Fwd: PhD Dissertation Defense: Shawqi Mohammed Farea

Developing Data-Driven Models for Anomaly Detection in Automotive and Additive Manufacturing Applications

Shawqi Mohammed Farea
Mechatronics Engineering, PhD Dissertation, 2025

Thesis Jury

Prof. Dr. Mustafa Ünel (Thesis Advisor), Prof. Dr. Bahattin Koç, Assist. Prof. Dr. Melih Türkseven, Assoc. Prof. Dr. Ali Fuat Ergenç, Assist. Prof. Dr. Abdurrahman Eray Baran

Date & Time: 1st July, 2025 – 10:00 AM

Place: FENS G029

 

Abstract

Anomaly detection is a fundamental yet inherently challenging task in machine learning and statistics, with wide-ranging applications spanning domains such as healthcare, manufacturing, automotive, and aerospace. Unlike conventional classification problems, anomaly detection must contend with intrinsic difficulties including class imbalance, anomaly heterogeneity, and the scarcity of labeled anomalies. Addressing these challenges requires thoughtfully designed, domain-aware frameworks capable of operating under limited supervision while maintaining robustness and interpretability. 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 architectures is developed, effectively mitigating the inherent challenges of anomaly detection. In addition, the thesis adopts an interpretable framework grounded in Explainable Boosting Machine (EBM), offering transparency and domain-aligned insights without sacrificing performance. A domain-guided preprocessing pipeline is integrated into all frameworks to systematically incorporate expert knowledge, facilitate robust anomaly discrimination, and improve interpretability by aligning feature representations with meaningful physical phenomena.

Two real-world industrial applications were considered in this thesis: (1) failure detection in air pressure systems (APS) of heavy-duty vehicles using operational sensor data, and (2) defect detection in directed energy deposition (DED) using thermal imaging. The APS plays a vital role in ensuring the proper functioning of vehicle subsystems such as braking and suspension, where failures can pose significant safety risks and economic consequences. Meanwhile, DED, an effective additive manufacturing technology, offers a promising pathway for fabricating complex, large-scale components; however, it suffers from recurring in-situ defect formation, compromising part reliability and quality. The data-driven models yielded promising results in both applications. Remarkably, for APS failure detection, the semi-supervised transformer-based approach—although trained using only a small portion of non-anomalous data—led to strong predictive performance on par with the fully supervised models, attaining 91.4% accuracy and an F1 score of 0.79. In parallel, the interpretable EBM-based framework achieved similarly competitive performance (an F1 score of 0.80) while providing meaningful insights into feature contributions and potential root causes, corroborated by domain knowledge. For DED defect detection, semi-supervised models exhibited strong performance, with an accuracy and F1 score up to 95% and 0.88, respectively.

These findings demonstrate that combining domain-specific feature engineering with data-efficient learning paradigms enables effective anomaly detection across diverse settings. The thesis underscores the practical utility of semi-supervised learning—specifically for scenarios with limited anomaly labels—and highlights the growing importance of explainability, particularly in high-stakes applications, where transparent models such as EBM can provide actionable insights without sacrificing accuracy. The frameworks developed in this thesis are readily adaptable to other industrial contexts, depending on the nature of the underlying datasets (balanced vs imbalanced) and desirable characteristics (e.g., highly interpretable). Furthermore, they can be extended to incorporate multi-defect classification, closed-loop control integration, and real-time decision-making.