Levenberg-Marquardt algorithm for nonlinear principal component analysis neural network through inputs training

Nonlinear principal component analysis (PCA) through inputs training neural networks (IT-nets) based on gradient descent algorithm is effective in coping with the intrinsic nonlinearity in realistic processes. However, the gradient-based method suffers from the slow convergence behavior after the first few iterations and thus greatly affects its practicability in many cases. In this paper, Levenberg-Marquardt algorithm is introduced to accelerate the training of inputs of the IT-nets. Its efficiency is demonstrated through application to the nonlinear dimensionality reduction of data from an industrial fluidized catalytic cracking (FCC) plant.