With the advancement of intelligent manufacturing and the Industrial Internet, continuous casting production lines place increasingly stringent requirements on real-time task processing and energy-efficiency. In multi-edge-node collaborative environments, substantial disparities in computational capability exist among terminal devices, edge servers, and cloud platforms, while communication links exhibit strong temporal variability. These factors make it challenging for task offloading strategies to achieve an effective balance between latency and energy consumption.To address these challenges, this study proposes a dynamic offloading framework tailored for multi-edge-node collaboration and develops a dual-layer Q-network offloading strategy referred to as Two-Stage Q-Network Offloading (TS-QNO). By constructing a latency–energy joint optimization reward function, the method decomposes the offloading decision into two stages—node selection and allocation-ratio optimization—thereby significantly reducing the complexity associated with high-dimensional action spaces. This enables the offloading strategy to adaptively select optimal computing nodes and adjust task distribution ratios, enhancing the precision and stability of decision-making during the offloading process.Validation is performed using the straightening-machine control task within the continuous casting process as a representative application scenario. Experimental results demonstrate that TS-QNO consistently achieves lower task latency and reduced energy overhead under various task densities, node loads, and communication conditions, while maintaining robust optimization performance in high-load environments. The findings indicate that the proposed method effectively addresses device heterogeneity and dynamic scheduling demands in industrial production lines, providing an efficient and scalable offloading solution for cloud–edge–terminal collaborative computing in intelligent manufacturing.
Chen et al. (Thu,) studied this question.