Structural Health Monitoring With Deep Learning

Structural health monitoring(SHM)have gained rapid development in building and bridge of civil engineering over the decades.To proceed damage inspection of civil infrastructure,continuous structural damage variation has been detected and recorded by various pre-embedded sensors at the very begging of construction stage of some new structures.However,as for practical engineering,environmental noise,incomplete measurement data and high computing complexity,etc.challenge the algorithm design of damage identification.Damage identification algorithm with accuracy and broad applicability is thus of crucial importance and has tended to be one of the research hotspots.Currently,effective measurement for evaluating the accuracy and applicability of the damage identification algorithm is rare.In addition,although the neural network based on deep learning method can achieve damage identification by simulating an arbitrary complicated function,the complex function such as mode and frequency projecting to damage values is only performed theoretically due to its high training cost for the real project.Besides,structural damage localization based on the neural network damage is a typical classification of 0-1,which cannot effectively identify continuous damage values.Therefore,the neural network model needs to comprehensively consider the calculation efficiency and the prediction error beyond real value.Moreover,elimination of environmental noise inducing differences between training samples and real values is also required to be considered carefully.In view of the above challenges,thorough research and practical work regarding evaluation of damage detection method,modal and temporal data processing,deep-learning classification model building are explored in this study.The main studies are listed as follows.1.The evaluation index of damage detection method is proposed.The evaluation index can indicate the identification accuracy of local unit damage and entire structural damage.Three damage identification indexes are proposed: unit accuracy,single point accuracy and record accuracy,which can effectively identify the stability and adaptability of damage identification method.As for a trial test,the ant colony algorithm is evaluated.Results show that the ant colony algorithm is feasible for evaluating the single damage detection while prone to fall into local optimum in the case of multiple damage cases.The ant colony algorithm is hence to prove as imprecise algorithm with low record accuracy and inadequate stability.2.The reasonable granularity discretization of the damage value and the high-level feature extraction method based on convolution neural network are proposed in this study.In view of the difficult training and low generalization performance of the regression of continuous damage by neural network,the damage value is discretized according to reasonable granularity.The damage values from continuous numerical prediction effectively reduce to 401 class classification problem,which attains a good balance between calculation efficiency and damage value error.Distinguishing from the traditional methods regarding dynamic calculation and deduction of damage,this study explores a damage detection model based on convolutional neural networks and residual network structure.The high-level characteristics of vibration mode and frequency through the hidden layer of neural network are learned cross entropy loss function,and a damage detection model with high accuracy and stability is achieved.3.A SVM machine learning noise reduction model and the time domain feature extraction method based on cyclic neural network are proposed.Under noisy environment,the convolution neural network model,based on finite element simulation as the prior data,has disadvantages,since the training model has difficulty adjusting the change of environmental factors.In view of that,this paper firstly analyzes the measured acceleration signal of a structure.Regularity of noise distribution is obtained by supporting vector machine model.The noises consistent with the distribution rule of the noise are added to the simulation data base of finite element model.Then,the cyclic neural network is used for feature extraction of the acceleration signal,and the damage prediction is conducted through weighted calculation by the convolution neural network model.Experimental results show that the model can significantly improve the generalization performance of input with noise.