Sensor Correlation Network Based Anomaly Detection for Thermal Systems on Ships

In this paper, we propose an approach to handle the anomaly detection for the thermal system on ships by the sensor associated network method. A large number of sensors are placed in different positions of the thermal system. These sensors form a topological network which can represent the operation state of the thermal system. There are both linear correlation and nonlinear correlation between the operating parameters reflected by these sensors. The MAS index from MINE is utilized to represent the correlation information between sensors when the thermal system is in dynamic operation condition. Based on the MAS correlation coefficient, the sensor correlation network is constructed to represent the dynamic operation process of thermal system. Using DBSCAN clustering algorithm, the large topological network is divided into different subnetworks. When the system is running dynamically, the Manhattan distance between the sub networks can reflect the running state of the system. Based on the distance of sub networks, the similarity of sensor networks with continuous changes is calculated and the similarity correlation sequence is formed. Through the matching between the similarity correlation sequence and the historical experience sequence, we can judge whether the system dynamic condition is abnormal. By the simulation experiment data, we verify the effectiveness of the proposed method.