Abstract:
This thesis presents a quality-forecasting model based on neural network for the paper making industry with different source data transaction processes. The paper quality test and control plays an essential role in the paper making industry, which affects the whole operation process and the future paper market. Compared with other paper quality indexes, paper curl is closer to terminal clients and more difficult to pretest and control in the actual working environment. Large-scale data from production database, which would potentially affect final paper quality, have been cleansed and abstracted. Modeling based on MLP neural network was designed to compare between Quasi-Newton algorithm and Double Dogleg with early stopping regularization in different source data sets. With bootstrap accuracy estimation, the final result has been evolved which would annotate the relationship between workflow data and paper curvature in a more constructive way.