Abstract:
Industrial processes are monitored continuously to meet the standards dictated by the market and environmental regulations. There are many process variables like temperature, pressure, flow rate which can be easily measured online through mechanical sensors and lots of data can be stored thanks to the advancements in data storage technology. At the same time, there are variables that show the quality of a product, process safety or some other restricted chemical composition. These are called quality variables and they are measured less often due to requirement of a detailed laboratory analysis. Measuring quality variable ones in a shift is a weak link in process control and monitoring. Implementing a soft sensor is a very efficient way to predict quality variables through statistical learning methods. Traditional soft sensors are built in an offline manner and used for online prediction while it requires maintenance periodically as process shifts to another state. Just-in-time learning is an adaptive method in which a local model is built when a new sample is obtained, and the model is discarded after a prediction is made. JITL outperforms traditional methods in terms of efficiency and predictive ability. The prediction performance of a soft sensor is also affected by the quality of the training data stored in the data base. Data reduction methods are used to eliminate data that weaken prediction quality and to store meaningful data for increasing prediction performance and model efficiency. In this thesis, JITL models are trained with Lasso and least squares support vector regression and three different data reduction algorithms using four different data sets. It is shown that the effect of each data reduction method changes from data set to data set, and prediction accuracy of JITL using all data can be attained using a smaller training sets. Additionally, results show that prediction accuracy of nonlinear models trained by LSSVR outperforms that of Lasso.
