The optical signal is easy to be absorbed and scattered during transmission with Underwater Optical Wireless Communication (UWOC) technology, resulting in serious optical power attenuation and further affecting the signal quality. In order to realize long-distance data transmission, it is very important to recognize, enhance and extract weak light signal under low Signal-to-Noise Ratio (SNR). Stochastic resonance produces synergistic effect through nonlinear system, weak driving signal and appropriate amount of noise under certain conditions, which not only improves the output signal-to-noise ratio, but also detects useful signals. However, the current parameter selection of stochastic resonance system depends on artificial setting, which is not flexible enough to give full play to the advantages of stochastic resonance signal detection. In this paper, an adaptive stochastic resonance detection scheme based on multi-strategy fusion particle swarm optimization is proposed by analyzing the characteristics of weak underwater light signals and the conditions of stochastic resonance generation. It solves the problem that traditional particle swarm optimization is easy to fall into local optimization resulting in low convergence accuracy and difficult convergence. By introducing adaptive inertia weights to dynamically adjust the local search ability and global search ability of particles, the convergence speed of the algorithm is accelerated. In the process of particle evolution, neighborhood detection is used to strengthen the detection of local extremum location neighborhood, which makes the search radius of the algorithm larger in the initial stage of evolution, and gradually decreases with the increase of iteration times, which increases the refinement ability of the algorithm. Using Cauchy variation and reverse learning interactive strategy to mutate the optimal solution, the local optimal solution in Particle Swarm Optimization is broken, and the ability of the algorithm to escape from local space is effectively improved. In order to evaluate the feasibility and effectiveness of the proposed algorithm, simulation is carried out under the established UWOC weak signal detection system. Considering the special property of pilot signal, that is, some known data is inserted at the sending end and can be accurately extracted at the receiving end, it can be used as a reliable reference signal for parameter estimation. Therefore, this paper selects a specific number of code elements for parameter optimization. By taking the output SNR of the system as the selection index, the optimal system parameter which makes the output SNR maximum is searched and iterated continuously within the preset algorithm parameter range. The optimal system parameters are substituted into the fourth-order Runge-Kutta equation, the output response is obtained by discretization, and the weak light signal is detected. Finally, the error performance of bipolar non-return-to-zero signal with white Gaussian noise is compared under four detection schemes: non-stochastic resonance, fixed parameter stochastic resonance, adaptive stochastic resonance based on particle swarm optimization algorithm and multi-strategy fusion particle swarm optimization algorithm. The simulation results show that the bit error rate performance of the non-stochastic resonance system is worse than that of the other three detection schemes, and the bit error rate performance of the fixed parameter stochastic resonance system has limitations. Adaptive stochastic resonance can significantly improve the bit error rate performance of the system, especially above -6 dB, and the improvement effect is very obvious. Compared with the adaptive stochastic resonance based on particle swarm optimization algorithm, the proposed algorithm has faster convergence speed, more accurate optimization results and less error performance. In order to verify the effectiveness and feasibility of the proposed method, a UWOC experimental system is established. The experimental results show that when the received signal-to-noise ratio is -1.7 dB, the bit error rate of the proposed algorithm can reach 2×10^-4, and its performance is better than that of NO-SR and F-SR, which once again verifies the effectiveness of the proposed algorithm.