Study On Seismic Performance Evaluation Methods For Small-and Medium-span Girder Bridges Based On Machine Learning

Small-and medium-span girder bridges possess the characteristics of large quantity and wide distribution,accounting for more than 90% of the total number of highway bridges in China.The structural types of the bridge piers,bearings and superstructures(mostly T-beams or box girders)are highly similar.It is of large workload and high repeatability to complete earthquake-resistant design analysis and seismic performance evaluation for each bridge across the transportation network.On the other hand,research achievements have been obtained in the seismic performance evaluation of a specific bridge,but there is a lack of systematic research on the seismic performance evaluation methods of bridge portfolios in the highway network.Given this fact,under the framework of performance-based earthquake engineering analysis,this work suggested a machine learning(ML)-based methodology that matures in computer science to evaluate bridge seismic performance,and achieved the parametric modelling of the finite element models of the bridge portfolios.Also,an intelligent optimization procedure was developed to improve seismic performance of bridge portfolios based on ML and genetic algorithm.The main research contents are as follows:1.The status of bridge seismic fragility analyses was investigated,and three sections in the application of ML method in bridge seismic performance evaluation were summarized based on the characteristics of ML,including data acquisition,ML model selection and model performance evaluation.The key hyperparameters for constructing and debugging the supervised learning models in classification and regression tasks were investigated.And on this basis,a basic framework for the application of ML methods in the seismic performance evaluation and optimization design of bridges was developed.2.The statistics of the design information of 157 bridges including 599 frames of the small-and medium-span girder bridge portfolios within the newly-built Wenchuan to Maerkang highway in Sichuan province were carried out.The structural,geometry and material behaviors of the bridge portfolios were reflected in the probability distributions.The batch parameterized finite element models for dynamic analyses of bridge samples generated by Latin Hypercube Sampling(LHS)were established.The selected ML models were used to study and analyze the fundamental dynamic characteristics of the bridge portfolios.The fundamental dynamic characteristics of the bridges were classified and regressed,respectively,and the fundamental dynamic characteristics identification of the bridge portfolios based on ML were proposed.3.Based on the statistical results of the bridge design information and ML methods,the capacity models of the bridge bearings and piers were developed.According to the results of Pushover analyses of the double-column piers,the position of each damage state and the critical factors that determine the damage positions were identified by random forest(RF)algorithm.The regularization techniques were adopted to estimate the quantitative relations between the damage state thresholds and the pier attributes,the seismic capacity models of the piers were established,and the influence of the capacity models on the pier fragility was analyzed in comparison with the traditional capacity models.4.The bridge seismic fragility analyses applying ML were investigated to assess the seismic performance of the small-and medium-span girder bridges in the examined highway network.In line with a set of dynamic analyses results of the bridges,the probability seismic demand models(PSDMs)of the bridge piers and bearings were developed by the selected ML algorithms.The error propagation caused by the traditional assumptions was traced and analyzed in the calculation of seismic fragility.The multi-dimensional fragility models of the components were derived from Lasso-logistic regression and the influence rule of the key bridge properties on the seismic performance were achieved.With the consideration of the relative importance of components on carrying capacity and repair cost,the system-level composite damage state indicators were derived,and the reasonable ones were suggested in accordance with the analysis results.5.An optimization procedure for the seismic performance of small-and medium-span girder bridges portfolios was proposed based on the ML algorithm and genetic algorithm.The uniform performance indicators of the portfolios,namely,repair cost ratio(RCR)surfaces were derived,which acted as a target function of the seismic performance optimization procedure.The optimal design parameter assemblings of the restraint systems of the simply supported girder bridges were determined by the application of neural network and genetic algorithm,and the practicability and effectiveness of the proposed optimized framework were further verified by RCR.