Data Validity Diagnosis Study In Bridge Health Monitoring Based On Deep Learning

Bridge health monitoring has always been an important guarantee for bridge safety.Real-time detection can help to detect bridge damage in time and ensure the safety of national travel.However,the basis of all is to be able to collect accurate and effective data.Only by collecting the corresponding sensor data that can truly reflect the nodes of the bridge and can finish the output work well for bridge health monitoring.On one hand,effective data segment collection is difficult,on the other hand,it's always difficult to delineate valid data and invalid data.In addition,the division of data requires a lot of manpower and energy.Deep learning networks are becoming more and more perfect and widely used in the field of bridge health monitoring..Based on this,this paper makes the following main research:(1)For the structural characteristics of cable-stayed bridges,the same section or adjacent nodes maintain a high degree of correlation,so the effective data segment collected by the sensor also should be accompanied by a node with a high degree of correlation,and vice versa,an invalid data segment.In this paper,the characteristics of the continuous data segment of deflection sensor of Masangxi in March~December2006 are analyzed.The gray correlation degree between nodes is calculated,and the data samples are divided into valid data segments,general valid data segments and invalid data segments.The combination of 3?,local anomaly factor LOF and isolated forest algorithms is used to analyze the reliability of tag data with high confidence.(2)Through screening and comparison,DNN and DBN networks are selected to train the tag data.The learning rate and activation function of the DNN network are selected,and the number of hidden layers and the number of corresponding neurons in the DNN are determined experimentally.In the prediction of the sample data,the accuracy rate reaches 82.5%.The learning rate and activation function are performed on the DBN network.Etc.,and determine the number of RBM and the number of corresponding neurons through experiments.In the prediction of the sample,the accuracy rate reaches 85.59%,and finally chooses the DBN network model.(3)For the DBN,the parameters that appear in the data segment of the cable-stayed bridge deflection sensor may be partially optimal,and it causes poor learning results.The advantages of GSO are analyzed.The improved network model of GSO-DBN is proposed.After experiment,the convergence effect of each layer ofRBM is better,and the prediction accuracy of sample data is reached 92.47%,6.88%higher than the DBN network model.(4)Randomly extract 20 data samples of Masangxi Bridge in January and February 2006,and put them into the designed GSO-DBN network model,the accuracy rate is 95%.The effect is better compared with other models.