Abstract:
A novel modeling based on deep learning framework which can exactly manifest the
characteristics of nonlinear system is proposed in this paper. Specifically, a Deep Reconstruction
Model (DRM) is defined integrating with the advantages of the deep learning and Elman neural
network (ENN). The parameters of the model are initialized by performing unsupervised pretraining
in a layer-wise fashion using Restricted Boltzmann Machines (RBMs) to provide a faster
convergence rate for modeling. ENN can be used to manifest the memory effect of system. To
validate the proposed approach, two different nonlinear systems are used for experiments. The first
one corresponds to the Class-D power amplifier (CDPA) which operates in the ohmic and cut-off
regions. According to error of time domain and spectrum, Back Propagation Neural Network
model improved by RBMs (BP-RBMs) and ENN are compared of different input signals which
are the simulated two-tone signal and actual square wave signal. The second system is a permanent
magnet synchronous motor (PMSM) servo control system based on fuzzy PID control strategy. In
terms of simulated and actual speed curves, BP-RBMs, DRM and ENN model are adopted on
comparison respectively. It is shown by experimental results that the proposed model with fewer
parameters and iteration number can reconstruct the nonlinear system accurately, and depict the
memory effect, the nonlinear distortion and the dynamic performance of system precisely.