Improved Neural Network Performance Using Principal Component Analysis on Matlab

Most of the real-world data samples used to train artificial neural networks (ANNs) consist of correlated information caused by overlapping input instances. Correlation in sampled data normally creates confusion over ANNs during the learning process and thus, degrades their generalization capability. This paper proposes the Principal Component Analysis (PCA) method for elimination of correlated information in data. Since it has been wellknown that Electrical Capacitance Tomography (ECT) data are highly correlated due to overlapping sensing areas, the PCA technique has been examined on the ECT data for oil fraction estimation from gas-oil flows. After application of PCA, the uncorrelated ECT data were used to train a Multi-Layer Perceptron (MLP) ANN system. Then, the trained MLP was tested upon unseen ECT data. The results demonstrated that the elimination of correlated information in the sample data by way of the PCA method improved the MLP’s estimation performance and reduced the training time.