AUTOMATED FAULT DETECTION AND TIME-TO-FAILURE PREDICTION OF INDUSTRIAL SYSTEMS

<p dir="ltr">This thesis presents a comprehensive, data-driven framework for fault detection and time-to-failure prediction in HVAC systems, with a specific focus on AHUs and zone-level temperature behavior. The study begins with a system-level characterization, detailing the control sequences and fault-prone elements within various AHU configurations. Emphasis is placed on diagnosing temperature faults in different thermal zones, conditioned by an AHU in a campus lecture hall. Operational data, collected at 5-minute intervals over one month, undergoes rigorous preprocessing, including missing value imputation, normalization, and removal of redundant features. Statistical analyses identify sensor interdependencies and redundant signal groups. Feature reduction is implemented using PCA, SPCA, PLS-R, LDA, TAA, and DCCA, with each technique evaluated based on variance retention, classification separability, and interpretability. A suite of nine deep learning models: LSTM, GRU, BiLSTM, CNN, DNN, TCN, CNN-BiLSTM, Transformer, and an ensemble method is trained for binary fault classification using reduced feature spaces. Models are validated via stratified k-fold cross-validation and evaluated using accuracy, F1-score, ROC-AUC, PR-AUC, and Brier Skill Score. The zone with localized control dominated by its reheat valve demonstrates strong classification performance, achieving ≥95% accuracy and maintaining stable time-to-failure prediction up to 60 minutes prior to fault occurrence. In contrast, the zone characterized by distributed control dependencies and transient thermal behavior requires deeper temporal models, such as BiLSTM and TCN, to achieve robust detection and calibration under conditions of poor synchronization control and noisy occupancy patterns. Calibration diagnostics indicate that zones with more balanced class distributions provide more reliable probability estimates compared to imbalanced cases. Time-to-failure prediction further confirms that SPCA-reduced features preserve early detection capacity, while the unique degradation patterns of each zone strongly influence how far in advance faults can be accurately predicted. The study concludes that HVAC fault diagnostics must be tailored to zone-level control architecture. Univariate, interpretable models are suitable for isolated control zones, while multivariate, sequence-aware models are essential for zones with systemic coupling. The methodology enables accurate, calibrated, and generalizable fault detection suitable for complex HVAC and industrial systems.</p>