Application of Neural Networks for Estimation of Paper Properties Based on Refined Pulp Properties

Abstract

The main objective of the work presented was to determine the possibility of the prediction of paper properties based on refined chemical pulp properties using the neural network approach. Three main parameters related to basic refining effects were used: pulp and fibre WRV, the amount of fines and the average fibre length. These parameters were used for prediction of the following paper parameters: apparent density, breaking length and tear resistance. The classical multilayer perceptron with one hidden layer was used. The number of inputs and outputs was related to that of input and output variables. The size of the hidden layer (number of hidden neurons) was determined experimentally. The Levenberg-Marquardt algorithm was used as a training method. The available dataset was divided into two groups: 90% of experimental results were applied as training data and 10% for model verification. As a result of the trials conducted, a satisfactory level of the correlation between simulation data and experimental data was obtained. Results allow to presume that the method presented could be adapted for other papermaking pulp grades as a general control system in the industrial refining process. In such a case, the accuracy of the presented method could be even higher because of the large number of data available on-line. These data could be used as in a real-time training procedure, which would significantly improve the precision of the whole system. The lack of other effective methods of paper property prediction makes the method proposed an attractive solution to the problem presented.