A Lightweight Defect Detection Algorithm of Insulators for Power Inspection

The rapid development of smart grid causes a large increase of power equipments. Power insulator is one of the most important infrastructures in the power transmission lines which is vital to ensure the safe operation of power system. As a common defect, the missing-cap issues will affect the structural strength of power insulators and the safe operation of power lines. Consequently, the monitoring and assessment of abnormal power insulators are of extreme importance for the safe power transmission lines. Due to the good feature expression ability, machine learning algorithms have got much applications on power line inspection, and they could be divided into two categories: shallow learning and deep learning algorithms. Nevertheless, the defect recognition issues on power insulators are always against complex power inspection environment. It will bring a certain effect to the handcrafted feature design of shallow learning algorithms. Meanwhile, the small-scale defect data set will affect the model training of the deep network model. To address the above issues, aimed at the feature of the limited processing power of airborne processor, a novel lightweight defect detection algorithm is proposed for the defects of power insulators which fuses the advantages of shallow learning and deep learning models. Firstly, an insulator location algorithm based on lightweight deep convolutional neural network (DCNN) model is proposed to remove the disturbance of complex backgrounds and serve the high-precision defect detection. On the basis, the high-level image feature based on the improved transfer learning is acquired to effectively distinguish the normal and abnormal power insulators. Finally, a defect recognition method based on Support Vector Machine (SVM) is proposed to solve the small-scale defect detection issue. Experiments show that the proposed method could well meet the precision and speed demands of power system compared with other inspection methods.