Hybrid modelling in a digital-twin context for driveline damage detection and prevention within a mechanised, digitalising mining industry

Abstract

Modern mining operations rely on large fleets of mechanised mobile equipment to ensure productivity. However, harsh duty cycles make the drivelines on these machines particularly susceptible to failure, establishing them as a primary cause of expensive, unplanned downtime. Against this backdrop, the sector is undergoing a rapid digital transformation, creating an ideal opportunity to leverage digital twins to minimise the impact of these failures. However, industry surveys conducted in this work reveal that the digital twin concept is often misunderstood, confirming the critical need for a standardised definition and approach. To address this ambiguity, this thesis first proposes a digital twin framework tailored for the mining industry. This strategic contribution defines five dimensions of a digital twin, providing a standardised structure for integrating various models to improve decision-making in mining.

With this framework established, the work presents two technical contributions that address driveline damage detection and prevention. Both contributions rely on hybrid modelling methods that combine physics-based and data-driven models to provide advanced insights.

The first technical contribution addresses reactive damage detection in gearboxes, specifically where historical fault data is scarce. A Simulation-Driven Domain Adaptation (SDDA) methodology was developed to overcome this “few-shot” challenge. In this approach, a generalised physics-based model generates abundant, labelled synthetic fault data, which is used to train a data-driven model. A transfer learning pipeline calibrates and adapts the data-driven model to the real gearbox data domain using limited real-world vibration signals. Validated through controlled numerical investigations and two case studies on real gearboxes, SDDA achieves high fault detection accuracy even with minimal real-world data, successfully classifying faults that purely data-driven and physics-based models fail to detect.

The second technical contribution focuses on proactive damage prevention through wheel-slip prediction for skid-steered vehicles. The study finds that standard isotropic friction models are sub-optimal for modelling heavy vehicles’ behaviour on mining terrain, as they miss the directional physics of soil deformation. Consequently, an anisotropic traction model was formulated to explicitly separate the longitudinal and lateral traction coefficients. Integrated with a vehicle state estimator, this method generates real-time available torque envelopes for each wheel or track of a skid-steered vehicle. These envelopes dictate when the wheel is predicted to slip. Field experiments confirmed that this anisotropic approach captures directional traction effects significantly better than the isotropic baseline, enabling timely warnings to prevent damaging wheel slip.

Ultimately, this research demonstrates the power of hybrid modelling, combining the structural interpretability of physics with the adaptive calibration of data. By framing these hybrid methods within the proposed digital twin framework, the thesis illustrates a concrete capability to transform raw sensor data into actionable insights for asset failure detection and prevention, paving the way for a more reliable, digitalised mining future.