PhD Oral Exam - Saba Sanami, Electrical and Computer Engineering

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

This thesis develops frameworks for reliable anomaly detection and fault diagnosis in multivariate time-series data. The research addresses challenges that limit the use of data-driven monitoring methods, including the scarcity of labelled anomalies, dependencies among sensor channels, contamination of nominal training data, variation in operating conditions, and the limited interpretability of purely data-driven diagnostic outputs. The methodological development begins with unsupervised representation learning based on Fisher autoencoders. By imposing structure on the latent representation of normal behaviour, the proposed approach improves the separation between normal and abnormal patterns. On the other hand, an informative latent representation does not by itself provide a reliable decision boundary when labelled anomalies are unavailable. This limitation is addressed through a framework that combines synthetic anomalies and uncertainty information to refine the separation between normal and abnormal regions. The analysis is subsequently extended from representation learning to the complete anomaly detection pipeline. In this setting, reconstruction models, anomaly scoring, and threshold selection methods are examined jointly to determine how their interaction affects detection reliability. The research then progresses from detecting abnormal behaviour to identifying the underlying fault condition. To support this transition, physical knowledge is incorporated through an inverse physics-informed neural network that estimates interpretable dynamic parameters from time-series measurements. These parameters are subsequently used as classification features. The proposed frameworks are evaluated on benchmark datasets selected according to the methodological requirements of each study, including the availability of suitable anomaly annotations or an estimated physical model.