Impact of protective window contamination on monitoring and quality detection in laser cutting processes

Camera-based monitoring combined with Machine Learning (ML) has shown strong potential for real-time quality assessment in laser cutting processes. In industrial environments, contamination of the protective window degrades image quality and may affect the reliability of ML-based monitoring models. This work investigates how the performance of ML models for surface roughness estimation and loss-of-cut classification is influenced by protective window conditions during reactive fusion cutting of mild steel. Three representative window states are considered: clean, moderately contaminated, and highly contaminated. Results show that the models perform reliably under clean and moderately contaminated condition, while high contamination causes an underestimation of the roughness and misclassifies loss of cut events. To address this issue, an image-based indicator is introduced to quantify image integrity associated with window contamination. This indicator enables assessment of image reliability and supports decisions on whether ML model predictions can be trusted. The proposed approach improves the robustness of quality monitoring in laser cutting processes.