Data-driven approach for control performance monitoring and fault diagnosis

Due to the industrial value, control performance and process monitoring have attracted increasing attention in recent years. However, there still exist challenges that restrict the industrial applications of monitoring technology. This dissertation presents some innovative solutions to the monitoring issues. To avoid the interactor requirement of minimum variance control (MVC) benchmark, a data-driven covariance monitoring framework is established. Relative to a user-defined benchmark, generalized eigenvalue analysis is employed to extract the directions with the worst performance. A statistical inference strategy is then developed to identify the worse or better performance directions and subspace. The covariance based indices are further derived to assess the performance degradation or improvement. To diagnose the controlled variables causing the performance change, two types of multivariate contribution methods are proposed. One is to evaluate the significance of the eigenvector loadings while the other to examine the angle between each variable and the worse/better subspace. Complementary to the data-driven performance monitoring scheme, a simplified solution to MVC benchmark is also developed. A right diagonal interactor is first factorized from process time delays and the corresponding MVC benchmark is derived with numerical simplicity. For more general MIMO processes, left and right diagonal interactors are integrated to characterize the more complex delay structure. The MVC estimation using the left/right diagonal interactors are presented. To further improve multivariable control performance, an iterative strategy of output weighting selection is proposed. Eigenvalue decomposition is implemented on the output covariance to find the directions with the largest variance inflation. A nondiagonal weighting matrix is then designed with respect to the eigendirections and more importance proportional to the corresponding eigenvalues is assigned. In addition to control performance monitoring, process monitoring is also investigated with focus on fault detection and diagnosis of multistage overlay lithography processes. In our work, a multistage state-space model for the misalignment errors is developed from the physical principles and then formulated into the general mixed linear model. Subsequently, variance component analysis is employed to estimate the mean and variance components of the potential fault sources. A hypothesis testing procedure is adopted to detect the active faults in different layers while the mean/variance estimates are used to diagnose their magnitude and orientation.