SF-DQN: A Cloud-Edge Collaborative Online Task Offloading Framework Based on SwiftFormer and Reinforcement Learning in Industrial Scenarios

With the rapid advancement of artificial intelligence, big data, and Internet of Things technologies, the global manufacturing industry is swiftly transitioning toward digitization and intelligentization. The novel cloud-edge collaborative industrial interconnect computing architecture effectively addresses the limitations of centralized cloud processing and distributed edge computing. However, its multi-layered design introduces challenges for efficient task collaboration, making computation offloading research crucial for enhancing system efficiency and optimizing performance. Existing computation offloading strategies typically focus on single optimization objectives, exhibit pro-longed decision-making times in large-scale edge architectures, and only consider current environmental conditions, thus failing to meet the demands of dynamic environments. To tackle these challenges, this paper proposes SF-DQN, a deep reinforcement learning-based serial task offloading algorithm tailored for a three-tier cloud-edge-end architecture. The approach includes a lightweight bandwidth prediction model based on SwiftFormer to provide proactive support for dynamic offloading decisions, the construction of a comprehensive objective function that accounts for resource competition, task latency, energy consumption, quality of service assurance, and link conditions, and the application of the Deep Q-Network (DQN) algorithm to define state space, action space, and reward function suited for serial tasks, thereby determining the optimal offloading strategy. Simulation experiments validate that the proposed algorithm significantly outperforms existing methods in terms of latency, energy consumption, and system resource utilization, demonstrating its superior performance in large-scale cloud-edge collaborative environments.