Embedded Operation Effort Estimation: A Machine Learning Performance-Efficiency Trade-off Analysis

Le résumé fourni par la source

Real-time monitoring of a vehicle’s operational effort is critical for enhancing dependability and enabling predictive maintenance in modern automotive systems. Implementing this on resource-constrained on-board computers, however, requires navigating a trade-off between prediction accuracy, data-label efficiency, and computational cost. This paper investigates this trade-off through a comparative analysis of four distinct methods: a simple heuristic baseline, a semi-supervised clustering model (K-Means), a semi-supervised hybrid model, and a fully supervised XGBoost benchmark. Using a tailored dataset from the CARLA simulator, we evaluate the prediction performance of each model. We validate their feasibility for on-board deployment by benchmarking their computational workload in C++ on a RISC-V-based platform, measuring execution time and energy consumption. Our results reveal a clear performance-efficiency spectrum, demonstrating that the hybrid model offers a compelling engineering compromise by achieving robust predictive performance while remaining label-efficient and computationally lightweight. This analysis provides a guide for model selection for on-device intelligence under the practical constraints of automotive systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.