• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.64 no.1
• /
• pp.128-135
• /
• 2015

In this study, we develop the very short-term precipitation forecasting model as well as classifier based on polynomial radial basis function neural networks by using AWS(Automatic Weather Station) and KLAPS(Korea Local Analysis and Prediction System) meteorological data. The polynomial-based radial basis function neural networks is designed to realize precipitation forecasting model as well as classifier. The structure of the proposed RBFNNs consists of three modules such as condition, conclusion, and inference phase. The input space of the condition phase is divided by using Fuzzy C-means(FCM) and the local area of the conclusion phase is represented as four types of polynomial functions. The coefficients of connection weights are estimated by weighted least square estimation(WLSE) for modeling as well as least square estimation(LSE) method for classifier. The final output of the inference phase is obtained through fuzzy inference method. The essential parameters of the proposed model and classifier such ad input variable, polynomial order type, the number of rules, and fuzzification coefficient are optimized by means of Particle Swarm Optimization(PSO) and Differential Evolution(DE). The performance of the proposed precipitation forecasting system is evaluated by using KLAPS meteorological data.

• Journal of IKEEE
• /
• v.2 no.1 s.2
• /
• pp.24-33
• /
• 1998

Independent Component analysis (ICA) is a new statistical method for extracting statistically independent components from their linear instantaneous mixtures which are generated by an unknown linear generative model. The recognition model is learned in unsupervised manner so that the recovered signals by the recognition model become the possibly scaled estimates of original source signals. This paper addresses the neural learning approach to ICA. As recognition models a linear feedforward network and a linear feedback network are considered. Associated learning algorithms for both networks are derived from maximum likelihood and information-theoretic approaches, using natural Riemannian gradient [1]. Theoretical results are confirmed by extensive computer simulations.

• Journal of Welding and Joining
• /
• v.15 no.6
• /
• pp.116-123
• /
• 1997

In GMA welding processes, bead contour and penetration patterns are criterion to estimate weld quality. Bead geometry is commonly defined with width, height and depth. When weaving is taken into account, selection of welding conditions is known to be difficult. Thus, empirical or trial-and-error method are usually introduced. This study examined the correlation of welding process variables including weaving parameters with bead geometry using srtificial neural networks(ANN). The main task of the Ann estimator is to realize the mapping characteristics from the sampled welding process variables to the actual bead geometry through training. After the neural network model is constructed, welding process variables for desired bead geometry is selected by inverse model. Experimental varification of the inverse model is conducted through actual welding.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.17 no.1
• /
• pp.105-110
• /
• 2022

Traffic sign recognition plays an important role in solving traffic-related problems. Traffic sign recognition and classification systems are key components for traffic safety, traffic monitoring, autonomous driving services, and autonomous vehicles. A lightweight model, applicable to portable devices, is an essential aspect of the design agenda. We suggest a lightweight convolutional neural network model with residual blocks for traffic sign recognition systems. The proposed model shows very competitive results on publicly available benchmark data.

• 제어로봇시스템학회:학술대회논문집
• /
• 2004.08a
• /
• pp.2033-2036
• /
• 2004

A novel approach to estimate the real time moving trajectory of an object is proposed in this paper. The object position is obtained from the image data of a CCD camera, while a state estimator predicts the linear and angular velocities of the moving object. To overcome the uncertainties and noises residing in the input data, a Kalman filter and neural networks are utilized. Since the Kalman filter needs to approximate a non-linear system into a linear model to estimate the states, there always exist errors as well as uncertainties again. To resolve this problem, the neural networks are adopted in this approach, which have high adaptability with the memory of the input-output relationship. Kohonen Network(Self-Organized Map) is selected to learn the motion trajectory since it is spatially oriented. The superiority of the proposed algorithm is demonstrated through the real experiments.

• Journal of the Korean Institute of Navigation
• /
• v.21 no.1
• /
• pp.109-118
• /
• 1997

In this paper, applications of multilayer neural networks to control of flexible robot beam are considered. The multilayer nerual networks can be used to approximate any continuous function to a desired degree of accuracy and the weights are updated by Gradient Method. When a flexible beam is rotated by a motor through the fixed end, transverse vibration may occur. The motor torque should be controlled insuch a way that the motor rotates by a specified angle, while simultaneously stabilizing vibration of the flexible manipulators so that is arrested as soon as possbile at the end of rotation. Accurate control of lightweight beam during the large changes in configuration common to robotic tasks requires dynamic models that describe both rigid body motions, as well as the flexural vibrations. Therefore, a linear dynamic state-space model of for a single link flexible robot beam is derived and PD controller, LQP controller, and inverse dynamical neural networks controller are composed. The effectiveness the proposed control system is confirmed by computer simulation.

• Proceedings of the Computational Structural Engineering Institute Conference
• /
• 2001.04a
• /
• pp.233-240
• /
• 2001

Artificial neural networks(ANN) have been exploited where the relationship among information is very complicated and nonlinear. It is appropriate to computerize the information and knowledge used in the preliminary design stage where it lacks of formality of representation of designers' experience and intuition. However, most designers start the preliminary design stage with very little information. Therefore, the ANN model for this stage must be designed to have input much less than output. This case usually causes big troubles such as in learning time, convergence and reliability of solutions. To address this problem, this paper proposes multi-level neural networks for progressive structural design considering that all the design information can not be obtained at a time but are growing gradually. The use of multi-level networks developed in this paper has been proved its validity by applying it to the preliminary design of cable-stayed bridges.

• Proceedings of the Korean Society of Machine Tool Engineers Conference
• /
• 1999.05a
• /
• pp.199-204
• /
• 1999

The robot inverse calibration method using a neural networks is proposed in this paper. A high-order networks has been used in this study. The Pi-Sigma networks uses linear summing units in the hidden layer and product unit in output layer. The inverse calibration model which compensates the difference of joint variables only between measuring value and analytic value about the desired pose(position orientation) of a robot is proposed. The compensated values are determined by using the weights obtained from the learning process of the neural networks previously. To prove the reasonableness, the selected compliance automatic robot arm type direct drive robot and anthropomorphic robot are simulated. It shows that the proposed calibration method can reduce the errors of the joint variables from ${\pm}$0.15$^{\circ}$to ${\pm}$0.12$^{\circ}$.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2019.05a
• /
• pp.401-403
• /
• 2019

In this paper, we propose a generative adversarial networks (GAN) based text-to-image generating method. In many natural language processing tasks, which word expressions are determined by their term frequency -inverse document frequency scores. Word2Vec is a type of neural network model that, in the case of an unlabeled corpus, produces a vector that expresses semantics for words in the corpus and an image is generated by GAN training according to the obtained vector. Thanks to the understanding of the word we can generate higher and more realistic images. Our GAN structure is based on deep convolution neural networks and pixel recurrent neural networks. Comparing the generated image with the real image, we get about 88% similarity on the Oxford-102 flowers dataset.

• Proceedings of the KIEE Conference
• /
• 2000.07d
• /
• pp.3001-3003
• /
• 2000

In this paper. we propose Fuzzy Self-Organizing Networks (FSON) using both Polynomial Neural Networks(PNN) and Fuzzy Neural Networks(FNN) for model identification of complex and nonlinear systems. The proposed FSON is generated from the mutually combined structure of both FNN and PNN. Accordingly it is possible to consider the nonlinearity characteristics of process and to get the better output performance with superb predictive ability. In order to evaluate the performance of proposed models. we use the nonlinear data sets. The results show that the proposed FSON can produce the model with higher accuracy and more robustness than previous any other method.