Abstract: The joint modeling scheme of several main nonlinear characteristics of piezoelectric actuator is studied, including hysteresis, creep and temperature drift. A cascaded single hidden layer feedforward neural network based on NARMAX model is proposed to eliminate the influence of hysteresis. The information criteria and error reduction ratio algorithm are used to determine the regression factors which have the greatest influence on the model error as the input nodes of the network. The experimental results show that the absolute error of the network on the test dataset can be reduced to no more than ±0.1 μm via the multinetwork generalization and regularization. By leveraging the running time, the measured value of the temperature sensor and the frequency of input voltage into the input nodes, the nonlinear factors caused by creep and temperature drift can be compensated, and the final absolute error on the test set can be limited to ±0.01 μm. The network has good generalization ability for different excitation voltage frequencies. The research results can be used as reference in the modeling of multiple nonlinear coupling piezoelectric actuators.

Key words: piezoelectric actuator, multiple nonlinear coupling, NARMAX model, neural network model