Model-Based and Data-Driven Contributions to Fault Detection
From Reaction Wheel Monitoring to Statistical Analysis
Time: Thu 2026-09-24 10.00
Location: Harry Nyquist, Malvinas Väg 10, SE-100 44
Language: English
Subject area: Electrical Engineering
Doctoral student: Alejandro Penacho Riveiros , Digital futures, Reglerteknik
Opponent: Associate Professor Francesca Boem, University College London
Supervisor: Professor Karl H. Johansson, Reglerteknik, Integrated Transport Research Lab, ITRL, Digital futures; Assistant Professor Matthieu Barreau, Digital futures, Reglerteknik
QC 20260901
Abstract
Ensuring the reliability of modern engineering systems requires the timely detection of abnormal behavior, as it allows prompt corrective action.Algorithms capable of detecting such behavior are therefore essential for the safe operation of wind turbines, aircraft and nuclear power plants, among other applications. The development of such algorithms is the focus of the field of \gls{fdi}. In this licentiate thesis, we cover some of its applied and theoretical aspects through three contributions.
In the first contribution, we develop an algorithm to detect abnormal friction behavior in \glspl{rwa}, actuators mounted on satellites that are critical for attitude control. The problem poses two challenges: normal and abnormal \gls{rwa} behavior are characterized only by a small labeled dataset, and \glspl{rwa} exhibit switching dynamics. The proposed algorithm addresses these issues by combining machine learning with a hidden semi-Markov model to capture the switching behavior.
The second contribution addresses the use of statistical testing for \gls{fdi} in a setting where the system's nominal behavior is known only from data. By exploiting the probabilistic foundation of Gaussian processes, we develop an algorithm that extends the notion of a $z$-testing to data-driven monitoring of dynamical systems. This algorithm determines whether a system's behavior is consistent with nominal data while maintaining a prescribed false-alarm rate, and has potential applications in detecting unmodeled changes in dynamical systems.
The third contribution tackles the question of how long it takes to determine whether a specific fault has occurred with a given confidence level. Using concepts from information theory, we provide a metric that quantifies how the probability of correctly detecting a fault in a binary hypothesis-testing scenario evolves over time. This metric can be used to guide the design of critical systems that require reliable and timely monitoring.
With these contributions, the thesis provides important insights into how models, data and statistical inference can be combined to design and analyze novel diagnosis algorithms.