Research And Application Of Signal Recognition Algorithm Based On Convolutional Neural Network

Modulation identification is an important technology in wireless communication,while burst signal identification is a technology for identifying more complex burst shortwave signals composed of modulated signals.This thesis mainly explores and improves the identification of burst signals and modulated signals.To identify modulated signals,this research proposes the OBN structure that has good performance under low signal-to-noise ratio.On this basis,for identifying the special waveform structure in the burst signal,the self-attention layer is introduced to build the OAN model.Then,an adaptive pooling method is proposed to improve the model and solve the problem of inconsistent length in burst signal datasets.Finally,aiming to deploy the network to the real environment,the RepSig network based on structural re-parameterization is proposed to speed up inference.The contributions of this research are as follows:1)To solve the noise problem,a convolutional neural network OBN for modulated signals is proposed.OBN is composed of Octave,BlurPool and Mish,and is trained by some anti-noise strategies.OBN achieves 95.07%accuracy on the RADIOML 2018.01 A dataset,beating the previous best method.2)Aiming at the problem of inconsistency in the length of burst signals,an adaptive pooling method is proposed,which can also perform well for burst signal with large signal length spans.Then,a high-precision end-to-end model named OAN with self-attention layer is proposed to extract the matching position of the waveform,and the influence of position in self-attention layer is discussed.3)Based on the consideration of speed in the real environment,a high-speed model RepSig using re-parameterization is redesigned.The model is accelerated by means of operator fusion,warmup,INT8 quantization,etc.,which can achieve an inference speed of 1.48 milliseconds per signal.In addition,a quantization-friendly multi-branch structure is proposed to alleviate the quantitative collapse phenomenon when multi-branch fusion is introduced into the model.