Research On Multi-tasking Wireless Signal Recognition Technology

Wireless signal recognition identifies the signal modulation mode,individual radiation source,equipment model and other information by extracting the characteristics of the received wireless signal,which has wide application value in the fields of wireless spectrum supervision,electronic countermeasures,wireless network security and so on.In this paper,wireless signal recognition based on multi-task learning method is studied,and the main research contents and results are as follows:(1)An individual recognition method of wireless signal radiation source based on MT-SVM is proposed.Firstly,based on the different modulation methods,the individual recognition of the radiation source is divided into two sub tasks,and the fusion features of the signal rectangular integral bispectrum and the selection bispectrum are used as inputs to MT-SVM to identify the individual radiation source.In the experimental testing,five routers of the same model with two modulation methods were validated,and the results showed that the average recognition rate of individual radiation sources reached 83.2%,which increased by 2.2% on the basis of single task SVM.(2)An individual recognition method of wireless signal radiation source based on MT-SVM is proposed.The method firstly builds a single-input multiple-output multi-task model based on CNN.Next,a differential differential constellation trace figure of the signal is generated and input into the MT-CNN,and features are extracted by a shared convolutional layer,and then three independent branching networks with different structures are used to classify and identify the modulation mode,individual radiation source and device model respectively.In the experimental tests,15 individual routers(including 2 modulation methods and 3 device models)were tested,and the results showed that the modulation method recognition rate reached 96%,individual recognition 91.5% and model recognition 98.3%,while the performance of individual recognition and model recognition improved by 4.8% and 1% respectively compared to single-task recognition.(3)A multi-task wireless signal open-set recognition method based on OCSVM is proposed.Firstly,the signal samples are input to a multi-task convolutional neural network trained in the closed set,the softmax maximum confidence of each task is calculated,and then the probability values are input to the OCSVM for open-set recognition respectively.In the experimental tests,the training set is 15 routers from the MT-CNN closed-set experiments,while the test set adds two new individual routers to the closed-set,one of which is a new model.The results show that the open-set recognition method proposed in this paper can achieve more than 89%open-set recognition accuracy in all three different test scenarios.(4)A multi task wireless signal recognition method based on incremental learning is proposed.On the basis of MT-CNN,an incremental learning method combining sample playback and fine-tuning can effectively achieve the recognition of newly added unknown samples.In the experiment,the OCSVM open-set was used to identify new individual categories and new model categories.The results showed that the incremental model can achieve recognition rates of 92% and 79% for newly added target individuals and models,respectively.Compared with traditional incremental learning methods for joint training,the proposed method can reduce the amount of data by about 64% and greatly save spatial resources.