面向车队云协同计算的深度神经网络推理任务拆分与卸载策略研究

PDF下载HTML

Title:

Task partitioning and offloading strategy for DNN inference in platooning cloud

作者:

邓贻香¹, 张玉林¹, 柏龙²

1. 重庆城市科技学院 电气工程与智能制造学院, 重庆 402160;
2. 重庆大学 机械与运载工程学院, 重庆 400044

Author(s):

DENG Yixiang¹, ZHANG Yulin¹, BAI Long²

1. College of Electrical Engineering and intelligent manufacturing, Chongqing metropolitan College of Science and Technology, Chongqing 402160, China;
2. College of Mechanical and Vehicle Engineering, Chongqing University, Chongqing 400044, China

关键词:

深度神经网络; 自动驾驶; 车队云; 任务拆分; 计算卸载; 双维度拆分; 多智能体强化学习; 资源分配

Keywords:

deep neural networks; autonomous driving; platooning cloud; task partitioning; computation offloading; dual-dimensional splitting; multi-agent reinforcement learning; resource allocation

分类号:

TP183

DOI:

10.11992/tis.202512029

摘要:

随着深度神经网络(deep neuralnetworks，DNN)在自动驾驶感知任务(如目标跟踪、语义分割等)中的广泛应用，DNN 推理对实时性的要求与车载计算资源受限之间的矛盾日益突出，已成为制约智能驾驶系统性能提升的重要瓶颈。针对该问题，本文提出一种面向车队云协同计算环境的 DNN 推理任务拆分与卸载策略。以最小化推理时延为目标，构建层间-层内双维度协同拆分与卸载模型：在层间维度，结合车辆异构计算能力与链路状态，对 DNN 任务进行自适应拆分；在层内维度，针对单层计算负载较重的部分，进一步进行空间划分并分配给多辆车并行处理，以缓解局部计算瓶颈。将动态任务拆分与卸载联合优化问题建模为分布式部分可观测马尔可夫决策过程(decentralized partially observable Markov decision process，Dec-POMDP)，并设计基于多智能体深度确定性策略梯度(multi-agent deep deterministic policy gradient，MADDPG)的求解方法，实现车队协同环境下的动态决策优化。仿真结果表明，在不同车队规模、计算能力和带宽条件下，所提方法均能有效降低 DNN 推理时延；与全本地计算、整体卸载和传统层间拆分方案相比，所提方法在资源受限场景下表现出更优的时延性能。

Abstract:

With the wide deployment of deep neural networks (DNNs) in autonomous driving perception tasks such as object tracking and semantic segmentation, the contradiction between stringent real-time inference requirements and limited onboard computing resources has become a major bottleneck restricting system performance. To address this issue, this paper proposes a DNN task partitioning and offloading strategy for platooning cloud environments. First, a dual-dimensional collaborative partitioning and offloading model is developed with the objective of minimizing inference latency. In the inter-layer dimension, DNN tasks are adaptively partitioned according to heterogeneous vehicle computing capabilities and link conditions. In the intra-layer dimension, the parts with heavy single-layer computational loads are further spatially partitioned and assigned to multiple vehicles for parallel execution, thereby alleviating local computational bottlenecks. Second, the joint optimization problem of dynamic task partitioning and offloading is formulated as a decentralized partially observable Markov decision process (Dec-POMDP), and a multi-agent deep deterministic policy gradient (MADDPG) based solution is designed to achieve dynamic decision optimization in platooning cloud environments. Simulation results show that the proposed method can effectively reduce DNN inference latency under different platoon scales, computing capacities, and bandwidth settings. Compared with full local computing, whole-model offloading, and conventional inter-layer partitioning schemes, the proposed method achieves superior latency performance, especially under resource-constrained conditions.