A Machine Learning Technique for Predictive Maintenance and Quality in Cut Glass Machinery

The study presented in this thesis work is based on the development of a machine learning project applied to the particular case of the 548 Lam machine, a cut glass machine produced by the company Bottero s.p.a., which collaborates with this thesis. The work describes how to develop a project based on machine learning and how it can be applied in a real case. The thesis is in fact divided into two distinct parts, the first more didactic in which are explained the various steps to be followed to prepare the data and apply different algorithms of machine learning to maximize the results, the second instead shows how to use in a concrete way the results of the study of machine learning in order to effectively increase productivity. To do this, two problems have been selected to be solved, indicated by the company. The first study tries to predict machine stops and anomalous failures. The work includes an accurate understanding of the machine and the starting dataset, and then concentrate on the data preparation. Therefore, all the various steps to be followed during the pre-processing phase are listed: how to perform data merging, correlational and statistical analysis to find important information from the data, feature selection, how to manage categorical features. When the initial data have been properly manipulated, we proceed to the evaluation phase of the selected algorithms based on the characteristics of the dataset, evaluating which model obtains the best results in predicting machine errors. The results obtained show that the probabilistic algorithms based on the tree classification are those that best predict errors. In details, the Random Forest Classifier proves to be the best model obtaining an F1 score of about 50% on the positive prediction. The second part deals with the prediction of machining times based on the characteristics of the work to be performed. This second part want to show how a machine learning model can be transposed from the mere field of study to a real application, so there is a reconstruction of the prediction function inside the machine itself, in order to make real time predictions based on the data selected by the user. The linear regression model was first trained on a PC thanks to the data contained in the dataset and then reconstructed in the software interface of the machine thanks to the coefficients listed in an orderly JSON file that is first created during the training of the model and then passed to the machine that is at this point able to make predictions. The results show that the time prediction based on this model achieves an error of less than 10% with respect to the actual measured values.