Rapid amplitude control of high-power ultrasonic transducers using deep reinforcement learning

Abstract Rapid response control of ultrasonic transducers is crucial for applications such as welding, machining, aeronautics, and semiconductor manufacturing. Traditional control methods, using voltage amplitude and frequency adjustments, perform well but often require complex circuitry to adjust the voltage, making them less practical. This study proposes a quick and accurate method for controlling current amplitude through frequency adjustments using deep reinforcement learning (DRL). The system was trained to adjust the frequency based on real-time feedback of the current states. Experimental validation with a Langevin transducer shows that the DRL system achieves near-optimal performance with faster and robuster response than PID control, effectively managing nonlinearities such as hysteresis and the jump phenomena. Future research may include extending this approach to multi-mode transducers and enhancing robustness to external condition change. These findings underscore the DRL system’s potential as a practical alternative to conventional methods, in managing the complex nonlinearity of high-power ultrasonic transducers.