Purpose and KPI approach
Load-haul-dump machines operate through uneven terrain, repeated acceleration and long periods of demanding engine duty. Yet the standard engine control unit exposes only part of the information needed to distinguish normal operation from developing faults. This dissertation develops a hybrid model of the Cummins QSB6.7 diesel engine that turns readily available ECU measurements into a broader view of engine condition and performance.
The target quantities span thermal behaviour, intake and boost pressure, indicated and brake power, fuel consumption and brake-specific fuel consumption. Rather than treating these as unrelated outputs, the study connects them through the physics of combustion, gas exchange and turbocharging.
Test programme
Controlled dynamometer tests were designed to represent the varying speed and load combinations encountered by mining vehicles. Twenty-seven operating conditions were selected using Latin hypercube sampling. ECU data were captured with a CANedge2 logger, while a QuantumX system recorded additional pressure and temperature measurements for model development and validation.
A separate fault experiment operated the engine with four of its six injectors. The combustion-event frequency amplitude rose to 2.59 times the healthy value. Some frequency patterns remained strongly correlated with healthy operation, but the standard deviation under the fault was approximately eleven times higher. This illustrates why multiple indicators and contextual models are more informative than a single threshold.
Hybrid model architecture
The physics layer represents turbocharger behaviour, airflow, valve timing, a dual-cycle in-cylinder process, air–fuel ratio, volumetric efficiency and valve-overlap backflow. Calibration used 243 data points and repeated global and local optimisation. Data-driven mappings then translated the limited set of accessible ECU inputs into the intermediate quantities required by those equations.
This division is deliberate. Physics gives the model interpretable internal relationships; empirical mappings allow it to operate when proprietary maps or additional sensors are unavailable.
Accuracy and evidence
After calibration, the physics model achieved mean relative errors of 4.16% for thermal quantities, 6.15% for pressures and 4.47% for power and efficiency quantities. The full hybrid model, using a reduced input set, achieved an average mean absolute percentage error of 5.36% across its KPI suite. The small loss of accuracy buys a substantial reduction in instrumentation requirements.
Brake-specific fuel consumption also indicated that efficient LHD operation should be interpreted around a torque-dominant regime rather than simply seeking maximum speed or power.
Maintenance significance
The work provides a practical foundation for an engine digital twin that can infer difficult-to-measure quantities, compare observed and expected behaviour and support condition-based maintenance. Further testing at extreme and transient conditions, higher-frequency data and field validation would strengthen its ability to detect anomalies on production machines.