Abstract: The measure method for sizing percentage is not real time method in practical production process. Therefore the yarn quality can not be guaranteed. A new modeling method is proposed for sizing percentage soft sensor real time based on multiple SVR fusion by bagging. Firstly, the main factors that influence sizing percentage are obtained by analyzing the sizing process. Then the basic SVR soft sensor models are built by different kernel functions, different loss functions and different parameters. Finally, the different basic SVR models are integrated by bagging. The basic SVR can learn from each other by this way. The new sizing percentage soft sensor model is established. The practical production data are used to test the new soft sensor model. And the traditional soft senor methods based on single BP neural network and single SVR are used to compare. The results demonstrate that the new method has the best performance. The accuracy of new soft sensor can meet the needs of practical production.

Key words: sizing percentage, sizing process, soft measuremtent, Bagging, support vector regression