Blade tip timing (BTT) is an effective non-contact measurement technology for rotating blade health monitoring. However, due to the mismatching between the high-speed rotation and the limited amount of probes, the signal collected from BTT system is severely under-sampled, which induces the difficulty in feature extraction. Multiple signal classification (MUSIC) has the potential to overcome the under-sampled problem once the probes are properly placed. Whereas, if traditional MUSIC is directly used in BTT, the accuracy of frequency identification cannot be high enough and the identified numbers of frequency components are also severely restrained. To address these two problems, an improved MUSIC is proposed as an alternative methodology to extract the blade vibration frequency for BTT. Based on the orthogonality of the signal subspace and the noise subspace from under-sampled signal, the presented method can effectively identify the vibration frequency components from the under-sampled signal of BTT.