Substation Equipment Defect Detection based on Temporal-spatial Similarity Calculation

Abstract

Infrared-image-based defect detection is a critical step in automatic inspection of substation power equipment. Problems in recent researches exist such as manual feature extraction lacks transferability and interpretability, and deep learning methods rely too much on object recognition performance. Therefore, a defect detection framework is proposed based on Mask RCNN network and temporal-spatial similarity calculation. The implementation process is to first extract features of original pictures which are used to detect objective equipment position. Then the possible defect region is given by maximum gray level gradient. It is finally determined by temporal-spatial similarity between possible defect equipment and normal equipment of the same kind. Experiments show that, compared with other popular methods, the proposed framework takes multi-factor similarity into consideration, and therefore has a much better precision, bringing a new idea to automatic power equipment inspection.