Research On Bridge Structure Abnormal Diagnosis Method Based On Multi-modal Monitoring Data Fusion

In order to ensure the safety and durability of the bridge structure in service,the structural health monitoring of the bridge structure combined with artificial intelligence and big data technology has become an effective technical means for the intelligent management of bridges.The bridge health monitoring system has accumulated a large amount of monitoring and perception data with the increase in the service life of the bridge.This type of data reflects the structural health of the bridge.Although the structural health monitoring technology under the data-driven paradigm has achieved fruitful research results,There are still certain shortcomings: First,due to the complexity of the bridge structure and the uncertainty of the service environment,the research on the joint extraction of the time series and multivariable spatial correlation features of the health monitoring perception data still has low feature extraction efficiency.,Pattern recognition accuracy is low,and at the same time,a joint model for spatio-temporal feature extraction of multivariate time series also needs to be constructed.Second,although the single-modal structural damage identification method has achieved good results in pattern recognition,the monitoring and perception data collected by the multi-type sensors deployed at the location of the key components of the bridge has not been effectively used.In view of the above two points,this thesis proposes a convolutional neural network(Convolutional Neural Network,CNN)with dual-channel data input and a joint model of Bidirectional Gated Recurrent Unit(Bidirectional GRU)to monitor and perceive data.Feature extraction is carried out in two dimensions of time and space,combined with the adaptive tensor fusion network proposed in this thesis for feature-level information fusion,and the joint representation of multi-source heterogeneous data improves the robustness and accuracy of model decision-making.The main research contents are as follows:(1)This thesis takes the data set of the scaled model of the Heiconggou Bridge in Yunnan Province as an example,and analyzes the characteristics of the acceleration and vibration data containing rich structural dynamic response information from the two dimensions of time and space,which proves that the acceleration data is used for the structure The effectiveness and feasibility of abnormal pattern recognition provide a basis for the extraction of spatiotemporal features with a joint neural network in the following.(2)Aiming at the spatiotemporal characteristics of the above-mentioned monitoring perception data,this thesis proposes a spatiotemporal feature extraction model(Parallel CNN and Bidirectional GRU Framework,PCBG)with dual-channel data input for parallel convolution and bidirectional recurrent neural network to extract monitoring The spatial correlation and time dependence of the data.In order to verify the recognition performance of this model,experimental verification and comparative analysis were carried out on the scale model data set of Heiconggou Bridge and the IASC-ASCE Benchmark data set.The experimental results show that the pattern recognition is accurate on the above two data sets.The rates are 94.92% and 85.11%respectively,which are better than other comparative models.(3)In view of the insufficient characterization ability of single data for the damage state of bridge structure,this thesis proposes an adaptive tensor fusion network to perform feature-level information fusion on multi-source heterogeneous monitoring and perception data,so as to improve the comprehensive analysis of structural abnormal patterns by monitoring data.Sexuality and robustness.In order to verify the effectiveness of the feature-level fusion method based on monitoring and sensing data,this thesis selects the acceleration vibration data and strain data of the Beili Bridge finite element model to jointly express the overall structural damage and local damage.The experimental results show that the adaptive tensor fusion network proposed in this thesis achieves 92.36% recognition accuracy in the six structural conditions of the Beili Bridge finite element model,which is better than other network structures with feature fusion.