Multivision Tasks Offloading of IoT Perception in Intelligent Manufacturing

Visual IoT perception is a cutting-edge technology for intelligent transformation of manufacturing industry, and the demand for handling vision tasks is extremely huge. However, resource-constrained visual perception devices hamper the real time requirement. Therefore, we propose a partial video frames edge offloading model for multivision tasks in IoT to minimize processing delay. First, computing and transmission processes of vision tasks is characterized by the queuing mechanism, considering both task sensitivity and queue computing load for resource allocation. Then, Markov decision process formalizes the high-dimensional, nonconvex, and nonlinear offloading problem. The deep reinforcement learning algorithm PRSAC is designed to explore the optimal offloading policy. Moreover, integrated prioritized experience replay stabilizes the training process. Finally, experiments demonstrate the feasibility and superiority of the model through a case study in stereo warehouse of a manufacturing enterprise. This research provides a novel and efficient edge offloading method of visual perception for smart factory, improving its agility and quality of service in intelligent manufacturing.