Applying deep reinforcement learning to minimize flow fluctuations in digital flow control

Abstract

This work addresses the non-linear and complex optimization challenges in digital hydraulic flow control, where piecewise behaviors and unpredictable interactions make traditional optimization methods impractical for presented system with continuous inputs. The study aims to promote real-time application of artificial intelligence algorithms for system optimization. Most off-highway construction and agriculture equipment use hydraulic valve manifolds, which offer unmatched power density and dynamics, excelling over electro actuators in high-capacity applications. The growing demand for more efficient and accurately controlled autonomous heavy machinery has driven the need for steady and precise flow control systems with reduced pressure drop. However, managing flow and pressure fluctuations when switching valves remains a significant challenge.The proposed agent effectively mitigates flow fluctuations by interactively refining valve-timing decisions over tens of thousands of possible actions. Validated under approximately 90 % of available conditions and tested against unseen pressure values, the agent achieved a median integrated flow error of less than 0.5 cm³, showcasing its potential for AI-driven optimization in digital hydraulic systems.