Currently, the maintenance process of undercarriage components relies primarily on manual inspections. This is time-consuming and prone to human error. This study focuses specifically on the sprocket component of the undercarriage. The purpose of this study is to develop a prescriptive maintenance framework for Volvo Construction Equipment, presented as a proof of concept given the constraints of data availability. The system aims to predict sprocket wear and translate these predictions into actionable maintenance recommendations.

Since empirical data was unavailable, a synthetic dataset was generated using a hybrid rule-based approach. This approach collected knowledge from domain literature and expert survey responses from VCE personnel. The dataset included operational and environmental variables to reflect real-world conditions. The machines were categorised into three weight classes: light, medium, and heavy. Random Forest and XGBoost were evaluated by multiple metrics for classification and regression.

The prescriptive engine translated predictive outputs into actionable recommendations through a combination of SHAP, DiCE, and LLM. Machines were classified into three operational zones: critical, warning, and safe. Each zone triggers a different response from the system. The system was developed across two experiments: an initial prototype using placeholder parameters and a refined prescriptive system informed by expert survey responses.

The refined prescriptive system performed considerably better than the initial prototype. The improvement between experiments confirmed that the newer dataset produced more consistent patterns for the models to learn from. The results demonstrate that both models achieved comparable performance in the refined prescriptive system. XGBoost consistently outperformed Random Forest across all metrics. In terms of LLM evaluation, Gemini produced more consistent recommendations than Gemma, achieving higher alignment with the reference text. The system was designed as a decision support tool rather than an automated decision-making system. This ensures that the final decision remains with the operator. The next iteration of this project should focus on collecting real sensor data in collaboration with VCE and expanding the scope to cover the entire undercarriage.

i