传统的水声信号识别方法是将特征提取和分类识别分开进行处理的，影响了水声信号识别的整体性能。本文根据水声信号的特点，结合一维卷积网络（1DCNN）的卷积运算、时间平移不变性和门控循环网络(GRU)内部充分考虑时序相关性的记忆能力等优势，将一维卷积网络和门控循环网络进行串联中并对网络参数和模型结构进行优化，自适应提取特征给出分类结果，并与单独使用1DCNN和GRU网络模型的分类性能进行对比。结果表明，本文提出的网络对水声信号的识别准确率最高。
The traditional underwater acoustic signal recognition method deals with feature extraction and classification separately, the overall performance of underwater acoustic signal recognition is affected. In this paper, based on the characteristics of underwater acoustic signal, we combine the advantages of convolution operation, time-shift invariance of one-dimensional convolutional network, and memory ability of gated loop network that fully consider temporal correlation, we connect the one-dimensional convolutional network and the gated loop network in series, and optimize the network parameters and model structure, adaptive feature extraction and give classification. Compared with the classification performance of 1DCNN and GRU network model alone, the result shows that the proposed network has the highest recognition accuracy for underwater acoustic signal.
2023,45(15): 107-110 收稿日期：2023-02-23
DOI：10.3404/j.issn.1672-7649.2023.15.020
分类号：TN911
作者简介：杜柏润(2002-),男,研究方向为机器学习应用
参考文献：
[1] 伊恩. 古德费洛, 约书亚. 本吉奥, 亚伦. 库维尔. 深度学习[M]. 北京: 人民邮电出版社, 2018.
[2] 阿斯顿. 张, 李沐, 扎卡里. C. 立顿, 亚历山大. J. 斯莫拉. 动手学深度学习[M]. 北京: 人民邮电出版社, 2019.
[3] 惠俊英, 生雪莉. 水下声信道[M]. 哈尔滨: 哈尔滨工程大学出版社, 2011.
[4] ABDOLI S, CARDINAL P, LAMEIRAS KOERICH A. End-to-end environmental sound classification using a 1D convolutional neural network[R]. arXiv, 2019.
[5] 王强, 曾向阳. 深度学习方法及其在水下目标识别中的应用[J]. 声学技术, 2022, 44(1): 136–141
[6] HUGHES Tyler W, WILLIAMSON Ia A. D, MOMCHIL M, et al. Wave physics as an analog recurrent neural network[R]. Science Advances, 2019.
[7] 权稳稳, 林明星. CNN特征与BOF相融合的水下目标识别算法[J]. 山东大学学报, 2019, 49(1): 107–113
