• The Journal of the Acoustical Society of Korea
• /
• v.38 no.4
• /
• pp.433-441
• /
• 2019

In this paper, a method of estimating speech parameters for ultrasonic Doppler signals reflected from the articulatory muscles using LSTM (Long Short Term Memory) RNN (Recurrent Neural Networks) was introduced and compared with the method using MLP (Multi-Layer Perceptrons). LSTM RNN were used to estimate the Fourier transform coefficients of speech signals from the ultrasonic Doppler signals. The log energy value of the Mel frequency band and the Fourier transform coefficients, which were extracted respectively from the ultrasonic Doppler signal and the speech signal, were used as the input and reference for training LSTM RNN. The performance of LSTM RNN and MLP was evaluated and compared by experiments using test data, and the RMSE (Root Mean Squared Error) was used as a measure. The RMSE of each experiment was 0.5810 and 0.7380, respectively. The difference was about 0.1570, so that it confirmed that the performance of the method using the LSTM RNN was better.

• Journal of the Korea Academia-Industrial cooperation Society
• /
• v.22 no.3
• /
• pp.234-241
• /
• 2021

By recognizing the importance of demand forecasting, the military is conducting many studies to improve the prediction accuracy for repair parts. Demand forecasting for repair parts is becoming a very important factor in budgeting and equipment availability. On the other hand, the demand for intermittent repair parts that have not constant sizes and intervals with the time series model currently used in the military is difficult to predict. This paper proposes a method to improve the prediction accuracy for intermittent repair parts of the Patriot. The authors collected intermittent repair parts data by classifying the demand types of 701 repair parts from 2013 to 2019. The temperature and operating time identified as external factors that can affect the failure were selected as input variables. The prediction accuracy was measured using both time series models and data mining models. As a result, the prediction accuracy of the data mining models was higher than that of the time series models, and the multilayer perceptron model showed the best performance.

• Journal of Digital Convergence
• /
• v.20 no.4
• /
• pp.389-396
• /
• 2022

Deep Learning is based on a perceptron, and is currently being used in various fields such as image recognition, voice recognition, object detection, and drug development. Accordingly, a variety of learning algorithms have been proposed, and the number of neurons constituting a neural network varies greatly among researchers. This study analyzed the learning characteristics according to the number of neurons of the currently used SGD, momentum methods, AdaGrad, RMSProp, and Adam methods. To this end, a neural network was constructed with one input layer, three hidden layers, and one output layer. ReLU was applied to the activation function, cross entropy error (CEE) was applied to the loss function, and MNIST was used for the experimental dataset. As a result, it was concluded that the number of neurons 100-300, the algorithm Adam, and the number of learning (iteraction) 200 would be the most efficient in deep learning learning. This study will provide implications for the algorithm to be developed and the reference value of the number of neurons given new learning data in the future.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.1
• /
• pp.67-74
• /
• 2023

The images acquired from the speaker's mouth region revealed the unique patterns according to the corresponding voices. By using this principle, the several methods were proposed in which speech signals were recognized or synthesized from the images acquired at the speaker's lower face. In this study, an image-based speech synthesis method was proposed in which the depth images were cooperatively used. Since depth images yielded depth information that cannot be acquired from optical image, it can be used for the purpose of supplementing flat optical images. In this paper, the usefulness of depth images from the perspective of speech synthesis was evaluated. The validation experiment was carried out on 60 Korean isolated words, it was confirmed that the performance in terms of both subjective and objective evaluation was comparable to the optical image-based method. When the two images were used in combination, performance improvements were observed compared with when each image was used alone.

• Journal of the Korea Society for Simulation
• /
• v.19 no.2
• /
• pp.91-98
• /
• 2010

Conventional path-finding has focused on finding short collision-free paths. However, as computer games become more sophisticated, it is required to take tactical information like ambush points or lines of enemy sight into account. One way to make this information have an effect on path-finding is to represent a heuristic function of a search algorithm as a weighted sum of tactics. In this paper we consider the problem of learning heuristic to optimize path-finding based on given tactical information. What is meant by learning is to produce a good weight vector for a heuristic function. Training examples for learning are given by a game level-designer and will be compared with search results in every search level to update weights. This paper proposes a learning algorithm integrated with search for tactical path-finding. The perceptron-like method for updating weights is described and a simulation tool for implementing these is presented. A level-designer can mark desired paths according to characters' properties in the heuristic learning tool and then it uses them as training examples to learn weights and shows traces of paths changing along with weight learning.

• Journal of the Korea Computer Graphics Society
• /
• v.29 no.4
• /
• pp.17-24
• /
• 2023

In computer graphics and computer vision research and its applications, various representations of 3D objects, such as point clouds, voxels, or triangular meshes, are used depending on the purpose. The need for animating characters using these representations is also growing. In a typical animation pipeline called skeletal animation, "skinning weight painting" is required to determine how joints influence a vertex on the character's skin. In this paper, we introduce a neural network for automatically performing skinning weight painting for characters represented in various formats. We utilize signed distance fields (SDF) to handle different representations and employ graph neural networks and multi-layer perceptrons to predict the skinning weights for a given point.

• Smart Media Journal
• /
• v.11 no.1
• /
• pp.58-66
• /
• 2022

As Internet technology develops, SNS users are increasing. As SNS becomes popular, SNS-type crimes using the influence and anonymity of social networks are increasing day by day. In this paper, we propose a fake account classification method that applies machine learning and deep learning to statistical and image data for fake accounts classification. SNS account data used for training was collected by itself, and the collected data is based on statistical data and image data. In the case of statistical data, machine learning and multi-layer perceptron were employed to train. Furthermore in the case of image data, a convolutional neural network (CNN) was utilized. Accordingly, it was confirmed that the overall performance of account classification was significantly meaningful.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2021.05a
• /
• pp.293-296
• /
• 2021

K-water에서는 다목적댐의 관리를 위해 실시간으로 댐수위, 하천 수위 및 강우량 등을 계측하고 있으며, 계측된 값들은 댐을 효과적으로 운영하는데 필요한 데이터로 활용되고 있다. 특히 댐수위 이상 데이터를 탐지하지 못한 채 그대로 사용할 경우 댐의 방류 시기와 방류량 등을 결정하는 중요한 의사결정을 그르칠 수 있으므로 이를 신속히 탐지하는 것이 매우 중요하다. 현재의 자동화된 이상 데이터 탐지방법 중 하나는 현재 데이터가 최댓값과 최솟값을 초과할 때, 다른 하나는 현재 데이터와 일정 시간 동안의 평균값 간의 차이가 관리자가 정한 특정 값을 벗어났을 때를 기준으로 삼고 있다. 전자는 상한과 하한의 초과 여부만 판단하므로 탐지가 쉬우나 정상범위 내에서 발생한 이상 데이터는 탐지가 불가하다. 후자는 관리자의 경험을 통해 판단 조건을 정하기 때문에 객관성이 결여되는 문제가 있다. 특히 방류와 강우가 복합적으로 댐수위에 영향을 미치는 홍수기에 관리자의 경험에 기초한 이상 데이터 판별은 신뢰성의 문제가 있을 수 있다. 따라서 본 연구에서는 기계학습을 최초로 적용하여 이상 데이터를 탐지하고자 하였다. 댐수위, 누적강우량 및 누적방류량 데이터와 댐수위데이터를 가공하여 생성한 댐수위차, 댐수위차평균, 댐수위평균 등 자질들의 다양한 조합을 만든 후 이를 Random Forest, SVM, AdaptiveBoost 및 다층퍼셉트론(MLP) 등과 같은 여러 가지 기계학습모델 등을 통해 이상 데이터를 판별하는 실험(분류)을 하였다. 실험결과 댐수위, 댐수위차, 댐수위-댐수위평균, 누적강우량, 누적방류량 및 댐수위차평균을 사용하였을 때 MLP에서 가장 우수한 성능을 보였다. 이 연구를 통해서 댐수위 이상 데이터를 기계학습의 분류기능을 통해 효과적으로 탐지할 수 있다는 것과 모델의 성능은 실험에 사용한 자질의 수뿐 아니라 자질의 종류에도 큰 영향을 받는다는 것을 알 수 있었다.

• Asia-pacific Journal of Multimedia Services Convergent with Art, Humanities, and Sociology
• /
• v.8 no.3
• /
• pp.57-67
• /
• 2018

The ensemble is a unified approach used for getting better performance by using multiple algorithms in machine learning. In this paper, we introduce boosting and bagging, which have been widely used in ensemble techniques, and design a method using support vector regression, radial basis function network, Gaussian process, and multilayer perceptron. In addition, our experiment was performed by adding a recurrent neural network and MOHID numerical model. The drifter data used for our experimental verification consist of 683 observations in seven regions. The performance of our ensemble technique is verified by comparison with four algorithms each. As verification, mean absolute error was adapted. The presented methods are based on ensemble models using bagging, boosting, and machine learning. The error rate was calculated by assigning the equal weight value and different weight value to each unit model in ensemble. The ensemble model using machine learning showed 61.7% improvement compared to the average of four machine learning technique.

• Proceedings of the Korea Water Resources Association Conference
• /
• 2021.06a
• /
• pp.317-317
• /
• 2021

효율적인 물관리를 위한 댐 유입량 대한 연구는 필수적이다. 본 연구에서는 다양한 머신러닝 알고리즘을 통해 40년동안의 기상 및 댐 유입량 데이터를 이용하여 소양강댐 유입량을 예측하였으며, 그 중 고유량과 저유량예측에 적합한 알고리즘을 각각 선정하여 머신러닝 알고리즘을 결합한 CombML을 개발하였다. 의사 결정 트리 (DT), 멀티 레이어 퍼셉트론 (MLP), 랜덤 포레스트(RF), 그래디언트 부스팅 (GB), RNN-LSTM 및 CNN-LSTM 알고리즘이 사용되었으며, 그 중 가장 정확도가 높은 모형과 고유량이 아닌 경우에서 특별히 예측 정확도가 높은 모형을 결합하여 결합 머신러닝 알고리즘 (CombML)을 개발 및 평가하였다. 사용된 알고리즘 중 MLP가 NSE 0.812, RMSE 77.218 m3/s, MAE 29.034 m3/s, R 0.924, R2 0.817로 댐 유입량 예측에서 최상의 결과를 보여주었으며, 댐 유입량이 100 m3/s 이하인 경우 앙상블 모델 (RF, GB) 이 댐 유입 예측에서 MLP보다 더 나은 성능을 보였다. 따라서, 유입량이 100 m3/s 이상 시의 평균 일일 강수량인 16 mm를 기준으로 강수가 16mm 이하인 경우 앙상블 방법 (RF 및 GB)을 사용하고 강수가 16 mm 이상인 경우 MLP를 사용하여 댐 유입을 예측하기 위해 두 가지 복합 머신러닝(CombML) 모델 (RF_MLP 및 GB_MLP)을 개발하였다. 그 결과 RF_MLP에서 NSE 0.857, RMSE 68.417 m3/s, MAE 18.063 m3/s, R 0.927, R2 0.859, GB_MLP의 경우 NSE 0.829, RMSE 73.918 m3/s, MAE 18.093 m3/s, R 0.912, R2 0.831로 CombML이 댐 유입을 가장 정확하게 예측하는 것으로 평가되었다. 본 연구를 통해 하천 유황을 고려한 여러 머신러닝 알고리즘의 결합을 통한 유입량 예측 결과, 알고리즘 결합 시 예측 모형의 정확도가 개선되는 것이 확인되었으며, 이는 추후 효율적인 물관리에 이용될 수 있을 것으로 판단된다.