Research On Intelligent Detection Of Concrete Structure Appearance Damage Based On Convolutional Neural Networ

With economic and social development and accelerated urbanisation,the demand for civil engineering construction is constantly increasing.Newly built engineering structures can suffer some damage due to temperature differences,earthquakes and other factors,while the pressure to maintain ageing buildings and public facilities has increased,which has brought challenges to maintaining infrastructure in China.Traditional damage methods are inefficient and subjective,and no longer meet the needs of today’s society for structural health testing.In recent years,convolutional neural networks(CNNs)have been rapidly developed in deep learningbased classification to accurately detect structural damage conditions.We conduct research related to structural damage identification and quantification based on convolutional neural networks.The main research content and results are as follows:(1)To address the above challenges in crack detection,Mobile Net V3-Large is employed as the backbone combined with CBAM(Convolutional Block Attention Module)to gain Mobile Net V3-Large-CBAM in this study.The classification and identification of crack are studied by using the open-source bridge crack dataset.Mobile Net V3-Large-CBAM is compared with cutting-edge CNNs,and it verifies that the proposed model combined with the preferred Focal Loss has good performance in dealing with imbalanced datasets and hard samples.To verify the generalization ability of the proposed model,this paper further studies the crack datasets with various material and huge-width cracks under different illumination issues.Finally,the sliding window is adopted to perform crack detection and localization on the three randomly reconstructed crack images.The research results show that,compared with other CNNs,the proposed lightweight Mobile Net V3-Large-CBAM combined with the preferred Focal Loss has better comprehensive performance,and the model size is 16.6 MB.I.For imbalanced datasets,the proposed model obtains the best results for crack classification.The Overall Accuracy(OA),F1-score,training speed,and classification speed of Mobile Net V3-Large-CBAM are 95.90 %,95.89 %,101 images/second and 48 images/second,respectively.The proposed model has a balanced recognition accuracy for different crack categories,and the recognition accuracy for hard samples-irregular crack reaches 94.40 %.II.The proposed model has excellent generalization ability.For two test sets-various material cracks and huge-width cracks under different illumination issues,the OA of Mobile Net V3-Large-CBAM reaches 99.66 % and 99.69 %,respectively,the accuracy of crack identification is 99.50 % and 100.00 %,and the accuracy of non-crack identification is 99.90 % and 99.50 %,respectively.III.For crack detection and localization,the proposed model combined with a sliding window,the accuracy of crack detection for three reconstructed images achieves 100 %,and the average crack localization accuracy achieves 98.40 %.(2)To address the problem,the framework of Siamese Res Net50 algorithm is improved:on the one hand,the input of labeled images is frozen after extracting features;on the other hand,the input of recognized images is continuously iterated.Contrastive loss is used to classify the labeled images and the recognized images.The results showed that the improved Siamese Res Net50 classification speed improved from 1 images/second to 11.24 images/second,with an improvement of 1024%;the highest OA and Macor-F1 were also obtained with 94.1% and94.17%,respectively.Siamese Res Net50 achieved a satisfactory accuracy of 88.4% for hard samples.The Axiom-based Gradient-weighted Class Activation Mapping(XGrad-CAM)showed that in the case of limited samples,the CNN cannot extract features correctly even if satisfactory accuracy is obtained;while the t-Distributed Stochastic Neighbour Embedding(tSNE)showed that the improved Siamese Res Net50 could use the Contrastive Loss for effective damage identification with limited samples.(3)A crack geometry parameter measurement algorithm is proposed.Firstly,image distortion is removed by image geometric distortion correction;secondly,crack morphology is extracted using semantic segmentation;followed by distance measurement and edge pixel point matching based on the crack skeleton.The Mobile Net V3-Large-CBAM and Siamese Res Net50 trained in Chapters 3 and 4 were also combined using Pyqt5 and Qt Designer to implement the damage identification system.The system has a crack identification and quantification function and a bridge damage assessment function.