By Nadeem Qazi, Hoi Yeung, Yi Cao
Keywords: Gas liquid cyclone; Separation Efficiency; Inverse Neural Network; Mutual Information; Stacked Neural Network.
The work presented in this paper demonstrates a method to obtain an inverse function of a trained neural network by connecting it with a PID controller in a close loop. This scheme is implemented to predict a process parameter controlling the efficiency of a compact axial flow gas liquid separator (I-SEP). The data is taken from an experimental study of I-SEP (compact separator) with air-water two phase flow. It was found during these experiments that by manipulating the pressure difference between the two outlets of separator and the inlet, the performance parameter i.e. Gas Carry-Under (GCU) and Liquid Carry-Over (LCO) could be controlled but nonlinearly. It requires a tedious job to set the differential pressure between tangential and axial outlet to control the GCU. A stacked neural network model consisting of several individual neural networks having different architecture is developed. The inverse function of the combined neural network was then determined by connecting this trained neural network with a PID controller in a closed loop, which is then used to predict the pressure at the two outlets of the I-SEP for a given GCU. The optimal weight determination techniques for stacked neural network is also studied and compared in this paper.