• Proceedings of the KIEE Conference
• /
• 1995.07b
• /
• pp.606-609
• /
• 1995

This paper presents discriminate between magnetizing inrush and internal faults of power transformer by artificial neural networks trained with preprocessing of fault discriminant. The proposed neural networks contain multi-layer perceptron using back-propagation learning algorithm with logistic sigmoid activation function. For this training and test, we used the relaying signals obtained from the EMTP simulation of model power system. It is shown that the proposed transformer protection system by neural networks never misoperated.

• Computational Structural Engineering
• /
• v.10 no.2
• /
• pp.273-281
• /
• 1997

In the structural analysis, artificial neural networks as a parallel computational model that is similar to the human brain and can self-organize complex nonlinear relationships without making assumptions is introduced. The purpose of this paper is to develop the Neural Network for Approximate Structural Analysis(NNASA) to predict the behaviour of the stub-girder system. As an initial stage, the paper presents the development of the prototype of NNASA based on the problem related to the deflection of a simple beam, and shows the verification of this model by two examples.

• Journal of Korea Society of Industrial Information Systems
• /
• v.4 no.2
• /
• pp.7-13
• /
• 1999

This paper propose a fuzzy regression method using fuzzy neural networks when a membership value is attached to each input -output pair. First, an architecture of fuzzy neural networks with fuzzy weights and fuzzy biases is shown. Next, a cost function is defined using the fuzzy output from the fuzzy neural network and the corresponding target output with a membership value. A learning algorithm is derived from the cost function. The derived learning algorithm trains the fuzzy neural network so that the level set of the fuzzy output includes the target output. Last, the proposed method is applied to the quality evaluation problem of injection molding

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 1998.10a
• /
• pp.324-329
• /
• 1998

In this paper, neural networks predictive PID (NNPPID) control system is proposed to reduce the vibration of structure. NNPPID control system is made up predictor, controller, and self-tuner to yield the optimal parameters of controller. The neural networks predictor forecasts the future outputs based on present input and output of structure. The controller is PID type whose parameters are yielded by neural networks self tuning algorithm. Computer simulations show displacements of multi-story structures applied to NNPPID system about environmental load-wind forces and earthquakes.

• Proceedings of the KIEE Conference
• /
• 1993.07a
• /
• pp.224-226
• /
• 1993

Multi-layered neural networks offer an exciting alternative for modelling complex non-liner systems. This paper investigates the identification of continuous time nonliner system using neural networks with a single hidden layer. The digital low - pass filter are introduced to avoid direct approximation of system derivatives from sampled data. Using a pre-designed digital low pass filter, an approximated discrete-time estimation model is constructed easily. A continuous approximation liner model is first estimated from sampled input-out signals. Then the modeling error due to the nonlinearity is decreased by a compensator using neural network. Simulation results are given to demonstrate the effective of the proposed method.

• Proceedings of the KIEE Conference
• /
• 1994.11a
• /
• pp.349-352
• /
• 1994

An object recognition and restoration using ultrasound sensors and neural networks are presented. The planar arrangement of the sensor is used to reduce the interference effects between sensors. The SOFM(Self-Organizing Feature Map) Neural Network and SCL(Simple Competitive Learning) method are learned with the acquired data. Lab experiments were performed that the object can be recognized ed the resolutions of the object can be enhanced by using the small number of the ultrasound array and neural networks.

• Proceedings of the KIEE Conference
• /
• 2003.11c
• /
• pp.657-660
• /
• 2003

Control of multi-link robot arms is a very difficult problem because of the highly nonlinear dynamics. Decentralized control scheme is developed for control of robot manipulators based on RBF(Radial Basis Function) Neural Networks. RBF Neural Networks is used to approximate the coupling forces among the joints, coriolis force, centrifugal force, gravitational force, and frictional force. The compensation controller is also proposed to estimate the bound of approximation error so that the chattering effect of the control effort can be reduced. The proposed scheme does not require an accurate manipulator dynamic, and it is proved that closed-loop system is asymptotic stable despite the gross robot parameter variations. Numerical simulations for two-link robot manipulator are included to show the effectiveness of controller.

• Transactions of the Korean Society of Mechanical Engineers A
• /
• v.20 no.4
• /
• pp.2752-2759
• /
• 1996

This paper presents crack growth analysis approach on the basis of neural networks, a branch of cognitive science to high temperature low cycle fatigue that shows strong nonlinearity in material behavior. As the number of data patterns on crack growth increase, pattern classification occurs well and two point representation scheme with gradient of crack growth curve simulates crack growth rate better than one point representation scheme. Optimal number of learning data exists and excessive number of learning data increases estimated mean error with remarkable learning time J-da/dt relation predicted by neural networks shows that test condition with unlearned data is simulated well within estimated mean error(5%).

• Journal of Sensor Science and Technology
• /
• v.30 no.5
• /
• pp.290-294
• /
• 2021

Experts have designed popular and successful model architectures, which, however, were not the optimal option for different scenarios. Despite the remarkable performances achieved by deep neural networks, manually designed networks for classification tasks are the backbone of object detection. One major challenge is the ImageNet pre-training of the search space representation; moreover, the searched network incurs huge computational cost. Therefore, to overcome the obstacle of the pre-training process, we introduce a network adaptation technique using a pre-trained backbone model tested on ImageNet. The adaptation method can efficiently adapt the manually designed network on ImageNet to the new object-detection task. Neural architecture search (NAS) is adopted to adapt the architecture of the network. The adaptation is conducted on the MobileNetV2 network. The proposed NAS is tested using SSDLite detector. The results demonstrate increased performance compared to existing network architecture in terms of search cost, total number of adder arithmetics (Madds), and mean Average Precision(mAP). The total computational cost of the proposed NAS is much less than that of the State Of The Art (SOTA) NAS method.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2019.06a
• /
• pp.98-101
• /
• 2019

At present, deep convolutional neural networks have made a very important contribution in single-image super-resolution. Through the learning of the neural networks, the features of input images are transformed and combined to establish a nonlinear mapping of low-resolution images to high-resolution images. Some previous methods are difficult to train and take up a lot of memory. In this paper, we proposed a simple and compact deep recursive residual network learning the features for single image super resolution. Global residual learning and local residual learning are used to reduce the problems of training deep neural networks. And the recursive structure controls the number of parameters to save memory. Experimental results show that the proposed method improved image qualities that occur in previous methods.