Abstract: The signals received by vibration sensors often contain vibration signals of different components and environmental noise. In order to identify the number of signal sources and various frequency components from the received signals of a small number of vibration sensors, an underdetermined blind source separation method based on sparse component analysis is proposed. This method first performs time-frequency transformation on the mixed signal, extracts the local principal components of the neighborhood of each time-frequency point through principal component analysis, and filters out the single-source domain feature data. Then use the cosine distance to improve the cluster verification technology and fuzzy clustering algorithm, identify the number of vibration sources, update the clustering parameters, and obtain the number of signal sources and the estimation of the mixing matrix. Finally, a series of least square methods are used to extract the source signal from the time-frequency points corresponding to the mixed signal. Simulation experiments and measured data experiments verify the effectiveness and robustness of the method in this paper. Compared with the time-frequency ratio method, more accurate and robust separation results are obtained, which is helpful for the identification and quantitative evaluation of mechanical vibration sources, so as to facilitate the subsequent mechanical condition monitoring and vibration and noise reduction processing.