试验环境水下声信号的特征提取方法

Feature extraction method for underwater acoustic signalsin an experimental environment

水下试验环境参数的反演是水声学研究领域的重要内容。而当前研究的关键是通过对水下声信号做特征提取从而获取参数信息。针对特征提取较难、模型很难拟合等问题。本文提出了一种试验环境水下声信号的特征提取方法。将水下声信号同时用梅尔频谱倒谱系数及线性预测系数处理,两者运用特征加权组合方法得到新的特征矩阵;再应用映射插值算法对特征矩阵进行处理,获得适应神经网络输入的三通道矩阵。本文选取的网络模型为残差神经网络。利用实验室所录制的对河口水库数据集测试表明,本文提出的特征提取方法普遍优于仅利用梅尔频谱倒谱系数或线性预测系数的特征处理方法。利用单频矩形脉冲信号对环境进行深度5分类,准确率平均提升2%。利用线性调频信号对环境进行深度5分类,准确率平均提升2.03%。本文提出的特征提取方法对线性调频信号在深度分类任务下处理的结果要优于单频矩形脉冲信号处理的结果。

The inversion of environmental parameters in underwater tests is an important topic in the field of hydroacoustics.With the breakthroughs in deep-learning technologies in many research fields,such technologies can be applied to environment parameters for underwater testing.The key to the current research is how to obtain parameter information by extracting the features of underwater acoustic signals.With the aim of addressing the difficult problems of feature extraction and model fitting,this paper proposes a feature extraction method for underwater acoustic signals in an experimental environment.The underwater acoustic signals are processed simultaneously by the Mel-frequency cepstral coefficient(MFCC)and the linear prediction coefficient(LPC),to which the feature-weighted combination method is applied to obtain a new feature matrix.Furthermore,the mapping interpolation algorithm is used to process the feature matrix,thereby obtaining a three-channel matrix that adapts to the input of the neural network.The residual neural network(ResNet)is also selected in this paper.Training with the laboratory recorded Duihekou reservoir dataset reveals that the proposed feature extraction method is generally better than the feature processing method that only uses MFCC or LPC.In addition,accuracy is increased by 2%on average by using single-frequency rectangular pulse signals(i.e.,continuous wave(CW)signals)to classify the environment in depth 5.At the same time,the accuracy is increased by 2.03%on average by using linear frequency modulation(LFM)signal to classify the environment in depth 5.The results treated by the proposed feature extraction method are better than those obtained by CW signals when applied to the depth classification task.