Almost precision laboratories and semiconductor production plants have vibration isolated design. But, there may be still some vibration out of limit around. It turns into micro-vibration interference source after passing through vibration isolated installation. To find the interference sources around the site, it is necessary to classify and identify the detected blind sources signals. Due to the characteristics of low frequency, low amplitude, and short duration of the transient micro-vibration signal, traditional vibration signal analysis methods have difficulty to handle this. Thus, an identification method of micro-vibration source based on K-medoids is proposed. The endpoint detection algorithm is used to cut off transient micro-vibration signals from long-term data after preprocessing. Then, the normalized Mel filter coefficient (EN-Fbank) feature is extracted and used to constitute feature matrix. In addition, the data are clustered by K-medoids with dynamic time wrapping (DTW) distance. Finally, Gaussian mixture models are created for clustered data to identify the inspection data of the suspected vibration source with model probability threshold to find serious interference sources. In the experiment with 24 h data, two types of vibration sources with the largest average amplitude and the highest frequency of occurrence are found, and the classification accuracy reaches 90. 57% besides the identification rate reaches 96. 8% , which proves the effectiveness and accuracy of the method.