This study aimed to develop a model for accurately predicting the acute aquatic toxicity (48h- EC50) of chlorine disinfection by-products (DBPs). DBPs have caused environmental risks, but experimental toxicity data are difficult to obtain due to time, cost, and ethical constraints. Therefore, a deep learning model was developed using actual concentration-based data. Toxicity data for 139 aliphatic chlorinated compounds were from the OECD QSAR Toolbox and from aquatic toxicity test results provided by the japan ministry of the environment. Various concentration criteria, including nominal and measured concentrations, were encoded as additional inputs, and EC50 values were augmented via log transformation and structural string modifications to overcome small data limitations. The directed message passing neural network (D-MPNN) model, which considers bond directionality, was applied to reflect structural complexity accurately. Also, this model effectively reflected subtle structural differences and showed stable performance even with limited data. Comparisons between models with and without concentration criteria revealed that the model considering all concentration criteria had superior predictive accuracy. This result shows that concentration criteria are a critical factor in toxicity prediction. This study suggests a baseline model that works reliably even with small datasets reflecting realistic concentration criteria, showing its potential use for replacing some experiments and for screening toxic substances.

This study compares the performance of various convolutional neural network (CNN) models for building an automated deep learning-based letter screening system targeting letters received by inmates in correctional institutions. The models evaluated include well-known architectures such as MobileNet, ResNet, and Inception, as well as recently proposed lightweight models such as ResMobileNet and IGSe, along with GroupConv and SE. Each model was trained on image data containing the Korean word for "knife" ("칼") to assess performance in terms of accuracy, processing time, and model compactness. A total of 1,197 letter image samples were used in the experiment, including 1,140 images with normal words and 57 images containing the target word. The experimental results showed that the MobileNet model had the shortest processing time, making it suitable for real-time applications, while the IGSe model achieved the highest accuracy, demonstrating optimal performance for letter screening tasks. This study suggests that deep learning-based screening techniques can be effectively applied to enhance digital security in the management of inmate correspondence within correctional institutions.

This study proposes a weighted ensemble deep learning framework for accurately predicting the State of Health (SOH) of lithium-ion batteries. Three distinct model architectures—CNN-LSTM, Transformer-LSTM, and CEEMDAN-BiGRU—are combined using a normalized inverse RMSE-based weighting scheme to enhance predictive performance. Unlike conventional approaches using fixed hyperparameter settings, this study employs Bayesian Optimization via Optuna to automatically tune key hyperparameters such as time steps (range: 10-35) and hidden units (range: 32-128). To ensure robustness and reproducibility, ten independent runs were conducted with different random seeds. Experimental evaluations were performed using the NASA Ames B0047 cell discharge dataset. The ensemble model achieved an average RMSE of 0.01381 with a standard deviation of ±0.00190, outperforming the best single model (CEEMDAN-BiGRU, average RMSE: 0.01487) in both accuracy and stability. Additionally, the ensemble's average inference time of 3.83 seconds demonstrates its practical feasibility for real-time Battery Management System (BMS) integration. The proposed framework effectively leverages complementary model characteristics and automated optimization strategies to provide accurate and stable SOH predictions for lithium-ion batteries.

Fault detection in electromechanical systems plays a significant role in product quality and manufacturing efficiency during the transition to smart manufacturing. Because collecting a sufficient number of datasets under faulty conditions of the system is challenging in practical industrial sites, unsupervised fault detection methods are mainly used. Although fault datasets accumulate during machine operation, it is not straightforward to utilize the information it contains for fault detection after the deep learning model has been trained in an unsupervised manner. However, the information in fault datasets is expected to significantly contribute to fault detection. In this regard, this study aims to validate the effectiveness of the transition from unsupervised to supervised learning as fault datasets gradually accumulate through continuous machine operation. We also focus on experimentally analyzing how changes in the learning paradigm of the deep learning model and the output representation affect fault detection performance. The results demonstrate that, with a small number of fault datasets, a supervised model with continuous outputs as a regression problem showed better fault detection performance than the original model with one-hot encoded outputs (as a classification problem).

This study developed a deep learning-based software module for classifying the ripeness of bananas in real time as they move along a conveyor belt. A total of 5,286 images annotated with three ripeness stages, namely unripe, ripe, and overripe, were divided into training, validation, and test datasets at a ratio of 88:8:4. The datasets were used to train YOLOv5s and YOLOv5l object detection models over 50 epochs. The model performance was evaluated using box loss, object loss, class loss, and mean average precision (mAP). Both models exhibited decreasing loss values approaching zero and achieved mAP, precision, and recall scores exceeding 90%, thus indicating a robust classification performance without overfitting. The software module integrated with the trained YOLOv5l model accurately identified the ripeness stage of bananas in motion on the conveyor system without misclassification. Collectively, these findings indicate that the proposed system can be effectively applied to banana-processing lines for automated and accurate ripeness-based sorting.

본 연구는 딥러닝 기술을 활용하여 공동주택 외벽의 균열 탐지를 효과적으로 하기 위한 다양한 데이터 전처리 방법을 비교 분석하 였다. 특히, 표준 균열 데이터셋에 일반적으로 나타나지 않는 오탐균열을 식별하는 데 중점을 두고 있다. 이 연구는 탐지 정확도를 최 적화하기 위해 여러 이미지 전처리 기법을 적용한 결과를 비교한다. 객체 탐지를 위한 엣지 필터링과 RGB 색 필터 등을 이용한 색상 정규화를 결합한 방법을 집중적으로 검증하였다. 이러한 기술들은 실제 균열과 오탐균열을 구분하기 위해 적용되었으며, 이들의 탐 지 성능에 미치는 영향을 철저히 조사하였다. 효율적인 균열 탐지 모델을 찾기 위해 EfficientNet V2s 기반 모델을 적용하였다. RGB, YUV, LAB, HSV 네 가지 이미지 필터가 원본 이미지와 CLAHE 정규화된 이미지에 적용되었는데, 그 결과 단색 콘크리트 균열 탐지 에 효과적인 전통적인 정규화 방법이 공동주택 외벽 균열 탐지에는 제한적인 효과를 보인다는 것을 확인하였다. 또한, 단일 색 필터의 적용이 일관된 탐지 결과 개선 효과를 주지 않는다는 것을 밝혔다. 결국, 본 연구를 통해 다양한 이미지 정규화와 색 필터 조합의 균열 탐지 성능을 검증하였으며, 실제 균열과 오탐균열을 구분하는 탐지 성능 향상을 위해 추가적으로 다양한 접근의 연구가 필요하다는 것을 확인하였다.

Germination of chili pepper seeds is critical for crop yield and resource utilization. A high germination rate increases yield and effectively reduces resource wastage. This study collected 450 macroscopic images of chili pepper seeds and constructed a dataset for deep learning training through standardized germination experiments. Six deep learning models were evaluated to improve the chili pepper seed classification accuracy and germination rate. After comparing the performance of the models, MobileNet_v2 performed the best, not only having the fewest number of parameters but also achieving a 98.89% accuracy and 97.82% F1 score. The model improved the original germination rate from 87.33% to 100% on the test set, significantly optimizing the seed selection process

As the E-commerce market grows, the importance of personalized recommendation systems is increasing. Existing collaborative filtering and content-based filtering methods have shown a certain level of performance, but they have limitations such as cold start, data sparseness, and lack of long-term pattern learning. In this study, we design a matching system that combines a hybrid recommendation system and hyper-personalization technology and propose an efficient recommendation system. The core of the study is to develop a recommendation model that can improve recommendation accuracy and increase user satisfaction compared to existing systems. The proposed elements are as follows. First, the hybrid-hyper-personalization matching system provides recommendation accuracy compared to existing methods. Second, we propose an optimal product matching model that reflects user context using real-time data. Third, we optimize Personalized Recommendation System using deep learning and reinforcement learning. Fourth, we present a method to objectively evaluate recommendation performance through A/B testing.

본 연구는 딥러닝 영상 재구성 기법을 적용한 8개의 뇌질환군의 감마나이프 수술 계획용 자기공명영상(magnetic resonance imaging, MRI)의 유용성을 알아보고자 하였다. 연구 방법은 전이성 뇌종양, 뇌동정맥 기형, 수막종, 뇌하수체선종, 삼차신경통, 청신경초종, 맥락얼기 유두종, 해면상 혈관종, 총 8개의 질병을 진단받은 사람들의 T2 강조 영상(T2 weighted imaging, T2WI), 조영증강 T1 강조영상(contrast enhancement T1 weighted imaging, CE-T1WI)의 방법으로 검사한 MRI 영상을 SwiftMR을 이용하여 딥러닝 영상 재구성 기법인 디노이징(denoising)과 초해상도(super resolution)가 적용된 영상을 획득하였다. 이에 대한 성능 평가는 최대 신호대잡음비(peak signal to noise ratio, PSNR), 구조적 유사도(structural similarity index measure, SSIM), 감마나이프 방사선수술(gamma knife radiosurgery, GKRS)의 좌표계로 평가하였다. 그 결과, 원본영상을 기반으로 영상 품질이 개선된 영상의 PSNR과 SSIM은 높은 수치를 나타냄으로써 MRI 영상의 재구성이 문제없이 이루어졌고, GKRS의 수술 좌표계 또한 변화를 보이지 않았다. 결론적으로 딥러닝 영상 재구성 기법은 영상 품질 향상과 영상 보존에서 뛰어난 성능을 보임과 동시에 좌표계도 변화를 보이지 않아서, 딥러닝 영상 재구성 기법은 감마나이프 수술 계획에 유용하게 사용할 수 있는 기법임을 확인하였다.