Research On CPM Signal Modulation Recognition

Continuous phase modulation(CPM)has been widely used in military field and satellite measurement and control due to its high efficiency in power utilization and frequency band utilization.At the same time,in order to establish the electromagnetic cognitive intelligent perception architecture and realize the intelligent receiving and processing of modulated signals,it must be combined with artificial intelligence technology(machine learning,deep learning)to enable the machine to have real-time dynamic perception ability.Therefore,it is extremely urgent to develop CPM system detection and reconnaissance technology based on artificial intelligence technology.In this paper,7 KINDS of CPM signals and 9 kinds of conventional phase discontinuous signals in MIL-STD-188-181 C protocol are taken as the research objects.The feature extraction method is used to complete the inter-class recognition of CPM signals,and the deep learning method is used to complete the signal denoising and CPM in-class recognition.Two new features,namely differential envelope continuous factor and phase locus kernel mapping feature,are analyzed,and the applications of attention mechanism in noise reduction and signal classification are discussed,as well as how to build an optimal network model.Firstly,for the recognition of CPM signals between classes,the phase transformation characteristics of adjacent symbols are extracted,and the differential envelope continuity factor is proposed for the first time to determine the phase continuity of signals and accurately separate CPM signals from phase discontinuous signals.The characteristics of anti-frequency deviation and anti-phase deviation are verified by simulation experiments.Then,a series of entropy features and phase trajectory features are constructed to realize the recognition of phase discontinuous signals.In this paper,the characteristics of phase trajectory kernel mapping are studied,and the phase change paths of adjacent symbols are found by interpolation.The problem of OQPSK signal and PSK signal,especially QPSK signal,is solved.Simulation results verify the effectiveness of the proposed feature,and the accuracy reaches 97% at 5d B.Then,aiming at the slow training speed of existing noise reduction encoders,the parallel computation in attention mechanism is introduced.By dividing the input signals,calculating the correlation in pairs and directly extracting the global information for modeling,the modeling efficiency is greatly improved.Speed Speeds up network integration.Simulation results show that the denoising effect of the proposed method on CPM signals is much better than that of conventional wavelet threshold denoising methods.Finally,the shortcomings of the existing attention mechanisms are analyzed,that is,only the primary distribution information and low-frequency components of the signal are concerned,and the high-order moment attention and multi-spectral attention modules are proposed to extract the signal from the time domain distribution and frequency domain distribution.The rich details are generated,the two-bit attention map is generated,and the original feature map is re-calibrated to achieve the purpose of highlighting the useful components and suppressing the interference components.Simulations verify the low complexity and high performance of the proposed module,which can be inserted into any convolutional neural network to improve network performance.