Real-time Analysis And Recognition Of Communication Signals Based On GPU

In today’s context,the time-frequency analysis and modulation identification of full-band communication signals are of real need and importance.Time and frequency domain analysis of communication signals enables data processing of the received signal to obtain the basic signal parameters,while modulation identification helps the receiver to accurately receive signals from complex electromagnetic environments and to determine the demodulation method to be used for subsequent processing.In order to meet the requirements of real-time processing,fast calculations are required for large data volumes.The focus of this paper is therefore to exploit the powerful parallel computing capabilities of GPU processors and to utilise the multi-threaded resources of the GPU for real-time processing of time-frequency analysis and modulation identification using a unified computing device architecture(CUDA).This paper designs a software system for signal detection and reconnaissance in complex electromagnetic environments from the perspectives of both time-frequency analysis and modulation identification,based on which the main challenge of the subject is reflected in real-time processing.Therefore,this paper is dedicated to a more in-depth study of some of the existing parallel algorithms such as STFT parallel algorithm,peak finding parallel algorithm,complex point multiplication/modulation parallel algorithm,etc.,and proposes new related parallel algorithms such as detector/hold parallel algorithm,pulse transition area outside/transition point data finding parallel algorithm,etc.in combination with some software functions.Through the analysis of the relevant signal processing theory,the software system is implemented on a GPU processor based on CUDA programming applying parallel algorithms for the time-frequency analysis function.The parallel algorithm applied or proposed for the calculation of this function is then described in detail,in particular the parallel process and the parameters in the kernel function are described in detail,and the parallel algorithm will be the focus of this part of the research.The parallel algorithm will be the focus of this part of the study.Finally,the validity of the scheme is further verified by analysing the results of the functional implementation,and the effectiveness of the parallel algorithm in terms of processing speed increase is verified by means of partial test experiments.In the modulation identification part of the software system,two methods are studied: the identification method based on the time-frequency map and the residual neural network,and the identification method based on the instantaneous feature parameters and the decision tree classifier.The first method uses the time-frequency analysis function of the software to obtain the time-frequency map of the signal in real time,and then loads the pre-saved residual neural network model which has been built,trained and tested to identify the modulation mode.The second method focuses on a parallel algorithm for instantaneous feature extraction,which can be used to quickly calculate the time domain data of the signal to obtain the feature parameters,and then design a decision tree classifier and adjust the parameter thresholds through extensive experiments to achieve fast identification of the modulation mode.This paper also presents time tests on both methods to analyse the performance of the software processing speed.