Wasserstein distributionally robust learning for predicting the cycle time of printed circuit board production

This paper proposes a Wasserstein distributionally robust learning (WDRL) model to predict the production cycle time of highly mixed printed circuit board (PCB) orders on multiple production lines. The PCB production cycle time is essential for optimizing production plans. However, the design of the PCB, production line configuration, order combinations, and personnel factors make the prediction of the PCB production cycle time difficult. In addition, practical production situations with significant disturbances in feature data make traditional prediction models inaccurate, especially when there is new data. Therefore, we establishe a WDRL model, derive a tight upper bound for the expected loss function, and reformulate a tractable equivalent model based on the bound. To demonstrate the effectiveness of this method, we collected data related to surface mounted technology (SMT) production lines from a leading global display manufacturer for our computational experiments. In addition, we also designed experiments with perturbations in the training and testing datasets to verify the WDRL model’s ability to handle perturbations. This proposed method has also been compared with other machine learning methods, such as the support vector regression combined with symbiotic organism search, decision tree, and kernel extreme learning machine, among others. Experimental results indicate that the WDRL model can make robust predictions of PCB cycle time, which helps to effectively plan production capacity in uncertain situations and avoid overproduction or underproduction. Finally, we implement the WDRL model for the actual SMT production to predict the production cycle time and set it as the target for production. We observed a 98–103 % achievement rate in the last 20 months since the implementation in September 2022. • We propose a Wasserstein distributionally robust learning model to predict the production time. • We derive a tight upper bound and reformulate a tractable equivalent model. • We demonstrate the robustness of the proposed model through industrial datasets. • We implement it as the production target and observe a 98–103 % achievement rate.