Data-Driven Insights and Predictive Analytics in Manufacturing: Machine Learning for Industry 4.0

In the continually evolving landscape of modern manufacturing, the fusion of data-driven insights, predictive analytics, and Machine Learning (ML) algorithms has emerged as a transformative and revolutionary force. This master's thesis embarks on an extensive journey through this field, delving deep into the intersection of ML algorithms and the principles of Industry 4.0. The focal point of this exploration lies in the analysis of manufacturing production data sourced from the proprietary software, Sandeza, utilized in some contemporary manufacturing operations. The thesis begins with an overview of the state-of-art of ML algorithms applied within the context of Industry 4.0. It analyzes their integration into manufacturing operations with a specific focus on classification and regression algorithms. Then, it presents some papers in which different regression and classification algorithms are compared. Furthermore, it introduces the innovative concept of imputing missing data using Autoencoders, a cutting-edge technique that holds immense potential. The thesis meticulously analyzes the huge delivered dataset, with the complexity of the relationships between tables and different entities. The purpose of our analysis leads to the creation of two distinct datasets. The first dataset revolves around Articles, while the second is centered on Production Options. Detailed statistical insights into these datasets reveal the characteristics and distributions of the features, with a particular emphasis on those characteristics that will be the target of the Machine Learning algorithms. Within the pages of this thesis, a comprehensive exposition of the methodology employed is presented. This encompassing methodology covers a range of crucial aspects, including data preprocessing techniques, like Term Frequency-Inverse Document frequency or Principal Component Analysis, the selection and implementation of ML algorithms, such as Random Forest, Support Vector Machines, and K-Nearest Neighbors, and the innovative application of denoising Autoencoders to impute missing data. The culmination of this research lies in the experimental results, which demonstrate the effectiveness of the applied methods in forecasting critical manufacturing data features. Among the ML algorithms, Random Forest emerges as a standout performer, excelling in both classification and regression problems. However, it's worth noting that the results are significantly influenced by the difficult distribution of the features and by the presence of inherent outliers within the dataset. However, by progressively deleting the outliers, the results improve up to reach an acceptable error, according to the company. The thesis lays a foundation for future work, with opportunities to further refine and optimize the results through the exploration of additional techniques and algorithms. The objective of this work is to empower manufacturing industries, particularly those leveraging the Sandeza software, by giving them tools and insights needed to drive efficiency, innovation, and success in the dynamic landscape of modern manufacturing.