반도체 제조공정 품질 향상을 위한 그래프 신경망 기반 이상탐지 및 근본 원인 시각화 사례 연구
A Case Study of Graph Neural Network-Based Anomaly Detection and Root Cause Visualization for Quality Improvement in Semiconductor Manufacturing
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
Purpose: This study evaluated the effectiveness of a Graph Neural Network (GNN)-based anomaly detection and Root Cause Analysis (RCA) method designed to enhance semiconductor manufacturing quality focusing specifically on Chemical Vapor Deposition (CVD) processes. Traditional anomaly detection methods often fail to accurately model complex interactions within multidimensional and time-series sensor data, limiting their applicability in highly integrated manufacturing environments. This research proposes an advanced approach to address these critical limitations. Methods: An unsupervised GNN model was developed utilizing three months of actual sensor data from a semiconductor company's real-world CVD production line. The relationships between sensors were explicitly represented through a graph structure, employing Long Short-Term Memory (LSTM)-based embedding and attention mechanisms to effectively capture temporal dynamics and sensor interdependencies. The model's performance was rigorously evaluated using F-1 score, precision, and recall metrics. Its performance was directly compared against established algorithms including Isolation Forest, One-Class Support Vector Machine (One-Class SVM), and k-Nearest Neighbor (k-NN). Results: The proposed GNN model significantly outperformed traditional methods, achieving an F-1 Score of 0.906, precision of 0.953, and recall of 0.864. Moreover, it generated intuitive visualizations of sensor interactions, facilitating rapid identification and understanding of root causes during anomaly events. Consequently, the time required for anomaly detection and root cause diagnosis was markedly reduced, leading to substantial reductions in quality management costs and equipment downtime. Conclusion: The GNN-based anomaly detection approach effectively addresses the complexity of sensor interactions, significantly improving the efficiency and accuracy of quality management in semiconductor manufacturing. Future research should extend this promising methodology to diverse manufacturing contexts.
Abstract
Purpose: This study evaluated the effectiveness of a Graph Neural Network (GNN)-based anomaly detection and Root Cause Analysis (RCA) method designed to enhance semiconductor manufacturing quality focusing specifically on Chemical Vapor Deposition (CVD) processes. Traditional anomaly detection methods often fail to accurately model complex interactions within multidimensional and time-series sensor data, limiting their applicability in highly integrated manufacturing environments. This research proposes an advanced approach to address these critical limitations. Methods: An unsupervised GNN model was developed utilizing three months of actual sensor data from a semiconductor company's real-world CVD production line. The relationships between sensors were explicitly represented through a graph structure, employing Long Short-Term Memory (LSTM)-based embedding and attention mechanisms to effectively capture temporal dynamics and sensor interdependencies. The model's performance was rigorously evaluated using F-1 score, precision, and recall metrics. Its performance was directly compared against established algorithms including Isolation Forest, One-Class Support Vector Machine (One-Class SVM), and k-Nearest Neighbor (k-NN). Results: The proposed GNN model significantly outperformed traditional methods, achieving an F-1 Score of 0.906, precision of 0.953, and recall of 0.864. Moreover, it generated intuitive visualizations of sensor interactions, facilitating rapid identification and understanding of root causes during anomaly events. Consequently, the time required for anomaly detection and root cause diagnosis was markedly reduced, leading to substantial reductions in quality management costs and equipment downtime. Conclusion: The GNN-based anomaly detection approach effectively addresses the complexity of sensor interactions, significantly improving the efficiency and accuracy of quality management in semiconductor manufacturing. Future research should extend this promising methodology to diverse manufacturing contexts.