The fault diagnosis of rotating machinery involves revealing the possible faults in advance to reduce dispensable breakdowns; the difficulty of this lies in identifying the periodic features of rotating machines contaminated by background noise. Time-domain synchronous averaging (TSA) has been studied to eliminate random noise in signals. However, TSA sometimes cannot extract useful signals more accurately because of the accumulative phase error caused by the discrete sampling process and the difficulties in obtaining accurate prior information. Hence, the moving interpolation and Kurtosis searching criterion are used for more accurate extraction of harmonics and transient impacts. Also, an improved compensation algorithm based on moving interpolation is proposed to overcome the amplitude attenuation caused by cumulative phase error for low signal to noise ratio (SNR) signal. To determine some parameters in the algorithm such as the number of periods and the time delay of windows which depend on a priori information relevant to the fault period, a searching method for the prior information in vibration signals including transient impacts and harmonics with Kurtosis and minimum mean square error criterion is proposed to optimize the algorithm in the process of feature extraction. Finally, the improved TSA (ITSA) is applied to extracting the fault features in a real factory, and the performance of fault feature extraction in low SNR signal conditions with the ITSA has been enhanced effectively.