• Clean Technology
• /
• v.7 no.2
• /
• pp.119-126
• /
• 2001

The prediction of ozone formation was studied using the neural network and the stochastic method. Parameter estimation method and artificial neural network(ANN) method were employed in the stochastic scheme. In the parameter estimation method, extended least squares(ELS) method and recursive maximum likelihood(RML) were used to achieve the real time parameter estimation. Autoregressive moving average model with external input(ARMAX) was used as the ozone formation model for the parameter estimation method. ANN with 3 layers was also tested to predict the ozone formation. To demonstrate the performance of the ozone formation prediction schemes used in this work, the prediction results of ozone formation were compared with the real data. From the comparison it was found that the prediction schemes based on the parameter estimation method and ANN method show an acceptable accuracy with limited prediction horizon.

• Journal of Institute of Control, Robotics and Systems
• /
• v.10 no.2
• /
• pp.112-124
• /
• 2004

A neural network based adaptive reconfigurable flight controller is presented for a class of discrete-time nonlinear flight systems in the presence of variations of aerodynamic coefficients and control effectiveness decrease caused by control surface damage. The proposed adaptive nonlinear controller is developed making use of the backstepping technique for the angle of attack, sideslip angle, and bank angle command following without two time separation assumption. Feedforward multilayer neural networks are implemented to guarantee reconfigurability for control surface damage as well as robustness to the aerodynamic uncertainties. The main feature of the proposed controller is that the adaptive controller is developed under the assumption that all of the nonlinear functions of the discrete-time flight system are not known accurately, whereas most previous works on flight system applications even in continuous time assume that only the nonlinear functions of fast dynamics are unknown. Neural networks learn through the recursive weight update rules that are derived from the discrete-time version of Lyapunov control theory. The boundness of the error states and neural networks weight estimation errors is also investigated by the discrete-time Lyapunov derivatives analysis. To show the effectiveness of the proposed control law, the approach is i]lustrated by applying to the nonlinear dynamic model of the high performance aircraft.

• Biotechnology and Bioprocess Engineering:BBE
• /
• v.7 no.5
• /
• pp.275-280
• /
• 2002

In order to control glucose concentration during fed-batch culture for antibiotic production, we applied so called “software sensor” which estimates unmeasured variable of interest from measured process variables using software. All data for analysis were collected from industrial scale cultures in a pharmaceutical company. First, we constructed an estimation model for glucose feed rate to keep glucose concentration at target value. In actual fed-batch culture, glucose concentration was kept at relatively high and measured once a day, and the glucose feed rate until the next measurement time was determined by an expert worker based on the actual consumption rate. Fuzzy neural network (FNN) was applied to construct the estimation model. From the simulation results using this model, the average error for glucose concentration was 0.88 g/L. The FNN model was also applied for a special culture to keep glucose concentration at low level. Selecting the optimal input variables, it was possible to simulate the culture with a low glucose concentration from the data sets of relatively high glucose concentration. Next, a simulation model to estimate time course of glucose concentration during one day was constructed using the on-line measurable process variables, since glucose concentration was only measured off-line once a day. Here, the recursive fuzzy neural network (RFNN) was applied for the simulation model. As the result of the simulation, average error of RFNN model was 0.91 g/L and this model was found to be useful to supervise the fed-batch culture.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• v.40 no.4
• /
• pp.264-274
• /
• 2003

This paper presents a direct generalized minimum-variance self tuning controller with a PID structure using neural network which adapts to the changing parameters of the nonlinear system with nonminimum phase behavior, noises and time delays. The self-tuning controller with a PID structure is a combination of the simple structure of a PID controller and the characteristics of a self-tuning controller that can adapt to changes in the environment. The self-tuning control effect is achieved through the RLS (recursive least square) algorithm at the parameter estimation stage as well as through the Robbins-Monro algorithm at the stage of optimizing the design parameter of the controller. The neural network control effect which compensates for nonlinear factor is obtained from the learning algorithm which the learning error between the filtered reference and the auxiliary output of plant becomes zero. Computer simulation has shown that the proposed method works effectively on the nonlinear nonminimum phase system with time delays and changed system parameter.

• Journal of Power Electronics
• /
• v.8 no.4
• /
• pp.354-362
• /
• 2008

This paper presents neural load torque compensation method which is composed of a deadbeat load torque observer and gains compensation by a parameter estimator. As a result, the response of the PMSM (permanent magnet synchronous motor) obtains better precision position control. To reduce the noise effect, the post-filter is implemented by a MA (moving average) process. The parameter compensator with an RLSM (recursive least square method) parameter estimator is adopted to increase the performance of the load torque observer and main controller. The parameter estimator is combined with a high performance neural load torque observer to resolve problems. The neural network is trained in online phases and it is composed by a feed forward recall and error back-propagation training. During normal operation, the input-output response is sampled and the weighting value is trained multi-times by the error back-propagation method at each sample period to accommodate the possible variations in the parameters or load torque. As a result, the proposed control system has a robust and precise system against load torque and parameter variation. Stability and usefulness are verified by computer simulation and experiment.

• Proceedings of the KIPE Conference
• /
• 1998.07a
• /
• pp.384-387
• /
• 1998

This paper presents a newly developed speed sensorless drive using RLS(Recursive Least Squares) based on Neural Network Training Algorithm. The proposed algorithm based on the RLS has just the time-varying learning rate, while the well-known back-propagation (or generalized delta rule) algorithm based on gradient descent has a constant learning rate. The number of iterations required by the new algorithm to converge is less than that of the back-propagation algorithm. The RLS based on NN is used to adjust the motor speed so that the neural model output follows the desired trajectory. This mechanism forces the estimated speed to follow precisely the actual motor speed. In this paper, a flux estimation strategy using filter concept is discussed. The theoretical analysis and experimental results to verify the effectiveness of the proposed analysis and the proposed control strategy are described.

• Journal of the Korea Society of Computer and Information
• /
• v.4 no.4
• /
• pp.94-100
• /
• 1999

This paper concerns the implementation of a digital neural network which performs the high speed operation of Hopfield model's arithmetic operation. It is also designed to use a look-up table and produce floating point arithmetic of nonlinear function with high speed operation. The arithmetic processing of Hopfleld is able to describe the matrix-vector operation, which is adaptable to design the array processor because of its recursive and iterative operation .The proposed method is expected to be applied to the field of real neural networks because of the realization of the current VLSI techniques.

• The Journal of the Acoustical Society of Korea
• /
• v.33 no.4
• /
• pp.268-272
• /
• 2014

This paper presents a method for speech recognition using multi-section vector-quantization (MSVQ) and time-delay recurrent neural network (TDTNN). The MSVQ generates the codebook with normalized uniform sections of voice signal, and the TDRNN performs the speech recognition using the MSVQ codebook. The TDRNN is a time-delay recurrent neural network classifier with two different representations of dynamic context: the time-delayed input nodes represent local dynamic context, while the recursive nodes are able to represent long-term dynamic context of voice signal. The cepstral PLP coefficients were used as speech features. In the speech recognition experiments, the MSVQ/TDRNN speech recognizer shows 97.9 % word recognition rate for speaker independent recognition.

• The Transactions of the Korean Institute of Power Electronics
• /
• v.5 no.2
• /
• pp.131-139
• /
• 2000

일반적으로 시스템 인식과 제어에 이용하는 다층 신경회로망은 기존의 역전파 알고리즘을 이용한다. 그러나 결선강도에 대한 오차의 기울기를 구하는 방법이기 때문에 국부적 최소점에 빠지기 쉽고, 수렴속도가 매우 늦으며 초기 결선강도 값들이나 학습계수에 민감하게 반응한다. 이와 같은 단점을 개선하기 위하여 확장된 칼만 필터링 기법을 역전파 알고리즘에 결합하였으나 계산상의 복잡성 때문에 망의 크기가 증가하면 실제 적용할 수 없다. 최근 신경회로망을 선형과 비선형 구간으로 구분하고 칼만 필터링 기법을 도입하여 수렴속도를 빠르게 하고 초기 결선강도 값에 크게 영향을 받지 않도록 개선하였으나, 여전히 은닉층의 선형 오차값을 역전파 알고리즘에 의해서 계산하기 때문에 학습계수에 민감하다는 단점이 있다. 본 논문에서는 위에서 언급한 기존의 신경회로망 알고리즘의 문제점을 개선하기 위하여 은닉층의 목표값을 최적기법에 의하여 직접계산하고 각각의 결선강도 값은 반복최소 자승법으로 온라인 학습하는 알고리즘을 제안하고 이들 신경회로망 알고리즘과 비교하고자 한다. 여러 가지 시뮬레이션과 실험을 통하여 제안된 방법이 초기 결선강도에 크게 영향을 받지 않으며, 기존의 학습계수 선정에 따른 문제점을 해결함으로써 신경회로망 모델에 기초한 실시간 제어기 설계에 응용할 수 있도록 하였다. 또한, 유도전동기의 속도추정과 제어에 적용하여 좋은 결과를 보였다.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.21 no.4
• /
• pp.475-480
• /
• 2011

In this paper, we propose a collision risk assesment system. First, using Kalman Filter, we estimate the information of pedestrian, and second, we compute the collision probability using Monte Carlo Simulations(MCS) and neural network(NN). And we update the collision risk using time history which is called belief. Belief update consider not only output of Kalman Filter of only current time step but also output of Kalman Filter up to the first time step to current time step. The computer simulations will be shown the validity of our proposed method.