基于深度流形学习的水中目标声信号特征提取

Feature extraction of underwater target acoustic signals based on deep manifold learning

随着海洋环境日益复杂,水中目标声信号观测数据呈现出高维数、非线性、非结构化等特点,无疑给水中目标声信号特征提取带来严峻挑战,该文提出了一种基于流形自编码器的水中目标声信号特征提取方法。首先,通过自编码器重建误差对原始数据进行全局优化,找出潜在的低维表示结果;然后,利用流形学习保持近邻重构权值的思想对潜在表示实施局部约束,保留其内在拓扑结构;最后,引入生成对抗网络架构进行正则化处理,使潜在表示服从特定分布,从而实现一种局部与全局的联合保持低维嵌入方法。在DeepShip深水船公开数据集上进行试验,使用该文方法提取4种深水船数据的本质特征,为评估该类特征的质量水平,利用经典分类器支持向量机进行分类识别,与现有深度学习以及流形学习特征提取方法对比,识别精度平均提高14.96%。

With increasingly complex marine environment,observation data of underwater target acoustic signals exhibit characteristics of high dimensionality,nonlinearity and unstructured to undoubtedly bring serious challenges for feature extraction of target acoustic signals.Here,a method for feature extraction of underwater target acoustic signals was proposed based on manifold autoencoder.Firstly,global optimization was performed for the original data by using autoencoder to reconstruct error,and seek out potential low dimensional representation results.Then,the idea of preserving neighboring reconstruction weights with manifold learning was employed to enforce local constraints on latent representation,and preserve its inherent topological structure.Finally,a generative adversarial network architecture was introduced for regularization processing to make latent representation obey a specific distribution,and realize a local and global joint preserving low dimensional embedding method.Experiments were conducted on DeepShip public dataset,the proposed method was used to extract essential features of 4 types of deepwater ships’data.To evaluate the quality level of these features,the classic classifier SVM was used for classification and recognition.It was shown that compared with existing deep learning and manifold learning feature extraction methods,the recognition accuracy of the proposed method increases by an average of 14.96%.