Research On Recognition Technologies Of Communication Signal Based On Feature Fusion

Detection and recognition of the communication signal is an essential part of signal situational awareness.The emitter can be effectively recognized by detecting the wanted signals and extracting multi-dimension and multi-domain features from signals to benefit the subsequent emitter tracking and monitoring.With the rapid development of software radio and integrated circuit manufacturing,the waveform design of the signal exhibits more flexibility and diversity,the hardware differences between emitter individuals are also getting smaller and smaller,the identification of the emitters are becoming more and more difficult.Therefore,it is becoming a challenging issue to detect the signal efficiently and map the signal itself to the emitter individual from multiple signal collections with little prior information,a low signal-to-noise ratio and complex channel effects.Considering the above-mentioned issues,the dissertation highlights the following topics:(1)Research on wideband spectrum detection under low signal-to-noise ratio(SNR)and complex channelA signal detection algorithm based on hidden Markov model is proposed via the probability distribution of the sub-band signal spectrum,it requires no prior information and any threshold parameters,the detection probability can achieve 94%at a-10 d B SNR.It can also be robust to the Rayleigh flat fading channel to some extent.The analysis of the real acquired satellite signal indicates that it has a moderate complexity and a stale carrier frequency estimation.(2)Research on high-order modulation recognition under limited bandwidth and non-ideal channel effectsAiming at the inadequate recognition accuracy of high-order modulation types under limited bandwidth and non-ideal channel conditions,a feature extraction algorithm based on cyclic spectrum statistical feature and the denoised high order statistics is proposed to improve its discrimination among high-order modulations.An improved particle swarm weighting strategy is also formulated in view of the optimization criteria of maximizing the correlation between classes and features and minimizing the degree of conflict between classes,which increases the average classification accuracy by 8.67% compared with the multi-domain features,the average recognition accuracy on high-order modulation formats also increases 6.4% at6 d B SNR.(3)Research on emitter subtle feature identificationFor the inadequate recognition accuracy of the emitter subtle features with the same modulation and similar modulation parameters under a small number of samples,a feature fusion method based on a pretrained model is proposed.The method fuses the deep features extracted by the deep network and the secondary features of the smoothed pseudo Wilgner-Wille distribution and the statistical features of the spectrum based on morphilogical filtering.The analysis results of 8 classes emitter signals indicate that: the recognition accuracy is increased by 5.13% and 2.47%respectively compared with the supervised canonical correlation analysis algorithm.(4)Research on multimodality emitter identificationAiming at the loose network structure of the feature extraction network due to the inconsistent feature dimensions of different modal inputs and the inadequate improvement on recognition accuracy by the fused feature at different positions of the network.A compact and lightweight network based on Transformer is designed.The recogniton accuracy on time series and time-frequency matrix is basically the same as that of the convolutional neural networks with fewer parameters and shorter inference time.A multimodality feature fusion based on sparse cross modality is also proposed.Under the dual modality input,the recognition accuracy of this method is improved by 3.06% compared with the fusion method of feature concatenation in different positions of the network.