Using a Neural Network Controller to Minimize the Pressure Peaks in Binary Coded Digital Valve Systems
Elsaed, Essameldin (2026)
Tampere University
2026
Teknisten tieteiden tohtoriohjelma - Doctoral Programme in Engineering Sciences
Tekniikan ja luonnontieteiden tiedekunta - Faculty of Engineering and Natural Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Väitöspäivä
2026-09-18
Julkaisun pysyvä osoite on
https://urn.fi/URN:ISBN:978-952-03-4745-1
https://urn.fi/URN:ISBN:978-952-03-4745-1
Tiivistelmä
Heavy-duty mobile and industrial hydraulic machines face growing demands for efficiency and controllability at high flow rates. Digital hydraulics, and in particular digital flow control units (DFCUs) based on parallel-connected on/off valve branches, are promising approaches because they can provide low leakage and flexible discrete flow levels with scalable capacity. However, when several branches switch between discrete states, unavoidable differences in valve response times and fluid inertia create brief flow mismatches that excite pressure peaks lasting only a few milliseconds. These transients stress components and complicate high-flow DFCU design and tuning.
This work combines hydraulic design, modelling, simulation, and experimental validation, with machine learning used as an enabling method where transient behavior is difficult to model or optimize directly. A dynamic model of a two-stage pilot-operated poppet valve is developed to quantify how pilot flow paths and stroke limiting affect response time. A novel six-branch semi-binary DFCU with stroke-limited cartridges and Dual-Mode piloting is then designed and simulated to provide high flow capacity with fine discrete flow resolution. Numerical and machine-learning models are used to develop pressure and temperature aware transient flow predictors from sparse data. Finally, valve-delay selection is formulated as a one-step contextual-bandit problem, first in simulation and then on a physical test rig using measured pressure impulse as feedback.
The prototype DFCU delivered about 480 LPM at Δp = 5 bar, with a maximum flow-step difference of about 10 LPM and a discretization error of 2.1%. In simulation, the learned timing policy reduced the integrated flow error by approximately 60 %. On the test rig, timing-only control reduced pressure impulse in tested transitions by about 45%. Overall, the thesis shows that a stroke-limited, high-flow DFCU can be physically implemented, and that its switching-induced pressure spikes can be mitigated on a test rig, with model-free real-time learning used to select the valve-timing adjustments.
This work combines hydraulic design, modelling, simulation, and experimental validation, with machine learning used as an enabling method where transient behavior is difficult to model or optimize directly. A dynamic model of a two-stage pilot-operated poppet valve is developed to quantify how pilot flow paths and stroke limiting affect response time. A novel six-branch semi-binary DFCU with stroke-limited cartridges and Dual-Mode piloting is then designed and simulated to provide high flow capacity with fine discrete flow resolution. Numerical and machine-learning models are used to develop pressure and temperature aware transient flow predictors from sparse data. Finally, valve-delay selection is formulated as a one-step contextual-bandit problem, first in simulation and then on a physical test rig using measured pressure impulse as feedback.
The prototype DFCU delivered about 480 LPM at Δp = 5 bar, with a maximum flow-step difference of about 10 LPM and a discretization error of 2.1%. In simulation, the learned timing policy reduced the integrated flow error by approximately 60 %. On the test rig, timing-only control reduced pressure impulse in tested transitions by about 45%. Overall, the thesis shows that a stroke-limited, high-flow DFCU can be physically implemented, and that its switching-induced pressure spikes can be mitigated on a test rig, with model-free real-time learning used to select the valve-timing adjustments.
Kokoelmat
- Väitöskirjat [5374]