Lester Lik Teck Chan

Active learning dynamic soft sensor with forward-update scheme

Soft sensors are used to infer the quality variable from easy-to-measure process variables. The conventional static soft sensor is incapable of handling the dynamic of processes. For data-based soft sensor development, with abundance of the raw sensor data, the problem of variable correlations and large number of sample are encountered. This work presents a latent variable model (LVM) based active learning strategy to select representative data for efficient development of the dynamic soft sensor model. In order to carry out data selection the uncertainty information is provided by Gaussian process (GP) model. The LVM with auxiliary GP model is developed under a dynamic framework which is suitable for dynamic processes. A forward-update scheme for updating the soft sensor model in advance is proposed so that the soft sensor is able to reflect the current status of the process and to improve the soft sensor model without waiting for the quality measurements. The proposed method is applied to an industrial fluid catalytic cracking process data.

Identification of LTV systems with cascade control loops using basis function approach

In this work, the subspace identification method (SID) based on the linear time varying (LTV) state space model is developed to identify a nonlinear process under cascade control loops. By using the basis functions, the LTV subspace identification can describe a nonlinear process mathematically the same as the LTI one. Then, like the conventional LTI-SID approach, the same two-step procedure is developed. The subspace spanned by the columns of the extended observability matrix is estimated first from the input-output data of the system. Second, the system matrices are determined directly by the extended observability matrix. Also, to deal with closed-loop data, an innovation estimation method is adopted. Finally, level cascade control for reflux drum is presented to illustrate the features of the proposed LTV-SID method in comparison with the past LTI-SID method.

Application of Gaussian processes with variable shrinkage method and just-in-time modeling in the semiconductor industry

In semiconductor manufacturing, as some of the variables in processes cannot be easily measured during or after the production, virtual metrology (VM) is employed to predict metrology outputs using ancillary process variables. However, because of changeable processes and high-dimensional inputs, VM can be expensive or difficult to implement. In this work, just-in-time (JIT) modeling is used to cope with changes in process characteristics and to automatically update the statistical model. In addition, owing to the non-uniform data distribution, Gaussian process regression (GPR) as a probabilistic approach is a typical method to enhance the robustness of the system in the probability density space. With high-dimensional input variables in deposition processes, a variable shrinkage and selection method for GPR is proposed. It is superior to the conventional methods. The features of the proposed method are shown by way of illustrative examples and the proposed method is compared to conventional work based on real semiconductor process data.

Probabilistic uncertainty based simultaneous process design and control with iterative expected improvement model

Abstract The simultaneous design and control aims to achieve economic profits and smooth operation of the process even under uncertainties. However, the over-estimation of the uncertainties leads to conservative design decisions. Because of the disturbance inputs, the cost is not easily evaluated. Unlike the past work of design and control, the proposed probabilistic approach framework directly uses the Gaussian process (GP) model to represent the uncertainty in the input. The GP model that acts as the cost function model is trained by an iterative approach. The variability can be evaluated statistically by the GP model. In addition, the expected improvement optimization is employed to select the representative data, so no redundant data are used in the modeling. The expected improvement searches for the most probable operating condition for improvement based on the predictive distribution from the GP model. The applicability of the proposed method is tested on a mixing tank.