In nonlinear industrial processes, some product qualities or key variables are usually difficult to measure online automatically due to the lack of sensors. In this work, a novel data-driven soft sensor technology based on multilayer perceptron neural network (MLP) with a double least absolute shrinkage and selection operator (dLASSO) approach, named dLASSO-MLP, is developed with a two-step procedure. Firstly, an MLP model is constructed through the process data set. Secondly, a double LASSO algorithm is integrated into the model to solve two redundancy problems, i.e., the input variables redundancy and model structure redundancy. The proposed method not only selects input variables that are the most sensitive to the model, but also simplifies the MLP structure by deleting redundant hidden nodes to avoid the model overfitting. In addition, the method is validated by data from simulation examples as well as an industrial application. Compared to the other neural network methods, the proposed method requires fewer neurons and presents better prediction performance.

Metrology delay is a natural problem in the implementation of advanced process control scheme in semiconductor manufacturing systems. It is very important to understand the effect of metrology delay on performance of advanced process control systems. In this paper, the influences of metrology delay on both the transient and asymptotic properties of the product quality are analyzed for the case when a linear system with an initial bias and a stochastic autoregressive moving average (ARMA) disturbance is under an exponentially weighted moving average (EWMA) run-to-run control. Tuning guidelines are developed based on the study of numerical optimization results of the analytical closed-loop output response. In addition, effective metrology delay of a variable time delay system is analyzed based on the resampling technique implemented to a randomized time delay system. A virtual metrology technique is a possible solution to tackle the problem of metrology delay. The tradeoff between additional error of virtual metrology and reduction in time delay is studied. The results are illustrated using an example of control of the tungsten deposition rate in a tungsten chemical-vapor deposition reactor. The basic conclusion is that metrology delay is only important for processes that experience nonstationary stochastic disturbance. In such a case, use of virtual metrology is justified if the error of the virtual metrology method is less than the error caused by stochastic process noise. The accuracy of the virtual metrology noise with respect to the traditional metrology is not critical, provided that the error due to metrology is much less than that due to process disturbances.

Soft-sensors attempt to predict key quality variables that are infrequently available using sensor and manipulated variables that are readily available. Since only limited amount labelled data are available, there is always the concern whether underlying physics were captured so that the model can be reasonably extrapolated, A sequence-to-sequence model in the form of a nonlinear state-observer/encoder and predictor/decoder was proposed. The observer can be trained using large amount of unlabeled data, but in a supervised manner in which the process dynamics was tracked. The encoder output and manipulated variables were used to train the quality predictor. The model was applied to product impurity predictions of an industrial column. Results show that good predictions and excellent consistency in the sign of estimated gains can be achieved even with limited amount of data. These findings indicated that the proposed sequence-to-sequence data-driven approach is able to capture the underlying physics of the process.