• Journal of Mechanical Science and Technology
• /
• v.20 no.7
• /
• pp.917-928
• /
• 2006

In recent decades, Artificial Neural Networks (ANNs) have become the focus of considerable attention in many disciplines, including robot control, where they can be used to solve nonlinear control problems. One of these ANNs applications is that of the inverse kinematic problem, which is important in robot path planning. In this paper, a neural network is employed to analyse of inverse kinematics of PUMA 560 type robot. The neural network is designed to find exact kinematics of the robot. The neural network is a feedforward neural network (FNN). The FNN is trained with different types of learning algorithm for designing exact inverse model of the robot. The Unimation PUMA 560 is a robot with six degrees of freedom and rotational joints. Inverse neural network model of the robot is trained with different learning algorithms for finding exact model of the robot. From the simulation results, the proposed neural network has superior performance for modelling complex robot's kinematics.

• Nuclear Engineering and Technology
• /
• v.32 no.5
• /
• pp.485-494
• /
• 2000

In this work, a fuzzy neural network is used to estimate the relevant sensor signal using other sensor signals. Noise components in input signals into the fuzzy neural network are removed through the wavelet denoising technique . Principal component analysis (PCA) is used to reduce the dimension of an input space without losing a significant amount of information. A lower dimensional input space will also usually reduce the time necessary to train a fuzzy-neural network. Also, the principal component analysis makes easy the selection of the input signals into the fuzzy neural network. The fuzzy neural network parameters are optimized by two learning methods. A genetic algorithm is used to optimize the antecedent parameters of the fuzzy neural network and a least-squares algorithm is used to solve the consequent parameters. The proposed algorithm was verified through the application to the pressurizer water level and the hot-leg flowrate measurements in pressurized water reactors.

• IEMEK Journal of Embedded Systems and Applications
• /
• v.16 no.1
• /
• pp.9-15
• /
• 2021

Recently, it is increasingly necessary to run high-end neural network applications with huge computation overhead on top of resource-constrained embedded systems, such as wearable devices. While the huge computational overhead can be alleviated by distributed neural networks running on multiple separate devices, existing distributed neural network techniques suffer from a large traffic between the devices; thus are very vulnerable to communication failures. These drawbacks make the distributed neural network techniques inapplicable to wearable devices, which are connected with each other through unstable and low data rate communication medium like human body communication. Therefore, in this paper, we propose a distributed neural network partitioning technique that is resilient to communication failures. Furthermore, we show that the proposed technique also improves the inference accuracy even in case of no communication failure, thanks to the improved network partitioning. We verify through comparative experiments with a real-life neural network application that the proposed technique outperforms the existing state-of-the-art distributed neural network technique in terms of accuracy and resiliency to communication failures.

• The Transactions of the Korean Institute of Electrical Engineers
• /
• v.39 no.12
• /
• pp.1306-1316
• /
• 1990

A controller using a multilayer neural network is proposed to the dynamic control of a PUMA 560 robot arm. This controller is developed based on an error back-propagation (BP) neural network. Since the neural network can model an arbitrary nonlinear mapping, it is used as a commanded feedforward torque generator. A Proportional Derivative (PD) feedback controller is used in parallel with the feedforward neural network to train the system. The neural network was trained by the current state of the manipulator as well as the PD feedback error torque. No a priori knowledge on system dynamics is needed and this information is rather implicitly stored in the interconnection weights of the neural network. In another experiment, the neural network was trained with the current, past and future positions only without any use of velocity sensors. Form this thim window of position values, BP network implicitly filters out the velocity and acceleration components for each joint. Computer simulation demonstrates such powerful characteristics of the neurocontroller as adaptation to changing environments, robustness to sensor noise, and continuous performance improvement with self-learning.

• Proceedings of the Korean Society of Machine Tool Engineers Conference
• /
• 2002.04a
• /
• pp.463-468
• /
• 2002

In this paper, we proposed a recurrent time delayed neural network controller which compensate a output of neural network controller. Even if learn by neural network controller, it can occur an bad results from disturbance or load variations. So in order to adjust above case, we used the fuzzy compensator to get an expected results. And the weight of main neural network can be changed with the result of learning a inverse model neural network of plant, so a expected dynamic characteristics of plant can be got. As the results of simulation through the second order plant, we confirmed that the proposed recurrent time delayed neural network controller get a good response compare with a time delayed neural network controller.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.33B no.10
• /
• pp.130-138
• /
• 1996

In this paper, we propose a neural network which detect edges of different orentation and spatial frequency in arbitrary image data. We constructed the proposed neural network iwth two different types neural network. A diffusion network performs the gaussian operation efficiently by the diffusion process. And the spatial difference network has specially designed connections suitble to detect the contours of a specific oriention. Simulation results showed that the proposed neural network can extract the edges of selected orientation efficiently by applying the neural network to a test pattern and the real image.

• Magazine of the Korean Society of Agricultural Engineers
• /
• v.36 no.2
• /
• pp.88-95
• /
• 1994

This paper is to evaluate the applicability of neural networks to the estimation of evapotranspiration. Two neural networks were developed to forecast daily evapotranspiration of the rice crop with back-propagation and counter-propagation algorithm. The neural network trained by back-propagation algorithm with delta learning rule is a three-layer network with input, hidden, and output layers. The other network with counter-propagation algorithm is a four-layer network with input, normalizing, competitive, and output layers. Training neural networks was conducted using daily actual evapotranspiration of rice crop and daily climatic data such as mean temperature, sunshine hours, solar radiation, relative humidity, and pan evaporation. During the training, neural network parameters were calibrated. The trained networks were applied to a set of field data not used in the training. The created response of the back-propagation network was in good agreement with desired values and showed better performances than the counter-propagation network did. Evaluating the neural network performance indicates that the back-propagation neural network may be applied to the estimation of evapotranspiration of the rice crop. This study does not provide with a conclusive statement as to the ability of a neural network to evapotranspiration estimating. More detailed study is required for better understanding and evaluating the behavior of neural networks.

• 제어로봇시스템학회:학술대회논문집
• /
• 1997.10a
• /
• pp.1557-1560
• /
• 1997

In this paper, we have built an experimental magnetic levitation system for a possible use of control education. We have give a mathermatical model of the nonlinear system and have shown the stability region of the linearized system when it is controlled by a PD controller. We also proposed a neural network control system which uses a neural network as a feedforward controller thgether with a conventional feedback PF controller. We have generated a desired output trajectory, which was designed for the benefit of the generalization of the neural network controller, and trained the desired output trajectory, which was desigend for the benefit of the generalization of the neural netowrk controller, and trained a neural network controller with the data of the actual input and the output of the system obtained by applying the desired output trajectroy. A good tracking performance was observed for both the desired trajectiories used and not used for the neural network training.

• Proceedings of the IEEK Conference
• /
• 2003.07d
• /
• pp.1565-1568
• /
• 2003

In this paper, we propose modular network to solve difficult and complex problems that are seldom solved with multi-layer neural network. The structure of modular neural network in researched by Jacobs and Jordan is selected in this paper. Modular network consists of several expert networks and a gating network which is composed of single-layer neural network or multi-layer neural network. We propose modular network structure using recurrent neural network, since the state of the whole network at a particular time depends on an aggregate of previous states as well as on the current input. Finally, we show excellence of the proposed network compared with modular network.

• The Transactions of The Korean Institute of Electrical Engineers
• /
• v.57 no.11
• /
• pp.2095-2101
• /
• 2008

In this paper, we consider a performance improvement of neural network for predicting defect size of steam generator tube using early stopping. Usually, neural network is trained until MSE becomes less than a prescribed error goal. The smaller the error goal, the greater the prediction performance for the trained data. However, as the error goal is decreased, an over fitting is likely to start during supervised training of a neural network, which usually deteriorates the generalization performance. We propose that, for the prediction of an axisymmetric defect size, early stopping can be used to avoid the over-fitting. Through various experiments on the axisymmetric defect samples, we found that the difference bet ween the prediction error of neural network based on early stopping and that of ideal neural network is reasonably small. This indicates that the error goal used for neural network training for the prediction of defect size can be efficiently selected by early stopping.