Abstract
In this paper, the development of an adaptive neural network for the calibration of novel velocity measurement instrumentation is presented. The backpropagation-based algorithm, PROBENET, was initially developed for the calibration of multi-hole pressure probes, although at this point it has evolved to a generic-application code. The code offers distinct advantages over commercial packages (such as the Matlab Neural Network Toolkit, Demuth and Beale, 1994) in terms of maximum allowable network size, training convergence rates, flexibility in network architecture design and network optimization capabilities. PROBENET incorporates multiple activation functions per layer, as well as heuristics-based procedures for network-architecture optimization. Techniques for local minima avoidance and convergence rate improvement, incorporated into the algorithm, are: momentum, variable learning rate and Levenberg-Marquardt optimization methods.
The present study documents the performance of PROBENET when applied to the calibration of a novel 18-hole probe developed to overcome the flow angularity measurement limitations of traditional 5- and 7-hole probes. The study compared the performance of two types of network architectures: single activation function per layer and multiple activation functions per layer, and showed that the latter consistently produced a better solution in terms of convergence rate and accuracy. The use of multi-function layers and network optimization resulted in very good prediction accuracy of the flow parameters.
