基于多类别特征融合的水声目标噪声识别分类技术

Underwater Target Noise Recognition and Classification Technology based on Multi-Classes Feature Fusion

目标噪声信号作为当前水声目标识别的主要信号源之一,由于目标信号来源单一,难以像多传感器探测不同角度表征目标特性,导致目标识别分类正确率低、虚警率高,严重制约水声探测系统功能的发挥。针对这一问题,采用长短时记忆网络,建立多层LSTM水声目标噪声特征提取模型,学习提取目标噪声时域包络、DEMON线谱、梅尔倒谱系数等信息特征,构建多类别特征子集;在此基础之上,建立了基于多类别特征子集的特征级融合识别分类模型和基于D-S证据理论的决策级融合识别分类模型;利用样本库数据对上述模型进行了测试,对比多类别特征融合判别与单一类别特征识别分类的差异,并使用港池相关试验数据对上述模型进行了测试验证。测试结果表明,提出的基于多类别特征融合的水声目标噪声智能识别分类方法判别效果更好,对水声目标噪声信号识别分类的正确率和虚警率等相关指标均优于单一类别特征判别方法。

As one of the main signal sources of underwater acoustic target recognition,the target noise signal is difficult to characterize the characteristics of the target from clearly comparing with the multi-sensor detection technology,which may lead to lower recognition rate and higher false alarm rate and seriously restricts the function of underwater acoustic detection system.In order to solve this problem,a multi-layers LSTM underwater acoustic target noise feature extraction model is established by using the long short term memory network.The information features such as time-domain envelope of target noise,Demon line spectrum and Mel frequency cepstrum coefficient are extracted,and a subset of multi-classes features is constructed.On this basis,the feature level fusion recognition and classification model based on the multi-classes features subset and the decision level fusion recognition and classification model based on the D-S evidence theory are established,and the above-mentioned models are tested by using the sample database.The difference of classification result between the multi-classes feature fusion and the single class feature recognition classification is compared,and the above model is tested and verified by using the relevant test data of port basin verification experiment.The correlation results show that the present intelligent recognition and classification method of underwater target noise based on the multi-classes feature fusion is more robust,and the recognition rate and false alarm rate of underwater target are better than those of single category feature discrimination method.