Efficient Control Of Full-Bridge Oscillation Transformer
Author
Mortensen, Kasper Juhl
Term
4. term
Education
Publication year
2025
Abstract
This thesis investigates how to control the Full-Bridge Oscillation Transformer (FBoT) efficiently. The FBoT is a novel hydraulic transformer intended to improve fluid power efficiency by replacing throttling proportional valves. It transfers energy by oscillating a free-floating piston using On/Off valves assisted by pressure-controlled check valves. A high-fidelity simulation model is developed, and a 12-step pump-mode sequence reduces the control problem to six valve timing decisions. Because the dynamics are nonlinear and strongly coupled, reinforcement learning is adopted; the device’s symmetry is exploited to double data collection per oscillation, addressing a training bottleneck. The policy is trained with a reward designed to minimize valve losses while achieving a centered oscillation at the desired amplitude. Simulations show that the highest-average-reward policies often resemble near open-loop timing schedules and react weakly to error signals, leading to steady-state errors in piston position and chamber pressures and increased losses when valves switch. Experimental validation confirms that the trained policy can drive oscillations but with large cycle-to-cycle amplitude variation and estimated low efficiency. To address these limitations, a control framework based on transforming control signals and errors is proposed to reduce coupling and potentially enable classical PI feedback for error correction combined with an intuitive or RL-derived feedforward.
Dette speciale undersøger, hvordan Full-Bridge Oscillation Transformer (FBoT) kan styres effektivt. FBoT er en ny hydraulisk transformer, der skal forbedre virkningsgraden i hydrauliske systemer ved at erstatte droslende proportionalventiler. Energioverførsel sker via et fritflydende stempel, som oscillerer ved hjælp af On/Off-ventiler assisteret af trykstyrede kontraventiler. Der udvikles en højfidelitets-simuleringsmodel, og en 12-trins sekvens i pumpetilstand reducerer styringsproblemet til seks ventil-tidspunkter. På grund af ikke-lineære og stærkt koblede dynamikker anvendes forstærkningslæring; systemets symmetri udnyttes til at fordoble datatilgangen pr. oscillation, hvilket afhjælper en træningsflaskehals. Politikken trænes med en belønningsfunktion, der minimerer ventiltab og søger en centreret oscillation med ønsket amplitude. Simulationer viser, at politikker med højst gennemsnitsbelønning ofte opfører sig som næsten åben kreds og reagerer svagt på fejl, hvilket giver stationære fejl i stempelposition og kammertryk samt øgede tab ved ventilskift. Eksperimenter bekræfter, at den lærte politik kan generere oscillationer, men med store amplitudeforskelle fra gang til gang og en anslået lav effektivitet. For at afhjælpe dette foreslås en ny styringsramme, der via signal- og fejltransformationer reducerer kobling og muliggjør klassisk PI-feedback til fejlkorrigering kombineret med en intuitiv eller RL-baseret forudstyring.
[This apstract has been generated with the help of AI directly from the project full text]