Research On Bridge Damage Identification Based On Deep Learning

Bridges are the arteries of China’s economic development,and with the increasing service life of bridges,the problem of bridge damage is becoming increasingly prominent.Due to the impact of rainfall,wind,temperature,vehicle load and other force majeure factors during the service of the bridge,its monitoring data often has a certain noise.To a certain extent,the traditional bridge damage identification methods have problems such as insufficient utilization of damage characteristic information,high analysis cost,and difficulty in massive data processing.At the same time,problems such as data loss,abnormality,and sparseness of damage information are also widely present in bridge monitoring data,and traditional damage identification methods are difficult to form effective countermeasures for these problems.Since the application of intelligent identification methods to bridge damage recognition research in the nineties of last century,it has significant advantages in solving the problems of long operation time and improper data processing of traditional methods,but the application of intelligent identification methods to actual bridge damage detection engineering is still insufficient.Aiming at the low efficiency of traditional bridge damage recognition methods in massive data processing and the difficulty of model hyperparameter adjustment for different types of data of bridge damage intelligent identification model,this paper proposes a bridge damage recognition method based on SSA-LSTM-Attention,and verifies the feasibility and effectiveness of the proposed method based on bridge damage numerical simulation data and on-site monitoring data.The details of this article are as follows:(1)The effectiveness of intelligent models based on BP neural network,SVM,XGBoost and LSTM on bridge damage recognition was studied: based on the numerical simulation data of bridge damage and on-site monitoring data,the damage recognition accuracy of the above intelligent model was compared,and the significant advantages of the LSTM neural network model in bridge damage recognition were determined.(2)The modeling method based on the combination of LSTM neural network,sparrow search algorithm and self-attention mechanism was studied,and the SSA-LSTM-Attention bridge damage recognition model was established.(3)Bridge damage recognition was carried out on bridge numerical simulation data and bridge monitoring data before and after PCA data compression,and the damage recognition accuracy and hyperparameter combination applicable to different data were output,and the single improvement strategy models LSTM-Attention and SSA-LSTM were used to compare the damage recognition accuracy and model performance with the proposed model,which verified the effectiveness and generalization ability of the bridge damage recognition method proposed in this paper.(4)The influence of the variation of deflection data and the size of data dimension on the accuracy of bridge damage recognition was studied,and the model running time and damage recognition accuracy were compared experimentally,and the correlation between deflection data change,data dimension and accuracy was explored.(5)The AUC-ROC curve principle,calculation formula and model performance evaluation method are studied,and the AUC value is used as the model performance evaluation index to verify the performance of the proposed model and the two single improvement strategy models.