應用於封包交換網路時間同步問題之適應性非線性卡爾曼濾波器

Abstract

IEEE 1588網路協議定義主動端與被動端借由時間戳封包交換，進行雙方振盪器時間同步。但由於網路流量難以預測與複雜特性，封包延遲時間會隨時間改變，卡爾曼濾波器是用來濾除封包延遲的一種方法，然而現有文獻顯示封包延遲分佈並非高斯分佈，所以卡爾曼濾波器並非最佳解，此外，網路流量突然的改變，使得卡爾曼濾波器效能更為不佳。在這篇論文中我們提出三種方法解決此問題，主要的想法是使用多模型非線性濾波器來解決同步問題，更精確地說，我們提出多模型馬斯雷立濾波器，並且透過最大期望值演算法不斷地修正適應性濾波器模型的參數。模擬結果顯示我們提出的方法同步精準度明顯優於卡爾曼濾波器，並且同步誤差範圍也在IEEE 1588網路協議規範之內。

IEEE 1588 defines a protocol for master and client to exchange information and synchronize each other’s clock time. Due to unpredictability and complication of the Ethernet traffic, packet delay is highly time-variant. One way to solve the problem is the use of a Kalman filter to smooth the packet delay. However, it has been shown that the distribution of the packet variation is non-Gaussian, and the performance of the Kalman filter is not optimal. Also, the network traffic may experience abrupt changes, further degrading the performance of the Kalman filter. In this thesis, we propose an approach to solve the problems. The main idea is to employ a multiple-model nonlinear filter in the synchronization problem. Specifically, we propose using the interacting-multiple-model (IMM) Masreliez filter. We propose three online methods for adaptively tuning the required parameters estimated by expectation-maximization algorithm (EM) algorithm. Simulation results show that our methods significantly outperform the standard IMM Kalman filter and the time estimation accuracy can meet the IEEE 1588 requirement.