Learning State Transition Functions of a Hydraulic Excavator using Data Driven Deep Learning

In this thesis, a neural network model is trained to learn the state transition function of a hydraulic excavator arm or a crane using data driven approach. The excavator arm or crane exhibits highly non linear behaviour due to hydraulic coupling between its actuators, non linearities in valves, parallel kinematics, cylinder friction, control input dead zones and also due to non linearities inherent in hydraulics. Hence, it is difficult to accurately represent the analytical model of these systems and it cannot be used to determine the state transitions. To achieve the automation of the hydraulic excavator, it is important to develop a neural network model to accurately predict the state transitions. This neural network model can be used to train a control policy in a reinforcement learning (RL) to learn the inverse kinematics operation of the
hydraulic excavators. After training the control policy, it can be deployed on a real hydraulic excavator to achieve the automation. The data required for training the neural network has to be recorded from the excavator during its operation. During neural network training, the model learns the non linearities of the system and once the training is completed it should be able to accurately predict the state transition of the system.