Identification of secondary path in ANC using diagonal recurrent neural networks with EKF algorithm

This paper presents theoretical and experimental modeling of secondary path of an active noise control system in free space by using recurrent neural networks. A learning algorithm for diagonal recurrent neural networks is developed based on Extended Kalman Filter and is referred to as Diagonal Recurrent Extended Kalman Filter algorithm. The neural network structure and its algorithm are applied to handle nonlinearity of the secondary path. To put the neural identification task within the context of ANC, a new control algorithm based on DREKF is also presented. The real-time experiment, however, is performed only for identification task. Experimental results using a floating point DSP show that the number of neurons in neural network can be reduced by introducing the diagonal recurrent elements, without deteriorating the identification system performance.