Prediction-decision collaborative optimization method for planning problem in mobile-phone assembly line

As the competition in the electronics manufacturing industry becomes increasingly fierce, the product update cycle keeps shortening. It becomes increasingly important for electronics manufacturing enterprises to accurately predict future demands. However, the existing research on demand prediction mainly relies on statistical indicators to evaluate the accuracy of the predicted result, and ignores the impact of different levels of prediction errors on production scheduling decisions. Therefore, this paper proposes a “prediction-decision collaborative optimization” (PDCO) method and constructs an end-to-end joint training framework. The framework designs a new error transmission mechanism, incorporates the quality of the production scheduling decision into the evaluation indicators, while training the demand prediction process. Moreover, since the production planning problem in the assembly line is a discrete mixed integer programming problem, we design a gradient approximation method to connect the back-propagation channel of the demand prediction neural network. We conducted experiments using the designed method in a scenario of a mobile phone assembly enterprise and compared our method with the traditional "predict first, then decide" method. The results show that our method improves the decision quality and can be effectively applied to the assembly line production scenarios of electronic products.