Personalized AI Driver Coaching in Heavy-Duty Trucks for Fuel-Efficient Driving - From Vehicle Telemetry to Eco-Driving Feedback through SHAP and Retrieval-Augmented LLMs

Abstract

Fuel-efficient driving is important for reducing both operational costs and environmental impact in heavy-duty transport. Although existing driver coaching systems can support eco-driving, they often rely on predefined rules, generic feedback, or numerical scores that provide limited explanation of how drivers can improve. This thesis investigates how vehicle telemetry data can be used to generate personalized eco-driving feedback through an AI-based conversational driver coaching prototype. Vehicle telemetry data from Volvo heavy-duty trucks operating in the Latin American market was analysed within two operational segments. Segment-specific Light GBMmodels were trained to predict fuel efficiency from driving behaviour variables, while SHAP values were used to identify the behaviours that most strongly influenced each prediction in each segment-specific model. These explanations were then translated into conversational coaching feedback using a large language model. To support trustworthy and personalized interaction, the prototype further incorporated retrieval-augmented generation for domain grounding and a persistent se mantic memory layer for adapting responses to user-specific preferences, goals, and constraints. The modelling results show that features related to engine torque usage, driving speed, and acceleration behaviour influence predicted fuel efficiency the most, with some variation between operational segments. The LightGBM models outperformed a ridge regression baseline, achieving mean cross-validated R2 scores of 0.796 and 0.744 for the two analysed segments. SHAP analysis further showed that segment specific models can interpret the same driving behaviour differently, supporting the relevance of context-aware modelling. The conversational prototype was evaluated through an expert questionnaire with participants from Volvo Group. The results indicate that personalized and context aware responses were generally preferred over non-personalized alternatives. How ever, the evaluation also showed that retrieval-augmented generation did not consistently improve perceived response quality, partly due to overly detailed or technically ambiguous feedback. These findings highlight both the potential and the challenges of combining explainable machine learning and large language models for driver coaching. Overall, the thesis demonstrates that telemetry-based explanations can be transformed into personalized conversational feedback, while emphasizing the need for careful grounding, concise presentation, and domain-specific validation.