In modern manufacturing, defect inspection plays a crucial role in ensuring product quality and compliance with specifications. However, traditional manual inspection methods are increasingly inadequate due to human error, subjectivity, and scalability limitations. This paper presents an intelligent system for automated detection and reasoning of time-varying samples in inspection videos. The proposed system comprises five interconnected modules: object detection, temporal aggregation, case retrieval, semantic labeling, and case retention. A YOLOv8-based model is employed to detect individual instances across video frames. Detections are then grouped using HDBSCAN clustering based on spatial and temporal continuity to form distinct defect cases. These cases are compared against a structured knowledge graph containing historical defect data using similarity measures. A semantic labeling process further refines the interpretation by leveraging a large language model to generate defect pattern descriptions, identify likely causes, and recommend appropriate actions. A real-world case study involving relay gross leak testing demonstrates the system’s ability to achieve high detection accuracy, effectively classify defect patterns, and generate interpretable and actionable outputs. The system supports continuous learning by retaining new defect cases and their labeled interpretations in the knowledge graph. Overall, the proposed framework enhances consistency, traceability, and decision-making in visual inspection tasks, paving the way for scalable and intelligent quality control in manufacturing environments.