Integrating XGBoost, SHAP, and DiCE for Prescriptive Maintenance of Undercarriage Sprockets

This work presents a proof‑of‑concept prescriptive maintenance framework for predicting sprocket wear in Volvo Construction Equipment undercarriages. To address the absence of empirical sensor data, a synthetic dataset was generated using a hybrid rule‑based method informed by domain literature and expert surveys. The dataset incorporated operational and environmental variables and grouped machines into light, medium, and heavy classes. Random Forest and XGBoost models were evaluated for both classification and regression tasks, with XGBoost achieving the strongest overall performance. A prescriptive engine combined SHAP, DiCE, and an LLM layer to translate model outputs into actionable maintenance recommendations. Machines were assigned to critical, warning, or safe zones, each triggering tailored guidance. Two experimental iterations were conducted, and the refined system—built on expert‑validated parameters—showed substantially improved predictive consistency. LLM evaluation indicated that Gemini produced more reliable recommendations than Gemma. Designed as a decision‑support tool, the framework keeps operators in control while demonstrating the potential of prescriptive analytics for undercarriage maintenance. Future work will focus on integrating real sensor data and expanding the system to additional components.