Neural networks and gradient boosting for predictive maintenance of a proton therapy machine

A company developed a machine in order to perform proton therapy. The machine includes a component which is subject to failures. The company wants to predict efficiently any upcoming failure of the component with the aim of doing predictive maintenance which may reduce time and costs of reparations. The company has decided to implement machine learning algorithms for doing prediction of failure events. In this thesis we show how we faced this task. First, we describe basic notions of machine learning and we introduce the three models which have been used in this project: multi-layer perceptron, extreme gradient boosting and recurrent neural networks. We show how we faced a task related to predictive maintenance: we describe the history-based feature aggregation, the trigger, the business score and a cross-validation procedure adapted for this problem. We describe how we faced the task on different steps, starting from a smaller set of data and then adapting it to new datasets. We show and comment the final results for the data coming from the different versions of the component. Finally, we make some proposals of improvement for the future.