A study on optimum electrical capacitance tomography data for intelligent system recognition of flow regime

Artificial Neural Network (ANN) has been shown to be a robust tool for intelligent recognition. In every intelligent recognition task, the first dilemma would be the correct size of training data, since too few is definitely not sufficient while too many may lead to over training problem. This paper presents a study on the dimensionality of training data for a Multi-layer Perceptron (MLP) neural network. Training data size is an important criterion in ensuring a well-developed MLP. In the study, thousands sets of simulated Electrical Capacitance Tomography (ECT) data had been used to trained several MLPs using the Levenberg-Marquardt (LM) algorithm to recognize gas-oil flow regimes. The performance of the MLPs had been assessed based on the training time and percentage of correct recognition. The results reveal that the number of training data used significantly affect the performances of a MLP. In addition, an optimum number of training data ensures an optimum MLP size.