Research On Bridge Crack Detection Method Based On Convolutional Autoencoder And Distributed Strain

Crack detections are of great significance for safe operation of Bridges.The distributed fiber sensor can perceive strain distributed along the length of the fiber over the bridge surface,it is the distributed strain and sensitive to structural surface cracks.The cracks can be located by detecting abnormal changes in the distributed strain.However,due to the low signal-to-noise ratio and unbalanced volumes of positive and negative samples,it brings great challenges to the distributed strain-based crack detection.Deep learning can automatically extract noise-robust and discernable features from noisy data.Therefore,it is of great practical significance to carry out research on the detection method of bridge crack based on deep learning and the distributed strain.A method for bridge crack detections is designed and implemented based on convolutional autoencoder(CAE)and the distributed strain.The effectiveness of the proposed method is verified by experiments.It is assumed that the distributed strain is composed of two classes of subsequences,one are measured from elements with cracks,the other are from elements without cracks.Therefore,the crack detection problem is transformed into an anomaly detection problem of the distributed strain.Firstly,the distributed strain are normalized and divided into equal-length subsequences.And then,the subsequence is sent to the CAE to extract feature vectors.Finally,a local outlier factor algorithm is designed to implement the bridge crack detections based on the extracted feature vectors.The method is verified on a simulated bridge and an in-field box girder bridge.It can detect cracks with crack opening displacements exceeding to 23 μm in the simulated bridge,and can also accurately detect surface cracks on the in-field bridge.The main work includes the following three parts:1.A feature representation extraction method is proposed based on CAE.A CAE model is constructed to automatically extract feature representations from the distributed strain.The noise-robust feature representations are obtained.Then,the final feature vectors of the strain subsequences are calculated based the original subsequences,the extracted feature representations and the reconstruction errors.2.An anomaly detection approach on the distributed strain is designed and implementedbased on a local outlier factor algorithm.The cracks are detected and located by anomaly detections on the calculated feature vectors.3.Experimental programs are carried out on a simulated bridge and an in-field bridge.The experimental results verify the proposed method.