面向复杂工业视频数据的流程发现方法OA

Process mining relied on structured event logs at a high level of abstraction,but more than 80%of the available data existed in unstructured forms such as videos in practical applications.However,existing technologies lacked adaptive event abstraction capabilities when handling the complexity of multi-objective interactions in dynamic environments,occlusion problems,and chaotic processes in industrial scenarios.To address this problem,this paper proposed a method to automatically abstract high-level event logs from industrial videos under occlusion and multi-objective interaction conditions,and conducted process discovery on the abstracted logs.In this method,YOLOv7 and Detectron2 were fused to detect key objects,and ByteTrack and OC-SORT were adopted to generate multi-target tracking trajectories and unique tracking IDs.Based on the SlowFast dual-stream network,the spatiotemporal features of tracking trajectories were extracted,and the multi-activity confidence distribution of objects within a fixed time window was calculated.Through confidence threshold filtering,sliding window smoothing,and time aggregation techniques,low-level activities were abstracted into high-level event instances,and structured event logs were generated via case association.Validations on the real video dataset of the collaboration process between the robotic arm and the press brake demonstrated that the proposed method achieved better Petri net model matching performance and activity recognition accuracy than traditional methods.Especially in occlusion scenarios,it significantly improved event continuity and log quality.