Memory-Efficient Hydraulic Pump Thermodynamic Efficiency Calculation for Non-Road Mobile Machinery

Mobile hydraulic machines operate in remote areas, so pumps installed on them must calculate efficiency on the controller using available sensors. Direct measurement with torque sensors and flow meters is costly. The analytical methods based on thermodynamic models allow calculating efficiency from temperature and pressure measurements: multiple formulations exist with different sensor requirements and accuracy trade-offs. Alternatively, lookup tables built from experimental data give high accuracy through interpolation, but a four-dimensional table covering pressure, speed, displacement, and temperature requires significant memory on resource-constrained microcontrollers. To fill the gap between analytical models and lookup tables, we train a feedforward neural network on the same experimental efficiency data used to build the lookup table. In total six estimation approaches are evaluated in MATLAB for a variable displacement axial piston pump: four thermodynamic formulations, a lookup table serving as ground truth, and the proposed network. The network reaches a mean absolute percentage error of 0.19% against the lookup table, with 94.2% of predictions within 0.5% and 99.9% within 2.0%, while cutting memory use by 3.5 times, from 1.50 kB to 0.43 kB, which makes it a good fit for embedded controllers where both accuracy and memory are limited.