• Smart Structures and Systems
• /
• v.15 no.4
• /
• pp.1139-1158
• /
• 2015

This paper proposes an efficient system identification method for modeling nonlinear behavior of civil structures. This method is developed by integrating three different methodologies: principal component analysis (PCA), artificial neural networks, and fuzzy logic theory, hence named PANFIS (PCA-based adaptive neuro-fuzzy inference system). To evaluate this model, a 3-story building equipped with a magnetorheological (MR) damper subjected to a variety of earthquakes is investigated. To train the input-output function of the PANFIS model, an artificial earthquake is generated that contains a variety of characteristics of recorded earthquakes. The trained model is also validated using the1940 El-Centro, Kobe, Northridge, and Hachinohe earthquakes. The adaptive neuro-fuzzy inference system (ANFIS) is used as a baseline. It is demonstrated from the training and validation processes that the proposed PANFIS model is effective in modeling complex behavior of the smart building. It is also shown that the proposed PANFIS produces similar performance with the benchmark ANFIS model with significant reduction of computational loads.

• Journal of Broadcast Engineering
• /
• v.25 no.4
• /
• pp.620-628
• /
• 2020

Automatically predicting video highlight is an important task for media industry and streaming platform providers to save time and cost of manual video editing process. We propose a new ensemble model that combines multiple highlight detectors with each focusing on different parts of highlight events. Therefore, our model can capture more information-rich sections of events. Furthermore, the proposed model can extract improved features for highlight detection particularly when the train video set is small. We evaluate our model on e-sports and baseball videos.

• Journal of Power Electronics
• /
• v.14 no.5
• /
• pp.1038-1046
• /
• 2014

Lithium-ion batteries are widely used in hybrid and pure electric vehicles. State-of-charge (SOC) estimation is a fundamental issue in vehicle power train control and battery management systems. This study proposes a novel model-based SOC estimation method that applies closed-loop state observer theory and a comprehensive battery model. The state-space model of lithium-ion battery is developed based on a three-order resistor-capacitor equivalent circuit model. The least square algorithm is used to identify model parameters. A multi-state closed-loop state observer is designed to predict the open-circuit voltage (OCV) of a battery based on the battery state-space model. Battery SOC can then be estimated based on the corresponding relationship between battery OCV and SOC. Finally, practical driving tests that use two types of typical driving cycle are performed to verify the proposed SOC estimation method. Test results prove that the proposed estimation method is reasonably accurate and exhibits accuracy in estimating SOC within 2% under different driving cycles.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.10
• /
• pp.3668-3684
• /
• 2021

Video action recognition is widely used in video surveillance, behavior detection, human-computer interaction, medically assisted diagnosis and motion analysis. However, video action recognition can be disturbed by many factors, such as background, illumination and so on. Two-stream convolutional neural network uses the video spatial and temporal models to train separately, and performs fusion at the output end. The multi segment Two-Stream convolutional neural network model trains temporal and spatial information from the video to extract their feature and fuse them, then determine the category of video action. Google Xception model and the transfer learning is adopted in this paper, and the Xception model which trained on ImageNet is used as the initial weight. It greatly overcomes the problem of model underfitting caused by insufficient video behavior dataset, and it can effectively reduce the influence of various factors in the video. This way also greatly improves the accuracy and reduces the training time. What's more, to make up for the shortage of dataset, the kinetics400 dataset was used for pre-training, which greatly improved the accuracy of the model. In this applied research, through continuous efforts, the expected goal is basically achieved, and according to the study and research, the design of the original dual-flow model is improved.

• Structural Engineering and Mechanics
• /
• v.70 no.6
• /
• pp.723-735
• /
• 2019

Track-bridge interaction (TBI) problem often arises from the adoption of modern continuously welded rails. Rail expansion devices (REDs) are generally required to release the intensive interaction between long-span bridges and tracks. In their necessity evaluations, the key techniques are the numerical models and methods for obtaining TBI responses. This paper thus aims to propose a preferable model and the associated procedure for TBI analysis to facilitate the designs of long-span bridges as well as the track structures. A novel friction-spring model was first developed to represent the longitudinal resistance features of fasteners with or without vertical wheel loadings, based on resistance experiments for three types of rail fasteners. This model was then utilized in the loading-history-based TBI analysis for an urban rail transit dwarf tower cable-stayed bridge installed with a RED at the middle. The finite element model of the long-span bridge for TBI analysis was established and updated by the bridge's measured natural frequencies. The additional rail stresses calculated from the TBI model under train loadings were compared with the measured ones. Overall agreements were observed between the measured and the computed results, showing that the proposed TBI model and analysis procedure can be used in further study.

• Smart Media Journal
• /
• v.10 no.1
• /
• pp.79-87
• /
• 2021

In recent years research on automatic font generation with machine learning mainly focus on using transformation-based methods, in comparison, generative model-based methods of font generation have received less attention. Transformation-based methods learn a mapping of the transformations from an existing input to a target. This makes them ambiguous because in some cases a single input reference may correspond to multiple possible outputs. In this work, we focus on font generation using the generative model-based methods which learn the buildup of the characters from noise-to-image. We propose a novel way to train a conditional generative deep neural model so that we can achieve font style control on the generated font images. Our research demonstrates how to generate new font images conditioned on both character class labels and character style labels when using the generative model-based methods. We achieve this by introducing a modified generator network which is given inputs noise, character class, and style, which help us to calculate losses separately for the character class labels and character style labels. We show that adding the character style vector on top of the character class vector separately gives the model rich information about the font and enables us to explicitly specify not only the character class but also the character style that we want the model to generate.

• Nuclear Engineering and Technology
• /
• v.54 no.8
• /
• pp.3027-3033
• /
• 2022

Background: In this study, various types of deep-learning models for predicting in vitro radiosensitivity from gene-expression profiling were compared. Methods: The clonogenic surviving fractions at 2 Gy from previous publications and microarray gene-expression data from the National Cancer Institute-60 cell lines were used to measure the radiosensitivity. Seven different prediction models including three distinct multi-layered perceptrons (MLP), four different convolutional neural networks (CNN) were compared. Folded cross-validation was applied to train and evaluate model performance. The criteria for correct prediction were absolute error < 0.02 or relative error < 10%. The models were compared in terms of prediction accuracy, training time per epoch, training fluctuations, and required calculation resources. Results: The strength of MLP-based models was their fast initial convergence and short training time per epoch. They represented significantly different prediction accuracy depending on the model configuration. The CNN-based models showed relatively high prediction accuracy, low training fluctuations, and a relatively small increase in the memory requirement as the model deepens. Conclusion: Our findings suggest that a CNN-based model with moderate depth would be appropriate when the prediction accuracy is important, and a shallow MLP-based model can be recommended when either the training resources or time are limited.

• Computers and Concrete
• /
• v.32 no.3
• /
• pp.233-246
• /
• 2023

Accurate prediction of concrete compressive strength can minimize the need for extensive, time-consuming, and costly mixture optimization testing and analysis. This study attempts to enhance the prediction accuracy of compressive strength using stacking ensemble machine learning (ML) with feature engineering techniques. Seven alternative ML models of increasing complexity were implemented and compared, including linear regression, SVM, decision tree, multiple layer perceptron, random forest, Xgboost and Adaboost. To further improve the prediction accuracy, a ML pipeline was proposed in which the feature engineering technique was implemented, and a two-layer stacked model was developed. The k-fold cross-validation approach was employed to optimize model parameters and train the stacked model. The stacked model showed superior performance in predicting concrete compressive strength with a correlation of determination (R²) of 0.985. Feature (i.e., variable) importance was determined to demonstrate how useful the synthetic features are in prediction and provide better interpretability of the data and the model. The methodology in this study promotes a more thorough assessment of alternative ML algorithms and rather than focusing on any single ML model type for concrete compressive strength prediction.

• Nuclear Engineering and Technology
• /
• v.55 no.1
• /
• pp.100-108
• /
• 2023

In this study, a neural network model inspired by a one-dimensional convolution U-net is developed to automatically accelerate edge localized mode (ELM) detection from big diagnostic data of fusion devices and increase the detection accuracy regardless of the hyperparameter setting. This model recognizes the input signal patterns and overcomes the problems of existing detection algorithms, such as the prominence algorithm and those of differential methods with high sensitivity for the threshold and signal intensity. To train the model, 10 sets of discharge radiation data from the KSTAR are used and sliced into 11091 inputs of length 12 ms, of which 20% are used for validation. According to the receiver operating characteristic curves, our model shows a positive prediction rate and a true prediction rate of approximately 90% each, which is comparable to the best detection performance afforded by other algorithms using their optimized hyperparameters. The accurate and automatic ELM-burst detection methodology used in our model can be beneficial for determining plasma properties, such as the ELM frequency from big data measured in multiple experiments using machines from the KSTAR device and ITER. Additionally, it is applicable to feature detection in the time-series data of other engineering fields.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.18 no.6
• /
• pp.1171-1176
• /
• 2023

In this paper, we predicts the estimated price using the DNBP (Deep learning Network to predict Budget Price) model with bidding data obtained from the bidding websites, ElecNet and OK EMS. We use the DNBP model to predict four lottery preliminary price, calculate their arithmetic mean, and then estimate the expected budget price ratio. We evaluate the model's performance by comparing it with the actual expected budget price ratio. We train the DNBP model by removing some of the 15 input nodes. The prediction results showed the lowest RMSE of 0.75788% when the model had 6 input nodes (a, g, h, i, j, k).