Maryam Ahang

Topic

Intelligent Condition Monitoring of Industrial Plants: Managing Uncertainty

Department of Electrical and Computer Engineering

Date & location

• Monday, September 14, 2026

• 7:00 A.M.

• Virtual Defence

Reviewers

Supervisory Committee

• Dr. Homayoun Najjaran, Department of Electrical and Computer Engineering, University of Victoria (Supervisor)

• Dr. Hong-Chuan Yang, Department of Electrical and Computer Engineering, UVic (Member)

• Dr. Keivan Ahmadi, Department of Mechanical Engineering, UVic (Outside Member) 

External Examiner

• Dr. Farnoosh Naderkhani, Cybersecurity and Intelligent Systems Engineering, Concordia University 

Chair of Oral Examination

• Dr. Jordan Stanger-Ross, Department of History, UVic

   

Abstract

Reliable fault detection and diagnosis in industrial systems is complicated by a persistent gap between the idealized conditions assumed by most machine learning methods and the realities of industrial deployment: fault samples are scarce, operating conditions vary, new failure modes emerge, and the consequences of an erroneous or overconfident diagnostic decision can be severe. This thesis addresses these challenges through four interconnected contributions that collectively advance the accuracy, adaptability, and trustworthiness of intelligent condition monitoring. The first contribution is a structured survey of machine learning and deep learning methods for fault detection and diagnosis, with particular emphasis on chemical process systems and the Tennessee Eastman Process (TEP) as a unifying benchmark. The survey synthesizes findings from over 460 publications, systematically reviewing autoencoders, convolutional and recurrent networks, deep belief networks, generative adversarial networks, and attention-based architectures, while also examining strategies for managing class imbalance, unlabeled data, and unseen fault modes. The second contribution addresses data scarcity across operating conditions. A novel signal-to-signal translation framework, designated N2FGAN (Normal-to-Fault Gener ative Adversarial Network), is developed by adapting a conditional GAN architecture to the one-dimensional vibration domain. N2FGAN is trained on paired normal and fault signals from a reference operating condition and learns to synthesize realistic fault signals for new conditions where only normal data are available. Validated on the Case Western Reserve University bearing dataset across multiple speeds, N2FGAN consistently outperforms standard CGAN, WGAN-GP, and classical augmentation baselines in downstream fault classification accuracy. The third contribution targets scenarios in which severe fault data are entirely absent from training. A variational autoencoder (VAE) is trained on normal and mildly degraded data and used to de fine a distance-based health index in the resulting latent space. This index tracks continuous system degradation and, through threshold-based classification, detects previously unseen severe faults with 99.51% accuracy on the IMS run-to-failure bearing benchmark, while the computed health indices closely follow established models of bearing wear evolution. The framework is further validated on a real industrial heat exchanger, achieving 93.33% fault detection accuracy and demonstrating generalizability beyond rotating machinery. 

The fourth contribution develops a hybrid, uncertainty-aware framework that integrates primary sensor measurements, lagged temporal features, and physics-informed residuals derived from lightweight nominal surrogate models. Two fusion strategies are evaluated: feature-level integration, which augments the classifier input space with residual and temporal information, and model-level ensemble fusion, which combines predictions at the decision stage. On the CSTR benchmark, both strategies achieve approximately a 3% accuracy improvement over data-only baselines, reaching 98.94% and 99.00%, respectively. On the larger-scale TEP benchmark, spanning 20 fault scenarios plus normal operation across 52 process variables, the best hybrid feature representation achieves 93.55% single-classifier accuracy and 93.78% with stacking. Conformal prediction is applied across both benchmarks to produce calibrated prediction sets and evaluate uncertainty through empirical coverage and prediction-set size, revealing that classification accuracy and uncertainty calibration are distinct properties that must be assessed jointly for trustworthy fault diagnosis. 

Together, these contributions form a coherent progression from surveying the field and identifying its limitations to generating missing fault data, detecting unseen failure modes, and ultimately delivering reliable, uncertainty-aware hybrid predictions suited to real industrial environments.