• 산업기술연구
• /
• 제18권
• /
• pp.303-308
• /
• 1998

In this paper a neural network controller called "Feedback-State Learning" for control of the attitude of a wheeled inverted pendulum is presented. For the controller the design of a stable feedback controller is necessary, so the LQR is used for the feedback controller because the LQR has good performance on controlling nonlinear systems. And the neural networks are used for a feed forward controller. The designed controller is applied to the stabilization of a wheeled inverted pendulum. Because of its nonlinear characteristics such as friction and parameter variations in the linearization, the wheeled inverted pendulum is used for demonstration of the effectiveness of the proposed controller.

• 제어로봇시스템학회:학술대회논문집
• /
• 제어로봇시스템학회 1998년도 제13차 학술회의논문집
• /
• pp.195-200
• /
• 1998

Since the neural network was introduced, significant progress has been made on data handling and learning algorithms. Currently, the most popular learning algorithm in neural network training is feed forward error back-propagation (FFEBP) algorithm. Aside from the success of the FFEBP algorithm, polynomial neural networks (PNN) learning has been proposed as a new learning method. The PNN learning is a self-organizing process designed to determine an appropriate set of Ivakhnenko polynomials that allow the activation of many neurons to achieve a desired state of activation that mimics a given set of sampled patterns. These neurons are interconnected in such a way that the knowledge is stored in Ivakhnenko coefficients. In this paper, the PNN model has been developed using the plasma enhanced chemical vapor deposition (PECVD) experimental data. To characterize the PECVD process using PNN, SiO$_2$films deposited under varying conditions were analyzed using fractional factorial experimental design with three center points. Parameters varied in these experiments included substrate temperature, pressure, RF power, silane flow rate and nitrous oxide flow rate. Approximately five microns of SiO$_2$were deposited on (100) silicon wafers in a Plasma-Therm 700 series PECVD system at 13.56 MHz.

• 한국정보통신학회:학술대회논문집
• /
• 한국해양정보통신학회 1999년도 추계종합학술대회
• /
• pp.224-229
• /
• 1999

음성 인식 연구는 유사음소 단위의 인식시스템을 구축하는 방향과 단어 단위의 인식시스템에서의 효율을 최대화하는 방향으로 이루어지고 있다. 이중 유용한 유사음소 단위의 인식시스템 구현을 위해서는 음소의 경계 검출 문제와 검출된 음소에 대한 인식률 향상 문제가 해결되어야 한다. 기존의 LPC(Linear Predictive Coefficient) 방법들은 기준 음소데이터의 LPC와 입력 음성프레임의 LPC 사이의 거리를 Itakura-Saito 방법으로 구하여 음소의 경계를 검출하였으며, 근래에는 MFCC(Mel-Frequency-Cepstrum Coefficient)를 이용하여 스펙트럼의 천이부분을 음소의 경계로 검출하는 방법들이 제안되어왔으나 이러한 방법들은 공통적으로 적응성이 미비하다는 단점이 있다. 본 논문에서는 이러한 단점을 극복하기 위해 음소경계검출을 위해서는 auto-correlation을 이용하고 음소인식을 위해서는 적응성이 뛰어난 다층 Feed-Forward 신경망을 사용하는 새로운 인식시스템을 제안하였다 제안하는 시스템은 기존의 방법들보다 적응성이 뛰어나고 특징추출부분과 인식 부분의 알고리듬이 독립적이라는 장점을 가지며 프레임단위의 음소인식시스템의 구현 가능성을 확인해 주었다.

• 전자공학회논문지CI
• /
• 제41권3호
• /
• pp.55-65
• /
• 2004

신경회로망을 이용하여 주어진 문제를 해결할 때, 문제의 복잡도에 맞는 구조를 찾는 것이 중요하다. 이것은 신경회로망의 복잡도가 학습능력과 일반화 성능에 크게 영향을 주기 때문이다. 그러므로, 문제에 적합한 신경회로망의 구조를 자기 구성적으로 찾는 알고리즘이 유용하다. 본 논문에서는 시그모이드 활성함수를 가지는 전방향 다층 신경회로망에 대하여 주어진 문제에 맞는 구조를 결정하는 알고리즘을 제안한다. 개발된 알고리즘은 구조증가 알고리즘과 연결소거 알고리즘을 이용하여, 주어진 학습 데이터에 대해 가능한 한 작은 구조를 가지며 일반화 성능이 좋은 최적에 가까운 신경회로망을 찾는다. 네 가지 함수 근사화 문제에 적용하여 알고리즘의 성능을 알아본다. 실험 결과에서, 제안한 알고리즘이 기존의 알고리즘 및 고정구조를 갖는 신경회로망과 비교하였을 때 최적 구조에 가까운 신경회로망을 구성하는 것을 확인한다.

• 정보과학회 컴퓨팅의 실제 논문지
• /
• 제21권4호
• /
• pp.338-342
• /
• 2015

시계열 데이터를 다룰 수 있는 기계학습모델인 회귀 신경망은 되먹임 연결을 허용하기 때문에 앞먹임 신경망에 비해 훨씬 다양한 구조를 가질 수 있다. 본 연구에서는 은닉 뉴런 간의 네트워크 구조에 초점을 맞추어 그것이 회귀 신경망의 정보처리 능력에 미치는 영향을 탐구하고자 한다. 이를 위해 회귀신경망 모델 중 하나인 Echo State Network을 기준으로 하여, 여러 가지 잘 알려진 네트워크 모델에 따라 은닉 뉴런 간 연결을 구성하고 각각의 경우에 시계열 학습 능력과 동역학을 분석하였다. 그 결과, 은닉 뉴런의 네트워크 구조에 따라 모델의 성능이 큰 폭으로 변하는 것이 관찰되었으며, 그러한 현상은 신경망 동역학이 가지는 임계도(criticality)의 변화와 잘 일치했다. 본 연구의 결과는 기존 회귀 신경망 연구에서 주된 관심사였던 신경망 연결 가중치뿐만 아니라 신경망의 연결 구조가 모델의 성능에 중요한 영향을 미친다는 사실을 보여주며, 성능 향상을 위한 중요한 단서가 될 수 있다.

• 한국재료학회지
• /
• 제24권7호
• /
• pp.381-387
• /
• 2014

The ${\beta}$-transus temperature in titanium alloys plays an important role in the design of thermo-mechanical treatments. It primarily depends on the chemical composition of the alloy and the relationship between them is non-linear and complex. Considering these relationships is difficult using mathematical equations. A feed-forward neural-network model with a back-propagation algorithm was developed to simulate the relationship between the ${\beta}$-transus temperature of titanium alloys, and the alloying elements. The input parameters to the model consisted of the nine alloying elements (i.e., Al, Cr, Fe, Mo, Sn, Si, V, Zr, and O), whereas the model output is the ${\beta}$-transus temperature. The model developed was then used to predict the ${\beta}$-transus temperature for different elemental combinations. Sensitivity analysis was performed on a trained neural-network model to study the effect of alloying elements on the ${\beta}$-transus temperature, keeping other elements constant. Very good performance of the model was achieved with previously unseen experimental data. Some explanation of the predicted results from the metallurgical point of view is given. The graphical-user-interface developed for the model should be very useful to researchers and in industry for designing the thermo-mechanical treatment of titanium alloys.

• Structural Engineering and Mechanics
• /
• 제63권4호
• /
• pp.429-438
• /
• 2017

Structural design has an imperative role in deciding the failure possibility of a Reinforced Concrete (RC) structure. Recent research works achieved the goal of predicting the structural failure of the RC structure with the assistance of machine learning techniques. Previously, the Artificial Neural Network (ANN) has been trained supported by Particle Swarm Optimization (PSO) to classify RC structures with reasonable accuracy. Though, keeping in mind the sensitivity in predicting the structural failure, more accurate models are still absent in the context of Machine Learning. Since the efficiency of multi-objective optimization over single objective optimization techniques is well established. Thus, the motivation of the current work is to employ a Multi-objective Genetic Algorithm (MOGA) to train the Neural Network (NN) based model. In the present work, the NN has been trained with MOGA to minimize the Root Mean Squared Error (RMSE) and Maximum Error (ME) toward optimizing the weight vector of the NN. The model has been tested by using a dataset consisting of 150 RC structure buildings. The proposed NN-MOGA based model has been compared with Multi-layer perceptron-feed-forward network (MLP-FFN) and NN-PSO based models in terms of several performance metrics. Experimental results suggested that the NN-MOGA has outperformed other existing well known classifiers with a reasonable improvement over them. Meanwhile, the proposed NN-MOGA achieved the superior accuracy of 93.33% and F-measure of 94.44%, which is superior to the other classifiers in the present study.

• Journal of Electrical Engineering and Technology
• /
• 제4권3호
• /
• pp.389-395
• /
• 2009

Industrial motors are subject to incipient faults which, if undetected, can lead to motor failure. The necessity of incipient fault detection can be justified by safety and economical reasons. The technology of artificial neural networks has been successfully used to solve the motor incipient fault detection problem. This paper develops inexpensive, reliable, and noninvasive NN based incipient fault detection scheme for small and medium sized induction motors. Detailed design procedure for achieving the optimal NN model and Principal Component Analysis for dimensionality reduction is proposed. Overall thirteen statistical parameters are used as feature space to achieve the desired classification. GFFD NN model is designed and verified for optimal performance in fault identification on experimental data set of custom designed 2 HP, three phase 50 Hz induction motor.

• Journal of Power Electronics
• /
• 제8권4호
• /
• 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.

• 대한의용생체공학회:학술대회논문집
• /
• 대한의용생체공학회 1997년도 춘계학술대회
• /
• pp.358-361
• /
• 1997

This paper describes a multichannel epileptic seizure detection algorithm based on wavelet transform(WT), artificial neural network(ANN) and expert system. First, through the WT, a small number of wavelet coefficients is used to represent the single channel epileptic spike. Next, 3-layer feed-forward network employing the error back propagation algorithm is trained and tested using parameters obtained above. Finally, 16 channel expert system which is based on clinical experience is introduced as a artifact rejection and reliable detection. The suggested algorithm was implemented on personal computer(PC). Two main events i.e., epileptiform and normal activities, were selected from 32 person's EEGs(normal: 20, seizure disorder: 12) in consensus among experts. The result was that WT reduced data input size and ANN detected 97 of the 100 EEGs containing definite spike - sensitivity of 97%. Expert rule system was capable of rejecting a wide variety of artifacts commonly found in EEG recordings. It also reduced false positive detections of ANN.