PhD Oral Exam - Maryam Mirnaghi, Building 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 presents a data-driven framework for fault detection and cross-loop fault co-occurrence analysis in HVAC chiller plant systems, combining Matrix Profile (MP) anomaly detection with Association Rule Mining (ARM). The methodology is validated on a Modelica-based Integrated Primary-Secondary chiller plant model with Water-Side Economizer, simulated under Paris-Orly weather for a full year using the Building Optimization Performance Test (BOPTEST) framework. A segment-based fault injection approach introduces simultaneous fault episodes preserving realistic controller dynamics. The multivariate Matrix Profile achieves F1 = 0.869 with oracle thresholding and F1 = 0.719 with label-free Peaks-Over-Threshold (POT) from Extreme Value Theory, using a separate healthy simulation (S00) as reference. Same-period counterfactual analysis shows that compound fault thermal impact is operating-mode dependent: winter/transition episodes produce +0.5 to +1.0 K condenser water temperature excursions consistent with parasitic-flow effects during valve cycling. A fault intensity binning scheme maps continuous MP scores to discrete severity levels per hydraulic loop, enabling ARM analysis that reveals strong cross-loop co-occurrence patterns (lift up to 241.5, interpreted as statistical co-occurrence ratio, not physical amplification). The condenser loop is identified as a dataset-specific cross-loop co-indicator under certain operating modes (confidence = 1.0 in this simulated dataset), while the bypass loop acts as a latent co-factor: weakly observable in isolation but statistically associated with elevated severity in other loops (lift = 35.3). Temporal ordering and Granger analysis provide evidence of directional predictability consistent with thermal coupling (primary-to-condenser), though this establishes predictive precedence rather than physical causality. The framework is further validated on real building data from the HIKARI positive-energy district cooling system, where MP analysis on binary and continuous valve signals confirms system-level mode-switching faults. Event-based evaluation shows zero isolated false alarm events across the seven primary scenarios (event precision = 1.000, event recall = 0.939), with mean detection delay of 4.1 hours after fault onset in batch processing mode.