Are you an EPFL student looking for a semester project?
Work with us on data science and visualisation projects, and deploy your project as an app on top of GraphSearch.
Production quality and process efficiency are the two main drivers that lead any industrial strategy. To ensure product quality, a duality historically existed between two approaches, namely batch sampling and systematic sampling. In batch sampling, the batch homogeneity is assumed, and a subset is measured, whereas, in systematic sampling, every part of the batch is measured independently. The latter approach is too expensive to be considered. Therefore, batch sampling is implemented, which, however, involves sacrifices product quality, as the homogeneity assumption cannot always be true. Today, virtual metrology (VM) is on the brink of enabling a new standard for quality control by offering a third approach that combines the benefits of the others. VM estimates product quality while preserving process efficiency. It enables continuous systematic monitoring of production in hidden time. To do so, it leverages process data and relates them to product quality through machine learning algorithms. This thesis proposes a proof of concept and guidelines on how to design, tune, implement, and maintain a VM solution for one of the most common industrial operations, namely milling.The literature lacked a comprehensive summary and a systematic review of the state of the art in VM. This motivated the realization of a systematic, complete, structured literature review of VM based on a content analysis of all research articles in the field published before 2021. The papers are classified thanks to a new framework highlighting shortcomings and future research directions. Finally, VM's use in various industrial fields is discussed, underlining its potential for every manufacturing industry. Two of the various findings of this literature review are the lack of development of VM for machine tool operations and the lack of research on the maintenance of data-based models to ensure their long-term performance. These two research gaps are addressed in the present thesis.Thereafter, a significant step toward VM for milling operations is made, with a specific focus on the estimation of the dimensional quality of single-pass milled parts. To do so, two experiments were conducted in an industrial environment. Various combinations of recorded process variables, data synchronization algorithms, dimensionality reduction algorithms, and regression algorithms were tested to maximize model accuracy. A new baseline for dimensional quality estimation in an industrial production environment is set with a mean absolute error of 14.4 µm. Besides the proof of concept, this part of the present thesis provides directions on how to design and tune a VM algorithm for milling operations.Finally, a methodology is proposed for implementing a problem-oriented complete solution to ensure the maintenance of industrial data-based models. This solution is structured based on an operational framework, including a sampling decision system and an updating system. Each methodology step is described thanks to guidelines, while research gaps are highlighted. The methodology starts with a concept drift identification phase; thereafter, solutions are pre-selected based on the identified concept drifts. An optimization problem is then designed to select the solution that most respects costs and constraints. This section of the present thesis provides a methodology to implement a solution for the maintenance of industrial data-based models such as VM.