PhD. Dissertation:Mehmet Emin Mumcuoğlu

Anomaly Detection and Root-Cause Determination for Automotive Applications Using Deep Learning and XAI Models

Mehmet Emin Mumcuoğlu
Mechatronics Engineering, PhD Dissertation, 2025

Thesis Jury

Prof. Dr. Mustafa Ünel (Thesis Advisor)

Prof. Dr. Kemal Kılıç

Assoc. Prof. Dr. Kemalettin Erbatur

 Assoc. Prof. Dr. Ali Fuat Ergenç

Assoc. Prof. Dr. Hüseyin Üvet

Date & Time: 21st July, 2025 – 10:00 AM

Place: FENS L056

Keywords : Anomaly Detection, Predictive Maintenance, Explainable AI, Heavy-Duty Vehicles, Fuel Efficiency, Air Pressure System, LSTM Autoencoder, LLM

Abstract

Anomaly detection in heavy-duty vehicles (HDVs) is crucial for predictive maintenance and efficient fleet management, yet it poses considerable challenges due to the complex interplay between mechanical systems, diverse operational conditions, and limited labeled data. Traditional diagnostic approaches often fall short, struggling with false alarms and lacking interpretability, which can undermine user trust and delay critical interventions. Addressing these challenges necessitates robust, data-driven anomaly detection frameworks that combine precision with explainability, informed by domain knowledge and human expertise. This thesis develops two tailored anomaly detection frameworks specifically designed for critical HDV applications: (1) detecting excessive fuel consumption under varying operational conditions, and (2) early detection of air pressure system (APS) failures. Excessive fuel consumption significantly impacts operational efficiency and regulatory compliance, whereas APS failures frequently result in costly breakdowns and down time. Each application demands unique methodological considerations due to the inherent variability and complexity of the underlying data.

For fuel consumption anomaly detection, a novel quartile-based labeling method was introduced, considering weight-normalized fuel consumption and multi-level road slope segmentation. Utilizing bagged decision trees, this supervised approach classifies operational anomalies at high accuracy across diverse driving datasets from Turkey and Germany. An interactive fleet monitoring dashboard provides actionable insights for fleet operators by visually identifying anomalous trips and facilitating targeted interventions. For APS failure detection, the thesis explores semi-supervised learning through Long Short-Term Memory (LSTM) Autoencoders, enhanced by a human in-the-loop framework incorporating expert analysis. These models effectively identify subtle temporal deviations preceding mechanical failures. Additionally, the Explainable Boosting Machine (EBM) model achieved an excellent balance of predictive accuracy and interpretability, complemented by a Large Language Model (LLM)-based agentic system that provides expert-level diagnostic reasoning and transparency.

This thesis emphasizes interpretability by integrating explainable AI techniques alongside human expertise, thus enhancing diagnostic reliability and user trust. These interpretable frameworks enable clear root-cause analysis, reduce false alarms, and improve practical decision-making across diverse operations. The developed methodologies offer versatile and adaptable solutions for sustainable fleet management, with potential future expansions toward real-time anomaly detection, multi-fault classification, and integration into automated, closed-loop predictive maintenance systems.