Transform domain adaptive processing with neural network controlled resonator banks

Most of the adaptive processing systems are typically designed with a predefined structure. In case of abruptly changing and unstable environments this type of fixed structures pose serious limitations in processing of signals. This paper describes a hybrid adaptive processing system which can dynamically change its transfer characteristics under changeable environment. The proposed architecture basically consists of two parts, a resonator-bank digital filter (RBDF) and a neural network (NN). The RBDF provides effective implementation of orthogonal transforms recursively. The NN is utilized to perform a nonlinear mapping between its input vector U(n) and the H(n), the complex coefficients determining the transfer characteristics of the RBDF. The authors present a performance comparison of the NN controlled RBDF under different orthogonal transforms