“数字+”与统计数据工程系列讲座（六十八）10月14日悉尼大学孙秋壮来中心讲座预告

讲座题目：Optimal Abort Policy for Mission-Critical Systems under Imperfect Condition Monitoring

主讲人：孙秋壮

讲座时间：2024年10月14日 15:30-16:30

讲座地点：综合楼615会议室

主讲人简介：

孙秋壮博士于上海交通大学取得工业工程和计算机科学双学士学位，于新加坡国立大学取得工业系统工程博士学位。他现为悉尼大学数学统计学院助理教授。主要研究方向包括可靠性安全工程，数据驱动决策问题，工业统计等。曾获国际运筹与管理科学协会（INFORMS）、工业系统工程协会（IISE）、IEEE等知名学术协会的多项最佳论文、热点文章、以及数据分析竞赛奖项，如2023年INFORMS QSR会议最佳论文奖。

讲座摘要：

Mission-critical systems, such as unmanned aerial vehicles (UAVs), chemical reactors, and paper mills, are required to operate continuously for a period to complete a mission. During operation, unexpected system failure is possible. When there is a sign of imminent failures, the mission can be aborted to increase the system survival probability and minimize damage to the system. We consider the abort decision-making problem for systems with three states, i.e., healthy, defective, and failure, where the healthy and defective states are unobservable before failure. Condition-monitoring sensors are installed on the system and periodically generate signals indicating the system health state. Mission abort can be made based on these signals. However, possible measurement errors and environmental noises result in imperfect sensitivity and specificity of the sensors, which significantly complicates the decision-making. Furthermore, the random time from system defect to failure typically has an increasing hazard rate, leading to a non-Markovian transition between the defective and failure state. This study meets these challenges by adopting the Erlang mixture distributions to approximate the non-Markovian failure process as a continuous-time Markov chain in a new state space. A partially observable Markov decision process (POMDP) is then formulated for decision-making. We show that the optimal policy of the POMDP follows a control-limit structure in a spherical coordinate system, and the sequence of optimal policies converges to the true optimal policy for the original problem when the phase of the Erlang mixture goes to infinity. A modified point-based value iteration algorithm is developed to deal with the curse of dimensionality. We further investigate two special cases of our model that can be exactly solved after discretizing the state space. Through a case study on a UAV, we demonstrate the capability of real-time implementation of our model, even when the condition-monitoring signals are generated with high frequency.