| During the long-term operation of the structure,due to external loads,earthquakes,environment,fatigue and other factors,it will inevitably cause damage to the interior,affect the service life of the structure,and threaten the safety of the structure in serious cases.Structural Health Monitoring(SHM)technology is widely used at home and abroad.Considering how to analyze the massive monitoring data and tap the intrinsic value,it is the core of judging the current structural damage state.Based on the traditional data-driven method,this thesis deeply analyzes the time dependence of sensor measurement points before and after the timing and the correlation characteristics of sensor measurement points,and proposes a Convolutional Neural Network(CNN)and Gated Recurrent Unit Neural Network,GRU)joint neural network damage identification method,the main research content is as follows.(1)A data enhancement method based on sliding windows is proposed.In order to make the neural network model have better generalization ability,a large number of samples are needed to train the model,but the sample data in the experimental environment is usually limited.In this thesis,based on the fundamental frequency of the structure and the sampling frequency of the sensor,a method for dividing the time series samples based on a fixed sliding window is constructed.A finite element simulation model was used for the experimental analysis,and the effect of the window length and step value of the sliding window on the damage identification results was studied.(2)Extraction method of correlation feature of acceleration sensor measurement point.Acceleration sensors are installed at key positions of structural members to achieve real-time perception and analysis of structural dynamic and static response and operating environment information.Under the action of external load and environment,the structure needs to satisfy geometric equations,physical equations,equilibrium equations,forces and boundary conditions,etc.There must be some correlation between each monitoring point.Taking the vibration response data of a bridge in Chongqing as an example,this thesis uses the correlation indexes of Pearson,Spearman and Kendall ’s to characterize the correlation of its acceleration sensor points,and uses the convolutional neural network to extract the correlation characteristics of the sensor points.(3)Time-dependent feature extraction method of structural vibration response data.The acceleration time series data collected by structural health monitoring has timedependent dependence.In this thesis,autocorrelation coefficients are introduced to characterize the time-dependent dependence of acceleration time series,and the monitoring data of a bridge in Chongqing are analyzed.Considering the superiority of the gated recurrent neural network in time feature extraction,GRU is used to extract the frontrear dependency characteristics of structural vibration response data.(4)Proposed a CNN and GRU-based joint network model structure damage identification method.In this thesis,structural pattern recognition is reduced to the problem of multi-variable time series structural damage recognition,combined with CNN’s advantages in feature correlation of sensor measurement points and GRU’s advantages in the feature extraction of time-series data.The joint neural network model of CNN and GRU was constructed and tested separately in the Benchmark finite element simulation model and the scale model of the Heichonggou Bridge.The loss recognition results were 97.42% and 94.84%,compared with mainstream deep learning.The method has better recognition effect.In summary,the deep learning method represented by the joint neural network can effectively improve the accuracy of structural recognition,and has a good application prospect in the field of structural health monitoring.Considering the complexity of the actual project,how can it be widely used in the structural field the application also requires a step of exploration. |
PDF Full Text Request