[ ]水下大型航行体远场低频声学特征提取鉴别方法综述

Review of Far-field Low-frequency Acoustic Feature Extraction and Identification Methods for Underwater Large Vehicles

摘要:

针对水下大型航行体的小样本远场低频特征提取与鉴别问题，本文从三个方面综述了目前国内外小样本低频特征提取与鉴别的传统方法和智能方法。时频域单独、时频域结合和视听感官特征提取的传统方法需要一定的先验知识与假设，易受环境干扰；专家系统、统计类方法和BP神经网络等早期的智能方法存在可移植性差、学习能力差、上限低、梯度消失等问题；深度置信网络（DBN）、卷积神经网络（CNN）、循环神经网络（RNN）、自编码器（AE）神经网络、生成对抗网络（GAN）、迁移学习深度网络等深度智能方法对先验知识依赖度低，可以提取深度不变特征，较其他方法更稳定，识别精度更高，但是也存在依赖数据量、可解释性不足的缺点。未来，传统方法与智能方法更深度的互补融合有望减少当前方法对数据量的依赖，提高深度特征的可解释性。

Abstract:

Aiming at the problem of small-sample far-field low-frequency feature extraction and identification of large underwater vehicles, this paper reviews the traditional methods and intelligent methods of small-sample low-frequency feature extraction and identification at home and abroad from three aspects. The traditional methods of time-frequency domain alone, time-frequency domain combination and audiovisual sensory feature extraction require certain prior knowledge and assumptions, and are susceptible to environmental interference. Early intelligent methods such as expert systems, statistical methods and BP neural networks have problems such as poor portability, poor learning ability, low upper limit, and gradient vanishing. Deep intelligence methods such as deep confidence network (DBN), convolutional neural network (CNN), recurrent neural network (RNN), autoencoder (AE) neural network, generative adversarial network (GAN), and transfer learning deep network have low dependence on prior knowledge, can extract deep invariant features, are more stable than other methods, and have higher recognition accuracy, but also have the disadvantages of relying on the amount of data and insufficient interpretability. In the future, the deeper complementary integration of traditional methods and intelligent methods is expected to reduce the dependence of current methods on data volume and improve the interpretability of deep features.