Research On Intelligent Recognition Method Of Inner Surface Diseases Of Concrete Bridge Tower Based On Deep Learning

As the main component of the cable-stayed bridge,the safety and stability of the bridge tower is an important core to ensure the normal use of the whole bridge project.However,with the gradual accumulation of service time of the whole bridge project and the combined influence of a large number of internal and external factors,it is easy to form a variety of disease characteristics on the inner surface of the bridge tower.If these disease characteristics are not dealt with in time,the durability and impermeability of the bridge tower will decline sharply,and even irreversible damage will be caused to the whole bridge project.Therefore,periodic inspection,accurate and robust identification and timely maintenance and repair of the internal diseases of the bridge tower are important measures to ensure the long-term safe and stable operation of the whole bridge project.Due to the complex internal environment and high-risk coefficient of the bridge tower,there is a lack of image data with multiple disease categories,multiple sample sizes and high imaging quality to analyze various disease types.Meanwhile,conventional identification methods cannot meet the needs of accurate identification and quantification of multiple types of diseases inside the bridge tower.Focusing on the above problems and requirements,this paper takes the multi-type diseases of the inner surface of the long-span cable-stayed bridge tower as the research object,and uses computer vision technology as the carrier to carry out in-depth research on the key technologies such as disease image preprocessing optimization,multi-category disease identification,crack semantic segmentation,and post-processing of segmentation results.A disease diagnosis system for the inner surface of long-span cable-stayed bridge towers was constructed.The specific research work is as follows:1)In view of the lack of relevant typical image data in the research on the identification method of the inner surface disease of the long-span cable-stayed bridge tower,this paper uses the overlapping rotating multi-section telescopic manipulator disease acquisition system equipped with high-definition industrial cameras and other sensors to collect the full-coverage disease image of each area inside the bridge tower.Aiming at the problems of small amount of original effective disease data,high image resolution,insufficient image brightness,and various types of noise interference,this paper analyzes the formation mechanism and evolution law of various types of bridge tower diseases,and carries out the research of preprocessing technology based on image enhancement and image augmentation.The functions of reducing image resolution,improving the contrast of disease area,and alleviating noise interference are realized.Combined with the fast image classification technology based on source domain migration,the accurate and fast classification of crack image,stomatal image,peeling image and rust image is realized,and the classification accuracy is 99.00 %.A multi-category,multi-sample and high-quality bridge tower inner surface disease data set was constructed.2)Aiming at the problems of less sample data and complex texture features,large loss of disease information in the process of feature extraction and insufficient accuracy of disease positioning,this paper studies the grid-level recognition technology of bridge tower inner surface disease for feature reuse.By introducing the DenseNet network based on dense connection structure,the disease features are extracted more effectively,and the shallow features and deep features are fully integrated to reduce the effective information loss in the feature extraction process.At the same time,combined with the global average pooling module and the multi-attention module,the extracted high-dimensional abstract features are enhanced to enhance the network ’s attention to disease information;a triple joint loss function suitable for multi-category disease identification of the inner surface of the bridge tower is designed,including the disease confidence loss part,the disease classification loss part and the disease location loss part.The performance of the network model is evaluated from three different angles and used in the training process.The final average precision rate,average recall rate,average F1 score and category average accuracy reached 89.19%,75.44%,81.06% and 86.04%,respectively.3)Aiming at the problems of serious loss of feature information,poor crack positioning accuracy and inability to effectively solve the imbalance of inter-class samples in the current semantic segmentation method of concrete cracks based on machine vision,as well as the characteristics of complex texture features and unstructured distribution of crack images in bridge towers,this paper studies the pixel-level segmentation technology of bridge tower cracks driven by knowledge and data.The implicit knowledge mining of crack image is carried out by statistical theory.Based on the statistical results,a multi-stage and multi-resolution fracture feature extraction network is constructed.With the strong feature transfer flow,the fracture feature information is learned from multi-dimension and multi-scale,and the loss of fracture feature information in the transfer process is reduced.Based on the statistical results,a category-based crack feature enhancement network is proposed.According to the segmentation task,the similarity between the target region and other regions is quickly calculated to gather the context features of different regions,so as to improve the segmentation performance of the model for crack features.Based on the statistical results,a loss function suitable for semantic segmentation of cracks is proposed.According to the proportion of crack pixels,the number of observable samples in the model is changed,and the contribution rate of crack pixels and non-crack pixels to the loss function is reset to improve the positioning accuracy of the segmentation model for crack features.The final crack pixel accuracy rate,recall rate,intersection ratio and total pixel accuracy rate reached 91.54 %,88.88 %,82.31 % and 99.82 %,respectively.4)Aiming at the problems of large number of parameters in the crack segmentation model of the inner surface of the bridge tower,some noise points in the prediction results,discontinuous crack characteristics and lack of quantitative information for various types of diseases,this paper studies the health diagnosis technology of the inner surface of the bridge tower based on the disease identification results.Through the knowledge distillation strategy,the high-precision crack segmentation network is used as the teacher model to guide the training of the lightweight crack segmentation student model.The crack pixel accuracy,crack pixel recall rate,crack pixel intersection ratio,crack pixel F1 index,and total pixel accuracy are 87.16 %,83.54 %,74.47 %,85.31 %,and 99.74 %,respectively.At the same time,the floating-point operation amount and parameter quantity of the model are only 25%and 20% of the original model,respectively.Combined with image optimization algorithm based on mathematical morphology opening and closing operation,noise interference is reduced and crack area is smoothed.The skeleton information of the crack area is extracted by mathematical morphology correlation algorithm,and the pixel size of the crack area is calculated.Combined with the coordinate transformation relationship of the sensor ’s own parameters,data acquisition process parameters and camera imaging principle,the pixel-level size of each type of disease area is mapped to the actual physical size,and the damage assessment method of the inner surface of the bridge tower is proposed,which provides data support for the risk assessment and construction requirements of the bridge project.Finally,the health diagnosis system of the inner surface disease of the concrete bridge tower is integrated.