Nonlinear Fault Identification Index Based on Variable Reconstruction

This paper describes a new nonlinear fault identification procedure that could be more accurate than a linear one when dealing with fault identification in chemical processes. A Principal Component Analysis based model is proposed to describe the process behavior and the fault detection and identification tasks are based on the Squared Prediction Error statistic and a variable reconstruction method, respectively. The nonlinear mapping of the fault signature is made by nonlinear PCA based on principal curves and neural networks, as proposed by Dong and McAvoy (1995). The well-known Tennessee Eastman problem is used as a benchmark in order to simulate realistic faults and compare the proposed nonlinear approach to a linear one. Results have shown the potentiality of the nonlinear approach as it easily identified fault that could not be identified by the linear one.