Research On Intelligence Algorithm Of Radar Signal Intra-pulse Modulation Recognition

As a key technology in electronic warfare,modulation type recognition of radar signals has essential applications in radar reconnaissance.Researchers have widely adopted timefrequency distribution(TFD)to obtain the signal time-frequency image(TFI)and used convolutional neural network(CNN)to identify radar signal modulation types.However,there are still three problems.Firstly,to improve the recognition rate,researchers often consider improving the structure of the CNN used.Because CNN's are nonlinear systems with a large number of adjustable parameters,it is difficult to find a suitable improvement direction.Secondly,since the performance of CNNs varies for different types of signals,it is incomplete to evaluate the performance of CNNs by recognition rate only.Finally,the existing studies only consider the recognition results of CNN on radar signals under a single TFD.They do not analyze the effect of different TFDs on the recognition effect of CNN.Therefore,this thesis proposes the new radar signal intra-pulse modulation recognition method to solve these three problems.Each problem and corresponding improvement measures are shown below.1.To avoid the problem of the unclear direction of CNN improvement and improve the recognition rate simultaneously,this thesis proposed a new radar intra-pulse modulation recognition method based on contour extraction.The key of the method lies in limiting the range of CNN extracted features and discard the signal energy attributes.It can prompt the CNN to learn the morphological features of the signal in the TFI.The difference between the TFIs of radar signals and conventional images is that the different modulation types of radar signals are clearly distinguished in the TFI signal contours.It means that learning morphological features can retain the differences between signals to the maximum extent.It is verified through experiments that after the introduction of contour extraction,there is no need to improve the CNN structure,and the recognition rate is significantly improved,avoiding CNN improvement.2.A new CNN recognition performance evaluation criterion is proposed to address the problem that the recognition performance of CNNs is not fully evaluated by recognition rate only.The criterion contains two indicators,which measure the ability of CNN to extract correct features and its anti-noise performance.At the same time,the criterion also implies the difference in the performance of CNN to identify different types of signals,which can evaluate CNN performance more comprehensively.3.To address the problem that existing studies have not explored the effect of different TFDs on CNNs,this thesis discusses short-time Fourier transform(STFT)and Choi-Williams distribution(CWD)on the recognition effect of CNNs through recognition rate and the proposed novel evaluation criterion.4.Also,since the CNN recognition performance criterion reflects CNN's recognition performance,it can be used as a reference indicator for CNN's improvement and provide a new basis for researchers to optimize CNN.Finally,based on the above research,this thesis combines a novel evaluation criterion with contour extraction to form a self-assessable intelligent algorithm for radar intra-pulse modulation identification.