X. Du, J. Yan
University of Massachusetts, Amherst,
United States
Keywords: web tension measurement, R2R, vibration, moving speed
Summary:
Tension measurement and control of a moving web span are critical for the quality control of a roll-to-roll (R2R) printing process. However, in many practical R2R printing systems, it is not easy to install and accurately measure and control web tension straightforward using a load cell. Instead of measuring the web tension by a physical sensor, an alternative approach is to use a contactless measurement of out-of-plane web vibrations to derive tension in the moving web for control. The coupled web moving speed, tension and out-of-plate natural frequency makes it difficult to analytically model and accurately calculate the tension of a web span for control from moving speed and out-of-plate natural frequency, especially when the web span length-width ratio is low (<=3). In this project, we propose to use an optical displacement sensor to measure the out-of-plane position of the moving web wherefrom a set of web natural frequency can be extracted for each setup of web material, thickness, width, and length, by a Fast Fourier Transform (FFT) algorithm. These measured tensions, moving speeds, and frequencies are used to train a learning model for predicting a natural frequency of any given pair of moving speed and measured frequency. We experimentally test two tension prediction methods, including polynomial model and artificial neural network (ANN) prediction, for a variety of moving webs of low length-width ratio. Our experimental web handling part consists of two motorized rolls and two idler rolls. The motorized rolls are used to control web tension and web speed, respectively. A speed encoder and a load cell are mounted separately on both idler rolls for measurement of the web moving speed and tension for training and testing purposes. Meanwhile, using the real-time web moving speed and web tension, fully-closed loop controls of web speed and tension are applied for data acquisition. The web span length-width ratio is 2 in the experiments. First, we measured the web’s natural frequencies in 90 different combinations of web tension in the range of [20N, 40N] with the increment of 2.5N, and web speed in the range of [1 in/s,10 in/s] with the increment of 1 in/s. Second, we used the acquired data to train the two prediction methods. The polynomial model of degree 3 was applied. For the ANN prediction, a feedforward neural network was used, including one hidden layer of 10 neurons. Finally, we used 10-fold cross-validation to evaluate the performance of each method. The results show that the web tension can be accurately estimated from the measured frequency and web moving speed. For the polynomial model, we obtained a 1.85% tension measurement error. For ANN prediction, the minimum measurement tension error we achieved is 0.88%. Both results are far above the desired precision values of industrial tension measurement, which is 10%-30% error.