Exploration Of Bridge Safety State Assessment Method Based On Sample Data Of Steel-concrete Composite Beams

This paper relies on the project of National Natural Science Foundation of China(NSFC)sponsored by the tutor,“Research on the Method of Obtaining the Bridge Deformation and Deducing the Structural State Based on the Differential Analysis of Image Contour(51778094)”.Using machine learning to excavate a large number of sample data,this paper attempts to combine machine learning with long-term health monitoring of bridge structure,and studies the application of machine learning in bridge security status recognition.The main research content is as follows.1.Study on the propagation method of steel truss concrete composite beam training data based on the original data similarity matrix.The steel truss-concrete composite beam specimens are designed and manufactured in the laboratory.Static load tests are carried out on steel truss-concrete composite beams under various damage conditions to obtain the deflection data of the specimens and the strain data of the members.A training sample similarity algorithm based on covariance matrix is proposed.According to the distribution characteristics of training sample similarity matrix,a training sample data reproduction criterion based on confidence interval is proposed.The GAN(Generative Adversarial Network)with discrete data processing ability is studied,and generative adversarial network system composed of generating network and discriminant network is established.The deflection and strain data under various damage conditions are manually generated by using the data multiplication criterion and the generative adversarial network.2.Research on bridge damage recognition method based on deformation(strain,deflection)data samples.According to the characteristics of deflection and strain data of composite beam specimens,a BP(Back propagation)neural network architecture is constructed,and sigmoid activation function suitable for machine understanding of non-linear behavior of structural mechanics is selected.According to the data size and characteristics of training samples,the number of hidden layers and the number of nodes in each layer suitable for training samples are determined.A new output labeling system,which is different from the conventional labeling system,is proposed by assigning the damage status of the specimen bar.Under the supervised learning of the labeling system,the neural network can simultaneously predict the damage location and the number of damage elements of the test beam.The prediction results of the test set show that the generalization ability of the neural network meets the requirement of structural damage identification.The relative error of the prediction for any unknown sample data is less than 10%,which can provide us with scientific reference for damage location.3.Application Research on Safety State Assessment of Real Bridge StructuresTwo-year structural deformation monitoring on the Daning River Bridge of Shanghai-Rongzhou Expressway.By using the method of structural safety state assessment based on long-term health monitoring data proposed in this paper,the temperature effect neural network of Daning River Bridge is constructed,the monitoring data for two years are trained,and the mapping relationship between environmental temperature and structural deflection of Daning River Bridge is established.The output label of the safety state evaluation neural network is changed,and the safety state evaluation index is introduced to participate in the safety state evaluation of the structure,and better safety state prediction results are obtained.The Application Research of real bridge also exposes that the adjustment of network output label leads to the lack of obvious linear relationship between input deflection and output security state index,which leads to abnormal prediction results of monitoring data under some relative extreme conditions.Four engineering application experiences are summarized through the study of practical engineering,and the practical engineering application method is improved.