KJCKorea
Journal Central

Obstacle detection in blind spots has emerged as a critical challenge due to the increasing adoption of Automated Guided Vehicle (AGV) and Autonomous Mobile Robot (AMR) systems in modern industrial environments. While conventional ultrasonic and single-bounce LiDAR sensors exhibit limitations in complex environments, Two-Bounce LiDAR offers higher detection precision but faces challenges such as high cost and limited diversity in training data. To address these challenge, this study proposes a framework that combines Two-Bounce LiDAR with Digital Twin simulation to generate synthetic data under various scenarios. Object detection models trained with this data achieved an average detection accuracy of 92.5% in environments including blind spots, with an average improvement of 14.67 percentage points in accuracy and 15.69 points in F1-score compared to models trained only on normal-environment data. These findings demonstrate that the proposed framework effectively overcomes data collection limitations and significantly enhances the safety and reliability of AGV and AMR operations.

In this study, we developed a model to support nurse decision-making using Korean nursing record data and explored methods to enhance performance by applying data augmentation techniques. Previous research primarily focused on English medical data, resulting in a lack of studies on Korean medical data. To address this gap, we utilized electronic medical record (EMR) data from abdominal surgery patients and developed a KoBERT-based model for predicting nursing actions. Additionally, we applied techniques such as up/down sampling, few-shot augmentation, back-translation, and synonym replacement to mitigate data imbalance and compared their performance. Experimental results show that the Few-shot Augmentation achieved the highest performance, confirming that data augmentation is effective in increasing the diversity of EMR data.

In this study, we propose the Hi-BERT model to identify effectively and to learn semantic associations between recruitment ontology generated automatically from recruitment information and academic papers using generative AI. The proposed model can learn semantic associations between ontology-transformed sentences, which are structural information of recruitment ontology converted into natural language, and academic papers. In addition, domain-specific characteristics are learned from recruitment information and academic papers, and optimized by the contrastive learning technique. The proposed model can improve the limitations of knowledge transfer between different domains of existing sentence embedding models such as KLUE and KoSimCSE-BERT. In particular, the learning method utilizing ontology-transformed sentences proposed in this study can provide richer semantic understanding than existing models trained by simple text corpus. In the future, the proposed model is expected to be used as a base model for developing expert search systems.