Making Industrial Processes More Cost Efficient
New techniques of nonlinear modelling have come up in the last ten twelve years which have opened up new possibilities in empirical process modelling. Secondary coating is a typical process which cannot be modelled adequately by physical modelling. The effects some of the process variables are clearly nonlinear, which makes it imperative for us to take into account the nonlinearities rather than ignore them. Neural network techniques are therefore more suitable for process modelling of secondary coating. Neural networks have been used in various process modelling applications in steel industries, pulp and paper industries, chemical industries, plastics industries, etc.
Secondary coating is a plastics extrusion process, followed by controlled cooling and winding under tension (Figure 1). The properties of secondary coatings like excess fibre length depend to a large extent on the process variables and the material properties of the plastic. For a given product, the plastic material, the jelly, the external and internal diameters, and the number of optical fibres in it are fixed. The properties of the secondary coatings, then depend on the process variables, starting from tension on the optical fibres, extrusion variables, jelly temperature, cooling water temperature, line speed, capstan location, winding tension, etc.
In this work, feed-forward neural network models (Figure 2) were developed based on experimental data with process variables as inputs. The nonlinearities are visible in the neural network model (Figure 5). The neural networks used logistic sigmoid activation functions, and were found to be effective. This is a typical situation where the conventional linear statistical techniques are not effective.