• The Journal of the Acoustical Society of Korea
• /
• v.30 no.8
• /
• pp.461-467
• /
• 2011

Accurate classification of input signals is the key prerequisite for variable bit-rate coding, which has been introduced in order to effectively utilize limited communication bandwidth. Especially, recent surge of multimedia services elevate the importance of speech/music classification. Among many speech/music classifier, the ones based on support vector machine (SVM) have a strong selling point, high classification accuracy, but their computational complexity and memory requirement hinder their way into actual implementations. Therefore, techniques that reduce the computational complexity and the memory requirement is inevitable, particularly for embedded systems. We first analyze implementation of an SVM-based classifier on embedded systems in terms of execution time and energy consumption, and then propose two techniques that alleviate the implementation requirements: One is a technique that removes support vectors that have insignificant contribution to the final classification, and the other is to skip processing some of input signals by virtue of strong correlations in speech/music frames. These are post-processing techniques that can work with any other optimization techniques applied during the training phase of SVM. With experiments, we validate the proposed algorithms from the perspectives of classification accuracy, execution time, and energy consumption.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.15 no.2
• /
• pp.291-298
• /
• 2020

Sign language was designed for the deaf and dumb people to allow them to communicate with others and connect to the society. However, sign language is uncommon to the rest of the society. The unresolved communication barrier had eventually isolated deaf and dumb people from the society. Hence, this study focused on design and implementation of a wearable sign language interpreter. 6 inertial measurement unit (IMU) were placed on back of hand palm and each fingertips to capture hand and finger movements and orientations. Total of 28 proposed word-based American Sign Language were collected during the experiment, while 156 features were extracted from the collected data for classification. With the used of the long short-term memory (LSTM) algorithm, this system achieved up to 99.89% of accuracy. The high accuracy system performance indicated that this proposed system has a great potential to serve the deaf and dumb communities and resolve the communication gap.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.5
• /
• pp.39-46
• /
• 2021

AI is deeply applied to various algorithms that assists us, not only daily technologies like translator and Face ID, but also contributing to innumerable fields in industry, due to its dominance. In this research, we provide convenience through AI categorization, extracting the only data that users need, with objective classification, rather than verifying all data to find from the internet, where exists an immense number of contents. In this research, we propose a model using LSTM(Long-Short Term Memory Network), which stands out from text classification, and compare its performance with models of RNN(Recurrent Neural Network) and BiLSTM(Bidirectional LSTM), which is suitable structure for natural language processing. The performance of the three models is compared using measurements of accuracy, precision, and recall. As a result, the LSTM model appears to have the best performance. Therefore, in this research, text classification using LSTM is recommended.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.8
• /
• pp.1-10
• /
• 2023

Recently, NAND flash memories have replaced magnetic disks due to non-volatility, high capacity and high resistance, in various computer systems but it has disadvantages which are the limited lifespan and imbalanced operation latency. Therefore, many page replacement policies have been studied to overcome the disadvantages of NAND flash memories. Although it is clear that these policies reflect execution characteristics of various environments and applications, researches on the foundation-policy decision for disk buffer management are insufficient. Thus, in this paper, we propose a foundation-policy recommendation model, called FRM for effectively utilizing NAND flash memories. FRM proposes a suitable page replacement policy by classifying and analyzing characteristics of workloads through machine learning. As an implementation case, we introduce FRM with a disk buffer management policy and in experiment results, prediction accuracy and weighted average of FRM shows 92.85% and 88.97%, by training dataset and validation dataset for foundation disk buffer management policy, respectively.

• Journal of Internet Computing and Services
• /
• v.24 no.2
• /
• pp.11-18
• /
• 2023

Recently, in the field of Speech Emotion Recognition (SER), many studies have been conducted to improve accuracy using voice features and modeling. In addition to modeling studies to improve the accuracy of existing voice emotion recognition, various studies using voice features are being conducted. This paper, voice files are separated by time interval in a time series method, focusing on the fact that voice emotions are related to time flow. After voice file separation, we propose a model for classifying emotions of speech data by extracting speech features Mel, Chroma, zero-crossing rate (ZCR), root mean square (RMS), and mel-frequency cepstrum coefficients (MFCC) and applying them to a recurrent neural network model used for sequential data processing. As proposed method, voice features were extracted from all files using 'librosa' library and applied to neural network models. The experimental method compared and analyzed the performance of models of recurrent neural network (RNN), long short-term memory (LSTM) and gated recurrent unit (GRU) using the Interactive emotional dyadic motion capture Interactive Emotional Dyadic Motion Capture (IEMOCAP) english dataset.

• Journal of Korean Society of Industrial and Systems Engineering
• /
• v.46 no.1
• /
• pp.105-111
• /
• 2023

Recently, research on prediction algorithms using deep learning has been actively conducted. In addition, algorithmic trading (auto-trading) based on predictive power of artificial intelligence is also becoming one of the main investment methods in stock trading field, building its own history. Since the possibility of human error is blocked at source and traded mechanically according to the conditions, it is likely to be more profitable than humans in the long run. In particular, for the virtual currency market at least for now, unlike stocks, it is not possible to evaluate the intrinsic value of each cryptocurrencies. So it is far effective to approach them with technical analysis and cryptocurrency market might be the field that the performance of algorithmic trading can be maximized. Currently, the most commonly used artificial intelligence method for financial time series data analysis and forecasting is Long short-term memory(LSTM). However, even t4he LSTM also has deficiencies which constrain its widespread use. Therefore, many improvements are needed in the design of forecasting and investment algorithms in order to increase its utilization in actual investment situations. Meanwhile, Prophet, an artificial intelligence algorithm developed by Facebook (META) in 2017, is used to predict stock and cryptocurrency prices with high prediction accuracy. In particular, it is evaluated that Prophet predicts the price of virtual currencies better than that of stocks. In this study, we aim to show Prophet's virtual currency price prediction accuracy is higher than existing deep learning-based time series prediction method. In addition, we execute mock investment with Prophet predicted value. Evaluating the final value at the end of the investment, most of tested coins exceeded the initial investment recording a positive profit. In future research, we continue to test other coins to determine whether there is a significant difference in the predictive power by coin and therefore can establish investment strategies.

• Computers and Concrete
• /
• v.30 no.5
• /
• pp.301-310
• /
• 2022

Cement-based sensors have been widely used as structural health monitoring systems, however, their long-term sensing performance have not actively investigated. In this study, a deep learning-based methodology is adopted to predict the long-term piezoresistive properties of cement-based sensors. Samples with different multi-walled carbon nanotube contents (0.1, 0.3, and 0.5 wt.%) are fabricated, and piezoresistive tests are conducted over 10,000 loading cycles to obtain the training data. Time-dependent degradation is predicted using a modified long short-term memory (LSTM) model. The effects of different model variables including the amount of training data, number of epochs, and dropout ratio on the accuracy of predictions are analyzed. Finally, the effectiveness of the proposed approach is evaluated by comparing the predictions for long-term piezoresistive sensing performance with untrained experimental data. A sensitivity of 6% is experimentally examined in the sample containing 0.1 wt.% of MWCNTs, and predictions with accuracy up to 98% are found using the proposed LSTM model. Based on the experimental results, the proposed model is expected to be applied in the structural health monitoring systems to predict their long-term piezoresistice sensing performances during their service life.

• Korean Journal of Remote Sensing
• /
• v.40 no.1
• /
• pp.57-69
• /
• 2024

Crop classification plays a vitalrole in monitoring agricultural landscapes and enhancing food production. In this study, we explore the effectiveness of Long Short-Term Memory (LSTM) models for crop classification, focusing on distinguishing between apple and rice crops. The aim wasto overcome the challenges associatedwith finding phenology-based classification thresholds by utilizing LSTM to capture the entire Normalized Difference Vegetation Index (NDVI)trend. Our methodology involvestraining the LSTM model using a reference site and applying it to three separate three test sites. Firstly, we generated 25 NDVI imagesfrom the Sentinel-2A data. Aftersegmenting study areas, we calculated the mean NDVI values for each segment. For the reference area, employed a training approach utilizing the NDVI trend line. This trend line served as the basis for training our crop classification model. Following the training phase, we applied the trained model to three separate test sites. The results demonstrated a high overall accuracy of 0.92 and a kappa coefficient of 0.85 for the reference site. The overall accuracies for the test sites were also favorable, ranging from 0.88 to 0.92, indicating successful classification outcomes. We also found that certain phenological metrics can be less effective in crop classification therefore limitations of relying solely on phenological map thresholds and emphasizes the challenges in detecting phenology in real-time, particularly in the early stages of crops. Our study demonstrates the potential of LSTM models in crop classification tasks, showcasing their ability to capture temporal dependencies and analyze timeseriesremote sensing data.While limitations exist in capturing specific phenological events, the integration of alternative approaches holds promise for enhancing classification accuracy. By leveraging advanced techniques and considering the specific challenges of agricultural landscapes, we can continue to refine crop classification models and support agricultural management practices.

• Convergence Security Journal
• /
• v.23 no.5
• /
• pp.65-72
• /
• 2023

This paper proposes a method for detecting malicious domains considering human habitual characteristics by building a Deep Learning model based on LSTM (Long Short-Term Memory). DGA (Domain Generation Algorithm) malicious domains exploit human habitual errors, resulting in severe security threats. The objective is to swiftly and accurately respond to changes in malicious domains and their evasion techniques through typosquatting to minimize security threats. The LSTM-based Deep Learning model automatically analyzes and categorizes generated domains as malicious or benign based on malware-specific features. As a result of evaluating the model's performance based on ROC curve and AUC accuracy, it demonstrated 99.21% superior detection accuracy. Not only can this model detect malicious domains in real-time, but it also holds potential applications across various cyber security domains. This paper proposes and explores a novel approach aimed at safeguarding users and fostering a secure cyber environment against cyber attacks.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.19 no.3
• /
• pp.151-158
• /
• 2024

Although on-device artificial intelligence (AI) has gained attention to diagnosing machine faults in real time, most previous studies did not consider the model retraining and redeployment processes that must be performed in real-world industrial environments. Our study addresses this challenge by proposing an on-device AI-based real-time machine fault diagnosis system that utilizes continual learning. Our proposed system includes a lightweight convolutional neural network (CNN) model, a continual learning algorithm, and a real-time monitoring service. First, we developed a lightweight 1D CNN model to reduce the cost of model deployment and enable real-time inference on the target edge device with limited computing resources. We then compared the performance of five continual learning algorithms with three public bearing fault datasets and selected the most effective algorithm for our system. Finally, we implemented a real-time monitoring service using an open-source data visualization framework. In the performance comparison results between continual learning algorithms, we found that the replay-based algorithms outperformed the regularization-based algorithms, and the experience replay (ER) algorithm had the best diagnostic accuracy. We further tuned the number and length of data samples used for a memory buffer of the ER algorithm to maximize its performance. We confirmed that the performance of the ER algorithm becomes higher when a longer data length is used. Consequently, the proposed system showed an accuracy of 98.7%, while only 16.5% of the previous data was stored in memory buffer. Our lightweight CNN model was also able to diagnose a fault type of one data sample within 3.76 ms on the Raspberry Pi 4B device.