面向工业互联网的PTP多源融合时钟同步与守时补偿算法

Multi-Source Fusion Clock Synchronization and Holdover Compensation Algorithm for PTP in Industrial Internet

摘要: 在工业互联网、分布式采样和同步触发系统中，IEEE 1588精密时间协议(Precision Time Protocol, PTP)能够为多节点设备提供统一时间基准，但网络传输时延抖动(Packet Delay Variation, PDV)、链路不对称、报文间歇丢失以及本地晶振温漂会显著降低同步精度和触发一致性。针对上述问题，本文提出一种面向工业现场的PTP多源融合时钟同步与守时补偿算法。该算法以相位偏差和频率漂移为状态量，构建自适应卡尔曼时钟伺服模型；在观测更新前引入幸运包过滤机制，利用滑动窗口内最小时延报文抑制交换机排队噪声；同时结合片载温度采样与三阶多项式拟合，并在同步稳定阶段在线学习温度–频率关系，实现PTP中断阶段的置信度融合守时补偿。算法在同步触发端加入固定硬件延迟预补偿，使时间同步结果能够直接服务于脉冲输出和分布式采样。基于可复现仿真平台的对比实验表明，在轻载、中载和重载网络条件下，本文方法的稳态均方根同步误差分别为6.2 ns、8.4 ns和12.5 ns；在重载快速温变场景中，60 s、180 s和300 s断链守时的最大时间误差分别为18.7 ns、895.8 ns和1901.9 ns，与传统PI伺服相比，最大误差分别降低约97.4%、67.7%和60.7%。实验结果验证了该方法在复杂网络和硬件漂移共同作用下的同步精度、鲁棒性和工程适用性。

Abstract: In industrial Internet, distributed sampling and synchronous triggering systems, the IEEE 1588 Precision Time Protocol (PTP) provides a common time base for multiple nodes. However, packet delay variation, link asymmetry, intermittent message loss and oscillator temperature drift still degrade synchronization accuracy. This paper proposes a multi-source fusion clock synchronization and holdover compensation algorithm for PTP. The phase offset and frequency drift are modeled as clock states in an adaptive Kalman servo. Before measurement update, a lucky packet filter selects low-delay packets in a sliding window to suppress switch queuing noise. A temperature-frequency polynomial model is learned during normal synchronization, validated by a confidence estimator, and then used for holdover compensation when PTP packets are unavailable. In addition, a deterministic delay pre-compensation module is introduced for synchronous trigger outputs. Reproducible simulation experiments show that the proposed method achieves RMS offset errors of 6.2 ns, 8.4 ns and 12.5 ns under light, medium and heavy network loads, respectively. Under heavy load with a fast temperature transient, the maximum holdover errors during 60 s, 180 s and 300 s PTP outages are limited to 18.7 ns, 895.8 ns and 1901.9 ns, respectively. Compared with PI, these values are reduced by approximately 97.4%, 67.7% and 60.7%. The results demonstrate the accuracy, robustness and practical value of the proposed method under combined network jitter and hardware drift.