Abstract

In seafloor mining environments, ultrasonic detection is often used for underwater target recognition, and a large number of suspended particles reduce the recognition rate of single-classifier cobalt-rich crusts. In this paper, we propose a recognition method based on multi-classifier fusion (MCF). First, the kernel Fisher discriminant analysis (KFDA) method is used to reduce the multi-class feature dimension of the signal, which improves the computational efficiency. Then, the probabilistic neural network (PNN), support vector data description (SVDD), and K-nearest neighbors (KNN) classifier are designed, and the input features of the different classifiers are selected by the genetic algorithm (GA) to improve the recognition rate of classifiers. Finally, the basic probability assignment (BPA) is recalculated in combination with the accuracy of the classifier, and conflicting evidence is synthesized to realize the MCF decision. The experimental results indicate that MCF recognition surpasses single classifier recognition. The Dempster-Shafer (D-S) theory MCF identification method proposed in this paper can be effectively applied to the identification of cobalt-rich crusts in seafloor mining environments.

摘要

在海底采矿环境中，超声波检测通常用于水下目标识别，并且大量的悬浮颗粒会降低单分类器富钴结壳的识别率。本文提出了一种基于多分类器融合的识别方法。首先，采用核Fisher判别分析（KFDA）方法来减小信号的多类特征维，从而提高了计算效率。然后，设计了概率神经网络（PNN），支持向量数据描述（SVDD）和K近邻（KNN）分类器，并通过遗传算法（GA）选择了不同分类器的输入特征，以改进分类器的识别率。最后，结合分类器的准确性重新计算基本概率分配（BPA），并综合矛盾的证据以实现MCF决策。实验结果表明，MCF识别优于单分类器识别。本文提出的DS理论MCF识别方法可以有效地应用于海底采矿环境中富钴结壳的识别。