一种基于VGGish神经网络的水声目标识别方法

水声目标智能识别是水声装备智能化的重要组成部分，深度学习则是实现水声目标智能识别的重要技术手段之一。当前水声目标智能识别经常面临数据集较小带来的训练样本量不足的情况，针对小数据集识别中存在的因过拟合导致模型泛化能力不足，以及输入的水声信号二维谱图样式不统一的问题，文章提出了一种基于VGGish神经网络模型的水声目标识别方法。该方法以VGGish网络作为特征提取器，并在VGGish网络前部加入了信号预处理模块，同时设计了一种基于传统机器学习算法的联合分类器，通过以上措施解决了过拟合问题和二维谱图样式不统一问题。实验结果显示，该方法应用在ShipsEar数据集上得到了94.397%的识别准确率，高于传统预训练-微调法得到的最高90.977%的准确率，并且在相同条件下该方法的模型训练耗时仅为传统预训练-微调方法的0.5%左右，有效提高了识别准确率和模型训练速度。

A target recognition method of underwater acoustic signal based on VGGish neural network model

Underwater acoustic target intelligent recognition is one of the important parts of intelligent underwater acoustic equipment, and deep learning is one of the important technical means to realize underwater acoustic target intelligent recognition. At present, underwater acoustic target intelligent recognition often faces the problem of insufficient training sample size caused by small data sets. Aiming at the problem of insufficient model generalization ability caused by over fitting in small data set recognition and the problem that the pattern of the two-dimensional spectrums of the input underwater acoustic signals are not unified, a target recognition method of underwater acoustic signal based on VGGish neural network model is proposed in this paper. This method takes VGGish network as the feature extractor, adds the signal preprocessing module in front of VGGish network, and designs a joint classifier based on traditional machine learning algorithm. Through the above measures, the problems of over fitting and the inconsistency of the patterns of the two-dimensional spectrums are solved. The experimental results show that this method achieves 94.397% recognition accuracy on Shipsear dataset, which is higher than the best accuracy of 90.977%achieved by traditional transfer learning method on this dataset. In the same operating environment, the model training time of this method is only about 0.5% of that of the traditional pre-training & fine-tuning method, which effectively improves the recognition accuracy and model training speed.