Quality-Related Fault Detection for Nonlinear Processes Based on Neural Orthonormal Subspace Analysis

Abstract

Quality-related fault detection is very important for the safe operation of modern industrial processes. Orthonormal subspace analysis (OSA), as an advanced quality-related fault detection method, has achieved excellent detection results. Although OSA effectively divides process data into quality-related and quality-unrelated subspaces by analytic solution, its application is often limited by the nonlinear feature of actual industrial processes. Therefore, a algorithm of nonlinear process and quality-related fault detection based on neural orthogonal subspace analysis (NOSA) is proposed in this paper. First, the radial basis function (RBF) neural network linearizes the data through the mapping of Gauss function. Among them, the back propagation training algorithm is used to train the center and width by minimizing the prediction error. The hidden layer output is used as the extension of the initial process data. Then, OSA is performed on the extended matrix containing the original process data, RBF hidden layer output, and bias term to achieve quality-related fault detection. Finally, the simulation results in Tennessee Eastman benchmark demonstrate that the presented algorithm has good performance in fault detection.