Bi-Sphere Anomaly Detection With Learnable Centroid for Active Sonar Classification

Abstract

Machine learning (ML)-based approaches are desirable for discriminating targets from clutter signals to enhance the performance of active sonar systems. However, a small dataset and imbalanced data samples between the target and clutter hinder ML applications in active sonar classification. Anomaly detection (AD), which effectively exploits the imbalance, is adopted to enhance the generalization of ML-based active sonar classifiers for small and imbalanced datasets. Generally, deep AD focuses on learning a representation of normal data samples (clutter) and finding a sphere embracing normal data samples in latent space. However, abnormal samples from artificial objects (underwater targets) have similar physical experiences as normal clutter samples from geological and biological scattering objects. Therefore, it is difficult to discriminate between the target and the clutter using conventional deep AD. To overcome the problem of active sonar classification, we propose semi-supervised learning-based bi-sphere anomaly detection (BiSAD) to find two spheres, embracing target and clutter samples each, by modifying conventional deep AD. Simultaneously, BiSAD searches for the latent space where two sphere centroids locate distantly to promote generalization. In the generalization test, the receiver operating characteristic (ROC) curve of BiSAD indicates a detection probability of 0.8 at a false alarm rate of 0.01, and the area under the ROC curve was 0.989, which was superior to the conventional deep AD and supervised learning-based approaches.