AI-Guided Real-Time Detection of Flow Irregularities and Bottlenecks in Pharmaceutical Vial Filling Lines Using Vision-Based Models

Ensuring smooth and consistent vial flow on pharmaceutical filling line conveyors is essential for maintaining sterility, reducing downtime, and avoiding costly disruptions. However, irregular vial movement such as spacing gaps, clustering, and vial flipping is often seen, leading to filling delays, broken glass, machine stoppages, and extended batch times. Existing systems lack the ability to detect such irregularities in real time and provide actionable alerts. A lightweight, AI-guided framework was developed for detecting vial flow irregularities and classifying bottleneck risks using computer vision. A high-speed camera positioned above the conveyor captured video frames processed by YOLOv8 for real-time object detection. OpenCV was used to extract vial positions and calculate flow metrics (average spacing, standard deviation, inter-vial gap thresholds). A rule-based classification system assigned bottleneck risk levels: low, medium, or high. Operational conveyor footage from a pharmaceutical filling line was analyzed; to protect confidentiality, frames were de-identified and re-rendered as schematic visualizations that preserve vial geometry and spacing statistics. Across 15 operational frames, the classifier achieved accuracy 93.3%, F1-macro 95.21%. On throughput, the end-to-end pipeline processed frames at ∼18 FPS on CPU and ∼25 FPS on Jetson Nano. Detector evaluation yielded mAP@0.5 = 100.00%, mAP@[0.5:0.95] = 70.09%, precision = 98.65%, recall = 100.00%, F1 = 99.32% at confidence 0.50. This scalable, non-invasive solution can be integrated into existing pharmaceutical lines to improve operational efficiency and product integrity. The rule-based approach offers interpretability, making it suitable for GMP-regulated environments. Across the operational validation set, the risk classifier achieved macro average precision 95.24%, recall 95.83%, and F1 95.21% .

确保制药灌装生产线输送机上药瓶输送流畅且稳定，对于维持生产无菌性、减少停机时间以及避免代价高昂的生产中断至关重要。然而，药瓶输送异常（如间距偏差、扎堆、翻倒）现象频发，进而引发灌装延误、瓶身破损、机器停机以及批次时长延长等问题。现有系统无法实时检测此类异常并提供可执行警报。本研究开发了一款轻量级AI驱动框架，依托计算机视觉技术实现药瓶流动异常检测与瓶颈风险分级。通过安装于输送机上方的高速相机采集视频帧，借助YOLOv8完成实时目标检测；利用OpenCV提取药瓶位置并计算流动指标：平均间距、标准差及瓶间间隙阈值。基于规则的分类系统将瓶颈风险划分为低、中、高三级。本研究分析了某制药灌装生产线实际运行的输送机影像；为保护数据保密性，所有影像帧均经过去标识化处理，并重绘为保留药瓶几何形状与间距统计特征的示意图可视化结果。在15组实际运行帧中，分类器的准确率达93.3%，宏F1值为95.21%。在吞吐量方面，端到端流程在中央处理器（Central Processing Unit，CPU）上的处理帧率约为18帧每秒（Frames Per Second，FPS），在Jetson Nano平台上可达约25 FPS。检测器评估结果显示，在置信度为0.50时，交并比阈值为0.5时的平均精度均值（mAP@0.5）为100.00%，交并比区间0.5至0.95的平均精度均值（mAP@[0.5:0.95]）为70.09%，精确率为98.65%，召回率为100.00%，F1值为99.32%。该可扩展、非侵入式方案可集成至现有制药生产线，以提升运营效率与产品完整性。基于规则的分类方法具备可解释性，适用于受药品生产质量管理规范（Good Manufacturing Practice，GMP）监管的生产环境。在本次操作验证集上，风险分类器的宏平均精确率为95.24%，宏平均召回率为95.83%，宏平均F1值为95.21%。