2025, Vol.30, No.2

Journal of The Korea Society of Computer and Information 2023 KCI Impact Factor : 0.65

pISSN : 1598-849X / eISSN : 2383-9945
https://journal.kci.go.kr/jksci

Semiconductor production management becomes increasingly challenging due to the growing complexity of products, raising issues with the inefficiency of on-site scheduling. The complexity of the production environment and the multitude of variables make it difficult for traditional scheduling methods to effectively respond to real-time production changes, limiting the ability to achieve optimal productivity and quality. In this context, production experts are demanding solutions that provide insights into production direction and operational status. Based on the requirements, I conducted research on a solution that uses machine learning to analyze actual semiconductor Fab log data, identifying and visualizing decision-making factors that impact scheduling. This study is significant in that it suggests a direction for future schedulers and provides a foundation for building an Autonomous Fab.

This study proposes a multi-layer perceptron based regression-classification model for identifying the relevance between a given input question and its corresponding answer using data from the defense domain. First, an embedding vector method and pre-processing method are introduced to effectively handle input data. In the pre-processing method, if a class imbalance problem arises during the model input stage, a re-weighted sampling process is applied. And an additional algorithm is incorporated to identify and filter out contradictory data in advance. Furthermore, the relevance identification regression-classification model adopts a regression task structure while simultaneously transforming it into a classification task by introducing a regularization term in the model architecture. The experiments were conducted using a pre-constructed simulated dataset from the defense domain, and the inference results demonstrate the effectiveness and performance of the proposed model.

In this study, we investigate how to recognize children's emotions from their face images. Children's face image datasets are fewer in number than adult face image datasets and are often not publicly available on the Internet due to child research ethics. In this study, we propose a new data augmentation technique to improve the emotion recognition accuracy of the convolutional neural network (CNN) model in a situation where the number of samples in the child face image dataset is limited. To demonstrate the effectiveness of the proposed data augmentation technique, we conducted experiments using the child face image dataset of the National Institute of Mental Health, which is publicly available on the Internet. The experimental results showed that the CNN model achieved the best performance in the emotion recognition task when the proposed data augmentation technique was applied.