Use of input selection techniques to improve the performance of an artificial neural network during the prediction of yarn quality properties

Abstract

The performance of an artificial neural network (ANN) is affected by the number and types of inputs. The aim of this article is to study the performance of ANN algorithms, used for the prediction of cotton yarn strength, elongation, and evenness, as the input units are subtracted (skeletonized) and added to the input layer. Nineteen factors, consisting of fiber properties, processing parameters, and yarn quality properties, were used as the main source of inputs. The initial sets of inputs, which were selected on the basis of their relationship with the output factors, were 13, 13, and 12 for yarn strength, elongation, and evenness, respectively. The final sets of inputs were 14 factors for the three yarn quality properties being predicted, and the new ANN algorithms showed performance improvement of 40, 37, and 47% for strength, elongation, and evenness, respectively, when compared to the algorithms with 19 factors. Yarn twist, fiber length, and fiber length uniformity were common among the five most influential factors affecting yarn strength, elongation, and evenness, accounting for 40, 37, and 37% for the prediction of yarn strength, elongation, and evenness, respectively.