Anomaly Detection on Process and Equipment data through Machine Learning

In recent times, automation has become a crucial driver of efficiency and productivity across the industry, which strives to streamline operations and reduce human intervention. In this scenario, anomaly detection plays an essential role in maintaining systems’ integrity and ensuring the smooth functioning of automated processes and the quality of the resulting products. The present Master Thesis develops the topic of automatic anomaly detection, applying it to real-world case studies, derived from the specific needs of the Procter & Gamble Company, with which this work is in collaboration. The objective of the project is to develop algorithms able to perform anomaly detection on two different levels: in online data, coming from sensors and actuators installed on the lines, and in offline data, specifically in a database, containing all the centrelines of the working machines. The research methodology begins with an exhaustive review of the existing literature about anomaly detection in time series and static data, which serves as the foundation for selecting algorithms, that best suit the specific case studies. Then, the most promising models are assessed against the available data, leading to a refined selection process. Ultimately, two dedicated algorithms are developed for the two different scenarios. For real-time data derived from sensors and actuators, the Median Absolute Deviation (MAD) algorithm is employed to identify anomalous behaviours across various machines within the production line. To detect outliers in the centrelines database, a robust linear regression with Huber loss is applied to the data, followed by the computation of the absolute error between the predicted and actual values of the machine centreline. The comprehensive algorithm framework for both scenarios includes key stages, like data querying, data filtering, data cleaning, model application, and results visualization.