Developing Data-Driven Models For Anomaly Detection In Automotive And Additive Manufacturing Applications

Anomaly detection is a fundamental yet inherently challenging task in machinelearning and statistics, with wide-ranging applications spanning domains such ashealthcare, manufacturing, automotive, and aerospace. Unlike conventional classificationproblems, anomaly detection must contend with intrinsic difficulties includingclass imbalance, anomaly heterogeneity, and the scarcity of labeled anomalies. Addressingthese challenges requires thoughtfully designed, domain-aware frameworkscapable of operating under limited supervision while maintaining robustness andinterpretability. This thesis develops several data-driven anomaly detection frameworks,spanning supervised, semi-supervised, and unsupervised learning paradigms.In particular, an efficient semi-supervised framework built upon Transformer architecturesis developed, effectively mitigating the inherent challenges of anomalydetection. In addition, the thesis adopts an interpretable framework grounded inExplainable Boosting Machine (EBM), offering transparency and domain-alignedinsights without sacrificing performance. A domain-guided preprocessing pipelineis integrated into all frameworks to systematically incorporate expert knowledge,facilitate robust anomaly discrimination, and improve interpretability by aligningfeature representations with meaningful physical phenomena.Two real-world industrial applications were considered in this thesis: (1) failure detectionin air pressure systems (APS) of heavy-duty vehicles using operational sensordata, and (2) defect detection in directed energy deposition (DED) using thermalimaging. The APS plays a vital role in ensuring the proper functioning of vehiclesubsystems such as braking and suspension, where failures can pose significant safetyrisks and economic consequences. Meanwhile, DED, an effective additive manufacturingtechnology, offers a promising pathway for fabricating complex, large-scalecomponents; however, it suffers from recurring in-situ defect formation, compromisingpart reliability and quality. The data-driven models yielded promising resultsin both applications. Remarkably, for APS failure detection, the semi-supervisedtransformer-based approach—although trained using only a small portion of nonanomalousdata—led to strong predictive performance on par with the fully supervisedmodels, attaining 91.4% accuracy and an F1 score of 0.79. In parallel, theinterpretable EBM-based framework achieved similarly competitive performance (anF1 score of 0.80) while providing meaningful insights into feature contributions andpotential root causes, corroborated by domain knowledge. For DED defect detection,semi-supervised models exhibited strong performance, with an accuracy andF1 score up to 95% and 0.88, respectively.These findings demonstrate that combining domain-specific feature engineeringwith data-efficient learning paradigms enables effective anomaly detection acrossdiverse settings. The thesis underscores the practical utility of semi-supervisedlearning—specifically for scenarios with limited anomaly labels—and highlights thegrowing importance of explainability, particularly in high-stakes applications, wheretransparent models such as EBM can provide actionable insights without sacrificingaccuracy. The frameworks developed in this thesis are readily adaptable to otherindustrial contexts, depending on the nature of the underlying datasets (balanced vsimbalanced) and desirable characteristics (e.g., highly interpretable). Furthermore,they can be extended to incorporate multi-defect classification, closed-loop controlintegration, and real-time decision-making.