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

Abstract

Modern electro-hydraulic systems, such as those found in hydraulic excavators, exhibit highly nonlinear, coupled, and time-varying dynamics. Developing accurate mathematical plant models for these systems is a challenge in heavy machinery automation. Traditional system identification methods often fail to capture complex behaviors like valve dead-zones and hysteresis, while high-fidelity analytical simulation models are computationally heavy and difficult to calibrate. This thesis introduces a data-driven alternative developed in collaboration with CPAC Systems AB: an "Excavator Dynamics Predictor" (EDP) capable of providing fast, accurate real-time dynamic predictions. Using an automated optimization suite powered by Optuna, this research systematically evaluated a variety of deep sequence learning architectures, including Long Short-Term Memory networks, Temporal Convolutional Networks, and hybrid configurations. Testing revealed that a hybrid TCN-LSTM model, augmented with targeted differential pressure features, achieves the highest prediction accuracy while compressing the model size. The framework utilizes small, specialized machine learning modules to reduce composite prediction errors by 41.3% while keeping the same parameter footprint. To validate the practical utility of the EDP, a transfer learning framework was developed to bridge the gap between simulation and reality. While models trained purely on simulated data struggled with real-world noise and unpredictable soil-tool interactions, a two-phase fine-tuning schedule successfully adapted the model to a physical excavator using a limited amount of real operational logs. The transfer learning framework lowered the prediction mean absolute error from 8.0 mm/s to 2.9 mm/s, proving the model’s sim-to-real ability and generalizability across excavator models. Finally, to meet the strict computational and memory constraints of an industrial microcontroller, model compression via knowledge distillation and quantization-aware training was applied. The resulting compressed architecture achieved a model size of 24K parameters while only increasing prediction errors by ≈ 30%. This work demonstrates that hardware-aware deep learning models can serve as a fast, scalable foundation for digital twins and advanced model predictive control algorithms in autonomous construction applications.