A Surrogate Modeling Approach for Digital Hydraulic Valve Systems Using Model- Based Training Data

Abstract

In digital hydraulic systems, flow regulation is achieved by combining the discrete openings of multiple on/off valves. However, performing dynamic simulations for large digital flow control units (DFCU) is computationally intensive. A 16x4 DFCU, for instance, involves 65,536 possible valve states, each requiring separate dynamic evaluation. This work presents ValveNet, a data-based model that uses a feedforward artificial neural network (ANN)to estimate the cost value J, which reflects the pressure and velocity tracking errors. The network was trained on 2,048,000 physics-based samples (4000 random operating combination x 512 valve actions per state) and validated against a benchmark grid of 65,536 valve combinations derived from a simplified model-based type. The compact ANN architecture ([16–8], SCG training) achieved R2=0.998 on validation data and maintained strong correlation with the physical model (R2=0.78–0.99) while showing a 5.7x computational speed-up. ValveNet enables rapid evaluation of complex DFCU configurations. Also, it achieves real-time valve optimization and digital hydraulic control properly.