Abstract: Accurate extract of signal features of partial discharge (PD) is crucial to on-line monitoring of generator set insulation systems. This paper describes a new extraction method of the PD signals based on time-frequency manifolds. This method uses phase space reconstruction to convert a PD signal into multiple sub-sequences, calculates their respective time-frequency distributions, and constructs dynamic time-frequency manifolds of the raw PD signal. Then, using locally linear embedding, the high-dimensional data are mapped to a low dimensional space where feature parameters of the PD signals are extracted. The new method has been applied to identification of PD patterns of different generators using a K-nearest neighbor classifier (KNNC). Its failure recognition rate is higher than 95%.