1539 ¡∗(M) = arg min B(>) (–SF5IH(¡) + GY(¡) + oJ(¡)) where: • ¡∗(M)is the recommended optimal speed, • SF5IH(¡)represents hauling time, • Y(¡)represents energy or fuel consumption, • J(¡)is a safety-related penalty term, • –,G,oare weighting coefficients. This formulation reflects the trade-off between productivity maximization, fuel efficiency, and safe operation. The adaptive machine learning model is formulated as a multi-target regression problem, aiming to optimize both operational efficiency and productivity. Specifically, two target variables are defined: Recommended Vehicle Speed (km/h): This target represents the optimal speed recommendation for a given road segment, balancing safety, fuel efficiency, mechanical constraints, and driving smoothness. Predicted Cycle Time (minutes): This target estimates the expected traversal time for a road segment or haul cycle, enabling improved scheduling, dispatch planning, and productivity forecasting. Together, these targets provide actionable insights that are directly usable by fleet management systems and operators. 2.2 Model Algorithms The study evaluates multiple machine learning algorithms selected for their suitability to structured telemetry data and their balance between predictive performance and interpretability: • Light Gradient Boosting Machine (LightGBM) Employed as the primary model due to its superior performance on large-scale, treebased structured datasets and its ability to capture complex non-linear relationships efficiently. • Random Forest Regressor Used as a benchmark and robustness-oriented model, offering strong generalization performance and enhanced interpretability through feature importance analysis. • Online Stochastic Gradient Descent (SGD) Regressor (Optional) Considered for scenarios requiring real-time or near-real-time model updates, allowing incremental learning without full retraining when computational or latency constraints are present. 2.3 Adaptive Training Strategy Unlike conventional static machine learning approaches, this research implements an adaptive training strategy to maintain model relevance under evolving operational conditions such as changing road geometry, payload distributions, and vehicle performance characteristics. The adaptive strategy consists of the following components: 1. Rolling Training Window Model training data are restricted to a rolling window of the most recent 14–30 days of telemetry data, ensuring sensitivity to recent operational patterns while avoiding concept drift from outdated data. 2. Performance-Based Sample Weighting The top 10% of best-performing haul cycles, defined by faster cycle times, lower fuel consumption, and smoother driving profiles, are assigned higher weights during training to reinforce optimal operational behavior. 3. Scheduled Retraining Models are retrained on a daily basis using an automated workflow orchestration
RkJQdWJsaXNoZXIy MTM0Mzk2