【作者】 刘洁；

【导师】 陈劼；

【作者基本信息】 电子科技大学 ， 通信与信息系统， 2021， 硕士

【摘要】 随着20世纪50年代前后水声成像技术的出现,各国学者对于水下目标的研究脚步便从未停止,国外在这方面起步也是早于我国。但是早期的声呐图像全由人工读取,识别效果差。随着各类去噪技术、信号处理技术以及人工智能的发展,水下环境的研究也取得了越来越多的成绩。水下目标的研究一般不会直接输入未处理的时域信号,更多的是提取信号中的特征,将其输入系统来进行信号的识别。本文首先针对水下环境的特点,研究了水下目标的预处理方法,主要包含三种不同类型的常用特征提取方法。首先是计算相对简单,应用最广泛的短时傅里叶变换(Short Time Fourier Transform,STFT),然后是具有多分辨率特性的小波变换(Wavelet Transformation,WT),最后是一种Cohen类双线性变换。而Cohen类双线性变换又称核函数变换,其包含了许多种类,本文则主要选择了魏格纳分布(Wigner-Ville Distibution,WVD)来进行研究。三种时频变换各有其优缺点,其所适用的场景均不同,对于不同种类的信号其表达信号特征的能力也是不同的。其次本文针对实际水下目标信号研究其噪声特点,对实测信号进行建模,从信号去噪的角度出发来进行信号的预处理。主要包括基于经验模态(Empirical Mode Decomposition,EMD)的去噪方法、基于奇异值分解(Singular Value Decomposition,SVD)的去噪方法以及基于小波阈值(Wavelet threshold,WT)的去噪方法,同时在小波阈值去噪方法中提出了一种改进的变换小波阈值去噪方法。三种方法对于信号的噪声均有不同程度的抑制,其中改进的小波阈值去噪方法效果最佳,对于信号分析中常用的信号以及实测水声信号均可以较大幅度提升其信噪比,降低其均方误差。最后,在信号识别阶段,经过调研后发现传统的机器学习发展到现在已有了许多不同的改进种类,本文采用了卷积神经网络(Convolutional Neural Networks,CNN)的方法来对经过去噪处理的实测水声信号提取其不同类别时频特征并完成分类,随后通过准确率和混淆矩阵,将其性能与未经过去噪处理和基于多步判决的低频分析记录谱(Low Frequency Analysis Recording,LOFAR)输入网络的结果进行比较。实验结果表明经过去噪处理后的信号的平均识别率达到88.56%,而未去噪处理过的信号识别率最高为75.19%,由此证明了对信号的去噪预处理是有效的。与此同时,针对目前水下环境研究的数据样本较少的情况,采用生成对抗网络(Generative Adversarial Nets,GAN)的方法来增加信号样本,以此提升深度学习的准确率,经实验发现最终的信号识别率达到了96.673%。和此前对于信号进行增强特征的预处理再叠加卷积神经网络的方法相比,增加数据样本可以有效提升信号的识别率。更多还原

【Abstract】 With the emergence of underwater acoustic imaging technology around the 1950 s,scholars from various countries have never stopped researching the underwater targets.The foreign countries have started to research earlier than our country in this regard.However,the early sonar images were all read manually,and the target recognition effect was poor.With the development of various denoising technologies,signal processing and artificial intelligence,more and more achievements have been made in the research of underwater environment.The unprocessed time-domain signals will almost never be the input in the research of underwater targets generally.The scholars always extract features from the signals for signal recognition.This thesis firstly studies the preprocessing methods of underwater targets based on the characteristics of the underwater environment,and mainly studies three different feature extraction methods.The first is short-time fourier transformation(STFT)which can be simply calculated and most widely used.The second is wavelet transformation(WT)with multi-resolution characteristics.The third is Cohen-like bilinear transformation.The Cohen-like bilinear transformation is also called the kernel function transformation,which includes many types.And the Wigner-Ville Distribution(WVD)is selected for research in this thesis.The three kinds of time-frequency transforms are applicable to different scenarios,and their ability to express signal characteristics is different for different types of signals.Secondly,this thesis studies the noise characteristics of actual underwater target signals and preprocesses the signals from the perspective of signal denoising.It mainly includes denoising methods based on empirical mode decomposition(EMD),singular value decomposition(SVD)and wavelet threshold.At the same time,an improved transform is proposed in the wavelet threshold denoising method.The three methods have different degrees of suppression of signal noise.Among them,the improved wavelet threshold denoising method has the best effect.It can greatly improve the signal-to-noise ratio(SNR)and reduce the mean square error of measured underwater acoustic signals in signal analysis.Finally,in the signal recognition stage,with the development of traditional machine learning,it is found that there have been many different types of improvements after investigation.Convolutional neural networks(CNN)are used to complete the classification from different types of time-frequency features of the measured underwater acoustic signals after denoising.Then we compare its performance with the results of the multi-step decision low frequency analysis recording(LOFAR)spectrum which is without denoising by using the accuracy and the confusion matrix.We find that the average recognition rate of the signal after denoising is 88.56%,while the recognition rate of the signal without denoising is 75.19%,which proves that denoising the signal is effective.At the same time,in view of the fact that there are fewer data samples for underwater environmental research,the method of generative adversarial networks(GAN)is used to increase the signal samples to improve the accuracy of deep learning,it is found that the final signal recognition rate reaches 96.673% after the experiments.Compared with the previous method based on the multi-step decision LOFAR spectrum which is without denoising preprocessing,the recognition rate of the signal can effectively be improve after the network added the data samples by generative adversarial networks.更多还原