Aiming at the problem of low recognition rate of heterogeneous patterns caused by the characteristics of small sample size, strong impact, short response period and wide resonance frequency bandwidth of telemetry vibration signals, a method for sensitive feature extraction and anomaly detection of telemetry vibration signal based on referenced manifold spatial fusion learning was proposed. The multi-scale analysis method was used to decompose the signals orthogonally into each in the scale band; the multi-scale feature was extracted to construct the high-dimensional feature set. The same normal signal sample was combined with the same type of abnormal sample to establish the exclusive reference model unit, and the linear manifold learning was used to obtain the multi-scale manifold feature difference of each reference model unit to enhance the sensitivity of anomalous features. The projection matrix of each reference model unit was used to enhance the original feature set and obtain the low-dimensional multi-scale sensitive manifold feature. The input to the classifier was used to realize the state recognition of the unknown sample. The measured signal processing results verified the effectiveness of the algorithm.