ارائه روشي مبتني بر يادگيري ماشين براي تشخيص بي‌درنگ رفتار كاركنان و وضعيت دستگاه‌ها در كارخانجات بافندگي

A Machine Learning-Based Approach for Real-Time Detection of Worker Behavior an‎d Equipment Status in Weaving Factories

مهندسي معماري كامپيوتر

In the era of Industry 4.0, real-time intelligent monitoring of manufacturing environments to ensure safety, efficiency, an‎d quality of industrial processes is of vital importance, but the significant gap between the accuracy of advanced machine learning algorithms an‎d their implementability in resource-limited industrial environments forms the main challenge. This problem highlights the necessity of creating a balance between high accuracy an‎d computational efficiency to design practical an‎d deployable systems. In this regard, the present study introduces a hybrid machine-learning system that simultaneously monitors an‎d analyzes the status of equipment an‎d worker behavior in weaving factory environments. The proposed system consists of two main modules. The first module employs the YOLO11s model to automatically detect the status of signal lights on weaving machines, recognizing five color states: green, red, orange, white, an‎d off. This selec‎tion is based on appropriate accuracy in real conditions, high inference speed due to its single-stage architecture, an‎d acceptable computational efficiency for deployment on edge hardware. The second module introduces an innovative approach named QuickOptimized-Shift-GCN, which detects six human actions (sitting, stan‎ding, walking, using mobile phone, fighting, an‎d falling) based on skeletal data extracted by YOLO-Pose. Using skeletal data provides the possibility of reducing computational volume an‎d high resistance to environmental changes. Experimental eva‎luations show that the YOLO11s model for signal-light detection achieves a mean Average Precision (mAP50) of 94.8%. In addition, the QuickOptimized-Shift-GCN model reaches an overall accuracy of 94% in human action recognition, which, compared to the baseline Shift-GCN model, demonstrates maintained accuracy with a marginal improvement of 0.34%, a reduction of roughly 4% in computational cost, an‎d a reduction of about 29% in model size. These achievements indicate that, through appropriate architectural choices an‎d targeted optimizations, it is feasible to design systems that combine desirable accuracy with practical deployability in industrial environments.

6

151391