A set based probabilistic approach to threshold design for optimal fault detection

V. Rostampour Samarin, Riccardo Ferrari, Tamas Keviczky

Research output: Chapter in Book/Conference proceedings/Edited volumeConference contributionScientificpeer-review

7 Citations (Scopus)
91 Downloads (Pure)


Traditional deterministic robust fault detection threshold designs, such as the norm-based or limit-checking method, are plagued by high conservativeness, which leads to poor fault detection performance. On one side they are ill-suited at tightly bounding the healthy residuals of uncertain nonlinear systems, as such residuals can take values in arbitrarily shaped, possibly non-convex regions. On the other hand, they must be robust even to worst-case, rare values of the modeling and measurement uncertainties. In order to maximize performance of detection, we propose two innovative ideas. First, we introduce threshold sets, parametrized in a way to bound arbitrarily well the residuals produced in healthy condition by an observer-based residual generator. Secondly, we formulate a chance-constrained cascade optimization problem to determine such a set, leading to optimal detection performance of a given class of faults, while guaranteeing robustness in a probabilistic sense. We then provide a computationally tractable framework by using randomization techniques, and a simulation analysis where a well-known three-tank benchmark system is considered.

Original languageEnglish
Title of host publicationProceedings of the 2017 American Control Conference (ACC 2017)
EditorsJ. Sun, Z.-P. Jiang
Place of PublicationPiscataway, NJ, USA
ISBN (Electronic)978-1-5090-5992-8
Publication statusPublished - 2017
Event2017 American Control Conference, ACC 2017 - Seattle, United States
Duration: 24 May 201726 May 2017


Conference2017 American Control Conference, ACC 2017
Abbreviated titleACC 2017
CountryUnited States

Bibliographical note

Accepted Author Manuscript


Dive into the research topics of 'A set based probabilistic approach to threshold design for optimal fault detection'. Together they form a unique fingerprint.

Cite this