Data Visualization and Fault Detection Model for Sputtering Process Monitoring

Abstract

This paper introduces a real-time system for data visualization and fault detection specifically tailored for industrial sputtering processes, focusing on monitoring parameters such as deposition rate, film thickness, material types (Ti, Ag, Ni), and overall process status. The project's primary goals are to model multi-level anomaly outliers to detect potential OCR errors and process deviations, implement a real-time embedded vision system for automated data acquisition from the equipment's display (achieving high accuracy with YOLO and PaddleOCR), and deploy a complete monitoring application featuring real-time visualization and post-process analysis. Data acquisition is carried out using a high-resolution camera, where YOLO achieves 99.5% mAP@0.5 in supervised detection of visual indicators, and PaddleOCR attains 99.57% accuracy in extracting numerical parameters. Preprocessing incorporates a median filter to suppress noise, while DBSCAN identifies sudden OCR fluctuations and linear regression models parameter trends. The postprocessed data are stored in structured CSV files. By integrating robust supervised and unsupervised learning techniques with data science methodologies, the proposed solution ensures reliable operational monitoring, enables early anomaly detection, and supports predictive maintenance strategies in industrial settings.