An improved gradient optimization scheme for Hammerstein system with memory nonlinearity and quantized data

The typical Hammerstein model is widely used for system modeling various nonlinear systems due to its miserly structure and powerful modeling performance. Nevertheless, its performance in modeling real industrial processes with memory feature is inferior because the nonlinear sub-model in such Hammerstein system is memoryless. In order to enhance the dynamic description capability of the classic Hammerstein system, this paper investigates the parameter identification problem of quantized nonlinear Hammerstein systems with memory sub-model by proposing an improved gradient optimization scheme. For addressing the issue of the multi-innovation gradient algorithm (MI-GA) performance due to the innovation length limitations, this paper modifies the addition pattern of noise information by converting batch noise information to multi-single way one, improving estimation performance. Then, a triggering mechanism is employed to solve the local extremum problem in the MI-GA optimizer, making the estimator less sensitive to the initial values. By applying stochastic theory, the convergence performance of the parameter estimator is rigorously proven. Finally, validation examples are used to test and demonstrate the feasibility of the proposed improved gradient algorithm.