In the traditional sequential extraction of late exponentials method, the selection of the correlation function fitting window is the key to the bimodal particle size inversion by dynamic light scattering. The inversion results depend on the selection of the correlation function fitting window of the bimodal data, and the particle sizes retrieved by different fitting windows are different. Based on the traditional sequential extraction of late exponentials method, this paper analyzes the attenuation rate of the correlation function of bimodal particle samples and the difference of particle information distribution at different decay times, and proposes a new improved algorithm named attenuation characteristics sequential extraction of late exponentials method. By analyzing the absolute value of the difference between the derivative of the electric field correlation function and the derivative of the equivalent average particle size reconstruction electric field correlation function, the method defines it as the relative attenuation characteristics of the correlation function, and based on the relative attenuation characteristics of this correlation function, the starting point of the fitting window for the small particle correlation function, the interval point and the ending point of the fitting window for the large particle correlation function are defined respectively. These following three reference points are the selection criterion of the fitting window for the correlation function. Firstly, the interval point of the fitting window of the correlation function of large and small particles is defined as the delay time corresponding to the minimum value of the relative attenuation characteristics, the interval point corresponds to the starting point of the fitting window of large particles, and the previous delay time of the interval point corresponds to the termination point of the fitting window of small particles. Secondly, the starting point of the fitting window of the correlation function of small particles is defined as: the minimum delay time corresponding to the maximum value of the vertical coordinate of the relative decay characteristic graph of 0.003 times before the termination point of the fitting window of small particles. Thirdly, the termination point of the fitting window of the correlation function of large particles is defined as: the maximum value of the vertical coordinate of the relative decay characteristic graph of 0.02 times after the starting point of the fitting window of large particles the maximum delay time of the vertical coordinate value. Through the above steps, the problem of inaccurate positioning of the fitting window is solved, and the blindness of fitting window selection is reduced, thus improving the accuracy of the particle size inversion results. The inversion of the simulated data (95 nm/285 nm, 100 nm/400 nm, 105 nm/630 nm, equivalent average particle size of 140.9 nm; 48 nm/144 nm, 50 nm/200 nm, 50 nm/300 nm, equivalent average particle size of 70.6 nm) and the experimental data (60 nm/220 nm, 65 nm/450 nm) was performed by using the improved algorithm, simulated data was performed, in which for the 95 nm/285 nm simulated bimodal particles, the relative error of particle size for large particles was reduced by 11.7%~52.4%, and the relative error of particle size for small particles was reduced by 9.9%~95.1%; the relative error of peak position for large particles was reduced by 4.0%~19.3%, and the relative error of peak position for small particles was reduced by 23.9%~94.5%; the root mean square error of the correlation function was reduced by 36.5%~76.1%. For the 60 nm/220 nm measured bimodal data, the relative error of particle size of large particles was reduced by 2.1% and that of small particles was reduced by 1.7%; the relative error of peak position of large particles was reduced by 1% and that of small particles was reduced by 10.2%, and the root mean square error of the correlation function was reduced by 89.2%. The calculated results of both simulated and measured data show that the inversion results of the improved algorithm significantly reduce the relative error of particle size, relative error of peak position and root mean square error of correlation function. The proposed attenuation characteristics sequential extraction of late exponentials method is better than the traditional sequential extraction of late exponentials method.