Utilizing Neural Networks To Adaptively Demodulate And Decode Signals In An Impulsive Environment

Abstract: Electromagnetic disturbance can be detrimental to the performance of a radio communication system, and in today’s society where more and more electronic devices are present in our everyday life it is increasingly vital to consider man-made interference. A communication system can take into consideration the noise characteristics and by doing so will excel in such areas, however, this follows that the algorithms utilized in such systems are more computationally complex and are therefore slow. This master thesis aims to explore the possibility of a neural network-based solution that reaches the same accuracy, as existing methods, but more quickly. Numerous different existing model alternatives have been explored and a plethora of different improvement techniques have been outlined. Two models, Hannet and Lannet, have been designed and improved to enable adaptive demodulation both including or excluding decoding at the receiver in an end-to-end communication system. The evaluation results demonstrate that the proposed models are comparable and in some cases even more accurate than current standardized methods. However, the models are unable to fully learn the decoding algorithms present in the experiments. Thus even though demodulation by itself thrives, performing decoding in conjunction with demodulation is out of reach for these models.