Artificial neural networks for system modeling, monitoring and control

This dissertation introduces a new dynamic network architecture called the Dynamic System Imitator (DSI). It is especially designed to mimic the behavior of a wide range of dynamic systems. It is also designed such that its basic building neurons have a good approximation of the known functional organization of the human brain neuron. The DSI is a three layer network, with neurons in the hidden layer fully connected to every other neuron in the hidden and output layers and also to itself. Time delays are simulated in the network, using simple integrators at certain positions. Those time delays and feedback connections are very important to the dynamic behavior of the network. The DSI is a dynamic deterministic neural system that is designed to have enough flexibility to be arranged to have similar behavior to a wide class of other real deterministic systems. Two different training algorithms have been designed to train the DSI. One is based on a one-dimensional minimization. The other is based on a multi-dimensional minimization. The characteristics of both algorithms have been discussed. The feasibility of using the DSI to model linear and nonlinear systems has been studied. It has been adopted for check valve monitoring application in which the DSI was used to model a very complicated nonlinear relationship between two different vibration signals measured by accelerometers mounted on two different positions of the check valve. A control strategy using the DSI was developed. This control strategy, was applied to control the chaotic behavior of the Lorenz system. This application showed very interesting results that demonstrated the ability of the DSI for dynamic system control.