基于堆栈稀疏去噪自编码器神经网络的舰船辐射噪声目标识别算法研究

A stack sparse denoising autoencoder-based neural network approach for ship radiated noise target recognition

传统的特征提取算法往往依赖于算法设计者的先验知识,没有突出大数据的优势,所以在实际应用中分类正确率较差且对于不同应用场景的泛化能力也明显不足。使用深度学习算法进行舰船辐射噪声的特征提取,利用了大量无类标数据,使用堆栈稀疏自编码器算法训练特征提取神经网络,并使用Softmax分类器算法利用有类标数据对特征提取神经网络进行参数微调。应用SSDAE-Softmax算法以及主成分分析算法、线性判别分析算法以及局部线性嵌入算法三类机器学习算法对海试数据进行处理,SSDAE-Softmax算法能够从舰船辐射噪声中提取更加具有区分度的特征,能够提升舰船辐射噪声的分类识别正确率,试验结果表明在低信噪比以及少量训练样本的应用场景下分类效果均高于其他三类算法。

The traditional feature extraction algorithm relies on the prior knowledge.Because it does not have the advantage of highlighting big data,the classification accuracy in practical application is poor and the generalization ability for different application scenarios is also obviously insufficient.In this paper,a deep learning algorithm was used for feature extraction of ship radiated noise,and a large number of classless data was fully utilized.The stack sparse self-encoder algorithm was to train the feature extraction neural network,and the Softmax classifier algorithm was used to fine-tune the parameters of the neural network by using class-based data.By comparing with the principal component analysis algorithm,the linear discriminant analysis algorithm,and the local linear embedding algorithm,it can be seen that the SSDAE-Softmax algorithm proposed in this paper can extract more discriminative features from ship radiated noise and improve the classification and recognition accuracy to some extent.