Abstract: Time series are characterized by "high dimension and mass" in practical applications. The frequent local fluctuations caused by noise interference also render it non-conducive to the analysis of overall trends. To filter out noise and compress the original data, the trend characteristics of time series must be accurately extracted. Based on the piecewise linear representation of a time series, we propose a trend extraction algorithm based on the double evaluation factor of important points in the time series. First, we define the important points as candidate sets of the time series, and present quantitative calculation methods for two of its evaluation factors. Then, we measure the relative degree of difference of the important points as a distance factor, and use a trend factor to measure its degree of influence on the overall trend. Lastly, we comprehensively evaluate the importance of each important point to the overall trend, and select the segment point. The simulation results show that the proposed method can overcome the shortcomings of existing piecewise linearization methods, which have a single evaluate function and contain locality, and can effectively weaken the noise interference. In the case of the same number of segments, the extraction accuracy is higher than that of existing methods. In its application to the trend extraction of bubble gradation, we also validate its effectiveness.