Agentic AI Meets Deep LearningAn Anomaly Detection Pipeline for Fuel Cell Electric Vehicle X TelemetryData

Sammanfattning

Hydrogen fuel cell electric vehicles generate high-dimensional telemetry data from interacting sub-systems such as the fuel cell system, hydrogen supply, thermal management, power electronics, andvehicle control units. Detecting abnormal behaviour in this data is challenging because operatingconditions vary over time, labelled fault data is limited, and traditional rule-based monitoring canmiss subtle deviations or produce excessive false alarms.This thesis presents a data-driven anomaly detection workflow for logged telemetry data fromVolvo Vehicle X, a fuel cell electric truck monitored during real-world testing. Raw multivariatesensor signals are transformed into fixed-length time-window representations that summarize recentvehicle behaviour. Operating states are identified using unsupervised clustering, and anomalydecisions are made using state-conditioned thresholds rather than a single global threshold. Threepredictive reconstruction models are evaluated: a Transformer-based model, an LSTM baseline,and a graph-based GNN–LSTM model. The models are trained without labelled anomaly data andevaluated against labelled anomaly intervals documented during vehicle testing.The results show that the models capture different aspects of abnormal behaviour. The GNN–LSTM achieves the strongest file-level anomaly detection performance, with the highest F1-scoreand recall. The Transformer provides the strongest temporal overlap with labelled anomaly periods.The original LSTM threshold is overly sensitive and marks a large part of the test data as anomalous,but state-wise P99 threshold recalibration improves its practical usefulness. These results show thatthreshold calibration is a critical part of the anomaly detection workflow.To support interpretability, the thesis also introduces an AI-assisted explanation layer. Thislayer does not modify the anomaly detection output. Instead, it links anomaly scores to relevantsignals, components, subsystems, and operating states, allowing engineers to inspect detected eventsin a more structured way.Overall, the thesis demonstrates that combining window-based telemetry representation, operating-state-aware thresholds, predictive reconstruction models, and metadata-based explanation providesa practical workflow for anomaly detection in complex hydrogen fuel cell vehicle telemetry. Theapproach supports more efficient offline analysis of large-scale vehicle data and helps engineersinvestigate abnormal behaviour in a structured and interpretable manner.