Abstract
Accurate identification of duty-cycle activities (Transiting, Loading, Hauling, and Dumping) is essential for optimizing the performance of Battery-Electric Load-Haul-Dump (LHD) vehicles used in underground mining. The mining industry is increasingly adopting battery-powered equipment due to its lower emissions and reduced ventilation needs compared to diesel alternatives. Effective battery management and optimized charging schedules require reliable tracking of these operational activities.
This study is part of a broader collaborative initiative between the author and CanmetMINING. The project's aim is to develop advanced Artificial Intelligence (AI)-based systems for predicting fuel and battery consumption in mobile mining equipment. Improving duty-cycle classification accuracy could enhance the predictive capabilities of these AI models, resulting in more reliable forecasts of battery usage and equipment performance. Such systems could provide valuable decision support, helping mining operators select between traditional diesel machinery and electric alternatives by balancing environmental impact, cost, and operational efficiency.
While traditional methods for duty-cycle recognition rely heavily on manual labeling or supervised classification, this research explores an alternative unsupervised approach. Specifically, the Bisecting KMeans clustering algorithm was applied to approximately 65,000 rows of sensor data. This dataset represents two hours of vehicle operation, covering 20 full duty cycles of up-ramp hauling. Four sensor parameters were selected based on clear interpretability and direct relevance to vehicle activities: Traction Power, Auxiliary Power, Average Wheel Speed, and Rate of Change of State of Charge (SoC).
Clusters generated by the unsupervised method were then mapped to operational phases using rules based on domain knowledge. These mapped clusters were compared against manually labeled ground-truth data, classified by engineers familiar with the experimental conditions, to evaluate classification accuracy. Internal validation metrics confirmed reasonable clustering quality (Silhouette Score = 0.572, Calinski-Harabasz Index = 204,386).
The results demonstrated that the unsupervised approach effectively identified Hauling (95% accuracy) and Transiting (89% accuracy), but significantly struggled with distinguishing Dumping (49% accuracy) from Loading (10% accuracy). Occasionally, the algorithm switched rapidly between **Loading** and **Dumping** labels in a way that's physically impossible—such as predicting Loading immediately after Dumping, or vice versa—highlighting limitations of the current method.
These findings indicate both the potential and challenges associated with using unsupervised learning for duty-cycle classification. To address these limitations, future studies should incorporate additional descriptive sensor parameters, and consider supervised machine learning approaches utilizing labeled data to enhance accuracy and reliability.