Hierarchical vision-driven automatic control of quality and speed in fusion laser cutting

To address the growing need for higher productivity and operational efficiency, modern manufacturing is increasingly moving toward intelligent machines equipped with sensors and automation capabilities, driven by Artificial Intelligence (AI) techniques. In laser cutting applications, coaxial monitoring of the molten pool serves as a critical source of real-time process data. When coupled with Machine Learning (ML) techniques, this information enables the regulation of process parameters through closed-loop control, improving cut quality, stability, and maximizing productivity. This study proposes and experimentally validates a hierarchical double closed-loop control architecture designed for real-time quality assurance and productivity optimization in the laser cutting process of 5 mm thick AISI 304 stainless steel. In line with the increasing demand for intelligent manufacturing systems, the approach integrates machine learning-based estimation and classification with control techniques. A regression model was developed to monitor dross formation in real time, achieving R2 score above 90%, which is acceptable for industrial applications. A classification algorithm was employed to reliably detect plasma dominated cutting, loss of cut, and grid-related disturbances, maintaining accuracy levels above 90% under diverse operating conditions. The control architecture consists of an inner loop regulating the speed based on real-time dross estimation, and an outer supervisory loop that detects and suppresses major process anomalies. The system was tested under varying initial conditions to simulate real scenarios, demonstrating robust performance in minimizing dross, recovering optimal cutting states, and optimizing outcomes. The obtained results were validated on a complex case-study, proving applicability and feasibility in industrial manufacturing environments.