Application Of Machine Learning Approaches To Structural Reliability Analysis And Damage Identification

With the increase of scale and flexibility of modern engineering structures,the issues of structural safety and serviceability have become increasingly prominent.Structural reliability analysis and structural health monitoring(SHM)are two important stratigies to ensure the long-term service and safe operation of engineering structures.On the one hand,it is necessary to perform structural reliability analysis since the structure may be subjected to the extreme natural disasters that can lead to the entire structural failure.On the other hand,structures suffer from serious deterioration due to various factors such as corrosion and material aging over the service period.Therefore,it is of significant importance to implement structure safety assessment by using damage detection techniques.Unfortunately,structural reliability analysis and SHM still present many challenges at present,and there is still great space for improving the efficiency and accuracy of analysis.With the notice of the rapid development of various machine learning technologies,neural network featured by strong ability of nonlinear fitting,feature extraction and highly flexible structure form has been applied extensively in the field of structural engineering,exhibiting great engineering application potential.In this regard,this thesis focuses on the applications of machine learning technologies in structural reliability analysis and structural health monitoring with special emphasis on improving the efficiency and accuracy of reliability analysis and damage detection.In this thesis,a bridge flutter reliability analysis method based on radial basis function(RBF)neural network and a damage identification method based on transmissibility function and Bayesian convolutional neural network(BCNN)will be proposed.The accuracy and effectiveness of the proposed methods will be verified through numerical simulations and experimental studies.The main contributions and conclusions of this thesis are outlined as follows:1.The fundamental concepts,principles and mathematical models of artificial neural networks(ANN)and deep learning(DL)are systematically reviewed,and the relationship between ANN and DL as well as their application scopes are outlined.In particular,the functions and algorithms of the core network layers of RBF neural network and CNN applied to be applied in this thesis are investigated in more detail.Considering the shortcoming of overfitting of CNN in the procedure of training,Bayesian convolutional neural network which regularizes the neural network by probabilistically modeling of the weights within the Bayesian statistical inference scheme to quantify the uncertainty of parameters are introduced for effectively improving the generalization performance of the model.2.Regarding reliably analysis based on machine learning,a bridge flutter failure probability analysis method based on radial basis function(RBF)neural network is proposed.Due to the features of high computational cost,implicit expression and high dimensional nonlinearity of the flutter instability performance function,RBF neural network is employed as a surrogate model so as to avoid the complicated and time-consuming search process of flutter critical wind speed,which will enhance the computational efficiency of solving the failure probability of bridge flutter when integrating with Monte Carlo(MC)simulation.To further improve the efficiency,an adaptive stochastic sampling technique based on dynamic distance constraint and weighted sampling is proposed,which is able to ensure that the RBF neural network can fit the limit state surface as an alternative of fitting the flutter instability function to reduce the complexity of the calculation.In addition,in order to realize the iterative process of adaptive sampling,a parametric finite element model of flutter critical wind speed is established based on the full modal theory by interfacing MATLAB and ANSYS programs effectively,which will lead to “automatic” and “online” flutter analysis.3.The applicability and effectiveness of the proposed adaptive sampling strategy are verified by numerical examples of nonlinear undamped single degree of freedom system and two-dimensional model with small failure probability.The results show that this strategy has much higher computational efficiency than the conventional stochatic sampling methods without scarifying any preciseness.Based on a classical numerical example of a simply-supported beam bridge with an ideal flat plate section,the efficiency and accuracy of the proposed approach for failure probability due to bridge flutter are also validated.It turns out that,this method has a good adaptability.4.Regarding the structural damage detection based on deep learning,a transmissibility function-based Bayesian convolutional neural network method is proposed to detect structural anomaly by making full use of the advantages of dynamic transmissibility in eliminating the effect of excitation as well as the advantage of BCNN in automatic feature extraction.By collecting the acceleration responses of various damage scenarios,the proposed method can formulate the training data pool of transmissibility function of the BCNN model,thereby extracting the damagesensitive features.Subsequently,the unknown damage scenarios can be identified through the trained BCNN model using new tranasmissibility measurements corresponding to these scenarios.To address the critical issue of computational inefficiency in BCNN,an efficient stochastic gradient-based variational Bayesian inference method,i.e.,variational dropout method,is utilized to establish the BCNN model in a more rapid manner.The variances and the computational cost of training can be reduced significantly due to the use of a local reparameterization strategy in which the global uncertainty is fed into local uncertainty to improve the efficiency of statistical inference.5.The well-known case of the IASC-ASCE SHM Benchmark structure and an experimental cantilever beam are used to verify the proposed damage detection algorithm based on transmissibility function and Bayesian deep learning.Results indicate that a high accuracy can be acquired in this method.By noticing the fact that the transmissibility function as a mathematical representation of the output-to-output relationship can effectively eliminate the effects of external loading,the damage identification result of the neural network using transmissibility function as the input is capable of achieving much better performance than employing the structural acceleration response power spectral density.Compared with general CNN algorithms,the accuracy of BCNN is closer to the test set and training set,indicating that it has better generalization performance and better ability in detecting structural damage.