Research On Modulation Type Recognition Algorithm Based On Machine Learning

Modulation type recognition is an important part of the non-cooperative communication process.It is a prerequisite to complete demodulation and obtain information.It is often used in electronic reconnaissance,electronic jamming,and spectrum supervision.It has important military and civilian values.In recent years,machine learning has developed rapidly and received widespread attention.Machine learning has powerful classification performance.So this paper mainly studies modulation type recognition algorithms based on machine learning.Firstly,in order to improve the performance of modulation type recognition algorithms based on pattern recognition and single classifier,modulation type recognition algorithms based on ensemble learning for 16 APSK,32APSK,16 QAM,32QAM,BPSK,QPSK,and 8PSK signal recognition are studied.Feature parameters of multiple high-order cumulant combinations are proposed,and simulation results show that the proposed feature parameters are superior to the traditional feature parameters based on the ratio of two cumulants.In order to increase recognition rate and generalization of a single classifier,a HOCAB algorithm using Adaboost classifier and a HOCRF algorithm using random forest classifier are proposed.The simulation results show that the HOCAB and HOCRF algorithms with integrated learning have better recognition performance than the HOCDT algorithm using decision tree.The operation time of HOCRF is almost the same as HOCDT,which is much lower than that of the HOCAB algorithm.Secondly,in order to overcome the dependence of the performance of the modulation type recognition algorithm based on statistical pattern recognition on the feature parameters,the modulation type recognition algorithms based on deep learning are studied for its powerful signal feature extraction capability.LSTM algorithm based on LSTM network and CNN algorithm based on CNN network are studied respectively.Both algorithms take source signals data as input and use neural networks to automatically extract features to achieve intelligent classification.In order to fully extract the spatial and temporal feature of signals,a CLP algorithm is proposed,in which a CNN network and a LSTM network are connected in parallel.The CNN network and LSTM network are used to extract the spatial and temporal features of the signals,respectively,and the two features are fused and classified.Simulation results of recognition for 16 APSK,32APSK,16 QAM,32QAM,BPSK,QPSK,and 8PSK signals show that the recognition performance of CLP algorithm is better than CNN algorithm and LSTM algorithm.In order to further improve the recognition performance,a modulation type recognition algorithm(HNBagging)based on integrated heterogeneous neural networks is proposed.The models of CNN algorithm and CLP algorithm are used as the base classifier,and the random forest is used as the learner,and the output of each base classifier is learned and integrated.Simulation results show that the recognition performance of the HNBagging algorithm is better than that of a single classifier.Finally,in order to alleviate the dependence of deep learning performance on the number of samples and to break the assumption that the training and test sets are independent and identically distributed,a modulation type recognition algorithm based on model transfer is proposed.Based on the CLP algorithm network model,a TCLP algorithm which transfers the whole CLP algorithm network model,a TCLP-C algorithm which transfers only the CNN network model,and a TCLP-L algorithm which transfers only the LSTM network model is obtained,respectively.The three algorithms use the auxiliary data to train the CLP algorithm network model.After iterating to a certain number of times,the model is transferred to the target domain,and a small amount of data in the target domain is used to tune the network model finely to obtain the final classification recognition network.Simulation results of recognition for 16 APSK,32APSK,16 QAM,32QAM,BPSK,QPSK,and 8PSK signal show that the recognition performance of TCLP algorithm and TCLP-C algorithm is much better than that of CLP algorithm with a small number of samples,and the recognition performance of TCLP-L algorithm is slightly better than that of CLP algorithm.