AEBS Event Classification Under Data Heterogeneity and Limited Trusted Labels

Abstract

Advanced Emergency Braking System (AEBS) interventions can be classified according to whether the braking response was justified by the surrounding traffic situation. Accurate event classification is important for vehicle safety and driver trust, while fleet-scale assessment is challenging because most events are labeled by deterministic signal-based scripts and trusted video labels are scarce, costly, and privacy-sensitive. In addition, event data may exhibit distributional heterogeneity across operating contexts. This thesis investigates AEBS full-braking event classification under these challenges of data heterogeneity and limited trusted labels. The study uses logged full-braking events from Volvo heavy-duty vehicles from different countries under scarce trusted video labels. Within a Machine Learning framework, the task is formulated as binary multivariate Time-Series Classification. A one-dimensional CNN is used for supervised learning, unsupervised clustering is used to examine event structure, and Mean Teacher is used for semi-supervised learning. Federated Learning is evaluated with FedAvg and FedMeanTeacher to assess country-partitioned collaboration. Country partitions were strongly imbalanced, with the two largest accounting for 62% of the selected events and the three smallest accounting for 9%. Scriptgenerated outcomes agreed with 75% of the trusted FP events, while agreement with trusted TP events was 94%. Supervised classification achieved a Macro-F1 of 0.9995 on script-generated outcomes and 0.9374 on trusted video labels. FedMean- Teacher reduced the cross-country Macro-F1 standard deviation to 0.0343 compared with 0.1796 for local semi-supervised learning, while its trusted-label Macro-F1 remained lower than that of centralized learning, 0.7867 versus 0.8523. Overall, scriptgenerated outcomes enabled near-perfect reproduction of the deterministic labeling procedure, while scarce trusted labels, especially for FP events, limit correctnessoriented conclusions.