A Real-Time Textile Defect Inspection System for Industrial Polyester Reinforcement Layers Based on YOLO and PLC-Driven Closed-Loop Control

Manual inspection of textile fabrics remains common in industry, yet it is labor-intensive, subjective, and difficult to scale to high-speed production lines. This paper presents a real-time fabric defect inspection system for the middle reinforcement layer of high-strength industrial polyester filaments. The system integrates a YOLO-based detector with a Mitsubishi FX-series programmable logic controller (PLC) to enable closed-loop defect-triggered stopping of a rotary table or conveyor in an industrial setting. Five defect categories are considered: CL (disordered pattern), BJ (tight warp), TS (shuttle skipping), Short (short loose end), and Long (long loose end). To reduce production-specific false alarms, we propose a task-specific multi-stage post-filtering strategy for Short defects, combining an edge-based boundary constraint, class-specific confidence thresholding, and a minimum area constraint, with parameters selected on a held-out validation set and fixed for testing and deployment. The method achieves strong overall performance on the test set (precision 94.3%, recall 97.7%, mAP@0.5 98.9%, mAP@0.5:0.95 71.9%) and improves precision to 95.2% after applying the full post-filtering pipeline. We further benchmark YOLOv5s against recent lightweight variants (YOLOv8n and YOLOv8s) under the same protocol, reporting both detection metrics and PLC-integrated end-to-end latency (mean and P95) to quantify the accuracy–latency trade-off. The deployed system runs on an industrial PC (HIKROBOT MV-VC3704X-128G60) and maintains an average processing time of approximately 315 ms per frame on CPU-only inference. In on-line operation, the closed-loop system achieves an average end-to-end stop latency of about 420 ms with a low false stop rate of 0.35 per hour over 120 hours of continuous production, demonstrating robust long-term stability.