Colloquia

Summary

In the manufacturing industry of high precision computer numerical control (CNC) machining, there is a constant demand of higher precision and increase of production speed. In order to fulfill this demand, the present research aims to investigate the beneficial use of machine learning techniques in this industry. To be more specific, this research aims to predict the product quality during the making of the product based on the relevant measured process parameters. The motivation for this research is to get better insights in the process, reduce scrap, increase up time and realize predictive maintenance of the CNC machine. The most influential process parameters are temperatures in the machine, forces, positions of the machine axis, tool age and the machine state. The machine state which is the geometric error of the machine, is constantly changing over time due to wear and incidents like collisions. In order to map these process parameters to the product quality, two different machine learning algorithms are tested and analyzed, namley FNN and LSTM network. The used product of this thesis consists out of multiple holes and the FNN maps the process parameters to each of these holes individually. The LSTM network predicts the geometric feature of the holes in one, where the prediction of the hole is dependent on the prediction of the previous holes. The proposed neural networks showed equal performance where the prediction error of both networks had a standard deviation of 2 µm. Although both networks have equal performance it is expected that the LSTM network outperforms the FNN due to the low amount of training data available for the LSTM network. In the future the size of the data set can be increased and thus the LSTM network can reach higher performance after better training of the network. The main findings of this thesis are that it is possible to predict the product quality where the prediction error is has a standard deviation of 2 µm. Furthermore the most important process parameter is the machine state. This research showed the relation between the process parameters and the product quality. The next step is to elaborate the model to multiple products and machines. Another step is to find an inverse of the model to predict the machine state, to indicate when the machine needs to be re-calibrated.