Using machine learning to identify flow regimes from capacitance sensor data

Abstract

In this thesis the k-means clustering and a neural network is developed and used to classify capacitance data from multi phase flow in a horizontal tube. Theoretical background for the unsupervised machine learning algorithm: k-means clustering and for the supervised machine learning algorithm: Neural network with one hidden layer is presented. Data acquisition method and analysis of the multi-phase flow data is discussed. The machine learning algorithms are created in Matlab in a general manner so that the programs will work for input of varying sizes. The k-means algorithm is used as a method for clustering provided data examples in flow regime clusters. The algorithm fails to provide rigid clusters which match observations at phase transitions, but works well as a general indicator of ow regime clusters. Classifications from the k-means algorithm and a set of manual classifications is used as input in the neural network for training and testing. The neural network provides overall good results, and shows its ability to detect complex patterns.