OIL: A Closed-Loop Periodic Batch Adaptation Framework for Object Detection on Industrial Variable-Speed Conveyor Lines

To address motion blur and the degradation of object detection performance caused by variations in industrial conveyor-belt speed, this paper proposes OIL, a closed-loop periodic batch adaptation framework. First, the framework employs a YOLOv8 detector incorporating the Global Attention Mechanism (GAM) and uses the Classification Index Weight (CIW), which consists of detection confidence and the temporal consistency of predicted bounding boxes between adjacent frames. Second, the detection results are divided into three intervals according to the CIW and its constituent metrics: interval A, with a CIW below 0.3; interval B, with a CIW between 0.3 and 0.5; and interval C, with a CIW above 0.5. Results in interval A are treated as negative samples, those in interval B as ambiguous samples, and those in interval C as positive samples. Positive samples are used to train the candidate model, ambiguous samples are stored in a buffer and re-evaluated in subsequent cycles, and negative samples are excluded from the current model update. Finally, when the accumulated number of positive samples reaches a predefined threshold, the system trains a candidate model using the baseline data and newly collected samples. The candidate model is then compared with the currently deployed model on a fixed manually annotated evaluation set. The candidate model is deployed only when its overall performance improves; otherwise, a rollback is performed. Experiments on a coal conveyor line demonstrate that, at 25× speed, incorporating OIL into YOLOv8 + GAM improves Recall by 34.3 percentage points compared with the same detector without OIL. In the variable-speed experiments ranging from 5× to 25×, Recall remained above 64.3%, indicating that the proposed framework can effectively suppress missed detections under high-speed operating conditions.