Active Domain Adaptation for Safety Assessment:An Improved Energy-based Model
Online safety assessment of complex dynamic systems during operation is paramount and challenging.A large amount of labeled data is necessary to construct an effective data-driven model,which is difficult to obtain in practice.Furthermore,the safety assessment model should have a good generalization ability given the varying op-eration modes.Domain adaptation(DA)can transfer the model trained on a source domain with abundant labeled data to a target domain that has a different but similar data distribution.However,the task-related unknown scen-arios that have not appeared in the source domain will degrade the model performance,which remains an unsolved challenge at present.Active domain adaptation provides a potential solution to the aforementioned challenge by combining domain adaptation with active learning techniques.This paper investigates the problem of active do-main adaptation for safety assessment,specifically addressing task-related unknown scenarios within the target do-main.An active domain adaptation method with the improved energy-based model is proposed,and the out-of-dis-tribution detector is incorporated in the proposed method.On this basis,representative unlabeled samples from the target domain are actively selected for annotation,which are then used as training data to enhance the perform-ance of the domain adaptation model.At last,a case based on the bearing data is studied to demonstrate the effect-iveness and applicability of the proposed method.
万方数据
维普
