PAformer: Visually Indistinguishable Bolt Defect Recognition Based on Bolt Position and Attributes

Bolts are widely used in towers, insulators, and fittings. In transmission lines, their role is to fix fittings and maintain the stability of line structure. However, bolt defect is a visually indistinguishable problem that bolts located on different fittings have different definitions of defects. Aiming at the problem of the visually indistinguishable bolt defect in transmission lines, a bolt defects recognition method based on bolt position and attribute is proposed. Firstly, we use the target detection model YOLOX-S to detect several small-scale fittings, cut out the bolts and record the position of the bolts in the fittings. Then, we introduce the deformable patch module into the fine-grained transformer (TransFG) to enhance the learning ability of the model to fine-grained features and obtain a variety of bolt attributes. Finally, the bolt attribute and position are combined to further determine whether the bolt has defects. We conducted experiments in the Visually Indistinguishable Bolt Defects Dataset (VIBD Dataset) and the pre-training weight of vision transformer on the ImageN et dataset is used for transfer learning, in the training process. The experimental results show that compared with other bolt defects recognition methods, the proposed method solves the problem of visually indistinguishable bolt defect recognition, and has great significance for intelligent inspection of transmission lines.