Embeddable Convolutional layer-based Filter for Wireless Signal Detection

Abstract

The performance of traditional deep learning-based signal detection methods have been challenged by real world signal impaired by random noise and interferences through wireless channel, they quite often appeal to a prior filter for interference elimination before deep learning model. To deal with the problem, we introduce an embeddable convolutional layer-based filter (Conv-based filter), which takes raw time domain signal as input and can adaptively learn characteristics of the band-pass filter whose center frequency and bandwidth are compatible with the target signal. To enhance the performance of learned filters, the attention mechanism is introduced by using Squeeze-and-Excitation block (SE-block) after Conv-based filters. As an embeddable block, the filter is trained end-to-end in a deep learning network, no explicit assumptions about the raw data are made. For application of signal detection, compared with the way of signal preprocessing and then training separately, embedding the preprocessing as a block into the deep learning network works better.