Knowledge Graph based Visual Reasoning Method for Cabinets Connecting Defect Detection

Abstract Existing visual inspection algorithms for wiring defect detection in substation protection panels exhibit limited robustness when handling real-world conditions such as partial character occlusion and ambiguous symbols, achieving field accuracy rates of merely 60%. This paper proposes a novel defect detection algorithm that integrates knowledge graphs with visual reasoning. By constructing a dedicated knowledge graph for substation protection panel wiring systems and synergistically combining Yolov8's rotated object detection with Qwen-2.5 foundation vision model's image reasoning capabilities, the proposed method enables intelligent inference of missing or blurred characters while detecting and correcting ambiguous character recognition. Logical wiring rules are rigorously embedded throughout the process. The approach was validated through defect detection on over 5,000 panel images, with comparative analysis against conventional optical character recognition (OCR) algorithms. Statistical results demonstrate approximately 95% parsing accuracy for wiring images with incomplete local information elements. Compared to traditional PaddleOCR, our method achieves a 29.18% improvement in recognition accuracy for cable markings and significantly mitigates confusion between visually similar characters (e.g., "0" and "O"). This breakthrough substantially enhances the reliability of visual recognition for wiring defects in protection panels.