一种用于轨道交通桥梁的无监督异常检测方法研究OA

As bridges constitute critical infrastructure for elevated sections of urban rail transit systems,their operational safety has direct impacts on the reliability of the transportation systems and the life and property safety of the passengers.Unmanned Aerial Vehicles(UAVs)have increasingly become an important means for bridge inspections because of their flexibility and cost-effectiveness.In order to address multiple challenges such as difficulty in the modeling of diverse bridge defects like corrosions and cracks in images,and scarcity of anomaly samples,this paper proposes a lightweight Unsupervised Anomaly Detection Network(UADNet).This approach integrates a teacher-student network with a skip-connection autoencoder,leveraging the complementarity of local and global features to achieve the precise identification of multi-scale defects.The introduction of lightweight convolution effectively reduces the complexity of the proposed model.The experimental results demonstrate that the proposed UADNet achieves superior detection accuracy compared to existing methods on both steel bridge node plate and concrete bridge pier beam datasets.