최근 고도화된 딥러닝 모형을 이용하여 하천 수질에 영향을 줄 수 있는 과도한 조류(algae) 발생을 예측하는 연구에 대한 관심이 지속되고 있으며, 모형의 구축에 사용되는 현장 측정 자료의 특성상 다양한 이상치를 포함할 수 있어 데이터의 이상치 관리 필요성이 높아지고 있다. 본 연구에서는 현장 자료의 이상치가 딥러닝 모형의 성능에 미치는 영향을 분석하기 위해 딥러닝 Long Short-Term Memory(LSTM) 모형을 이용하여 하천 조류 발생을 정량적으로 평가하는 지표인 클로로필-a를 예측하는 모형을 구축하였으며, 10%의 이상치를 포함한 자료와 이상치가 포함되지 않은 원본 자료로 학습된 모형의 성능을 비교하였다. 또한 딥러닝 기반 이상치 탐지 알고리즘인 Autoencoder(AE)를 이용하여 이상치를 제거한 후 모형의 성능에 미치는 영향을 비교하였다. 분석 결과 이상치를 포함하지 않은 자료로 학습된 Base 모형과 10%의 이상치를 포함한 자료로 학습된 모형의 Nash-Sutcliffe efficiency(NSE)가 각각 0.882 및 0.858로 나타나 이상치가 모형의 성능을 저하시킬 수 있음을 확인하였다. 한편 AE를 이용하여 이상치를 다양한 비율로(5–20%) 제거한 자료로 학습된 모형의 성능을 분석한 결과 NSE가 0.883–0.896으로 이상치의 제거에 따라 모형의 성능이 Base 모형과 유사한 수준으로 개선되는 것으로 나타났다. 본 연구에서는 이상치가 딥러닝 모형에 미치는 영향을 분석하고 이상치 탐지 모형의 활용에 따른 조류 발생 예측 딥러닝 모형의 성능 향상이 가능함을 확인하였다.

High-entropy alloys (HEAs) are alloys that contain multiple principal elements, each in the range of 5–35%. HEAs exhibit excellent properties, however, even with conventional trial-and-error, high-throughput experimentation, and computational materials approaches, exploring their vast compositional space remains highly challenging. Accordingly, data-driven machine learning and generative-model-based inverse design methods are increasingly essential. In this study, we propose a generative-model-enabled HEA inverse design framework aimed at improving ultimate tensile strength (UTS). We first compiled 501 HEA data points from published literature and performed statistical analyses to understand their characteristics. Next, we tuned the hyperparameters of XGBoost and random forest (RF) models via Bayesian optimization, compared their performance with that of a deep neural network (DNN), and selected XGBoost as the optimal predictive model. In the subsequent stage, we trained a PyTorch-based variational autoencoder (VAE) on data from regions of the latent space associated with high-UTS probability. We randomly sampled 1,000 latent vectors, decoded them to generate candidate alloy compositions, and evaluated these candidates using the optimized XGB model. Finally, Shapley additive explanations (SHAP) interpretability analysis and a network plot were used to quantify the contributions and interactions of each feature variable, thereby assessing the physical plausibility of the model-suggested compositions.

Anomaly detection is a key technique for ensuring the reliability and stability of systems across various industrial domains. Autoencoder-based reconstruction models are particularly effective in learning normal patterns and detecting deviations. However, conventional loss functions such as Mean Squared Error (MSE) and Mean Absolute Error (MAE) are limited in capturing anomalies that follow heavy-tailed or asymmetric distributions, which are commonly observed in real-world industrial settings. To address this limitation, we propose a Mixture Negative Log-Likelihood (Mixture NLL) loss function based on a combination of Gaussian, Laplace, and Student-t distributions. The loss is constructed using the probability density functions of each distribution, with key parameters such as standard deviation, scale, and degrees of freedom learned during training. The mixture weights representing the contribution of each distribution are also jointly optimized. Experimental results on real-world time-series anomaly detection datasets demonstrate that the proposed MixtureLoss consistently outperforms conventional loss-based Autoencoder models, particularly in detecting tail-end anomalies.

Anomaly detection technique for the Unmanned Aerial Vehicles (UAVs) is one of the important techniques for ensuring airframe stability. There have been many researches on anomaly detection techniques using deep learning. However, most of research on the anomaly detection techniques are not consider the limited computational processing power and available energy of UAVs. Deep learning model convert to the model compression has significant advantages in terms of computational and energy efficiency for machine learning and deep learning. Therefore, this paper suggests a real-time anomaly detection model for the UAVs, achieved through model compression. The suggested anomaly detection model has three main layers which are a convolutional neural network (CNN) layer, a long short-term memory model (LSTM) layer, and an autoencoder (AE) layer. The suggested anomaly detection model undergoes model compression to increase computational efficiency. The model compression has same level of accuracy to that of the original model while reducing computational processing time of the UAVs. The proposed model can increase the stability of UAVs from a software perspective and is expected to contribute to improving UAVs efficiency through increased available computational capacity from a hardware perspective.

This study explores the use of a Deep Autoencoder model to predict depression among plant and machine operators, utilizing data from the Korean National Health and Nutrition Examination Survey (KNHANES, n=3,852). The Deep Autoencoder model outperformed the Logistic Regression, Naive Bayes, XGBoost, and LightGBM models, achieving an accuracy of 86.5%. Key factors influencing depression included work stress, exposure to hazardous substances, and ergonomic conditions. The findings highlight the potential of the Deep Autoencoder model as a robust tool for early identification and intervention in workplace mental health.

Anomaly detection for each industrial machine is recognized as one of the essential techniques for machine condition monitoring and preventive maintenance. Anomaly detection of industrial machinery relies on various diagonal data from equipped sensors, such as temperature, pressure, electric current, vibration, and sound, to name a few. Among these data, sound data are easy to collect in the factory due to the relatively low installation cost of microphones to existing facilities. We develop a real time anomalous sound detection (ASD) system with the use of Autoencoder (AE) models in the industrial environments. The proposed processing pipeline makes use of the audio features extracted from the streaming audio signal captured by a single-channel microphone. The pipeline trains AE model by the collected normal sound. In real factory applications, the reconstruction error generated by the trained AE model with new input sound streaming is calculated to measure the degree of abnormality of the sound event. The sound is identified as anomalous if the reconstruction error exceeds the preset threshold. In our experiment on the CNC milling machining, the proposed system shows 0.9877 area under curve (AUC) score.

Recently there was an incident that military radars, coastal CCTVs and other surveillance equipment captured a small rubber boat smuggling a group of illegal immigrants into South Korea, but guards on duty failed to notice it until after they reached the shore and fled. After that, the detection of such vessels before it reach to the Korean shore has emerged as an important issue to be solved. In the fields of marine navigation, Automatic Identification System (AIS) is widely equipped in vessels, and the vessels incessantly transmits its position information. In this paper, we propose a method of automatically identifying abnormally behaving vessels with AIS using convolutional autoencoder (CAE). Vessel anomaly detection can be referred to as the process of detecting its trajectory that significantly deviated from the majority of the trajectories. In this method, the normal vessel trajectory is gridded as an image, and CAE are trained with images from historical normal vessel trajectories to reconstruct the input image. Features of normal trajectories are captured into weights in CAE. As a result, images of the trajectories of abnormal behaving vessels are poorly reconstructed and end up with large reconstruction errors. We show how correctly the model detects simulated abnormal trajectories shifted a few pixel from normal trajectories. Since the proposed model identifies abnormally behaving ships using actual AIS data, it is expected to contribute to the strengthening of security level when it is applied to various maritime surveillance systems.

많은 사용자가 함께 즐기는 온라인 게임(MMOGs)에서 IoT의 확장은 서버에 엄청난 부하를 지 속적으로 증가시켜, 모든 데이터들이 Big-Data화 되어가는 환경에 있다. 이에 본 논문에서는 딥러 닝 기법 중에서 가장 많이 사용되는 Sparse Autoencoder와 이미 잘 알려진 부하분산 알고리즘 (ProGReGA-KF)을 결합한다. 기존 알고리즘 ProGReGA-KF과 본 논문에서 제안한 알고리즘을 이동 안정성으로 비교하였고, 제안한 알고리즘이 빅-데이터 환경에서 좀 더 안정적이고 확장성이 있 음 시뮬레이션을 통해 보였다.