基于改进能量模型的主动域自适应安全性评估方法

Active Domain Adaptation for Safety Assessment:An Improved Energy-based Model

复杂动态系统运行过程中的在线安全性评估至关重要且富有挑战性.构建有效的数据驱动模型需要大量有标注数据,但这在实际中通常难以获得.此外,考虑到系统不同的运行工况,安全性评估模型应该具有良好的泛化能力.域自适应(Domain adaptation,DA)可以将模型从数据标注丰富的源域迁移到具有不同但相似数据分布的目标域.然而,源域中没有出现过的任务相关未知情景会降低模型的性能,是目前尚未解决的挑战.主动域自适应通过结合域自适应与主动学习技术,为解决上述挑战提供了思路.本文研究目标域存在任务相关未知情景的主动域自适应安全性评估问题,提出一种基于改进能量模型的主动域自适应方法.在所提方法中融合分布外检测器,在此基础上主动选择目标域中具有代表性的无标注样本进行标注,作为训练数据以提高域自适应模型的性能.最后,通过基于轴承数据的案例研究,验证所提方法的有效性和适用性.

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 operation 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 scenarios 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 domain adaptation for safety assessment,specifically addressing task-related unknown scenarios within the target domain.An active domain adaptation method with the improved energy-based model is proposed,and the out-of-distribution 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 performance of the domain adaptation model.At last,a case based on the bearing data is studied to demonstrate the effectiveness and applicability of the proposed method.