A hybrid attention hierarchical network-based extreme event detection method for structural health monitoring

Large bridges, operating in harsh environments, generate a diverse array of anomalous data via their monitoring systems. In service time, the bridges may also be affected by extreme events such as earthquakes or ship collisions, resulting in data that exhibit similarities to anomalous data. This similarity complicates differentiation and compromises the accuracy of data analysis and early warning. This article proposes an extreme event and anomaly detection method based on a hybrid attention mechanism hierarchical network, which preserves the high-order tensor information of sensors. The designed Sensor-Spatial-Channel Hybrid Attention Hierarchical Network (SSCHA-HNet) conducts multilevel and multidimensional analyses: the sensor attention block constructs intersensor relationships, the spatial attention block captures long-range dependencies, and the channel attention block dynamically adjusts the representation of global and local features within feature maps. Additionally, it leverages features across different levels to execute hierarchical detection tasks. The proposed method is verified through actual structural health monitoring data and simulation data from a cable-stayed long-span bridge under conditions such as earthquakes and ship collisions. The results demonstrate that the approach can identify extreme event data mixed with diverse patterns of anomalous data efficiently and accurately.