Comparing the prediction capabilities of artificial neural network (ANN) and nonlinear regression models in pet-poy yarn characteristics and optimization of yarn production conditions

Abstract

In the manufacture of yarn, predicting the effect of changing production conditions is vital to reducing defects in the end product. This study compares, for the first time, non-linear regression and artificial neural network (ANN) models in predicting 10 yarn properties shaped by the influence of winding speed, quenching air temperature and/or quenching air speed during production. A multilayer perceptron ANN model was created by training 81 patterns using the Broyden-Fletcher-Goldfarb-Shanno (BFGS) algorithm. The hyperbolic tangent, or TanH, activation function and logistic activation functions were used for the hidden and output layers respectively. Results showed that the ANN approach exhibited a greater prediction capability over the non-linear regression method. ANN simultaneously predicted all of the 10 final properties of a yarn; tensile strength, tensile strain, draw force, crystallinity ratio, dye uptake based on the colour strengths (K/S), brightness, boiling shrinkage and yarn evenness, more accurately than the non-linear regression model (R-2 = 0.97 vs. R-2 = 0.92). These results lend support to the idea that the ANN analysis combined with optimization can be used successfully to prevent production defects by fine tuning the production environment.