Obtaining accurate projectile acceleration signals is essential for evaluating the performance of electromagnetic guns. However, the projectile is affected by different environmental factors in the chamber and out of the muzzle, which makes the acceleration signal have different modal characteristics in the bore and the muzzle stage, which leads to the failure of the conventional nonlinear non-stationary signal global processing method. Therefore, an adaptive analysis and reconstruction method of acceleration signal fusing maximum likelihood-wavelet and improved fully adaptive noise ensemble empirical mode decomposition (ICEEMDAN) is proposed in order to obtain accurate acceleration signals. Secondly, the partition signal was decomposed by ICEEMDAN to reduce the interference of harmful signals on signal parsing. Finally, the effective modal components were extracted based on the t-test for signal reconstruction to achieve accurate extraction of the effective acceleration signal. Correlation experiments show that the improvement rate of root mean square error is greater than 0, the correlation coefficient (ρ) is increased to 0.6731, and the signal-to-noise ratio (SNR) is increased to 3.8614, which avoids the problem of over-decomposition or incomplete decomposition of some regions compared with the conventional global processing methods, and realizes the accurate extraction of acceleration signals.