巴黎萨克雷大学-曾志国副教授-系列报告

(6月25日-6月30日）

嘉宾简介

Zhiguo ZENG received the Ph.D. degree in reliability engineering from Beihang university in 2016. He is currently an associate professor at CentraleSupélec, Université Paris-Saclay, France. His research focuses on the characterization and modeling of the failure/repair/maintenance behavior of components, complex systems and their reliability, maintainability, prognostics, safety, vulnerability and security. Dr. ZENG is an author/co-author of more than 100 papers in highly recognized international journals and conferences (including 43 journal papers indexed in Web of Science). His research has been funded by important government funding agencies like ANR and ERC, and also important industrial companies like EDF, SNCF, Orange and GE Healthcare. He is editorial board member of International Journal of Data Analysis Techniques and Strategies, and the leading guest editor of the special issue on “Dependent failure modeling” of the journal Applied Science. He is the co-head of the master program “Risk, Resilience and Engineering Management” in Universite Paris Saclay.

【报告一】

报告题目：1.《牛奶可乐可靠性》                                                                                                     2.巴黎萨克雷大学硕士、博士项目宣讲

报告人：曾志国 副教授

时间：6月25日  9：00-12：00

地点：学院路校区为民楼320会议室

报告摘要：

为什么汽车的油门踏板是塑料的，但是刹车踏板是金属的？火灾疏散的时候，如果在门口摆上一个障碍物，是会加快还是降低疏散的效率呢？瑞士的法律规定，饲养荷兰猪作为宠物的时候，最少必须养一对，这又是为什么呢？大部分的空客飞机采用两个并联的攻角传感器，但是大部分的波音飞机采用的是三中取二的表决设计，哪一种设计更加可靠呢？这些来源于生活的有趣问题看似简单，但其实蕴含着不少与故障和可靠性相关的深刻道理。在这个报告中，我们将共同探讨一些类似的，来自生活的有趣例子，并从中归纳出一些关于故障与可靠性的基本原理。

本报告的题目来源于Robert Frank 教授的著作《牛奶可乐经济学》。自2021年以来，本人开始为法国中央理工大学三年级工程师学生开始课程《故障的原理》。在本课程的教学过程中，本人尝试应用Frank教授的教学哲学：用直观的语言，教授基本的原理，然后让学生从日常生活寻找有趣的问题，尝试应用所学的原理加以分析，从而加深学生的理解。本报告中所列举的大部分例子，即来自于选课学生提交的课堂作业，在此向做出贡献的学生们表示由衷的感谢！

【报告二】

报告题目：《Dependent failure behavior modeling in risk and reliability: A systematic and critical literature review》

报告人：曾志国 副教授

时间：6月26日  9：00-11：00

地点：学院路校区为民楼320会议室

报告摘要：

Traditional risk and reliability models are based on the assumption of independence among the components or failure processes. In most practical cases, however, the failure behaviors are dependent on one another. How to properly consider dependency, then, becomes an important question in risk and reliability research. In this talk, we present a systematic and critical literature review on dependent failure behavior modeling in risk and reliability. The main results of the critical literature review include: First, we summarize the dependent failure behavior in different system hierarchies, i.e., failure mechanism level, component level, system level and systems-of-systems level. In each level, the main dependent failure behavior from literature is discussed with examples. Second, we develop a classification framework for the dependent failure behavior models. Depending on whether the dependency mechanism is explicitly considered, we broadly classified the existing models into statistical dependency models and mechanistic dependency models. Statistical dependency models do not explicitly consider the dependency mechanisms but model them in terms of statistical association among the variables, and can be further divided into lifetime distribution models, system state models and degradation process models. Mechanistic dependency models consider the dependency mechanisms explicitly, and can be further divided into failure interaction models and failure propagation models. The most frequently-used models in each category are critically reviewed, based on which we identify five challenging problems the current dependency models face, give our perspectives on their possible solutions, and discuss future research opportunities.

【报告三】

报告题目：《Why PHM has H and M? And how to do that?》

报告人：曾志国 副教授

时间：6月27日  9：00-11：00

地点：学院路校区为民楼320会议室

报告摘要：

Prognostics and Health Management (PHM) has been a hot topic in academia and industry for many years. However, most of the research effort is on developing more accurate and efficient algorithms for remaining useful life (RUL) prediction. How to make use of the predicted RUL to support health management is not well studied as compared to the RUL prediction problem. One of the most important use case for health management is maintenance planning, i.e., making decisions regarding when to do replace a component/system preventively. In this talk, we show some of our recent works on this topic. First, we show how to make opportunistic maintenance decisions in a closed-loop supply chain from health care industry, based on historical. Then, we show another work on predictive maintenance, i.e., maintenance planning based on the predicted RUL. Both single-component-level and fleet-level planning are discussed.

【报告四】

报告题目：《Resilience modeling and optimization for interdependent critical infrastructures》

报告人：曾志国 副教授

时间：6月29日  9：00-11：00

地点：学院路校区为民楼320会议室

报告摘要：

In this talk, we first review different perspectives on resilience definition from literature. Then, we propose a generic framework based on semi-Markov reward process to quantify different aspects of resilience. Computational issues of evaluating the semi-Markov reward process will also be discussed. In the second part of this talk, we will discuss some of our recent works on optimal design of critical infrastructures based on resilience. More specifically, we consider two interdependent critical infrastructures, and propose an optimization model for designing the coupling interface between the two infrastructures. Different formulation of the optimization model, including stochastic programming, robust optimization and distributionally robust optimization will be discussed to consider the uncertainty in the optimization model.

【报告五】

报告题目：《Can a language model understands reliability of a product and its impacts on sales?》

报告人：曾志国 副教授

时间：6月30日  9：00-11：00

地点：学院路校区为民楼320会议室

报告摘要：

Lack of failure data is one of the major challenges of traditional reliability engineering. On the other hand, one could have abundance of texts in natural language that contains failure information, e.g, customer reviews, accident reports, maintenance records. In the first part of this talk, we explore the possibility of developing a natural language processing model to automatically extract failure-related information from text data, and use this information to assess the reliability of a product. We design an experiment to compared the developed machine learning models with the performance of human assessors. The results find that the developed machine learning model can achieve comparable results as human assessors. In the second part of this talk, we continue to investigate the impacts from different contributing factors, e.g., the reliability, price, brand on the sales of the products. A major challenge here is that we could only collect observational data that might contain some confounders. We apply a causal inference model to adjust the effect of confounders and estimate the true impacts on sales. The results show that if we directly develop a machine learning model, like a language model, without considering the effect of confounders, the impacts on sales could be highly biased.