<![CDATA[Reliable operation and decision-making in industrial process monitoring increasingly depend on sensor networks. However, monitoring accuracy can deteriorate when sensor availability is limited by equipment failures, communication interruptions, or deployment constraints, particularly in complex industrial environments. This study proposes a Physics-Informed Graph Neural Network (PI-GNN) framework for industrial process monitoring under incomplete sensor conditions. The proposed approach combines graph-based representation learning, which captures relationships among process variables, with physics-informed constraints derived from mass and energy conservation principles to improve prediction reliability and robustness. The framework was evaluated using Supervisory Control and Data Acquisition (SCADA) data collected from a palm oil mill in Aceh Utara, Indonesia. Sensor unavailability was simulated through random masking at missing-data rates of 20%, 40%, and 60%. Performance was benchmarked against Long Short-Term Memory (LSTM), Graph Convolutional Network (GCN), Graph Attention Network (GAT), and Transformer models under identical experimental conditions. The results demonstrate that PI-GNN achieved the strongest prediction performance under moderate sensor loss, with an RMSE of 0.0785 and an MAE of 0.0532. Under severe sensor degradation, represented by a 60% missing-data rate, the proposed framework maintained an RMSE of 0.1297 and outperformed the Transformer baseline. Ablation analysis further demonstrated that the incorporation of physics-informed constraints contributed substantially to model robustness under incomplete sensor conditions. These findings indicate that integrating graph-based learning with domain-specific physical knowledge can provide a reliable approach for industrial process monitoring when sensor availability is constrained. The proposed framework therefore offers a promising foundation for resilient monitoring systems capable of maintaining predictive performance despite progressive sensor degradation and incomplete process observations in industrial environments]]>
Copyrights © 2026
