1997 Volume 33 Issue 8 Pages 759-765
Assurance of surface glossiness is one of the most important factors in the production of stainless steel strips, but the surface glossiness has been evaluated by off-line measurement and visual inspection. For the purpose of on-line evaluation of the surface glossiness, an optical measurement method of the surface glossiness and a neural network-based discriminant method have been developed.
For the optical measurement, the light reflection from a mercury lamp and an Ar laser are adopted. The former is concerned with information on relatively large irregularities and is effective in low glossiness areas. The latter is concerned with information on irregularities with the size of the wavelength of light and is effective in high glossiness areas. The combination of the two allows accurate classification of wide range glossiness on the stainless steel surface.
Because the relation between those optically measured data and the visual surface grade shows severe nonlinearity, the neural network is applied to derive the mapping function. The feedforward-type neural network with back-propagation learning and the LVQ neural network are compared by preliminary off-line experiments from the viewpoint of the ability to interpolate and extrapolate the distribution of training data. Performance of the former is dependent upon the decision of learning convergence and it does not guarantee the proper classification for unlearned data. On the other hand, the LVQ network can approximate the distribution by placing the reference vectors. The learning algorithm of the LVQ network is modified to get the proper classification, thereby improving the accuracy for unlearned data.
Experimental results show that the neural network classifier performed with higher accuracy than the linear discriminant function that uses a conventional statistical method. The developed system has been successfully applied to quality assurance in an actual stainless steel cold processing line.
