Real-Time Safety Assessment for Dynamic Systems With Limited Memory and Annotations

Abstract

Real-time safety assessment of dynamic systems has recently received increasing attention. However, the performance of existing advanced approaches is often negatively affected by realistic requirements such as limited annotations and memory. In this case, how to design reasonable query strategies to select valuable instances and exploit the memory space efficiently is extremely meaningful. This paper proposes a novel memory-triggered submodularity-guided active broad learning approach, termed MTSGABL, to deal with such issues simultaneously. Specifically, the broad learning system is introduced as the basic assessment model to update incrementally. A memory-triggered learning mechanism is then proposed based on the drift detection procedure, which controls the update process to exploit the latest sequential information. Furthermore, a submodularity-guided query strategy is introduced to select a small number of valuable samples sequentially, which is beneficial to alleviate the negative effects of the imbalanced data stream. Numerous comparison and ablation experiments with the realistic JiaoLong deep-sea manned submersible data are conducted to validate its effectiveness. Results show that the proposed approach is superior to the existing typical approaches subject to these constraints.