The invention belongs to the technical field of visible light communication, and particularly relates to a machine learning multi-band carrier-free amplitude phase modulation system based on independent component analysis (ICA). The system mainly comprises a matched filtering & down-sampling module, a first-stage CAP forming and cross-matched filtering module, an ICA sub-band signal purification module, a phase offset recovery module of each sub-band signal, a second-stage CAP forming and cross-matched filtering module, a subtraction sub-band interference module, a minimum mean square filtering module and the like. According to the invention, sub-band interference inversion is carried out at a receiving end; the signals and original receiving sub-band signals are sent to ICA for preliminary purification of each path of sub-band signals; secondly, performing second-stage sub-band interference inversion, and subtracting interference between sub-bands; and the second-stage equalization isadopted, so that the guard interval between the sub-bands can be eliminated, the tolerance degree of the system to aliasing between the sub-bands can be improved, the same transmission rate can be realized by using less bandwidth, the signal bandwidth can be utilized more effectively, and the spectrum utilization rate of the visible light communication system can be improved.