Prediction and optimization of joint quality in laser transmission welding using serial artificial neural networks and their integration with Markov decision process

Abstract

Laser transmission welding is a highly accurate method for joining plastics, but its diverse process parameters require effective modeling for optimal results. Traditional artificial neural networks (ANNs) typically establish predictive models between laser processing parameters and welding strength, neglecting the crucial role of welding morphology in feature extraction, thus diminishing accuracy. To address this, we developed a serial ANN model based on statistically evident correlations, which predicts joint morphology and strength sequentially, resulting in a 47% improvement in predictive accuracy and a mean error of just 7.13%. This two-layered approach effectively reduces the stepwise propagation of errors in ANNs, allowing the first layer to provide a refined data representation for the second layer to predict welding strength. Furthermore, finding the optimal laser parameter set is time-consuming and computationally demanding with traditional ANN-based optimization methods. To address this, we integrated the Markov decision process with the serial ANN for the first time and proposed a novel varying step strategy for the model, enabling a balance of swift convergence and avoidance of suboptimal solutions. Notably, the Markov-serial ANN model attained enhanced optimization results using only 15.5% of the computational resources required by a standard parameter interval optimization methodology. Welding experiments verified the reliability of the Markov-serial ANN, achieving a mean error of 4.54% for welding strength.