On‐Line Prediction of Delamination in Drilling of GFRP by Using a Neural Network Approach

The present contribution approaches the problem of predicting delamination in drilling of GFRP (Glass Fiber Reinforced Plastics) in order to reduce the damage on laminates. For this purpose, a software system has been developed and tested with the scope of performing an on‐line prediction of the entity of the damage both at the entry side of the tool (peel‐up damage) and at the exit side (push‐out damage). The prediction is made possible by means of feed‐forward neural networks, which are able to give a “measure” of the incipient damage process in function of the cutting conditions such as: feed rate, tool size, cutting forces, etc. In order to find the best solution in terms of performances, two architectures of neural networks have been proposed and investigated: the former is able to sort the predicted delamination in four categories (no damage, low, medium and high damage), the latter is able to directly predict an average value (in mm) of the damage. After a brief survey on related works, the paper describes the method used to determine the architectures of the neural networks, as well as the procedure used to train and validate them using the observations made in preliminary drilling tests. The results obtained in these investigations are also given and discussed. Furthermore, a description of a possible implementation of this neural system on a CNC machining center is briefly reported.