基于弱监督学习的不精确标注水下声音事件识别方法研究

Automatic recognition of underwater acoustic events has great significance for marine development and national defense safety, which has always been a hot research topic for many scholars at home and abroad. The problems such as the shortage of underwater acoustic data and the difficulty of annotation have severely limited the development of underwater acoustic event recognition. This project focuses on the important issue of underwater acoustic event recognition under the condition of inexactly labeled underwater acoustic dataset. Through establishing the experimental dataset, simulating rare data samples and designing recognition methods，this project aims at solving the basic problems such as the inexactly labeling of underwater acoustic data, data imbalance, multi-scale features and the design and training of deep learning models. Weakly supervised learning methods are employed to improve the accuracy and reliability of underwater acoustic event recognition and provide technical support for the further research on the automatic recognition of underwater targets and the development of intelligent underwater acoustic equipments.

水下声音事件自动识别对于海洋开发、国防安全有着重大意义，一直是国内外众多学者研究的热点问题。水声数据稀缺和标注困难等问题严重制约了水下声音事件识别研究的发展，本项目针对不精确标注水声数据条件下水下声音事件识别这一重要课题，采用建立实验数据集、仿真稀有数据样本与设计识别算法相结合的方法，重点围绕所涉及的数据不精确标注、数据不平衡、特征多尺度以及深度学习模型设计和训练问题展开研究，利用弱监督学习的有关思想提高水下声音事件识别的精度和可靠性，为水下目标自动识别以及水声装备智能化的深入研究提供技术支撑。

以复杂多变海洋环境条件下水下声音事件自动识别鲁棒性差的问题为切入点，本项目综合运用理论分析、建模仿真等方法，针对水声目标数据样本少、无精确标注和数据类别不平衡等问题，着力开展不精确标注水声数据条件下的水下声音事件识别研究，利用弱监督深度学习方法着力提高水下声音事件识别的精度和可靠性。聚焦面向声音事件特征多尺度分析的深度学习模型结构设计、基于不精确标注数据的声音事件识别弱监督学习算法、水下声音事件数据增强算法、基于人工合成方法的水下声音事件数据生成方法等四个方面研究内容。设计了具有不同时间尺度特征分析能力的多尺度多层级卷积神经网络架构，通过典型算例结果对比验证了模型架构的有效性；融合循环卷积神经网络、Mean Teacher协同训练策略、注意力机制等多种先进的模型结构和算法理念，研究新型水下声音事件的弱监督学习方法，提升了水下声音事件的弱监督学习效能；实践了Mixup策略等数据增强算法，进一步提高不精确标注数据条件下的声音事件识别模型的泛化能力；探究了利用声景合成等人工合成方法自主生成有标注水下声音事件数据的实现途径，为持续构建水下声音事件智能化识别数据集提供支撑。本项目的相关成果有望应用于未来海洋预警监测体系中声纳系统优化设计与水声目标识别系统开发构建。