Research On Structural Damage Identification Based On Dilated Convolution And Multi-task Learning

The civil engineering industry is transforming from large-scale construction to new construction,renovation and maintenance.So the importance of engineering diagnosis is becoming more and more prominent.Identification of structural damage by using the data of structural vibration monitoring is the focus of structural health monitoring.At present,the rapid development of artificial intelligence technology has greatly changed and enriched the structural diagnosis technology.More and more attention has been paid to the intelligent diagnosis of structural damage.The key to intelligent detection and diagnosis of structural damage using vibration response lies in the extraction of damage-sensitive features in the signal and the pattern classification.This paper proposes a structural damage identification method based on a one-dimensional dilated convolutional neural network,which uses dilated convolution to replace the traditional combination of convolution and pooling,and increases the receptive field while keeping the number of parameters unchanged.It can effectively solve the problems of large model parameters,loss of signal detail information,and poor generalization performance in the traditional deep two-dimensional convolutional neural network model when processing one-dimensional structural vibration signals.At the same time,global pooling is used to replace the traditional fully connected layer to reduce model parameters to prevent the occurrence of over-fitting.Furthermore,in view of the imbalance of data set categories in the actual acquisition of vibration signals,the cost-sensitive classifier is trained by setting penalty weights for different types of signals.This method can effectively extract damage sensitive features from unbalanced vibration signal samples,and improves the accuracy of structural damage detection under unbalanced samples.The verification and application analysis of the scale model damage experiment on the stands of Qatar University show that the method in this paper can automatically extract and classify the optimal features from the original acceleration signals while retaining the signal details to the greatest extent,with high classification accuracy.This method can be used for real-time damage detection.The vibration response signal of the actual monitored structure contains both structural damage location information and damage degree information.At present,single task learning algorithm is often used in structural damage identification using deep learning.In the process of feature extraction,this method often focuses on the sensitive feature information which is directly related to the task requirements,and ignores other useful information contained in the model input.In order to avoid the one sidedness of single task learning,this paper constructs a multi task deep learning model based on one-dimensional dilated convolution neural network.An adaptive shared feature extractor is constructed by using one-dimensional dilated convolution neural network to extract shallow shared features from the original vibration response.Secondly,two branch tasks of structural damage location identification and damage degree quantification are established.Through the complementary information between different tasks,the more comprehensive and deeper extraction of damage information is realized.This method improves the accuracy of damage identification and damage degree identification.The numerical example of five story frame structure and the damage identification of experimental model verify the advantages of multi task learning model in model convergence,identification accuracy and good generalization performance.