Calibration transfer methods for feedforward neural network based instruments
The calibration transfer problem examined by this thesis is that of attempting to exploit the knowledge of an initial instrument calibration model so as to obtain a second similar model, having acceptable accuracy, with less data than that used to obtain the initial model. This thesis considers instruments whose calibration model are based on a feedforward neural network, FFNN. The recalibration of these instruments has raised concerns regarding the significant quantity of data needed to perform their recalibration. Calibration transfer methods provide an alternative to recalibration which can reduce the data needed for a recalibration while maintaining acceptable levels of calibration accuracy. Currently no reported methods of calibration transfer exist for the FFNN based instrument. This thesis develops a number of calibration transfer methods that allow a recalibration using less data than that needed in a recalibration employing conventional backpropagation learning. First, a simple non-learning method is introduced. A new method is developed based on a supervised learning algorithm employing a measure that learns the 'n'th order partial derivatives of the desired calibration model provided by the calibration data. Finally, a simple unreported method of initialising the weights of a FFNN so as to begin learning from a point on the error surface that provides the approximation of a previously obtained calibration model is described. Using computer simulations, the calibration error associated with using these calibration transfer methods are compared to the error obtained from a recalibration using conventional backpropagation learning. The simulations varied the numbers of neurons, number of calibration points, and similarity between calibration models. The desired calibration models were selected from 8th order polynomials and bandlimited normal random processes. The simulations indicated that no one method of calibration transfer provides the least calibration error but it is possible to achieve a 2 to 1000 fold decrease in the median calibration error relative to that of the standard recalibration while using half the calibration data. The results revealed that it is difficult to predict whether a specific set of calibration conditions will achieve a reduction in calibration error.
Doctor of Philosophy (Ph.D.)