Fault diagnosis is essential for ensuring the safety and reliability of complex engineering systems, where faults in individual components can propagate into degradation of overall process performance and eventually failure. Traditional model-based diagnosis relies on mathematical models to generate residuals that capture deviations from expected behavior, while data-driven diagnosis uses machine learning on historical data to detect or classify faults. While both approaches aim for fault detection and isolation, achieving satisfactory performance is a non-trivial task. Model-based methods require detailed domain expertise and accurate system models, whereas data-driven approaches often struggle with scarce or unrepresentative fault data and unreliable generalization beyond training distributions.

A primary focus of this thesis is the investigation of hybrid fault diagnosis methods that integrate physical insight with deep learning architectures to reduce reliance on fault data. To address this, structural analysis is employed as a foundational tool for identifying redundancy, and computational sequences derived from the structural model are used to guide the design of neural network-based residual architectures. Additionally, the residuals are designed to respect the underlying differential relationships of dynamic systems, with careful consideration given to the numerical evolution of system states. Several sequential modeling approaches are implemented, and a methodology for generation, training and assessment of these hybrid diagnostic systems is provided.

Another aspect of this thesis is addressing model inaccuracies in data-driven diagnosis. Diagnosis systems must handle noisy sensor signals and incomplete training data, which can cause unreliable or overconfident diagnostic statements. To address this, the thesis explores different techniques to evaluate the reliability of the data-driven residuals. One approach is to model validity regions of residual models in which diagnostic conclusions remain reliable. The second approach uses probabilistic ensemble neural networks to quantify aleatoric and epistemic uncertainty and to handle model inaccuracies. In this thesis, it is shown how to integrate data-driven models and the assessed reliability measures, into a consistency-based diagnosis framework.

This thesis contributes to bridging the areas of model-based and data-driven fault diagnosis. The proposed methods have been validated using data from both simulations and real automotive case studies. The results show that the design of neural network-based residuals can reason about abnormal behavior in complex dynamic processes even when there is limited or no training data from faults.