Data-Driven Plant Models for Hydraulic Excavators: Development of Compact Machine Learning Architectures across Simulated and Physical Platforms

Sammanfattning

Modern electro-hydraulic systems, such as those found in hydraulic excavators, exhibithighly nonlinear, coupled, and time-varying dynamics. Developing accuratemathematical plant models for these systems is a challenge in heavy machinery automation.Traditional system identification methods often fail to capture complexbehaviors like valve dead-zones and hysteresis, while high-fidelity analytical simulationmodels are computationally heavy and difficult to calibrate. This thesisintroduces a data-driven alternative developed in collaboration with CPAC SystemsAB: an "Excavator Dynamics Predictor" (EDP) capable of providing fast, accuratereal-time dynamic predictions.Using an automated optimization suite powered by Optuna, this research systematicallyevaluated a variety of deep sequence learning architectures, including LongShort-Term Memory networks, Temporal Convolutional Networks, and hybrid configurations.Testing revealed that a hybrid TCN-LSTM model, augmented with targeteddifferential pressure features, achieves the highest prediction accuracy whilecompressing the model size. The framework utilizes small, specialized machine learningmodules to reduce composite prediction errors by 41.3% while keeping the sameparameter footprint.To validate the practical utility of the EDP, a transfer learning framework was developedto bridge the gap between simulation and reality. While models trainedpurely on simulated data struggled with real-world noise and unpredictable soil-toolinteractions, a two-phase fine-tuning schedule successfully adapted the model to aphysical excavator using a limited amount of real operational logs. The transferlearning framework lowered the prediction mean absolute error from 8.0 mm/s to2.9 mm/s, proving the model’s sim-to-real ability and generalizability across excavatormodels. Finally, to meet the strict computational and memory constraintsof an industrial microcontroller, model compression via knowledge distillation andquantization-aware training was applied. The resulting compressed architectureachieved a model size of 24K parameters while only increasing prediction errors by≈ 30%. This work demonstrates that hardware-aware deep learning models canserve as a fast, scalable foundation for digital twins and advanced model predictivecontrol algorithms in autonomous construction applications.