Towards Trustworthy Internal Due-Date Assignment in Manufacturing Systems: A Neutrosophic Neural Network-Based Approach

Internal due-date assignment begins by estimating the completion time of a possible order, typically based on the completion of its last job. However, most machine learning and deep learning approaches focus on overall estimation accuracy, which may result in inaccurate predictions for individual cases. Consequently, the trustworthiness of individual internal due-date assignments remains uncertain, an issue that has received limited attention. To address this gap, this study proposes a neutrosophic neural network approach. Initially, a fuzzy deep neural network is constructed and trained to estimate the completion time of a possible order. The resulting predictions are then converted into neutrosophic sets utilizing three membership functions. The truth and falsity memberships are determined from global estimation accuracy based on root mean squared error, whereas the indeterminacy membership incorporates local estimation accuracy and local learning sufficiency. An internal due-date assignment is considered trustworthy when its indeterminacy falls below a prespecified threshold. When users believe that AI applications are trustworthy, their cognition shifts, which can reduce the cognitive burden on users. In addition, both fuzzy logic and deep neural network (DNN) are well-known and widely used biologically algorithms. The proposed methodology is applied to a real-world case in a wafer fabrication facility (wafer fab). Experimental results indicate 44% of the internal due-date assignments were not trustworthy, primarily because of local learning insufficiency and local estimation inaccuracy. Furthermore, targeted collection of additional similar examples reduced untrustworthy predictions by 50%, providing practical guidance for subsequent data collection efforts. By identifying the reliability of individual predictions, the proposed approach can also improve user confidence in AI applications and reduce users’ cognitive burden.