Research>Pattern recognition using Neural Networks

Background

The research presented here is based on the application of Neural Networks (NN) for pattern recognition. In engineering terms, a NN may be thought of a 'black box' model that maps a set of input data to a set of output data. The model is developed by a process of training, in which the strength of connections between neurons in the NN is adapted such that the error between the actual response to input and the true response is minimised. In that sense, there is nothing 'magic' about NN - they are just a class of nonlinear regression models, albeit with some good marketing. NNs are best suited to problems where the process being modelled involves many inputs but few outputs, and there is ample example data available for training. Furthermore, if there is apriori knowledge of the model structure, e.g. from theory, then it likely that conventional parameterised modelling approaches are likely to be better. NN models are 'black box' in that the parameters they contain do not represent meaningful quantities in the process being modelled. This can lead to problems of trust and validation in NN models - it is very difficult to show analytically that a NN model is 'correct', only that it does the right thing for existing example data.

Publications

Crowther, W.J., Lamont, P.J., ‘A neural network approach to the calibration of a flush air data system’, Aeronautical Journal, Vol. 105, No 1044, pp. 85-95, February 2001.

Crowther, W.J. and Cooper, J.E, ‘Flutter speed prediction during flight testing using neural networks’, Proc. Instn. Mech. Engrs Part G (Journal of Aerospace Engineering), Vol. 215, pp37-47, November 2001.

Crowther, W.J., Edge, K.A., et al, ‘Fault diagnosis of a hydraulic actuator circuit using neural networks – An output vector space classification approach’, Proc. Instn. Mech. Engrs, Vol. 212 Part I, pp 57-68, February 1997.