test scenarios.