基于流形学习的单类分类算法及其在不均衡声目标识别中的应用

针对数据不均衡条件下的目标识别性能下降问题,首先讨论了目标声信号所包含低维流形的特点,在此基础上设计了基于流形学习的单类分类算法,通过比较测试样本与正类样本在流形上的符合程度判决其是否属于正类。将此分类算法应用于包含不均衡数据的声目标识别,三组不同环境和识别目标的实验数据集测试结果显示该算法可以有效地从多种目标中识别特定类别目标,与其他单类分类算法相比,提高数据不均衡条件下的识别性能,并对样本的混叠分布具有较好的鲁棒性。

One-class classification algorithm based on manifold learning and its application to imbalanced acoustic target recognition

Imbalanced data is one of the aspects that influence the performance of classification algorithm.Low- dimensional manifold embedded in acoustic signal spectrum was explored and a one-class classification algorithm was proposed based on manifold learning.This one-class classification algorithm recognizes the positive class target according to the error between the manifolds of the input sample and the positive class.This method was applied to acoustic target recognition problem with imbalanced data to verify its effectiveness.The experimental results show that,in comparison with other three one-class classification algorithms,this method can recognize the special target from multiple targets and achieve better recognition performance in the imbalanced data problem,and is more robust to the overlapping between the classes.