Bridge quality inspection: multiscale feature extraction and intelligent decision model

Abstract

Aiming at the problems of strong coupling, complex spatio-temporal correlation of multi-source sensor data in bridge structural health monitoring, and insufficient representation of local damage by traditional methods, an intelligent decision-making method based on multi-sensor fusion and graph neural network was proposed. By constructing a unified representation model of heterogeneous sensor information, the time sequence alignment and feature mapping of vibration, strain and displacement signals are completed. The graph structure modeling mechanism is used to describe the topological association between sensor nodes, and the attention decision strategy is used to realize the weighted aggregation of key features and state discrimination. Further, the data set construction, evaluation index and robustness analysis process are designed to verify the generalization ability and stability of the model under the condition of noise disturbance and node missing. The results show that the proposed method has better accuracy, real-time performance and engineering adaptability.