Neural network approach for linearization of the electrostatically actuated double-gimballed micromirror

In this paper, a hierarchical circuit based approach is used for the development of a reduced-order macro-model for a double-gimballed electrostatic torsional micromirror. The nonlinearity and cross-axis coupling of the micromirror subjected to the differential driving scheme are investigated using the proposed macro-model. The simulation results are used to train a feed-forward neural network which carries out a function approximation of the relation between the desired location and the required driving voltages. The trained neural network is then coded into MAST AHDL. System-level simulation of the micromirror together with the neural network is performed in the SABER(TM) simulator. It is found that using a feed-forward neural network, the linearity of the micromirror is greatly improved, the steady state of the cross-axis coupling is reduced to a negligible level and the transient response of the cross-axis coupling is also suppressed. This implies that introducing a feed-forward neural network would be useful to simplify the design of the feedback control system for the double-gimballed electrostatic torsional micromirror.