Position-aware Transformer-based one-stage torpedo can connecting device camouflage object detection algorithm

摘要

The production process of steel smelting is inherently perilous, particularly during the pouring molten iron phase, which heavily relies on manual execution for connecting torpedo cans to power sources. However, the presence of high voltage and excessively heavy connection devices during this operation poses significant safety risks to workers. Presently, machine learning technologies are being widely integrated into industrial manufacturing processes, offering the potential for automated solutions to replace manual labor. Nevertheless, the effectiveness of machine learning-based object detection combined with robotic arms is compromised by the challenges posed by the integration of sockets into the background and the complexity of the production environment during smelting operations. This paper proposes a Torpedo Can Connecting-Device Detection (TCCD) algorithm based on Position-Aware Transformer (PAT) to address the aforementioned issues, specifically tailored for the pouring molten iron process. Leveraging the Position-Aware Transformer, the TCCD algorithm employs position-guided queries to obtain location labels and instance-aware parameters, effectively integrating local features and long-range contextual dependencies to predict the presence of camouflage instance torpedo can connecting devices. Experimental validation of the proposed method demonstrates its significant advantages over state-of-the-art (SOTA) approaches.