As an important part of the steel structure, the bolt damage will affect the safety of the structure and even cause severe accidents. However, it is difficult to detect the bolt loosening from the perspective of the conventional dynamics, due to the complex vibration characteristics of the bolt joints. In order to detect structural damage intuitively, machine vision has been introduced into the field of structural health monitoring. Therefore, this paper combines deep learning and machine vision to propose a bolt loosening angle detection technology. First, the data sets with bolts were collected and divided into training sets, validation sets, and test sets. Second, the data sets were trained using Single Shot MultiBox Detector. And the recognition accuracy of the model was evaluated, which can reach 0.914. Thereafter, the images obtained from different angles and lighting conditions were detected by the training model;the results showed that this method still has high recognition accuracy and meets the requirements of engineering. Finally, the training model was migrated to the smartphone to achieve quick and simple bolt loosening monitoring.KEYWORDS bolt loosening angle, deep learning, machine vision, smartphone, SSD
| INTRODUCTIONWith the advancement of steel structures, it has developed rapidly in high-rise and supertall buildings. Compared with the concrete structure, the steel structure adopts steel plate instead of reinforced concrete, which has higher strength and better seismic performance. And its components can be prefabricated at the factory, which greatly reduces the construction period. As an important joint, bolts are widely used in steel structures. Thus, the bolt loosening damage not only affects the safety of the building, but it also causes severe accidents. The conventional structural health monitoring methods mainly collects acceleration, 1 displacement, 2 and strain. 3 Then these characteristics are analyzed through wavelet transform 4,5 and Bayesian transformation 6-8 to obtain the location and size of the damage. However, these methods rely on a large number of distributed sensors, and the feature extraction process need to intervention. The damage feature parameters directly determine the results of the damage detection. However, machine vision can complete damage monitoring from another perspective. Many damages on the structure can be detected more intuitively by vision. The bolt damage is mainly divided into rusty, shedding, and loosening, of which the most difficult to detect is loosening. To date, many researchers also used machine vision methods to monitor bolt loosening damage. For instance, Park et al. 9 proposed a bolt loosening angle detection method using Hough transformation, which detects bolt