Research On Data–driven Method And Equipment Development Of Structure Intelligent Inspection

Civil infrastructure is closely related to people’s life,and used to meet people’s most basic living and transportation needs.On the one hand,it is the lifeblood and pillar of the national economy and an important symbol to measure a country’s economic,scientific and technological development level and comprehensive national strength.On the other hand,the safety and longevity of infrastructure are directly related to the safety of people’s lives and property,which is also the core part of livelihood issues.The party and the state have always put ensuring the safety and longevity of infrastructure at the fundamental position of their work.Infrastructure inspection technology is the most straightforward and fundamental means to study and judge the security status and service reliability of infrastructure,which can provide strong support for the long–term safe service of infrastructure.Nevertheless,the existing detection technology mainly depends on manpower,with a heap of blind areas and low efficiency.It is difficult to achieve universal and comprehensive screening of a wide range of structural defects.Additionally,with the rapid development of China’s economy and society and the surge of traffic and people,the service environment of infrastructure has been further deteriorated,and a large number of structures is in the sub–health status.It is common for bridges or houses with degraded performance to have local damage or even collapse due to neglect of inspection and failure of timely reinforcement.Therefore,there is an urgent need for effective infrastructure security diagnosis technology.In recent years,the rapid development of novel technologies represented by data–driven methods and intelligent robots provides new development opportunities and directions for structure inspection.The new infrastructure diagnosis method enabled by artificial intelligence can significantly improve the intelligence level,enhance the detection efficiency,reduce the operation blind spots,and effectively ensure the safety and longevity of engineering structures.Considering the above background,the research core of this dissertation is the intelligent inspection of infrastructure,and this dissertation attempts to introduce novel computing methods and intelligent equipment into structural screening.The main innovations of this thesis are as follows:(1)The bottoms of an enormous number of small and middle span bridges often have low clearance,closed and complex environment,and are difficult for conventional detection methods to inspect effectively,which has led to these parts becoming the detection blind spots for many years.Aiming at these key areas that are difficult to detect,this dissertation puts forward a comprehensive solution covering an intelligent inspection algorithm for detecting multiple types defects on the underside of near–water bridges and an unmanned surface vehicle.Specifically,at the algorithm level,a novel Anchor–free object detection framework named Cen Whole Net is developed.Compared with the traditional Anchor–based object detection methods,it is more suitable for the inspection of infrastructure defects with complex shapes and various slenderness ratios.A lightweight parallel attention module called PAM is designed,which can enrich the representation ability of different convolutional neural networks on the premise of increasing negligible computational overheads.At the equipment level,an intelligent unmanned surface vehicle system using three–dimensional lidar navigation is developed.It does not need to rely on GPS signals and can be applied to the detection of complex bridge bottom environment.In addition,the proposed systematic solution has been applied to an inspection of a bridge group including five small and medium–sized bridges.The results show that the designed cenwholenet + PAM algorithm can obtain better recognition results than many famous object detection frameworks,and has the potential to become a new paradigm in the field of infrastructure damage detection.The stability and engineering practicability of the unmanned surface vehicle system are also verified,which has the potential of further popularization and application.(2)The previous research found that,for cracks,there are two prominent problems in the box–level detection method,namely,the division criteria of bounding boxes are chaotic,and the quantitative characterization of cracks cannot be realized.Pixel–level segmentation is more suitable for describing crack diseases,which can provide accurate representation of cracks and unified evaluation criteria.Besides,concrete cracks are also an important sign of the performance degradation of concrete structures.They are also the main manifestation of concrete structure damage and potential safety hazards that can not be ignored.The effective inspection and analysis of cracks can grasp the early deterioration details of structures.Consequently,this paper develops a lightweight segmentation framework for concrete apparent cracks enhanced by generative adversarial learning strategy.Specifically,the dense connection design is introduced,and a lightweight crack segmentation network called Crack Dense Net is proposed,which has the advantage of less parameters,and is suitable for deployment to intelligent devices with limited computing power.Inspired by "real and counterfeit currency recognition",the basic Generative Adversarial Network is improved and Mix GAN is proposed.It can introduce the soft perceptual constraint that emphasizes the visual perception consistency between the recognition results and the real defects into the original segmentation networks.Therefore,compared with the hard pixel–wise constraint,it pays more attention to the morphological differences and can improve the weights of subtle and fuzzy cracks.The model can predict the results more similar to the real crack shapes based the soft constraint.Moreover,this thesis also deduces and proves the effectiveness of Mix GAN from the theoretical level.In addition,it is also found that the conventional pixel–wise quantitative indicator can not accurately reflect the morphological differences of recognition results.In order to accurately estimate the morphological similarity between the prediction results and the real labels and evaluate the improvement effect of the perceptual constraint on the prediction morphology,the corresponding quantitative evaluation index: Fault–Tolerance indicator is further developed.The proposed framework is applied to the UAV crack inspection projects of a beam bridge and a cable–stayed bridge to verify the effectiveness of the method.The results show that Crack Dense Net is superior to some influential models in terms of qualitative visual perception and quantitative indicators with less parameters.After introducing Mix GAN,the morphological similarity of model recognition results is significantly enhanced,and Mix GAN can bring greater performance improvement to the models than other Generative Adversarial strategies.The proposed framework has broad engineering application prospects.(3)Further,this paper finds that the above research has three deficiencies,that is,it can only detect the cracks on the outer surface of structures,and can not inspect the internal areas,such as the interior of concrete box girders;Only the recognition results described in pixels can be provided,and the actual physical sizes of defects cannot be reported;Only off–line work can be carried out,namely,after collecting images,remote equipment is required to assist in calculation,which leads to long working time and complex steps in the whole process and also makes it impossible to give results in real time.Therefore,this thesis further develops the real–time monitoring method and on–line monitoring equipment for the dynamic development of cracks in key parts of concrete structures.Specifically,at the algorithm level,a lightweight segmentation network Crack Se U–Net integrating multi–level and multi–scale features and attention mechanism is proposed.Its multi–level feature fusion design can strengthen the information dissemination in the network,reduce the loss of information,realize more effective feature extraction and fusion,and improve the prediction accuracy.The experimental results corroborate that it can achieve better segmentation effect on the basis of significantly less parameters and computational complexity than other models.The gain effect of Layer Normalization on visual convolution neural network is studied and discussed in detail.The results show that LN＿He and LN＿VT strategies are more effective than traditional Batch Normalization,which breaks the traditional view that Layer Normalization is often not suitable for training visual convolution networks.A quantitative characterization algorithm of cracks’ actual size information is proposed,which can quantitatively characterize the physical information such as crack lengths,widths and areas on the basis of crack pixel–level recognition results.Then,the mapping of these information from pixel space to physical space is realized,so that the actual physical sizes in the real world of cracks can be accurately described.The effectiveness of the method is verified by a laboratory test in this thesis.The results substantiate that the proposed deep learning model + quantitative characterization algorithm can accurately capture and describe the subtle dynamic development process of cracks.At the equipment level,this paper develops an online crack intelligent monitoring system,which is composed of an online camera hardware and a monitoring software integrating the proposed algorithms.The system can be directly deployed inside structures and carry out the whole process of "detection–identification– analysis" through controlling software remotely to monitor the dynamic changes of damage online.So that engineers can have a more accurate grasp of the service condition and the performance degradation rate of infrastructure.(4)The above research discusses the related content of defect detection for two–dimensional data.This dissertation further discusses the related research on one–dimensional data processing and analysis based on deep learning method,mainly focusing on the detection of structural thermal performance,which is of great significance to the achievement of the national "Carbon Peak and Carbon Neutrality" strategic goal and the fire protection performance of structures under fire.The current solution is to use the conventional numerical calculation method,which has some problems,such as low intelligence,large difference in solving direct and inverse problem,low efficiency of inverse problem solving,cumbersome process and insufficient accuracy.Accordingly,this thesis introduces the physics and data–driven deep learning method into the structural thermal performance inspection,and proposes a novel solution framework Heat PINN,which can solve the direct and inverse heat conduction problems at the same time.Especially,in the direct analysis with Dirichlet boundary conditions,an adaptive activation function strategy embedded in variable network layers is proposed.In the direct problem of wood and steel heating with Neumann boundary conditions,three key improvement strategies are developed.In solving the inverse heat conduction problem,two kinds of inverse problems with fixed parameters and variable parameters are studied.The thesis makes full use of the inherent information of the inverse problem,two coupled neural network models with the skip connection are proposed.The experimental results corroborate that compared with the basic PINNs,Heat PINN can obtain higher accuracy of solving direct and inverse problems on the premise of ensuring the computational efficiency.Compared with the conventional numerical calculation methods,Heat PINN can obtain the same calculation accuracy of the direct problems within the processing time of the same order of magnitude.It is worth noting that the traditional calculation methods can not deal with the inverse problems directly.They must be combined with other algorithms and quite a few repeated calculations are involved in the solution process,which reduces the operation efficiency,increases the calculation overheads,and the inversion accuracy is not controllable.Heat PINN has great advantages in solving the inverse problem.It can solve the partial differential equation and estimate the parameters at the same time,so it can obtain the solution result of the inverse problem with high precision without repeatedly calculating the direct problem.Additionally,it also avoids the necessary and complex preprocessing modeling for conventional methods.These characteristics make the overall performance of Heat PINN much higher than that of traditional frameworks,which has great engineering application value.