Abstract: The randomness and complexity resulting from the process of rock failure could result in uncertainty in acquisition signals hence lower accuracy of acoustic emission (AE) signal feature extraction under external load. In order to effectively monitor the process of rock fracture, red sand rock test was studied for its fracture process and an innovative signal processing method based on combining integrating empirical mode decomposition (CEEMDAN) and wavelet threshold and convex optimization is proposed for feature extraction. This method can retain and extract features of effective signals. Firstly, the AE signals under different rupture states collected from the experiments were decomposed by CEEMDAN, Then the sensitive intrinsic mode function which can reflect characteristic of signal was selected using correlation coefficient method and deviation contribution rate, moreover, the high frequency energy demarcation point was determined according to the continuous mean square error criterion. Secondly, wavelet threshold was performed to eliminate noise and the result was compared to result from other noise eliminating methods such as direct wavelet threshold, discarding IMF1, discarding IMF1 and IMF2.Signal noise ratio and positioning precision were used for assessing those methods. Finally, the reconstructed signal was convexly optimized and the feasibility of proposed method was verified by experiments. Study results confirm that the proposed combining CEEMDAN and wavelet threshold method can effectively curb high frequency noise and render highest signal noise ratio and positioning precision. It is also revealed that as the increase of axial stress the sandstone uniaxial compression cracking process exhibited four stages. The number of AE signal events is consistent with stress-time curves, and ratios of average number of events at unstable failure stage could be up to 58.44%. This study can provide new basis and method for quantitatively monitor cracking process and unstability of rock.