Active Labeling Aided Semi-Supervised Safety Assessment With Task-Related Unknown Scenarios

Abstract

The open environment presents a challenging issue for the online safety assessment of dynamic systems, which means that unknown scenarios may arise unexpectedly. These unknown scenarios can be task-related and result in the within-class distribution mismatch. Addressing this challenge in the semi-supervised safety assessment task, where unlabeled training data contain task-related unknown scenarios, has not been explored. This article investigates this new semi-supervised safety assessment problem. A novel active-labeling-aided semi-supervised learning scheme is proposed to tackle the within-class distribution mismatch between labeled and unlabeled training data. The proposed scheme begins by detecting out-of-distribution unlabeled data through the construction of a deep support vector data description network for each class. Subsequently, an active labeling approach along with its kernel extension is introduced, taking into account both distribution mismatch degree and sample representativeness. The proposed active labeling approach can be seamlessly integrated with any semi-supervised learning algorithm to enhance its performance in handling task-related unknown scenarios. The effectiveness and applicability of the proposed method are demonstrated through case studies based on a bearing dataset and operation data from an actual deep-sea manned submersible.