Adversarial Weighted Active Domain Adaptation for Safety Assessment in Open Environments

Abstract

Ensuring the operational safety of complex systems often stands as a top priority, and employing data-driven safety assessment offers a promising way to achieve this goal. However, as systems often operate across various modes, models trained for one mode may not be applicable to others. Moreover, conducting the operational safety assessment task in open environments, where unknown scenarios can emerge unexpectedly, remains a challenging issue. Unsupervised domain adaptation allows for transferring models from a source domain with labeled data to a target domain with only unlabeled data. Yet, the effectiveness of such models diminishes when faced with unknown scenarios not observed in the source domain. Hence, this article introduces a novel problem termed open active domain adaptation for the safety assessment task, which addresses task-related unknown scenarios in the target domain and introduces a limited labeling budget to enhance model performance. To tackle this problem, an adversarial weighted active domain adaptation scheme is proposed, which incorporates an active labeling process and a weighting mechanism. This scheme leverages adversarial training in both the weighting mechanism and the domain adaptation process. Specifically, it identifies representative unlabeled data capable of approximating the target data distribution for label annotation. Furthermore, instance-level weights are generated for the target data based on an unknown separation module utilizing adversarial training, facilitating the adaptation to unknown scenarios and alleviating their adverse impacts on feature alignment. Experiments conducted on two bearing datasets illustrate the effectiveness and practicality of the proposed scheme.