Research On Bridge Health Assessment Method Based On Big Data Analysis

Bridge health monitoring is a monitoring technology developed to meet the needs of bridge safe operation,aiming at monitoring and evaluating the condition of bridge structure and ensuring the safe operation of bridge.At present,the bridge health monitoring system has been widely used in the operation and maintenance management of medium and large Bridges in China.However,the existing bridge health monitoring system has not yet met the requirements of real-time evaluation and early warning of structural performance and health status.In this paper,the Rugao Lieshi River Bridge of Yancheng-Nantong Coastal Expressway in Jiangsu Province is taken as the engineering background,and the nonlinear mapping model of multi-scale response is constructed by using the deep learning method,and the damage early warning method of bridge health state assessment based on LSTM neural network technology is realized.The main research contents and conclusions are as follows:（1）Based on the vehicle data collected by the WIN Bridge dynamic weighing system of Rugao Lieshi River Bridge,the vehicles are divided into six typical models according to the number of axles and wheelbase.The axle load and speed of each vehicle were statistically analyzed.The EM algorithm based on maximum likelihood estimation and K-S test method were used to analyze and fit the probability density distribution of measured axle load data and speed data,and the probability density distribution model of passing vehicles was obtained.In view of the structure form of Lieshi River Bridge,the finite element model of Lieshi River Bridge is built by using the general large finite element calculation software ANSYS.The ANSYS finite element model is applied to the independent transient analysis and calculation of the total number of random vehicle loads of 1000 times,and the local response data and global response data of 1000 random vehicles passing through the bridge at random speed are obtained.A large number of local response data and global response data were obtained to build a data set,which provided data support for the training and verification of LSTM neural network in the following paragraphs.（2）The dynamic strain data and dynamic displacement data of Lieshi River Bridge collected in 3 months were investigated.The method of wavelet packet decomposition and reconstruction was used to separate the thermal-induced strain and vehicle-induced strain,and the vehicle-induced strain was removed.Based on the LSTM neural network technology and its superior processing performance for time series data,the multi-scale response nonlinear mapping model of the Lieshi River Bridge is constructed.For multi-scale data,the longitudinal vehicle-actuated strain at the bottom of the Lieshi River Bridge is selected as the local response under vehicle-induced load structure input,and the vertical dynamic displacement is used as the global response under the same structure input.The single-dimensional vehicle actuated strain data was selected as input and the single-dimensional dynamic displacement as output.The training set,verification set and test set are constructed at a ratio of 90%:9%:1%for all data sets.Use the training set for training,pass the verification set to verify whether the model passes,if not,adjust the model hyperparameters to repeat the training,and finally get a model that passes the verification set verification,and then use the test set to test the model performance.Through the above process,the test set’s₍(6（6） of the one-to-one nonlinear mapping model remains at 0.0198,far lower than the control upper limit of 0.075.Using the same process to train the many-to-many mapping model,the training effect is better than that of the single-to-single mapping model.The regression prediction deviation of the many-to-many model was smaller than that of the single-to-single model,and the model performance was significantly better than that of the single-to-single model.The average₍(6（6） of the test set was maintained at 0.0102,and₍(6（6） was reduced by 48%compared with that of the single-to-single model.The² remained at 0.9610,and the² increased by 13%compared with the single-to-single model.This proves that the combined training of multi-dimensional input data can obtain a mapping model with higher accuracy and stability than the model trained only on one-dimensional data.（3）Based on the LSTM neural network method,the dynamic strain and dynamic displacement data obtained from the finite element numerical simulation of the Lieshi River Bridge under non-destructive condition were used to construct the LSTM multi-scale nonlinear mapping model under the numerical simulation condition.The₍(6（6） obtained by the model is only 0.0133,which verifies that the deep learning method is still applicable to the response data obtained from finite element numerical calculation.Then the stiffness of the finite element model of the Lieshi River Bridge is reduced by 20%in the span of 1 m,and the finite element model of the bridge damage is obtained.At this time,the decrease range of the natural frequency of the structure is only 0.02%,and the maximum increase range of the mid-span dynamic displacement of each beam after damage is only 2.87%.Multiple independent transient analyses were carried out on the damage model using random vehicle loads to generate numerical simulation datasets of local response and global response under the damage model.On the non-destructive LSTM model constructed by the local response and global response of the non-destructive numerical simulation model,input the local response under the damage model condition,and compare the predicted global response and the measured global response under the damage condition,the average increase of₍(6（6） can reach 256.03%.This can identify the huge deviation between the global response generated by the damaged structure and the predicted global response.This indicates that the LSTM neural network technology can be used to determine whether the structure is damaged.A PCA-based bridge health assessment damage warning feature index is constructed.When the simulated bridge structure is slightly damaged,a multi-scale nonlinear mapping model is generated based on the simulation data,and the global response is predicted through the local response of the five box girder mid-span sections.Calculate the absolute residual of the predicted global response and the measured global response.Principal component analysis was carried out on the absolute residuals,and the principal components whose contribution rate accounted for more than 85%were selected to calculate their²statistics and SPE statistics,and then their comprehensive indicators were calculated.When the structure is not damaged,the index fluctuates slightly,and the maximum value is 43.68.When the structure is slightly damaged,the index vibrates violently,and the maximum wave peak is 965.15.This indicates that the damage warning feature index can detect the slight damage of the structure sensitively.At the same time,the damage warning feature index under the measured condition was calculated,and it was found that the index only fluctuated slightly under the measured condition,and the maximum value of fluctuation was 6.93.It was verified that the index could remain stable when the actual bridge structure did not suffer damage that would affect the structure health.Therefore,in the framework of the bridge health state evaluation method proposed in this paper,the comprehensive index is taken as the bridge damage early warning feature index,which has the sensitive performance and stable performance that it should have.