Exploration of Railway Signal Unloading Task Based on Deep Reinforcement Learning Method

The conventional railway signal equipment maintenance platform typically gathers diverse monitoring data within cloud computing data centers, where the data is subsequently integrated, analyzed, and utilized. However, extant task offloading research primarily relies on a centralized architecture, which is susceptible to single points of failure and afflicted by issues of high energy consumption and time delays. Conversely, Deep Q-learning (DQN), as a pivotal deep reinforcement learning methodology, has demonstrated efficacy in addressing challenges within continuous state spaces. Drawing inspiration from this, this study introduces a cloud-edge collaborative offloading strategy grounded in deep reinforcement learning, denoted as Multi-Dimensional Deep Q-Network (MDQN). MDQN integrates Temporal Difference (TD) updates and experience replay mechanisms into reinforcement learning techniques, dynamically allocating tasks from users to multiple Mobile Edge Computing (MEC) servers, thereby optimizing offloading decision latency, reducing overall service latency, and minimizing user-perceived overall latency and device energy consumption. To validate its efficacy, experiments are conducted leveraging self-generated datasets. Experimental findings indicate that, compared to conventional greedy and round-robin offloading algorithms, MDQN is adept at substantially diminishing task execution overhead, yielding savings of up to 80% or more in resources encompassing energy and time consumption across varying server node densities, and attaining superior offloading strategies. The contributions of this study are significant for elevating the sophistication of intelligent railway signal maintenance, curtailing system maintenance expenditures, and enhancing system utilization efficiency.