Abstract: In Flooding Time Synchronization Protocol, several pairs of timestamps of the sender and receiver are collected to do linear regression at each synchronization period.Then it calculates related parameters including clock skew and clock offset, which are used to adjust the local clocks of the sensor nodes under synchronization.The choice of the linear regression algorithms has a direct influence on the accuracy of synchronization.The proposed algorithm replaces the common Least Square method with the Finite Weighted Least Square one to do linear fit, which realizes high-precision synchronization with observation noise of old timestamps considered.Besides, detection for abnormal time stamps is introduced, which avoids the influence of abnormal data points after filtering them.The experiment shows that the proposed algorithm effectively improves the synchronization accuracy.