Research On Defect Recognition Method For FAST Cable Detection Robot

To meet the operation and maintenance requirements of the 500 meter aperture spherical radio telescope(FAST)for environmental monitoring and remediation,a robot equipped with an auxiliary visual inspection system is developed to conduct safety inspection on six steel cables and pulleys that cannot be reached manually at a height of 100 meters.FAST’s feed cabin weighs 30 tons and is supported flexibly by6 steel cables after passing through a 100 meter tall tower.By changing the length of the six steel cables,the feed cabin can move on the virtual spherical crown surface with a diameter of 206 meters at a height of 140.Therefore,the operation and maintenance of the steel cables are particularly important to avoid catastrophic consequences such as cable breakage.At the same time,the service life of the steel cables can be determined more scientifically by testing the steel cables.Visual inspection is used to replace the way of regular replacement of steel cables,Thus,the observation time lost each time due to the replacement of the wire rope is reduced.In this thesis,several branch networks of YOLOX were tested under the same conditions.Through comprehensive analysis and comparison of different conditions such as parameter quantity,Gflops,and actual carrying environment,YOLOX-S was selected as the basic network,and the data set was collected and processed using the established simulation experimental site.Then,the number of data sets was expanded through methods such as rotation,splicing,adding noise,and saturation transformation,Then,it analyzes the shortcomings of current networks in data recognition,and makes targeted improvements to the network from different aspects.By introducing attention mechanisms,it enhances the ability to detect multi-channel information fusion and interaction in the network,better paying attention to the characteristics of the object itself.By improving the FPN structure,it enhances the receptive field of output features,improves the ability to extract deep semantics of the detected object,and replaces the original regression function,Enables the network to better ignore background information.Through ablation experiments,the network is trained using the collected data sets,and the effectiveness of the improvement is proven through comparative experiments with different additions.After training,the network model was evaluated using m AP50 and m AP50:95 as evaluation indicators,with m AP50 reaching 92.57 and m AP50:95 reaching 75.65.The experimental results show that the improved network has good detection performance and has certain reference value.