Optimization framework of laser oscillation welding based on a deep predictive reward reinforcement learning net

This research proposed a laser oscillation welding optimization system based on a deep reinforcement learning model with neural-network-based reward mechanism. A deep predictive reward reinforcement learning net (DPRRL-net) was developed to predict and improve the quality and efficiency of laser oscillation welding aluminum alloys by optimizing the process parameters, such as laser power, welding speed, oscillation amplitude, and oscillation frequency, with the porosity and penetration as the optimization target. A back propagation neural network (BPNN) prediction model optimized by differential evolutionary algorithm (DE) was established based on experimental results, and a linear weighting method was used to create a comprehensive evaluation system for weld quality and efficiency. The relative error between the predicted and experimental values of the DE-BPNN model was within 2.7%. The combination of reinforcement learning algorithm of twin delayed deep deterministic policy gradient (TD3) and comprehensive welding quality and efficiency prediction model was used to determine the optimal process parameters and avoid local optima or over-fitted solutions. Weld samples under these parameters showed a 7.04% increase in penetration and no porosity compared to traditional algorithm. The results demonstrated that the proposed method can effectively optimize the laser oscillation welding process parameters for aluminum alloys and significantly improve the weld quality and processing efficiency.