With the increasing control requirements for high-end equipment, the number of sensor components also increases. In order to ensure the normal operation of the control system, it is necessary to increase the sensor redundancy. However, too many sensors will increase system complexity and reduce reliability, and the analytical redundancy method can add redundancy to the system without increasing the number of sensors to ensure system reliability. While current analytical redundancy construction methods are usually only for Euclidean space data, ignoring the relationship and correlation strengths between sensor networks and limiting the ability to extract feature representations in non-Euclidean space. To address this problem, this paper proposes graph nodes masked autoencoder based method for constructing analytical redundancy of sensor networks and realizing sensor anomaly detection. By constructing a graph signal and randomly masking its nodes, it is input into the network to learn and compute the analytical redundancy. Experimental results demonstrate that the proposed method can effectively obtain the sensor analytical redundancy and realize accurate sensor anomaly detection.