ABSTRACT
Introduction The advances in the various automatic data acquisition and sensor systems continue to create tremendous opportunities for collecting valuable process and operational data for several enterprises including automobile
manufacturing, semiconductor manufacturing, nuclear power plants, and transportation. These technologies have also made it possible to infer the operating conditions of critical system parameters using data from correlated sensors. This approach is called inferential sensing. Inferential sensing is the prediction of a system variable through the use of correlated system variables. Most online monitoring systems produce inferred values that are used to determine the status of critical system variable. The values can also be used to monitor drift or other failures in the system, thereby reducing unexpected system breakdown. Several approaches have been used for inferential sensing including regularization methods (Hines et al., 2000) and support vector regression (Omitaomu et al., 2007). However, these methods assume that the training data are collected in a single batch. Therefore, each time a new sample is added to the training set, a new model is obtained by retraining the entire training data. This approach could be very expensive for online monitoring applications.
