Bridge Surface Crack Anomaly Detection Based On Variational Autoencoder And Distributed Strain

The detection of cracks on the bridge surface is of great significance to the safe operation of the bridge.The distributed optical fiber sensor can collect the strain distributed along the optical fiber axis,that is the distributed strain.However,the collected distributed strain often has the problem of low signal-to-noise ratio.The abnormal strain at the crack is easily "overwhelmed" by the noise.This causes huge problems for subsequent crack detection.Deep learning methods can automatically extract discriminable features and have robustness to noise.It is of great practical significance to develop bridge surface crack detection based on deep learning methods and distributed strain.Aiming at the problem of low signal-to-noise ratio in the detection of bridge surface cracks based on distributed strain,this paper designs and implements an anomaly detection method based on Variational Autoencoder(VAE).Distributed strain is represented as a sequence consisting of a set of strain points distributed in space,which is composed of normal subsequences and abnormal subsequences corresponding to fracture damage.Crack detection can be realized by detecting abnormal subsequences.Based on the high-quality features automatically extracted by VAE,the research on supervised and unsupervised distributed strain anomaly detection methods is carried out.And the effectiveness of the method is verified through experiments.The work of this paper mainly consists of the following three aspects:1.Propose a supervised anomaly detection algorithm based on VAE.First,a variational autoencoder model is designed to obtain the features of the distributed strain,and then,a support vector machine(SVM)is used to classify the abnormal-normal distributed strain sequence,and the crack detection is realized based on the classification result.2.Design and implement an unsupervised anomaly detection algorithm based on VAE.First,the normal subsequence is used to train the VAE,and then the distributed strain sequence to be detected is sent to the trained model for detection,and the bridge surface crack detection based on the distributed strain is realized based on the reconstruction error.3.Two experiments of laboratory steel structure and in-service bridges were carried out.The results show that using the proposed supervised method when the label is easy to obtain,and the proposed unsupervised method under the condition of no label can effectively detect cracks.