Design And Implementation Of Signal Identification Method Based On Light-Weighted Deep Learning Model

With the rapid development of wireless communication technology,the number of wireless communication devices has exploded,and for safty consideration,wireless communication devices authenticating is needed.Radio frequency fingerprint Identification(RFFI)can distinguish different signal transmitters according to the hardware differences hidden in radio frequency signals,which provides a new idea for device authentication,however,the deployment of RFFI on portable devices faces the problems of insufficient computing power,slow recognition speed,insufficient storage resources,and easily influenced by noise.In order to solve the above problems,based on the characteristics of radio frequency signals,this thesis conducts research on lightweighted model design,model compression and acceleration,and proposes a signal recognition method based on a lightweighted deep learning model,and builds a signal analysis system based on this method.The main contributions of this thesis are as follows:1.In order to reduce the impact of environmental noise on RFFI,a lightweighted noise reduction module based on matrix eigenvalue decomposition is designed,which has smaller computational overhead.Aiming at the complex-valued characteristics of signal data,complex-valued convolution is used to perform feature extraction.In order to optimize the computational overhead of complex convolution,this thesis uses group convolution to replace the convolution in complex-valued convolution,and studies the optimal number of groups in specific cases.In addition,this thesis proposes an optimized complex-valued max pooling with smaller computational overhead.In order to verify the effectiveness of the above method,signal identification model was constructed.In the data set with noises,the recognition accuracy of the signal identification model reached98.4%,and the signal identification model has fewer parameters than other models,most importantly,it has a faster convergence speed and inference speed.2.In order to obtain an optimal signal identification model in specific dataset,a channel pruning method based on the amount of feature map information is proposed.This method converts the amount of feature map information into the global contribution of the channel,and searches for the optimal conversion paramters through the Simulated Annealing Algorithm.Finally,once the optimal conversion paramters and the pruning threshold is given,the optimal sub-model is obtained.After experimental verification,the pruning method proposed in this thesis is feasible,and has better performance in specific data sets than other pruning methods.Using this pruning method to prune the signal identification model to obtain the optimal sub-model,and then quantize the weight of the optimal sub-model to low-bit integers to reduce the computational overhead.Finally,the pipeline model parallel method is used to accelerate model inference speed.In the experimental environment of this thesis,the number of trainable parameters of the optimal sub-model model is only 16092.As a result,under the condition that the recognition accuracy only drops by 5%,the model’s inference speed is nearly three times faster.3.Based on the above lightweighted signal recognition model,a signal analysis system is designed and implemented.The main functions of the system include visualizing statistic results of signal,signal retrieval,current active device detection,etc.providing a visual platform for signal analysis.Finally,the method of signal recognition based on lightweight deep learning model is proposed and verified by experiments.Based on this method,a signal analysis system is built.