Aiming at the problem that typical coating diseases and bolt defects on theinterior surface of steel box girders are difficult to be detected and identified rapidly，a broadlyapplicable orbital robot with large-stiffness multi-stage folding robotic arms was developed.Based on the typical disease classification of coating，disease impact weight，and coating deterioration level assessment，an expert decision-making system for typical disease identification is constructed. Additionally，a robust detection method for bolt loss and loosening defects is proposed based on computer vision and deep learning technology. Thismethod resolves the issue of inaccurate detection results caused by the poor robustness of keyfeature extraction in similar approaches. The results show that the accuracy of identifying coating defects on the interior surface of steel box girders is 97%，with a classification accuracy of 90. 3%. The accuracy of identifying lost bolts is 99. 0%，and for loose bolts，it is 99. 7%. Additionally，the detection method reduces the misjudgment of bolt looseness whenthe shooting angle is within 40 °. The rail robot of the unmanned platform achieves the fast andhigh-precision intelligent inspection of coating diseases and bolt defects on the inner surfaceof steel box girders.