Abstract: To address the issues of low sample efficiency and insufficient multi-objective optimization in traditional deep reinforcement learning (DRL) for vehicle active suspension control, a twin delayed deep deterministic policy gradient (TD3) algorithm integrated with a prioritized experience replay (PER) mechanism (TD3-PER) is proposed. This algorithm significantly improves sample efficiency and accelerates policy convergence in high-dimensional state spaces by prioritizing high-value samples. Additionally, an adaptive weighted objective function combining vertical acceleration and pitch acceleration is designed to achieve coordinated optimization of multi-degree-of-freedom vibrations. Simulation experiments demonstrate that, under class C random road conditions, the TD3-PER algorithm reduces the root mean square (RMS) of vertical body acceleration by 38.79% and the RMS of pitch angular acceleration by 45.24% compared to passive suspension. Within a vehicle speed range of 20～120 km/h, the optimization rate for body acceleration ranges from 36.26% to 52.46%, and for pitch angular acceleration from 20.99% to 48.92%, with performance improving as road smoothness increases. The algorithm exhibits superior efficiency, robustness, and multi-objective optimization capabilities.

Key words: active suspension; deep reinforcement learning; prioritized experience replay; multiobjective optimization

摘要： 针对传统深度强化学习（DRL）在车辆主动悬架控制中，样本利用效率低和多目标优 化效果不足的问题，提出一种融合优先级经验回放（PER）机制的双延迟深度确定性策略梯 度（TD3）算法（TD3-PER）。该算法通过优先筛选高价值样本，显著提高样本利用率，加速 高维状态空间中的策略收敛。同时，设计了一种综合垂向加速度与俯仰角加速度的加权目标 函数自适应调节，实现多自由度振动的协同优化。仿真实验表明，在 C 级随机路面下，相较 于被动悬架，TD3-PER 算法将车身垂向加速度均方根值降低 38.79%，俯仰角加速度均方根值 降低 45.24%。在 20～120 km/h 车速范围内，车身加速度优化率达 36.26%～52.46%，俯仰角 加速度优化率达 20.99%～48.92%，且性能随路面平滑度提升而增强，具有更好的高效性、鲁 棒性及多目标优化能力。

关键词: 主动悬架；深度强化学习；优先经验回放；多目标优化