An input-training neural network-based nonlinear principal component analysis approach for fault diagnosis

In this paper some problems existing in the linear principal component analysis methodology are discussed firstly. A nonlinear principal component analysis methodology based upon input-training neural network is presented for process fault diagnosis. The learning algorithm of input-training neural network is modified to improve its learning speed and avoid oscillation during learning. Then, input-training neural network and BP neural network are used to estimate the nonlinear principal component scores. Fault detection and diagnosis is performed by means of statistical methods like Hotelling's T-2 and Q. Finally, the simulation research to continuous stirred tank reactor (CSTR) is performed to show its advantages in extracting the nonlinear features compared with the linear principal component analysis methodology.