A Modified Bayesian Network Model to Predict Reorder Level of Printed Circuit Board

Identifying the printed circuit board (PCB) orders with high reorder frequency for batch production can facilitate production capacity balance and reduce cost. In this paper, the repeated orders identification problem is transformed to a reorder level prediction problem. A prediction model based on a modified Bayesian network (BN) with Monte Carlo simulations is presented to identify related variables and evaluate their effects on the reorder level. From the historically accumulated data, different characteristic variables are extracted and specified for the model. Normalization and principal component analysis (PCA) are employed to reduce differences and the redundancy of the datasets, respectively. Entropy minimization based binning is presented to discretize model variables and, therefore, reduce input type and capture better prediction performance. Subsequently, conditional mutual information and link strength percentage are combined for the establishment of BN structure to avoid the defect of tree augmented naïve BN that easily misses strong links between nodes and generates redundant weak links. Monte Carlo simulation is conducted to weaken the influence of uncertainty factors. The model’s performance is compared to three advanced approaches by using the data from a PCB manufacturer and results demonstrate that the proposed method has high prediction accuracy.