Negative matrix factorization (NMF), a tool for matrix factorization and nonlinear dimension reduction, has been widely used for factorization coding and feature extraction of vibration spectrograms from multiple simples. However, the systematic exploration and research on the NMF coding of single vibration spectrogram and the relationships between NMF factorization vector and vibration spectrogram components is still lacking. The basic principles of feature extraction using single spectrogram coding and demodulation are explained. The ability of the NMF to part-based characterize single spectrogram, the bandpass filtering amplitude-frequency characteristics (BFAC) of single spectrogram NMF basis vector, and the synchronous change characteristic between NMF coding vector and spectrogram component are mainly discussed. Two novel feature extraction methods using nonnegative matrix coding and demodulation of single spectrogram are developed, filtering demodulation based on NMF basis vector and demodulation directly performed on NMF coding vector. A BFAC index is defined for adaptive selection of NMF low dimensional parameter, likewise an optimal iteration rule with NMF basis vector normalization for optimization process of NMF coding. The proposed method is applied to analyze the simulated signal and gearbox vibration signal. Under the conditions of given factorization rank and NMF maximum iterations of 300, it takes about 3. 5 s to extract the features of data with the length of 6. 4 k points. Meanwhile, the fault characteristics of the simulated signal with signal-to-noise ratio of -10 dB and that of the gearbox vibration signal with compound faults are extracted.