Research On Distributed Strain Crack Detection Method Based On Stacked Convolution Autoencoder

Crack detection on bridges surface has great significance.The distributed fiber optic sensor can measure the distributed strain at a spatial resolution of centimeter on a full-scale surface of a bridge which is hundreds of meters or even thousands of meters long.The crack detection on bridge surface by detecting abnormal changes in distributed strain has broad prospect.However,mainly due to low signal-to-noise ratio(SNR)in measurements,crack detection based on the distributed strain has a large limitation.Deep learning can automatically extract high-quality feature representation from low SNR data and further achieve high-quality detection and classification.Therefore,the research on distributed strain crack detection based on deep learning has important theoretical and practical significance.A distributed strain crack detection method based on stacking convolutional autoencoder(SCAE)was designed and implemented.The effectiveness of the proposed method was verified by experiments.The proposed method assumes that distributed strain was composed of two categories of strain sequences.One was sampled on the surface of structure with crack.Another was sampled on the surface of structure without any cracks.Thereby,the crack detection problem was regarded as a binary-classification problem.Firstly,the distributed strain was normalized and segmented into a set of equal-length strain sequences.And then,the normalized strain sequences were input to a proposed one-dimensional SCAE to automatically extract highquality feature representation.Furthermore,the feature representation was binary classified based on a SoftMax classifier.Finally,laboratory experiment with an I-beam steel structure and an in-field experiment with an in-service bridge were carried out.The result show that proposed method can accurately detect the micro-cracks exceeding to 23 ?m of crack opening displacement(COD)in laboratory experiments and it also can accurately detect the micro-crack on surface of the in-service bridge.The work mainly has the following three aspects:1.Distributed strain data preprocessing.The strain sequence was normalized by z-score normalization with 0 mean and 1 standard deviation.For the problem of positive and negative sample imbalance,a sample amplification method was studied.2.Optimized design of the SCAE and comparisons with common methods.A SCAE based feature representation method was proposed which combine the advantage of convolutional neural network which can in parallel extract multi-feature with the advantage of stacked autoencoder which can extract noise robust feature.Through comparison between SCAE and traditional methods such as support vector machine and linear discriminant analysis,the advanced of the proposed method is verified.3.In-laboratory and in-field experiments.A testbed for the I-beam steel structure was built.Results of crack detection with different crack opening widths under the ideal noise conditions in the laboratory were obtained.The proposed method can accurately detect micro-crack on the bridge surface.