In order to reliably diagnose the open circuit fault of IGBT in active power filter, a fault feature extraction method of IGBT in active power filter is proposed based on multi  feature fusion. The method acquires the voltage across the clamped diode bridge arm in the threelevel APF main circuit as the measurement signal, to which wavelet decomposition is conducted. The energy coefficient, power spectrum entropy and singular spectrum entropy of each frequency band are extracted to compose the multifeature parameter matrix. Then, the feature dimension reduction is conducted to compose the eigenvector matrix. On the basis of theoretical analysis, corresponding experimental analysis was performed. Firstly, the measurement waveforms under different working states were obtained based on the above measurement signals, which were compared with those of other measurement signals; then, the kernel fuzzy Cmeans clustering method was used to analyze the distinguishing performance of the extracted features for the fault type, and the adaptive experiment of feature extraction on a threephase rectifier bridge harmonic source under load mutation and triggering angle change was conducted. Finally, the practical experimental platform of APF is built for further testing.The experiment results show that the measurement method based on the voltage across the diode can effectively distinguish different working states, and the adopted multifeature fusion extraction method overcomes the onesidedness of single feature extraction method, and has good distinguishing performance under various working conditions.