Interpretable Anomaly Detection for Automated Filling Systems through Machine Learning approaches

In the Industry 4.0 scenario, the rising adoption of new technologies such as Internet of Things (IoT), Artificial Intelligence and Big Data Computing is shaping the future of manufacturing and industrial automation. In this context, an increasingly central role is played by Machine Learning methodologies, whose application allows to perform highly complex tasks with impressive degrees of speed and accuracy. However, one of the major obstacles to the diffusion of these techniques is often represented by the lack of Interpretability, which generally accompanies the use of more complex and performing models. Therefore, the use of Machine Learning algorithms which are Interpretable, allowing the user to easily understand the reasons behind the returned outcome or prediction, appears to be increasingly necessary in order to enable a proper exploitation of their potentialities. The focus of this thesis is the application of Machine Learning techniques to achieve a task of considerable importance in the industrial scenario, namely Anomaly Detection. This thesis, in particular, will illustrate the development of an Anomaly Detection system for industrial filling machines. More specifically, the system has been designed for the analysis and monitoring of the automated processes that perform the sanitization of the machines’ components, whose correct execution plays an important role in guaranteeing the hygienic safety of the finite product. During the development, particular importance has been given to the concept of Interpretability, making the results returned by the system as comprehensible as possible for the end-user, also by designing a proper graphical representation, which provides useful information on the factors and events that could have induced the model to return a specific outcome.