Torque Control of Hydraulic Pressure Servo Valve Driven Actuator with Deep Neural Network

Non-linear dynamics, model uncertainties due to hydraulic fluid, and disturbances in hydraulic systems make it difficult to obtain accurate torque tracking performance. In this study, a learning-based torque-tracking method is proposed, which does not require approximating the torque dynamics. The proposed method can capture disturbances and model uncertainties of system. The applied neural network comprises a nonlinear autoregressive model with exogenous inputs (NARX) and a long shortterm neural network (LSTM). NARX is employed due to its ability to predict time series control input from the states of system, and LSTM is used to overcome the vanishing and exploding gradient, which causes long-term memory loss in NARX, leading to inaccurate torque tracking performance. LSTM with NARX achieved a better prediction performance with a mean square error and standard deviation of
$0.0015 \pm 0.4\times 103$
compared to only NARX with a mean square error of
$0.004 \pm 1.0\times 103$
at 10 K training data size.