Deep Learning and N8N-Based Soybean Defect Recognition and Traceable Quality Grading System

Abstract

Soybean is an important food commodity whose market value is strongly influenced by seed quality. However, quality inspection in many small and medium agro-industries is still performed manually, making the process slow, subjective, and poorly documented. This study develops an automated soybean inspection system that combines deep-learning-based defect recognition with traceability-supported quality reporting through workflow automation. A YOLOv11n model was trained on a self-collected dataset of 344 annotated images covering five classes, namely Intact, Broken, Skin Damaged, Spotted, and Immature, which was expanded to 1,032 images through augmentation. The model runs on a Raspberry Pi 5 inspection station, where accumulated detection results are converted into a defect rate that determines the quality grade of each batch. An n8n workflow then forwards every inspection result to the Gemini 2.5 Flash large language model to generate a narrative quality-control report, and each batch is stored in an SQLite database that can be accessed through a history viewer to support traceability. Validation results show a precision of 92.73%, a recall of 90.71%, an mAP@0.50 of 96.54%, and an mAP@0.50–0.95 of 88.67%. Functional testing demonstrates that the system produces consistent grades, accumulates detections from multiple captures into a single batch, and generates reports automatically, while still producing a rule-based report when the language model service is unavailable. The proposed system therefore offers an automated, low-cost, and traceability-supported approach to soybean quality inspection that is suitable for deployment on edge devices.