Recipe prediction of melange yarn using modular artificial neural network

Melange yarn is produced by mixing some pre-dyed fibers, as a part of melange yarn manufacturing process, predicting the recipe, i.e. a list of percentages of pre-dyed fibers, is the most important and difficult step. Artificial neural network (ANN) is considered as a more effective recipe prediction method than the traditional model. In order to improve the application performances of neural network in recipe prediction of melange yarn under the condition of larger data-set, this paper proposed a new framework to carry out recipe prediction of top dyed melange yarn from reflectance spectra using the concept of modular artificial neural network (MANN), which decomposed the whole data-set into different units by taking a dyed fiber as a module unit. Compared with ANN, MANN showed a clear superiority in correlation coefficient, training execution time, as well as root-mean-square error of recipe. The average CMC (2:1) color difference of the predicted spectrum obtained with MANN model was 1.26, which was much lower than that obtained with ANN (similar to 3.94), indicating that MANN model was more accurate than ANN. Even the color differences were not small enough in practice, it still could be used as a recognition method to find out the main compounds of melange yarn, which would be very helpful for accurate color matching.