Bridge Abnormal Data Detection And Damage State Identification Based On Encoded Image

Due to long-term effects such as traffic loads,material aging,environmental erosion,and accidental effects such as natural disasters,bridges will inevitably be damaged.As a key node in the traffic network,the safety status of bridges directly affects the operation of the entire road network.In order to ensure the safety of traffic road network operation,health monitoring systems are generally deployed on important bridge projects.In the bridge health monitoring system,the functions of sensor layout and data transmission and storage are relatively mature,but data processing analysis and damage status identification need further research.In this paper,aiming at the data of bridge health monitoring,the encoded image technologies are used to re-characterize the bridge vibration data,and the convolutional neural networks(CNNs)are used as a pattern classifier to detect and eliminate abnormal monitoring data,and further identify bridge damage state.The main research contents of this paper are as follows:1)The encoded image technologies are systematically introduced.The onedimensional vibration response of the bridge is converted into a twodimensional image by using the encoded images,and a method of fusing multitype of encoded images is proposed to fully characterize the inherent characteristics of the bridge response data.2)This article comprehensively introduces the principles of the algorithm principle,network construction,hyperparameter configuration and training strategy of the CNN.The CNNs are utilized as a pattern classification tool to identify the abnormal data and damaged state of bridge data respectively,and the performances of the networks are evaluated from various aspects.3)The measured acceleration data of a long-span cable-stayed bridge is classified and counted,and the multiple patterns of abnormal data and class imbalance problem are investigated.The encoded images are used to recharacterize the data,and CNN is used to detect the abnormal patterns of the data.The detection results have high accuracy and are not disturbed by the class imbalance problem.The representation abilities of different encoded images for each abnormal pattern are further discussed.4)The finite element model of a steel truss bridge is established by Open Sees.The different damage state are simulated,and the corresponding acceleration responses are obtained by using white noise as the excitation.The encoded images are used to re-characterize the data,and CNN is used to identify the damage state of the model.The proposed method can identify the location and degree of damage of the model with high precision,and has strong antinoise interference ability.5)The damage experiment of steel truss bridge model is carried out,and the acceleration responses under different damage state are measured.The acceleration encoded images and damage state of finite element simulation are used as training sets,and the acceleration encoded images measured by experiment are used as test sets.The accurate identification of damage location and degree of test model is realized,and the feasibility and effectiveness of the proposed method are verified.