Deep Learning-based Structural Surface Damage Detection And Monitoring Data Mining

Civil infrastructure is an essential indicator of the country's comprehensive national strength and scientific and technological development level.Its safe service is related to the national economy and people's livelihood.The Routine structural inspection based evaluation and the structural health monitoring are the main basis for decision-making of current bridge structure management.A large amount of data will be generated,whether it is the structural inspection-based evaluation or structural health monitoring.Simultaneously,the traditional data analysis methods are challenging to analyze and process the Massive Data efficiently.How to quickly analyze the Massive Data and apply the analysis results to help bridge management has become a research hotspot in civil engineering.With the visual inspection and health monitoring data,this thesis focuses on the classification problems and using deep learning technology to mine the inherent characteristics of Massive Data.In terms of research relying on visual inspection data,this thesis mainly studies the automatic crack segmentation method.This part of the study can be divided into three,gradually in-depth stages.In terms of research relying on health monitoring data,this thesis focuses on the first two stages of the big data problem processing framework: data preprocessing and data storage.The main innovations are as follows:(1)Aiming at the disadvantages of poor applicability of traditional digital image processing methods,a dual-scale Convolutional Neural Networks(CNN)based crack detection method is proposed.Based on CNN's classification ability,the detection of cracks in the image is divided into three levels: crack localization,crack segmentation and crack width measurement,respectively.At first,the proposed method detects and locates the cracks at two scales.Then the traditional digital image processing methods are utilized to generate the binary crack maps.Finally,CNN is used again to refine the crack maps and remove noise.Subsequently,a novel crack width estimation method based on the Zernike moment operator's use is further developed for thin cracks.The experimental results based on a laboratory loading test agree well with the direct measurements,which substantiates the proposed method's effectiveness for quantitative crack detection.(2)A pixel-level crack segmentation method based on a multi-scale feature fusion network is proposed,which can realize crack segmentation that entirely relies on deep learning to extract features instead of relying on the traditional digital image processing methods.The proposed framework can be divided into feature extraction layers and feature fusion layers.The feature fusion layers can be further divided into simple pixel-level fusion at different scales and deep fusion between dimensions of different scales.Totally 13 networks are designed with different kinds of feature fusion strategies and feature parameter selection methods.By comparing different networks,it is found that low-level features mainly represent the morphological information of cracks and are more effective for the detailed description of cracks.In contrast,high-level features are more useful for describing semantic features of cracks in the image.(3)A light-weight crack delineation network enhanced by Generative Adversarial Learning.The network proposed in Chapter 3 has difficulties in finding hairline cracks,and it is still timeconsuming.Therefore,a light-weight crack delineation framework improved from two aspects is further proposed.On the one hand,the dense block is utilized as the feature extraction structure,making the network more efficient with fewer parameters.On the other hand,the generative adversarial strategy is introduced in the training process,making the distribution of the network output approximate human labels' distribution and directly improving the prediction accuracy.Finally,the proposed light-weight crack delineation network is applied to the crack detection of a bridge's piers in China.(4)Based on monitoring data rapid preprocessing requirements,a one-dimensional CNN that extracts features directly from the input signals is designed to detect abnormal data with validated high accuracy.The proposed method is successfully applied to the monitoring data of a long-span suspension bridge in China.Compared with the traditional model-based approach,the proposed method has broad applicability and can detect various anomaly types with high precision.(5)A new SHM data compression and reconstruction framework based on Autoencoder structure is further developed,recovering the data with high-precision under a low compression ratio.The proposed framework can be divided into Compression Network and Reconstruction Network.The Compression Network extracts high dimensional features as the compressed data,while the Reconstruction Network is designed to recover the original signal with the compressed data input.It is found that,by using our proposed methods,under the lower compression ratio,the reconstruction errors are significantly less than that of the traditional way.Finally,the data compression and reconstruction framework is successfully verified with the real bridge health monitoring data.