Damage Identification Of Bridge Structures Based On Deep Learning Theory

Considering that the number of Bridges built in China is approaching saturation,and most Bridges have entered the normal operation stage,it is necessary to optimize the theoretical methods and technical means of bridge damage identification and health detection.With the proposal of smart bridge schemes in various cities in recent years,more research support is needed to improve the lack of continuous and real-time damage analysis of old and new Bridges in daily maintenance.Deep learning theory is developed and evolved from neural network,which is a framework widely used in the field of artificial intelligence.At present,the theory has been applied to image recognition,speech recognition and automatic driving.In addition,deep learning can also be used as a theoretical guide for bridge structural damage identification,and this conclusion has been reflected in many papers.However,at present,the analysis is usually carried out on simple finite element models or simple neural networks in deep learning theory.Therefore,on the basis of the existing research,the finite element model with more complex structure and the deep learning neural network with more difficult frame construction are further compared and analyzed comprehensively.The main work contents are as follows:(1)Study and analyze a large number of excellent literature at home and abroad,elaborate the current research background,point out the development status of the field of bridge structural damage identification at home and abroad,and put forward the difficulties and solutions of the main research problems.(2)Four finite element models of simply supported beam,continuous girder bridge,arch bridge and suspension bridge were established.For each model,four working conditions of natural frequency,frequency combination,curvature and curvature frequency combination were selected as input vectors,and the influence of input vectors on damage discrimination of various finite element models was compared and analyzed.(3)Based on the radial basis function theory in machine learning,a RBF neural network is built for bridge structural damage identification.According to the study,for simply supported beam,when the input vector is the natural frequency,the neural network can represent the damage degree of the structure.When the input vector is frequency combination,the neural network can represent the damage location of the structure.When the input vector is curvature,the neural network can identify the damage location and damage degree simultaneously.When the input vector is the combination value of curvature frequency,the accuracy of damage location and damage degree is decreased.It is proved that different input vectors of RBF neural network can be used for damage identification under different actual conditions in simply supported beam.(4)For the finite element models of continuous girder Bridges,arch Bridges and suspension Bridges,the damage location cannot be represented under the framework of RBF neural network,and the damage degree recognition accuracy is low.(5)A deep convolutional network suitable for damage identification of bridge structures is built by using the higher-order command stream in deep learning software.The analysis shows that the damage degree of the bridge structure can be identified well and the recognition accuracy is high when the natural frequency is taken as the input of the simple supported beam,continuous beam,arch and suspension bridge.(6)In addition,a deep autocoding network which is also suitable for damage identification of bridge structures is constructed.The study shows that the neural network can represent the structural damage of arch and suspension Bridges when the input vectors are natural frequency and curvature.The combined input vector cannot represent the damage in the framework of the neural network.In conclusion,the neural network based on deep learning theory has a deep theoretical foundation and a wide application prospect in the field of damage identification of bridge structures.The input vector,finite element model and neural network are compared and analyzed in order to provide reference for the further study of neural network in the damage identification of bridge structures.Through analysis,it is found that selecting original data from finite element model or actual situation can effectively improve the ability of neural network to extract relevant information for input vector.When the stress of the bridge structure is simple,the neural network with shallower layers and simpler network structure can be selected.When the force of bridge structure is complex,the neural network model with strong complex ability should be selected first.