Application of Neural Networks in the Detection of Breaks in a Paper Machine

Modem paper machines are very large and complex. The good runnability of a paper machine is very desirable but hard to achieve. Breaks are one of the most important problems considering the runnability and the productivity of a paper machine. In this paper we discuss the methods to detect situations, where the probability of a break is exceptionally high. Operators have usually some kind of feeling about situations, which can increase the probability of a break, but it is impossible to find a unique set of features which cause breaks. Our approach is to collect data from a production paper machine and then analyse the data. The measurement points have been selected in co-operation with the paper machine staff. Principal component analysis has been used to visualize the occurrence of breaks and the movements of the state of a paper machine. Because very little unique knowledge about breaks is available, we use unsupervised neural networks to classify the measurement data and to find situations which are sensitive to breaks. Kohonen self-organizing feature maps are suitable because they create classifications in which classes of similar features become near each other. This paper has two major contributions. First the paper describes the mathematical results for the detection of paper machine breaks and second the paper presents a concrete application, not a simulation study, of the use of artificial neural networks.