| Summary: | 碩士 === 國立中興大學 === 機械工程學系所 === 107 === In this thesis, a recurrent neural network (RNN) is used as a controller in an active noise control (ANC) system. Real-time recurrent learning algorithm (RTRL) is adopted as a learning algorithm to adjust weight parameters of the ANC controller. An error signal obtained from an error microphone is needed and used for the parameter adjustment. In practical applications, the error signal will contain outside interference and be affected by the dynamics of the secondary path. The interference and secondary path will degrade the weight updating of the algorithm and the ANC performance of the system. Therefore, three methods are implemented in this thesis, which are used to enhance the learning efficiency of the algorithm and ANC performance. One is adding commutation error (CE) to the NFxRTRL algorithm. Another is to use batch learning in the NFxRTRL algorithm. The other is to use an exponential decay learning rate (EDLR) for the NFxRTRL algorithm. The simulation results show that after incorporating the commutation error, the convergence is improved, but the level of noise attenuation decreases 7dB. Therefore, using batch learning and replacing the constant learning rate with exponential decay learning can increase the level of noise attenuation 11dB, and the convergence rate of the system can be improved as well. In the ANC experiment, using these methods can let the controller converge in 0.5 second, and the level of noise attenuation reaches 38 dB. Together, results of simulation and experiment support feasibility of the proposed method for effective ANC applications.
|
|---|
