Fuzzy CNN for job cycle time bounding in wafer fabrication

Abstract Job cycle time bounding is crucial for factories, as the upper and lower bounds of job cycle time can be used for various production control and management activities. However, precise job cycle time bounding remains a challenging task. Deep learning (DL) applications hold promise for overcoming this challenge. Convolutional neural networks (CNNs), as a highly attractive deep learning (DL) architecture, have achieved remarkable results in image and speech recognition, but have not yet been applied to job cycle time prediction or bounding. To explore the potential of CNNs in this field, this study proposes a fuzzy convolutional neural network (FCNN) approach to further enhance the effectiveness of job cycle time bounding. In the FCNN approach, a CNN with dropout is first constructed for job cycle time prediction. Job cycle time related data are then resized to be compatible for the CNN. Subsequently, network parameters are fuzzified to derive the lower and upper bounds of job cycle time. The FCNN approach has been experimentally applied to a wafer fabrication case. According to the experimental results, the CNN outperformed a deep neural network (DNN) in job cycle time prediction by reducing the mean absolute percentage error (MAPE) by up to 33%. In addition, the FCNN achieved an advantage of 19% over a fuzzy DNN (FDNN) in improving the job cycle time bounding performance in terms of cost for inclusion (CFI). Furthermore, compared to other deep, feedforward neural networks, the CNN required fewer parameters, making it a highly attractive DL architecture for job cycle time bounding.