Anomaly Detection And Root-Cause Determination For Automotive Applications Using Deep Learning And XAI Models

Anomaly detection in heavy-duty vehicles (HDVs) is crucial for predictive maintenanceand efficient fleet management, yet it poses considerable challenges due to thecomplex interplay between mechanical systems, diverse operational conditions, andlimited labeled data. Traditional diagnostic approaches often fall short, strugglingwith false alarms and lacking interpretability, which can undermine user trust anddelay critical interventions. Addressing these challenges necessitates robust, datadrivenanomaly detection frameworks that combine precision with explainability, informedby domain knowledge and human expertise. This thesis develops two tailoredanomaly 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 consumptionsignificantly impacts operational efficiency and regulatory compliance, whereasAPS failures frequently result in costly breakdowns and downtime. Each applicationdemands unique methodological considerations due to the inherent variability andcomplexity of the underlying data.For fuel consumption anomaly detection, a novel quartile-based labeling method wasintroduced, considering weight-normalized fuel consumption and multi-level roadslope segmentation. Utilizing bagged decision trees, this supervised approach classifies operational anomalies at high accuracy across diverse driving datasets fromTurkey and Germany, achieving up to 92.2% accuracy and an F1 score of 0.78.An interactive fleet monitoring dashboard further provides actionable insights forfleet operators by visually identifying anomalous trips and facilitating targeted interventions.For APS failure detection, the thesis explores semi-supervised learningthrough Long Short-Term Memory (LSTM) Autoencoders, enhanced by a human-inthe-loop framework incorporating expert analysis. These models effectively identifysubtle temporal deviations preceding mechanical failures with an overall F1 scoreof 0.75. Additionally, the Explainable Boosting Machine (EBM) model achieved anexcellent balance of predictive accuracy (91.4%, F1 score: 0.80) and interpretability,complemented by a Large Language Model (LLM)-based agentic system thatprovides expert-level diagnostic reasoning and transparency.This thesis emphasizes interpretability by integrating explainable AI techniquesalongside 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 developedmethodologies offer versatile and adaptable solutions for sustainable fleet management,with potential future expansions toward real-time anomaly detection, multifaultclassification, and integration into automated, closed-loop predictive maintenancesystems.