Prediction of Yarn Strength from Fiber Properties Using Fuzzy ARTMAP

The count-strength-product (CSP) of cotton yarn is a complex function of fiber properties and spinning performance. The traditional way of predicting yarn CSP is using linear multiple regression. The correlation coefficient between actual CSP and predicted CSP obtained from linear regression is almost always less than 0.9. In this paper, we used a Fuzzy ARTMAP network to predict yarn CSP from fiber properties and spinning performance. Fiber properties and spinning data were used as inputs to ARTa, and yarn CSP was used as ARTb input. Our objectives are: better prediction of the quality of the end product, and to determine the optimum set of fiber properties to make reliable predictions. Several experiments were designed with different combinations of fiber properties (based on the measuring instruments used in collecting those properties) as ARTa inputs. To improve relative accuracy of prediction, three voter networks were used in each experiment. During training, order of the training data was scrambled to create 3 ARTMAP networks. The ARTb templates in the voter networks indicates the range of CSP for any particular inputs to the ARTa. Since CSP is a continuous analog value, the boundary of ARTb templates is usually not fixed among the voters. To improve absolute accuracy of prediction, we took a Fuzzy OR function among the three chosen voter templates during recall to reduce the span of the range. When predicting, each ARTb template is represented by its center of gravity. In each experiment, the correlation coefficient between the actual and the predicted CSP was better than 0.95. A combination of all fiber properties from traditional and Advanced Fiber Information System (AFIS) tests made marginally better prediction than any other combination of fiber properties including when fiber properties from all the tests were fed into ARTa