Research On Modulation Signal Recognition Algorithm Based On Deep Neural Network

Modulation signal recognition technology is crucial as an auxiliary tool for identifying target categories and additional attributes in noncooperative communication scenarios.It has progressively become a focal point of research both nationally and internationally in recent years.Nevertheless,modulated signals pose numerous challenges due to a variety of factors such as diverse modulation methods,noise interference,non-stationary signals,lack of labeled data,and the continuous emergence of new signal types outside of established libraries.The identification of these signals presents significant difficulties and challenges.Additionally,the real-time performance of the algorithm and the cost of its application must also be thoroughly considered for successful integration into practical engineering.This thesis introduces a series of novel modulation signal recognition methods designed to address these challenges.The effectiveness of each technique is substantiated through rigorous experimentation,and the feasibility of their application in edge engineering is also explored.The main content of this study is divided into the following four aspects:(1)This thesis introduces a method for modulation signal recognition,based on multi-scale attention convolutional neural networks,designed to address challenges related to non-stationary signal performance and a variety of modulation methods.The method initially extracts the signal’s features in the time domain,frequency domain,and original sequence.It then employs multi-feature fusion to mitigate the issue of unstable feature performance in non-stationary signals.Furthermore,this approach constructs a multi-scale convolutional network and an efficient channel attention network for multi-scale feature extraction and adjustment of feature information channel importance.This allows the model to extract more subtle and highly separable portions,thereby adapting to various complex types of modulation signals.Experimental results indicate that this method surpasses previous ones in terms of recognition performance,thereby validating the model’s capacity to adapt to complex modulation signals.(2)In response to the problems of noise interference,difficulty in extracting time-varying signal features,and lack of annotation for massive signals,this thesis proposes a modulation signal recognition method based on comparative learning and Transformer.This method adopts wavelet threshold denoising technology to reduce the interference of noisy data on the model.In the model training stage,unlabeled data’s available features are obtained through an unsupervised pre-trained comparative learning encoder to improve the efficiency of unlabeled data utilization.Then,the encoder is transferred to a downstream recognition network based on a self-attention transformer to enhance its feature extraction ability for time-varying signals.The experimental results show that the model has good recognition performance in noisy environments and exhibits strong robustness under different training sample sizes,indicating reasonable application prospects.(3)To further obtain the specific types of radiation sources and improve the lack of labeled data in the training sample set and the recognition of new types of signals outside the library,this thesis proposes a radiation source fingerprint recognition method based on small sample learning,which introduces the small sample learning framework into the field of radiation source fingerprint recognition.The model combines a cavity convolution pooling pyramid and Transformer to extract the signal’s multi-scale and time series features.This can alleviate the overfitting phenomenon caused by the lack of training data.In the model training phase,the self-supervised pre-training method is first used to supplement the signal self-supervised characteristics to improve the overfitting phenomenon of the model on small sample data.Then the meta-transfer learning algorithm is used to learn the features and laws of new categories so that the model can quickly adapt to new types of signals.The experimental results show that this method can quickly learn and adapt to new types of signals without labeled data and has high recognition accuracy for these new types of signs.(4)In practical engineering,algorithms must have high real-time and low application cost characteristics when deployed at the edge.This article optimizes and adapts the algorithm through two hardware inference frameworks,Jetson TX1,and Itap-RK3588,to improve its engineering feasibility.The experimental results show that the adapted model deployed on this two low-power,low-cost Edge devices has the characteristics of high real-time,noise robustness,and low minor sample sensitivity,which provides a new idea for the optimal deployment of modulation signal recognition theory in hardware.In summary,the modulation signal recognition method proposed in this article has high recognition accuracy and environmental adaptability,improving the recognition difficulties caused by various modulation methods,noise interference,non-stationary signals,lack of labeled data,and recognition of new types of signs outside the library.In addition,this thesis realizes modulation signal recognition based on the hardware framework according to the actual engineering application requirements,which provides a new idea and method for learning modulation signal recognition technology on low-power,low-cost Edge devices.The research results of this article have essential references and reference value for promoting the development and application of modulation signal recognition technology.