Detection of drifts in defectivity: case study of an industrial production chain through data analysis and machine learning applications

The need to improve the production chains of industries more and more is increasingly important nowadays. Most of the times there is the necessity of analysing the current production pipeline and then try to anticipate future behaviours of the production chain, in order to save resources, such as time and money. This is mainly made thanks to big data, that are the fuel of the Industry 4.0: every step of a production chain can be strengthened by big data analysis, from the management of raw materials and suppliers, the production itself, to the distribution of finished product to the consumer throughout the retailer. The focus of this thesis work is the analysis of a multi-stage production chain, since a growing amount of steps, machines, and the related complexity make these type of systems prone to failures. In order to help the system to offset this downside and to achieve a zero-defect manufacturing, a data-driven approach is needed. Starting from some basic data visualizations, a little cleaning and preprocessing have been done, in order to better represent the data. Then, different aspects of this multi-stage production chain have been analysed, always as a function of the time variable: from the analysis of the frequency of faulty processes with the respect to all processes to see how it changes during the months taken in consideration, to the analysis of association rules between faulty steps, with the aim of checking if any of the steps are related between each other, and if yes, how this correlation evolves in different time slots. As a result, it was found that some of the combination step-machine had a higher fault frequency than others: sometime it was related to the period of time and sometimes they were coming from external causes, confirmed by company's people in charge. Regarding the second phase of analysis, it has been discovered that some steps were wrong further times together, rather than when they were failing alone, meaning that the antecedent outcome maybe influences consequent's one.