Statistical Modeling of Separator Processes - An Application of Gaussian Processes with Bayesian Optimization

Abstract: The separator is a machine with many applications, commonly used to separate liquids or solids into components with different density. Each application demands its own unique set of process parameters to achieve optimal results. Often the procedure of finding the best process parameters is conducted empirically, which can be very time consuming. This thesis aims to address this problem by providing a statistical model of separator processes, which can be used to find the optimal process parameters more efficiently. Different sensors are mounted on two separators used in regular production. The sensors measure values of the inputs, outputs and process parameters. A Gaussian process is used to model the regression relationships between the process parameters and outputs for two separators. Bayesian optimization is then used to find optimal process parameters, which are shown to be accurate in simulations. In four models, one for each output of the two separators, the optimal process parameters are seen to improve the outputs. In the first separator only small improvements can be seen, as the optimal process parameter is near the middle of the data used to build the model. In the second separator large improvements can be seen. Here, the optimal process parameter is at the upper endpoint of the interval, implicating that a higher value of the process parameter could further improve the outputs. Thus, further experiments with a higher value of the process parameter are needed in order to draw conclusions on the optimal process parameter for the second separator. These optimal process parameters will be used in the real separators to possibly improve the separator performance. The data used in this thesis is supplied by the manufacturer of the separators, Alfa Laval.