Accelerated clock drift estimation for high-precision wireless time-synchronization

Time synchronization is essential for most Wireless Sensor Network (WSN) applications and the required precision increases with the sampling- and communication rates. These algorithms require a significant amount of computational effort, especially for the clock drift estimation. In this paper, an improved Rolling Linear Regression (RLR) is proposed. By performing a coordinate transformation on each update of the regression table, the required arithmetic operates on smaller absolute numbers without loss of accuracy. On an 8 bit microcontroller (MCU), the improved algorithm is executed at least 22% faster than the conventional Linear Regression (LR) algorithm. In addition, a hardware-accelerated implementation of the improved RLR is proposed, which further reduces the time spent for LR by another 66% on the heterogeneous Hardware-Accelerated Low Power Mote (HaLoMote). Finally, the influence of the regression table size and the synchronization period on the accuracy of a multi-hop time synchronization protocol is investigated. For a 10 s synchronization period, the synchronization inaccuracy can be kept below 1 μs even at the fifth hop.