• Journal of Intelligence and Information Systems
• /
• v.25 no.1
• /
• pp.163-177
• /
• 2019

As smartphones are getting widely used, human activity recognition (HAR) tasks for recognizing personal activities of smartphone users with multimodal data have been actively studied recently. The research area is expanding from the recognition of the simple body movement of an individual user to the recognition of low-level behavior and high-level behavior. However, HAR tasks for recognizing interaction behavior with other people, such as whether the user is accompanying or communicating with someone else, have gotten less attention so far. And previous research for recognizing interaction behavior has usually depended on audio, Bluetooth, and Wi-Fi sensors, which are vulnerable to privacy issues and require much time to collect enough data. Whereas physical sensors including accelerometer, magnetic field and gyroscope sensors are less vulnerable to privacy issues and can collect a large amount of data within a short time. In this paper, a method for detecting accompanying status based on deep learning model by only using multimodal physical sensor data, such as an accelerometer, magnetic field and gyroscope, was proposed. The accompanying status was defined as a redefinition of a part of the user interaction behavior, including whether the user is accompanying with an acquaintance at a close distance and the user is actively communicating with the acquaintance. A framework based on convolutional neural networks (CNN) and long short-term memory (LSTM) recurrent networks for classifying accompanying and conversation was proposed. First, a data preprocessing method which consists of time synchronization of multimodal data from different physical sensors, data normalization and sequence data generation was introduced. We applied the nearest interpolation to synchronize the time of collected data from different sensors. Normalization was performed for each x, y, z axis value of the sensor data, and the sequence data was generated according to the sliding window method. Then, the sequence data became the input for CNN, where feature maps representing local dependencies of the original sequence are extracted. The CNN consisted of 3 convolutional layers and did not have a pooling layer to maintain the temporal information of the sequence data. Next, LSTM recurrent networks received the feature maps, learned long-term dependencies from them and extracted features. The LSTM recurrent networks consisted of two layers, each with 128 cells. Finally, the extracted features were used for classification by softmax classifier. The loss function of the model was cross entropy function and the weights of the model were randomly initialized on a normal distribution with an average of 0 and a standard deviation of 0.1. The model was trained using adaptive moment estimation (ADAM) optimization algorithm and the mini batch size was set to 128. We applied dropout to input values of the LSTM recurrent networks to prevent overfitting. The initial learning rate was set to 0.001, and it decreased exponentially by 0.99 at the end of each epoch training. An Android smartphone application was developed and released to collect data. We collected smartphone data for a total of 18 subjects. Using the data, the model classified accompanying and conversation by 98.74% and 98.83% accuracy each. Both the F1 score and accuracy of the model were higher than the F1 score and accuracy of the majority vote classifier, support vector machine, and deep recurrent neural network. In the future research, we will focus on more rigorous multimodal sensor data synchronization methods that minimize the time stamp differences. In addition, we will further study transfer learning method that enables transfer of trained models tailored to the training data to the evaluation data that follows a different distribution. It is expected that a model capable of exhibiting robust recognition performance against changes in data that is not considered in the model learning stage will be obtained.

• Journal of the Korean Recycled Construction Resources Institute
• /
• v.9 no.3
• /
• pp.303-310
• /
• 2021

Cracks in bridges are important factors that indicate the condition of bridges and should be monitored periodically. However, a visual inspection conducted by a human expert has problems in cost, time, and reliability. Therefore, in recent years, researches to apply a deep learning model are started to be conducted. Deep learning requires sufficient data on the situations to be predicted, but bridge crack data is relatively difficult to obtain. In particular, it is difficult to collect a large amount of crack data in a specific situation because the shape of bridge cracks may vary depending on the bridge's design, location, and construction method. This study developed a crack detection model that generates and trains insufficient crack data through a Generative Adversarial Network. GAN successfully generated data statistically similar to the given crack data, and accordingly, crack detection was possible with about 3% higher accuracy when using the generated image than when the generated image was not used. This approach is expected to effectively improve the performance of the detection model as it is applied when crack detection on bridges is required, though there is not enough data, also when there is relatively little or much data f or one class.

• The Journal of the Korea Contents Association
• /
• v.22 no.3
• /
• pp.81-93
• /
• 2022

The stock investing is one of the most popular investment techniques. However, since it is not easy to obtain a return through actual investment, various strategies have been devised and tried in the past to obtain an effective and stable return. Among them, the volatility breakout strategy identifies a strong uptrend that exceeds a certain level on a daily basis as a breakout signal, follows the uptrend, and quickly earns daily returns. It is one of the popular investment strategies that are widely used to realize profits. However, it is difficult to predict stock prices by understanding the price trend pattern of stocks. In this paper, we propose a method of buying and selling stocks by predicting the return in trading based on the volatility breakout strategy using a bi-directional long short-term memory deep neural network that can realize a return in a short period of time. As a result of the experiment assuming actual trading on the test data with the learned model, it can be seen that the results outperform both the return and stability compared to the existing closing price prediction model using the long-short-term memory deep neural network model.

• The Journal of the Convergence on Culture Technology
• /
• v.6 no.2
• /
• pp.449-454
• /
• 2020

This paper proposes a machine learning-based emotion analysis system that detects a user's depression through their SNS posts. We first made a list of keywords related to depression in Korean, then used these to create a training data by crawling Twitter data - 1,297 positive and 1,032 negative tweets in total. Lastly, to identify the best machine learning model for text-based depression detection purposes, we compared RNN, LSTM, and GRU in terms of performance. Our experiment results verified that the GRU model had the accuracy of 92.2%, which is 2~4% higher than other models. We expect that the finding of this paper can be used to prevent depression by analyzing the users' SNS posts.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.11
• /
• pp.491-500
• /
• 2021

In this study, we propose a method for detecting fall behavior using MS Kinect v2 RGBD Camera-based Human-Skeleton Keypoints and a 2-Stacked Bi-LSTM model. In previous studies, skeletal information was extracted from RGB images using a deep learning model such as OpenPose, and then recognition was performed using a recurrent neural network model such as LSTM and GRU. The proposed method receives skeletal information directly from the camera, extracts 2 time-series features of acceleration and distance, and then recognizes the fall behavior using the 2-Stacked Bi-LSTM model. The central joint was obtained for the major skeletons such as the shoulder, spine, and pelvis, and the movement acceleration and distance from the floor were proposed as features of the central joint. The extracted features were compared with models such as Stacked LSTM and Bi-LSTM, and improved detection performance compared to existing studies such as GRU and LSTM was demonstrated through experiments.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.05a
• /
• pp.397-400
• /
• 2022

본 연구에서는 소셜 네트워크 서비스(Social Network Service, SNS) 데이터로부터 약물 부작용 게시글을 추출하기 위한 순환 신경망(Recurrent Neural Network, RNN) 기반 분류 모델을 제안한다. 먼저, 처방 빈도가 높으며 게시글을 많이 확보할 수 있는 케토프로펜 약물에 대하여 국내 최대 소셜 네트워크 플랫폼인 네이버 블로그와 카페의 게시글(2005 년~2020 년)을 확보하고 최종 3,828 건을 분석하였다. 결과적으로 케토프로펜에 대한 3 종(약물, 부작용, 불용어)의 렉시콘을 정의하였으며 이를 기반으로 Bi-LSTM 분류모델 기준 87%의 정확도를 얻었다. 본 연구에서 제안하는 모델은 SNS 데이터가 약물 부작용 정보 획득을 위한 기존 (전자의무기록, 자발적 약물 부작용 보고 시스템 등) 자료원에 대한 보완적 정보원이 되며, 개발된 Bi-LSTM 분류모델을 통해 약물 부작용 게시글 추출의 편리성을 제공할 것으로 기대된다.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.2
• /
• pp.139-147
• /
• 2021

This paper introduces and implements a Sound Event Detection (SED) system based on weakly-supervised learning where only part of the data is labeled, and analyzes the effect of parameters. The SED system estimates the classes and onset/offset times of events in the acoustic signal. In order to train the model, all information on the event class and onset/offset times must be provided. Unfortunately, the onset/offset times are hard to be labeled exactly. Therefore, in the weakly-supervised task, the SED model is trained by "strongly labeled data" including the event class and activations, "weakly labeled data" including the event class, and "unlabeled data" without any label. Recently, the SED systems using the mean-teacher model are widely used for the task with several parameters. These parameters should be chosen carefully because they may affect the performance. In this paper, performance analysis was performed on parameters, such as the feature, moving average parameter, weight of the consistency cost function, ramp-up length, and maximum learning rate, using the data of DCASE 2020 Task 4. Effects and the optimal values of the parameters were discussed.

• Journal of Korea Water Resources Association
• /
• v.54 no.spc1
• /
• pp.1143-1154
• /
• 2021

The selection of activation function has a great influence on the groundwater level prediction performance of artificial neural network (ANN) model. In this study, five activation functions were applied to ANN model for two groundwater level observation wells in the middle mountainous area of the Pyoseon watershed in Jeju Island. The results of the prediction of the groundwater level were compared and analyzed, and the optimal activation function was derived. In addition, the results of LSTM model, which is a widely used recurrent neural network model, were compared and analyzed with the results of the ANN models with each activation function. As a result, ELU and Leaky ReLU functions were derived as the optimal activation functions for the prediction of the groundwater level for observation well with relatively large fluctuations in groundwater level and for observation well with relatively small fluctuations, respectively. On the other hand, sigmoid function had the lowest predictive performance among the five activation functions for training period, and produced inappropriate results in peak and lowest groundwater level prediction. The ANN-ELU and ANN-Leaky ReLU models showed groundwater level prediction performance comparable to that of the LSTM model, and thus had sufficient potential for application. The methods and results of this study can be usefully used in other studies.

• Information Systems Review
• /
• v.18 no.2
• /
• pp.173-198
• /
• 2016

Financial distress can damage stakeholders and even lead to significant social costs. Thus, financial distress prediction is an important issue in macroeconomics. However, most existing studies on building a financial distress prediction model have only considered idiosyncratic risk factors without considering systematic risk factors. In this study, we propose a prediction model that considers both the idiosyncratic risk based on a financial ratio and the systematic risk based on a business cycle. Ultimately, we build several IT artifacts associated with financial ratio and add them to the idiosyncratic risk factors as well as address the imbalanced data problem by using an oversampling technique and synthetic minority oversampling technique (SMOTE) to ensure good performance. When considering systematic risk, our study ensures that each data set consists of both financially distressed companies and financially sound companies in each business cycle phase. We conducted several experiments that change the initial imbalanced sample ratio between the two company groups into a 1:1 sample ratio using SMOTE and compared the prediction results from the individual data set. We also predicted data sets from the subsequent business cycle phase as a test set through a built prediction model that used business contraction phase data sets, and then we compared previous prediction performance and subsequent prediction performance. Thus, our findings can provide insights into making rational decisions for stakeholders that are experiencing an economic crisis.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.21 no.5
• /
• pp.171-182
• /
• 2022

The safe use of a structure requires it to be maintained in an undamaged state. Thus, a typical factor that determines the safety of a structure is a crack in it. In addition, cracks are caused by various reasons, damage the structure in various ways, and exist in different shapes. Making matters worse, if these cracks are unattended, the risk of structural failure increases and proceeds to a catastrophe. Hence, recently, methods of checking structural damage using deep learning and computer vision technology have been introduced. These methods usually have the premise that there should be a large amount of training image data. However, the amount of training image data is always insufficient. Particularly, this insufficiency negatively affects the performance of deep learning crack detection algorithms. Hence, in this study, a method of augmenting crack image data based on the image translation technique was developed. In particular, this method obtained the crack image data for training a deep learning neural network model by transforming a specific case of a asphalt crack image into a concrete crack image or vice versa . Eventually, this method expected that a robust crack detection algorithm could be developed by increasing the diversity of its training data.