Abstract
It is proposed that the Hough accumulator array technique, used in image processing to automate the extraction of geometrical primitives from image data, be used also to learn about structure present in symbolic data. By way of example, a generalization of the accumulator method is developed for use in logistic trend extraction. It differs from earlier generalizations of the Hough transform by incorporating early data fusion. It is shown that the method can provide unsupervised learning capability and deal with noisy data and data corrupted by gross errors. The multipoint accumulator is shown to be capable of generating an early warning of sudden change in trend, and the importance of this capability for condition-determined maintenance and control is stressed. An application of the method to real condition-monitoring data is outlined. This example highlights the necessity of carrying out good diagnostic feature extraction before attempting to automate the prognostic analysis of a trend.< >