In this article, we modeled a non-stationary time series by dividing it into several segments, then analyzed it by estimating the spectral envelope density function for each segments separately to reduce the effect of the trend. Three kernel functions are used, which are: Dirichlet, Modify Daniell, Féjer. We generated data to test our algorithm and compare the kernel functions to choose the best one by calculating Signal-to-noise ratio (SNR), we conclude that Dirichlet is the most efficient. a human DNA real data is used to apply our algorithm