INVITED SPEAKER

Moussab Orabi is currently a PhD Student in Data Science and Engineering at University of Lille and Data Scientist &Projectlead BI, Rosenberger Group, Germany. He got the Master degree in Data Science and Big Data at Higher Institute for Applied Science and Technology in 2015.

Up to now, he has had 6 publications including thesis, technical reports and conference papers.

Title: Innovations in Anomaly Detection for Catalyxing Operational Excellence in Complex Manufacturing Processes

Abtract:

In the evolving landscape of Smart Manufacturing and the increasing complexity of manufacturing processes, this thesis tackles the critical challenge of anomaly detection, a field gaining paramount importance due to increasingly complex customer requirements. As manufacturing systems become more intricate, the incidence of anomalies and deviations in processes escalates, posing significant risks to product quality, energy efficiency, and overall system reliability. This research delves into the realm of data-driven anomaly detection, emphasizing its indispensability in enhancing operational efficiency and effectiveness within manufacturing organizations.

The first part of the thesis introduces a novel approach for anomaly detection in electroplating processes. This methodology synergizes the ordering relationships of process steps with advanced machine learning techniques, specifically the XGBoost system. It incorporates Kernel principal component analysis for dimensionality reduction, effectively handling nonlinear phenomena through Gaussian kernels in a self-tuning procedure. This approach not only demonstrates high accuracy but also maintains the generalizability essential for tackling the challenges posed by data size and complexity. The methodology's efficacy is substantiated using a comprehensive dataset representing electroplating executions from the year 2021, showcasing its potential as a foundation for a generalized anomaly detection framework in electroplating.

The latter part of the thesis advances the field further by introducing the "Adaptive Adversarial Transformer" (AAT). This innovative architecture integrates the Transformer model, renowned for its proficiency in sequence modeling, with Adversarial Learning to enhance anomaly detection in manufacturing sequences. AAT excels in modeling the intricate temporal dependencies inherent in manufacturing data, while its adversarial component diligently identifies and differentiates between normal and anomalous patterns. This combination enables the detection of subtle anomalies that conventional methods might miss. Rigorous testing on multiple real-world manufacturing datasets establishes AAT's superiority over existing state-of-the-art methods, offering enhanced precision, recall, and F1-scores. Additionally, the approach demonstrates robustness against data noise and provides interpretability, allowing for the tracing of anomalies to specific stages in the manufacturing process.

Both approaches signify advancements in the field of anomaly detection in manufacturing, integrating deep learning and advanced data analysis techniques. The thesis highlights the potential of these methods in enhancing operational efficiency and product quality in complex manufacturing settings. By addressing both the technical and practical aspects of anomaly detection, the research contributes significantly to the field of smart manufacturing, offering robust and scalable solutions for industry practitioners. The work also paves the way for future research in the area, particularly in the application of these techniques to other manufacturing processes and scenarios.