Predictive Maintenance aims to anticipate breakdowns on industrial equipment using data analysis and machine learning methods. Therefore, industrial machines are designed to provide
a daily set of data in various formats. The goal of this PhD thesis research is to exploit log events data to predict the occurrence of several critical errors on machine tools.
In this thesis, we propose a supervised classification based approach dedicated to critical errors.
In this perspective, a history of critical and non-critical errors has been collected during one year on a fleet of textile industry machines. We explored different classification methods such as naive bayes, Decision Trees, SVM, and finally an artificial neural network, the latter allowed us to obtain the most reliable predictions. The model tackles the problem of imbalance classes in the data and produced high quality results for a specific critical error with a F1-score of 0.8.
To the purpose of labeling ambiguous data, we have developed a data structuration approach inspired by the Multiple Instance Learning paradigm. The data are organized in a structure
referred to as bags.
Finally, to capture and analyze the temporal error dependencies, we explored learning methods based on recurrent neural networks.