Abstract: Traditional quality detection methods have problems such as low accuracy,slow speed,poor flexibility and strong manual dependence,which has become a bottleneck in accelerating production line rhythm.To cope with them,a Lightweight object Detection model (LightDets) that could run on small wireless devices was proposed.LightDets Continuous Up-Down (CUD) sampling module was used to acquire multi-scale features and reduce feature redundancy and model parameters.Meanwhile,Dense skip connections and convolutions were adopted in LightDets for feature reuse and fusion to improve the accuracy and efficiency,while realize easy deployment.To meet the needs of online detection,a Distributed Task Scheduling strategy (DTS) for end-edge-cloud collaboration was designed to combine underlying controllers to achieve real-time early warning and control of defective products.The experimental results verified the efficiency of the proposed method in air conditioning production line:mAP@0.5=98.56%,and the detection time was reduced by about 70%.

Key words: product quality detection, object detection, end-edge-cloud collaboration, deep learning

摘要： 传统质量检测方法存在精度低、速度慢、部署成本高、灵活性差且人工依赖性强等问题,成为了产线节拍加速和效率提升的瓶颈。针对以上问题,基于所构建的空调外观质量检测数据集,提出一种可运行于小型无线设备的轻量化目标检测模型(LightDets)。LightDets通过连续上、下采样模块(CUD)快速获取多尺度特征,降低特征冗余度与模型参数量;同时,LightDets采用密集型跳过连接和卷积进行特征复用、集成与融合,提升产品质量检测的精度与效率,实现小型终端设备部署。其次,为满足产线在线检测需求,设计端边云协同的分布式任务调度策略(DTS),联合底层控制器,在空调外观产线上实现了不良产品高效精准预警与控制:mAP@0.5=98.56%,检测时间降低约70%。

关键词: 产品质量检测, 目标检测, 端边云协同, 深度学习