<![CDATA[Fuel consumption represents a significant operational cost in mining, where real-time eco-driving optimization is hindered by dynamic and non-stationary operating conditions. Variations in operator behavior and environmental factors often induce concept drift, which diminishes the reliability of static machine learning models and constrains the effectiveness of conventional drift detection methods. This study proposes a distribution-aware, event-triggered Explainable Artificial Intelligence (XAI) framework for detecting and diagnosing fuel consumption anomalies in streaming telematics data. A Hoeffding Tree Regressor was evaluated using a prequential scheme on 1,927,867 real-world observations, achieving a Mean Absolute Error (MAE) of 19.43 under non-stationary conditions. Concept drift was monitored using the Kolmogorov–Smirnov Windowing (KSWIN) algorithm, which detected 1,874 drift events. Upon detection, an event-triggered SHAP module identified contributing factors, indicating that behavioral features such as engine speed and accelerator position were dominant contributors in early drift events. The primary contribution of this study is the integration of distribution-based drift detection with event-triggered explainability within a unified streaming framework, facilitating both anomaly detection and interpretable root-cause analysis.]]>
Copyrights © 2026
