【作者】 张毅；

【导师】 高振斌；

【作者基本信息】 河北工业大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 摩尔斯电码(Morse code)是一种重要的短波通信方式,因编码简单、抗噪性能强等不可替代优点,在航空、航海、气象、消防、军事等领域,始终被广泛应用。其自动接收研究已进行多年,因电磁环境复杂、报务员发报手法不一等原因,接收算法的适用范围、准确率均存在很大的改善空间。本文以信号的时频图为基础,针对自动接收中的检测任务、识别任务,分别研究了基于深度学习的图像分类技术与文本识别技术,将其引入到相应的算法中,提高了信号检测与识别的准确率。另外,在发现检测与识别对时频图分辨率的矛盾要求后,首次提出了基于双重滤波的识别预处理任务,实现了检测与识别的顺利衔接,扩大了算法的适用范围。主要研究成果如下:1.基于卷积神经网络(Convolutional Neural Network,CNN)的Morse信号检测。本文首先利用改进系数的自适应阈值能量检测法,在原始时频图中盲选出各有效信号的时频图段,之后,将时频图段进行线性映射,并借鉴图像分类技术来对时频图片进行二分类。相较基于深度学习的目标检测方法,该方法无需手动标框,且对信号在原时频图中的可见性不做要求,实用性更强。在实录窄带音频上的实验结果表明,相比传统方法,提出的检测方法具有更好的性能,在满足实时性要求的同时,盲选漏选率降低了3%左右,分类F1值提高了10%以上。2.基于双重滤波的Morse信号识别预处理。由于识别任务中不适合直接应用检测到的时频图段,本文设计了双重滤波器进行Morse信号的识别预处理。首先,根据检测到的信号频率范围,将原音频信号进行带通滤波,然后,基于信号频点,将带通滤波后的信号进行卡尔曼滤波。实验结果表明,双重滤波算法滤除了段外干扰,减弱了段内干扰,解决了检测与识别对时频图分辨率的矛盾要求问题以及邻频干扰带来的时频图失真问题,保证了识别工作的顺利进行。3.基于改进卷积循环神经网络(Convolutional Recurrent Network,CRNN)的Morse信号识别。本文基于CRNN,设计了一个由CNN、卷积注意力机制模块(Convolutional Block Attention Module,CBAM)、双向门控循环单元(Bi-directional Gated Recurrent Unit,Bi GRU)、连接时序分类(Connectionist Temporal Classification,CTC)组成的识别网络。提取目标信号的时频图并伪彩色化,通过带有注意力机制的CNN层获取特征序列,再利用Bi GRU预测每个时间点对应的字符概率,最后通过CTC层转录输出报文。实验结果表明,提出方法较传统方法和CRNN方法,在不同信噪比、频率抖动、码长偏差情况下均有更好的识别效果,实录数据中字准确率较CRNN方法提高了10%以上,且满足实时识别要求。更多还原

【Abstract】 Morse code is an important short-wave communication method.Due to its irreplaceable advantages such as simple coding and strong anti-noise performance,it has always been widely used in aviation,navigation,meteorology,fire protection,military and other fields.Its automatic receiving research has been carried out for many years.Due to the complex electromagnetic environment and the different sending methods of the telegraph operator,there is a lot of room for improvement in the scope of application and accuracy of the receiving algorithm.Based on the time-frequency spectrum of the signal,this paper studies the image classification technology and text recognition technology based on deep learning respectively for the detection task and recognition task in automatic receiving,and introduces them into the corresponding algorithm to improve the accuracy of signal detection and recognition.In addition,after discovering the contradictory requirements of detection and recognition on the resolution of time-frequency spectrum,this paper proposes the recognition preprocessing task based on double filtering for the first time,which realizes the smooth connection between detection and recognition and expands the scope of application of the algorithm.The main research results are as follows:1.Morse signal detection based on convolutional neural network(CNN).This paper firstly uses the adaptive threshold energy detection method with improved coefficients to blindly select the spectrum segments of each valid signal in the original time-frequency spectrum.After that,the spectrum segment is linearly mapped,and the image classification technology is used to classify the segment.Compared with the target detection method based on deep learning,this method does not require manual marking,and has no requirement on the visibility of signals in the original time-frequency spectrum,so it is more practical.Experimental results on recorded narrowband audio show that the proposed method has better performance than the traditional method.The missed selection rate of blind selection is reduced by about 3% and the classification F1 value is increased by more than 10% while meeting the real-time requirements.2.Morse signal recognition preprocessing based on dual filtering.Because it is not suitable to directly apply the detected time-frequency spectrum segment in the recognition task,this paper designs a dual filter for Morse signal recognition preprocessing.First,according to the detected signal frequency range,the original audio signal is bandpass filtering,and then,based on the signal frequency point,the signal after bandpass filtering is Kalman filtering.The experimental results show that the dual filtering algorithm filters out the interference outside the segment,weakens the interference inside the segment,and solves the problem of the conflicting requirements for the resolution of the time-frequency spectrum between detection and recognition,as well as the distortion of the time-frequency spectrum caused by the adjacent interference,ensuring the smooth progress of the recognition work.3.Morse signal recognition based on improved convolutional recurrent network(CRNN).Based on CRNN,this paper designs a recognition network consisting of CNN,convolutional block attention module(CBAM),bi-directional gated recurrent unit(Bi GRU)and connectionist temporal classification(CTC).The time-frequency spectrum of the target signal was extracted and pseudo-colored,and the feature sequence was obtained through the CNN layer with attention mechanism.Bi GRU was used to predict the character probability corresponding to each time point,and finally the message was transcribed and output through the CTC layer.Experimental results show that the proposed method has better recognition performance than the traditional method and CRNN method under different SNR,frequency jitter and code length deviation.In the recorded data,the character accuracy of the proposed method is more than 10% higher than that of the CRNN method,and meets the requirements of real-time recognition.更多还原