Machine Learning Techniques to Improve Bending Machine Performance

The importance of a correct prediction is crucial to save time and money. For this reason in this thesis is developed a machine learning algorithm to solve a multi-output regression problem for the bending machine "TruBend Center 7030" for "Trumpf Macchine Italia". In particular the quantities that are predicted are: the correction angle with which is adjusted the trajectory computed during the bending process and two parameters used to counteract the crowning phenomenon. The first part of this work consists in creating a parsing algorithm needed to read the files containing the data and elaborate them to obtain a suitable dataset. In the second part are used different techniques like: linear least squares, decision tree, SVR, random forest and neural networks, to solve the regression problem. Then are compared and analyzed the results obtained from the various techniques applied. After having evaluated the best one (the one with the lowest root mean squared error), it is tackled the goal of implementing the regressor on the bending machines. For this purpose it is used a Python script to handle the update of the dataset and the training of the regression algorithm. Then it is implemented the best regression algorithm in a "C#" script where, given an input sample, it is obtained the prediction of the correction angle and crowning parameters. These are used to obtain an improved result of the bent piece.