The weak light signal of underwater laser communication system is interfered by spectrum and multipath, and the detection accuracy is not high. In order to improve the stability of underwater laser communication, a detection method of weak light signal of underwater laser communication system based on deep learning is proposed. The spatial beam forming method is used to construct the underwater laser communication acquisition and multi-layer spectrum decomposition model, and the beam characteristics of the signal are enhanced. The radio frequency tag distribution information and multipath information of the weak optical signal are estimated by combining the time reversal technology, and the weak optical signal of the underwater laser communication system is enhanced by using multipath fading suppression. The optimization control is carried out in the detection process of the weak optical signal by combining the deep learning algorithm, so that the whole signal detection process is guaranteed to have a constant false alarm probability. The simulation results show that this method has a good ability to suppress multipath fading when detecting weak light signals in underwater laser communication systems. The error rate is low in different signal-to-noise ratio environments. When the signal-to-noise ratio increases to 90 dB, the error rate is only 10-5, and the detection time is less than 10 ms, which is superior to the comparison method and has greater application value.