<p>Industrial manipulators need fault detection to reduce downtime, lower maintenance costs, and improve system reliability. Conventional methods for identifying infrequent errors, such as sensor or component failures, do not provide real-time anomaly detection, leading to costly operational interruptions. This paper presents a new Physics-Guided Graph Neural Network (PG-GNN) model that integrates advances in Graph Neural Networks (GNNs) with physics-based constraints to enhance the process of finding rare faults on industrial manipulators. The proposed model uses the multi-sensor data, including vibration, temperature, and pressure, and incorporates domain-specific physical information into the GNN architecture to steer the learning process. The combination of data-driven and physics-based methods not only helps to improve the predictive accuracy of the model but also makes it comply with the physical laws that guide the behavior of the system, thus making it more interpretable and robust. The PG-GNN model was tested on the Condition Monitoring of Hydraulic Systems dataset, which consists of normal and faulty operation data. The findings indicate that the PG-GNN model outperforms traditional fault-detection models, such as Support Vector Machines (SVMs) and conventional GNNs, across key evaluation metrics. The proposed model had a 95% accuracy, 93% precision, 91% recall and F1 score of 92% with the Mean Absolute Error (MAE) decreasing to 0.12. These findings support the fact that the PG-GNN model is highly efficient in detecting abnormal faults and abnormalities. As the physical knowledge is implemented in machine learning, the proposed solution will be more robust and efficient in predictive maintenance in a complex industrial environment, and fault detection systems will be improved in the future.</p>