Deep Learning Demodulation of Amplitude Noise Shift Keying Spread Spectrum Signals

Digital communications techniques based on random, chaotic, or noisy carriers are well known and successfully used in a number of applications. Simple on-off or amplitude shift noise keying modulation schemes are among the most popular. In this paper 1, we propose to use a classification model based on an artificial dense neural network and a deep learning approach for software-defined demodulation of amplitude noise shift keying spread spectrum signals. The main challenge with processing of such signals is that statistical properties of signal and interference are very similar. The aim of the research is to proof the feasibility of the proposed technique and to obtain the noise-immunity metrics. The methodology of the research is to evaluate the deep learning demodulation model pre-trained on the artificially synthesized dataset. The dataset contains the automatically labeled mixtures of noise carrier signals with additive white Gaussian interferences. The average signal-to-noise ratios in the dataset range from -30 dB to 0 dB. The numerical results from simulations are used to evaluate the demodulation performance. We present the demodulation performance as symbol and bit error rates and compare results with other well-known approaches. The paper reports noise immunity greater than immunity of “power reception” method at least for 3.5 dB. In addition, we assessed the time complexity and proved real-time processing capability.