Improved extension neural networks for lead-acid battery modeling

Official URL: http://192.168.1.20/record=b1306375 (Table of Contents)

There is an increasing demand for man-made dynamical systems to be reliable and safe. If a fault can be detected quickly, appropriate actions should be taken to prevent critical accidents, high cost malfunctions or failures. The key point in fault diagnosis is the assumption of the availability of good mathematical model of the plant. Mathematical modeling of non-linear dynamical systems may be computationally hard and time consuming. Therefore, modeling the plant using machine learning methods such as Neural Networks (NN), fuzzy logic, extension neural networks (ENN) can be more advantageous. Although a dynamical system is modeled via machine learning methods, there can be non-measurable states which are used in the system. Even though they are estimated with mathematical approaches, they can drift in time. Classification methods can be applied totally or to initialize the mathematical estimation. Although ENN is one of the promising classification methods, it sometimes gives poor results due to insensitivity to scatter of data-points. Its shifting and updating property takes more iterations than comparable methods to give an acceptable error rate. In this thesis, we propose improved extension neural networks (IENN) which improve on ENN's linear clustering method by using quadratic clustering and generating clustering criteria which depend on statistical properties of the training set. Rechargable Lead-Acid Battery is modeled via feed-forward NN approach and its state of charge is classified via proposed IENN method. The proposed method produces more accurate classifying results than ENN.