In order to investigate the influence of rock interface roughness on the characteristics of the charge induction signal during fault slip, the time-frequency characteristics of the multi-channel charge induction signal waveforms, the cumulative velocity of charge, the fractal dimension, and the primary and secondary frequency zones of the rock assemblage structure with different roughness during the slip process in the double-sided shear test under different vertical loads were investigated. The results show that: (1) The localized micro-rupture nucleation in the elastic deformation stage leads to multiple charge induction clusters with maximum values, which increase with the increase of interface roughness and vertical load, and then become dense and small-amplitude signals when entering into the start-slip stage. (2) With the increase of interface roughness and vertical load, the fluctuation of the accumulated charge velocity and fractal dimension are more obvious and highly correlated with the change of the waveform of the charge induction signal. In the elastic deformation stage, the accumulated charge velocity shows “slow increase in the main body and sudden increase in multiple points”, and each charge induction cluster is accompanied by the phenomenon of “first ascending and then descending” of the fractal dimension, with the main frequency area located in the low-frequency domain and the sub-main frequency area located in the high-frequency domain. In the start-slip stage, the accumulated charge velocity changes to an overall rapid increase and the fractal dimension fluctuates more obviously with the increase of fault interface roughness and vertical loading. During the start-slip stage, the charge accumulation rate changes to an overall rapid increase, and the fractal dimension is continuously downgraded, and the primary and secondary frequency regions show the phenomenon of “translational interchange”, with the primary frequency region shifted right to the high-frequency domain, and the secondary frequency region shifted left to the low-frequency domain, and the primary frequency of the charge signals at each slip stage falls into the frequency aliasing domain common to the whole process of slipping. (3) Comparing the time-frequency resolution and time-frequency focusing of the three time-frequency transform methods, wavelet transform, short-time Fourier transform and S transform, it is found that the wavelet transform performs the best in the low-frequency domain, the short-time Fourier transform the second, and the S transform the worst, while in the high-frequency domain, the S transform performs the best, the wavelet transform the second, and the short-time Fourier transform the worst. (4) Differences in charge signals of sensors at different locations during fault slip destabilization are mainly related to the aggregation of charges in specific regions caused by locally concentrated micro-ruptures before the start-slip phase, and are mainly caused by the change of misalignment of the relative positions between the slip surface and the sensors after the start-slip phase.