Online identification methods for a class of Hammerstein nonlinear systems using the adaptive particle filtering

Hammerstein structure is commonly used for describing nonlinear dynamic characteristics, and its identification is a basic premise of nonlinear system analysis and control. This paper investigates online identification methods for a class of Hammerstein nonlinear systems, which consists of a nonlinear memoryless element followed by a linear output-error subsystem. The unmeasurable noise-free output of the linear subsystem makes the model parameters cannot be directly estimated by traditional identification methods. To address this difficulty, by using a series of weighted particles to adaptively approximate the posterior probability density function of the unmeasurable noise-free output, this paper proposes a particle filter-based stochastic gradient algorithm. Moreover, to enhance the data utilization and estimation accuracy, a particle filter-based multi-innovation stochastic gradient algorithm is developed through the innovation expansion technique. The simulation results demonstrate that compared with the existing benchmark algorithms, the proposed algorithms need a little more computational time due to the introduction of the adaptive particle filter, but they have the improved identification accuracies.