• 제어로봇시스템학회:학술대회논문집
• /
• 제어로봇시스템학회 2004년도 ICCAS
• /
• pp.1319-1323
• /
• 2004

In this paper, a problem of Thai sign language recognition using a neural network is considered. The paper addresses the problem in classifying certain signs conveying quantitative meaning, e.g., large or small. By treating those signs corresponding to different quantities as derived from different classes, the recognition error rate of the standard multi-layer Perceptron increases if the precision in recognizing different quantities is increased. This is due the fact that, to increase the quantitative recognition precision of those signs, the number of (increasingly similar) classes must also be increased. This leads to an increase in false classification. The problem is due to misinterpreting the amount of quantity the quantitative signs convey. In this paper, instead of treating those signs conveying quantitative attribute of the same quantity type (such as 'size' or 'amount') as derived from different classes, here they are considered instances of the same class. Those signs of the same quantity type are then further divided into different subclasses according to the level of quantity each sign is associated with. By using this two-level classification, false classification among main gesture classes is made independent to the level of precision needed in recognizing different quantitative levels. Moreover, precision of quantitative level classification can be made higher during the recognition phase, as compared to that used in the training phase. A standard multi-layer Perceptron with a back propagation learning algorithm was adapted in the study to implement this two-level classification of quantitative gesture signs. Experimental results obtained using an electronic glove measurement of hand postures are included.

• Journal of Power Electronics
• /
• 제15권4호
• /
• pp.1139-1147
• /
• 2015

In this paper, an algorithm based on a model analysis of the online calculation of the root-mean-square (RMS) value of welding current for single-phase AC resistance spot welding (RSW) was developed. The current is highly nonlinear and typically non-sinusoidal, which makes the measuring and controlling actions difficult. Though some previous methods focused on this issue, they were so complex that they could not be effectively used in general cases. The electrical model of a single-phase AC RSW was analyzed, and then an algorithm for online calculation of the RMS value of the welding current was presented. The description includes two parts, a model-dependent part and a model-independent part. Using a previous work about online measurement of the power factor angle, the first part can be solved. For the second part, although the solution of the governing equation can be directly obtained, a lot of CPU time must be consumed due to the fact that it involves a lot of complex calculations. Therefore, a neural network was employed to simplify the calculations. Finally, experimental results and a corresponding analysis showed that the proposed algorithm can obtain the RMS values with a high precision while consuming less time when compared to directly solving the equations.

• 대한전자공학회논문지SP
• /
• 제45권4호
• /
• pp.10-21
• /
• 2008

본 논문은 실시간 감시 시스템을 위한 웨이블릿(wavelet) 기반의 신경망과 불변 모멘트를 이용한 이동물체 인식과 추적 방법을 제안한다. 제안한 방법의 첫 번째인 움직임 후보영역 검출 단계에서는 연속된 두 프레임간의 차영상 분석 방법을 기반으로 하여 물체의 움직임에 의해 화소값 변화가 발생한 후보영역을 검출한다. 두 번째인 물체 인식 단계에서는 검출된 후보영역에 웨이블릿 신경망(wavelet neural network: WNN) 기반의 인식 방법을 사용하여 추적하고자하는 물체가 포함되어 있는지를 판별한다. 세 번째인 물체 추적 단계에서는 인식된 물체에 웨이블릿 불변 모멘트(invariant moments) 기반의 매칭 방법을 사용하여 인식된 이동 물체를 추적한다. 영상내에서 이동물체를 검출하기 위해 본 논문에서는 이전 영상과 현재 영상간의 화소밝기 차이에서 적응적 임계값(adaptive threholding)을 사용하여 주위 환경 변화에 강인한 이동물체 검출이 가능하였다. 또한 물체의 인식과 추적을 위해 웨이블릿 특징값을 사용함으로써, 계산 시간의 감소와 영상의 잡음에 의한 영향을 최소화시킬 수 있을 뿐만 아니라, 물체 인식 정확도가 향상되었다. 제안한 방법을 일반 도로에서 획득한 영상에서 실험한 결과, 자동차 검출율은 92.8%, 프레임당 처리 시간은 0.24초이다. 이것을 통해 제안한 방법은 실시간 지능형 교통 감시 시스템에 유용하게 적용될 수 있음을 알 수 있다.

• 스마트미디어저널
• /
• 제5권3호
• /
• pp.35-41
• /
• 2016

본 논문에서는 심층 학습 모델 방법을 이용하여 EPS(Electronic Potential Sensor) 기반의 손동작 신호를 인식하는 시스템을 제안한다. 전기장 기반 센서인 EPS로부터 추출된 신호는 다량의 잡음이 포함되어 있어 이를 제거하는 전처리과정을 거쳐야 한다. 주파수 대역 특징 필터를 이용한 잡음 제거한 후, 신호는 시간에 따른 전압(Voltage) 값만 가지는 1차원적 특징을 지닌다. 2차원 데이터를 입력으로 하여 컨볼루션 연산을 하는 알고리즘에 적합한 형태를 갖추기 위해 신호는 차원 변형을 통해 재구성된다. 재구성된 신호데이터는 여러 계층의 학습 층(layer)을 가지는 심층 학습 기반의 모델을 통해 분류되어 최종 인식된다. 기존 확률 기반 통계적 모델링 알고리즘은 훈련 후 모델을 생성하는 과정에서 초기 파라미터에 결과가 좌우되는 어려움이 있었다. 심층 학습 기반 모델은 학습 층을 쌓아 훈련을 반복하므로 이를 극복할 수 있다. 실험에서, 제안된 심층 학습 기반의 서로 다른 구조를 가지는 컨볼루션 신경망(Convolutional Neural Networks), DBN(Deep Belief Network) 알고리즘과 통계적 모델링 기반의 방법을 이용한 인식 결과의 성능을 비교하였고, 컨볼루션 신경망 알고리즘이 다른 알고리즘에 비해 EPS 동작신호 인식에서 보다 우수한 성능을 나타냄을 보였다.

• 한국정보통신학회:학술대회논문집
• /
• 한국해양정보통신학회 2005년도 추계종합학술대회
• /
• pp.677-680
• /
• 2005

유체의 상변화를 이용하는 냉난방장치 등의 열장치에 대한 열역학적인 성능평가는 열역학적 성질들에 대한 구체적인 수치값을 필요로 한다. 그러나 이러한 열역학적 성질들을 제공하는 증기표를 그대로는 사용할 수 없기 때문에 효과적인 모델링이 필요하다. 본 연구에서는 신경회로망의 함수근사 특성을 이용하여 냉방장치의 매질로 사용되는 냉매(R12)의 포화증기 영역을 모델링하였다. 냉매 R12의 포화증기 영역의 함수근사 해석을 위하여 1개의 노드를 가진 입력층에 대하여 7개의 노드를 가진 출력층을 기본으로 하여, 각각 10개와 20개의 노드를 가진 두 개의 은닉층을 가진 회로망을 구성하였다. 또한 입력이 온도와 압력 두 가지의 경우에 대하여 검토하였다. 제안된 신경회로망을 사용한 결과 엔탈피, 엔트로피의 백분율오차가 대부분 ${\pm}$0.005%, 비체적은 ${\pm}$0.02%, 압력과 온도는 특별한 몇 개를 제외하고는 ${\pm}$0.02% 범위 내로 수렴되었다. 이 결과로부터 냉매를 함수근사하는데 있어서 신경회로망이 아주 강력한 수단이 될 수 있음을 확인하였다.

• 전자공학회논문지B
• /
• 제28B권9호
• /
• pp.739-746
• /
• 1991

This paper develops a modified multilayer perceptron architecture which speeds up learning as well as the net's mapping accuracy. In Phase I, a cluster partitioning algorithm like the Kohonen's self-organizing feature map or the leader clustering algorithm is used as the front end that determines the cluster to which the input data belongs. In Phase II, this cluster selects a subset of the hidden layer nodes that combines the input and outputs nodes into a subnet of the full scale backpropagation network. The proposed net has been applied to two mapping problems, one rather smooth and the other highly nonlinear. Namely, the inverse kinematic problem for a 3-link robot manipulator and the 5-bit parity mapping have been chosen as examples. The results demonstrate the proposed net's superior accuracy and convergence properties over the original backpropagation network or its existing improvement techniques.

• 동력기계공학회지
• /
• 제7권2호
• /
• pp.29-35
• /
• 2003

The monitoring and diagnostics of the rotating machinery have been received considerable attention for many years. The objectives are to classify the machinery condition and to find out the cause of abnormal condition. This paper describes a classification method of diagnosing the small reciprocating compressor for refrigerators using the artificial neural network and the wavelet transform. In order to extract salient features, the wavelet transform are used from primary noise signals. Since the wavelet transform decomposes raw time-waveform signals into two respective parts in the time space and frequency domain, more and better features can be obtained easier than time-waveform analysis. In the training phase for classification, self-organizing feature map(SOFM) and learning vector quantization(LVQ) are applied, and the accuracies of them ate compared with each other. This paper is focused on the development of an advanced signal classifier to automatize the vibration signal pattern recognition. This method is verified by small reciprocating compressors, for refrigerator and normal and abnormal conditions are classified with high flexibility and reliability.

• 대한임베디드공학회논문지
• /
• 제18권6호
• /
• pp.277-283
• /
• 2023

Vital signals provide essential information regarding the health status of individuals, thereby contributing to health management and medical research. Present monitoring methods, such as ECGs (Electrocardiograms) and smartwatches, demand proximity and fixed postures, which limit their applicability. To address this, Non-contact vital signal measurement methods, such as CW (Continuous-Wave) radar, have emerged as a solution. However, unwanted signal components and a stepwise processing approach lead to errors and limitations in heart rate detection. To overcome these issues, this study introduces an integrated neural network approach that combines noise removal, demodulation, and dominant-frequency detection into a unified process. The neural network employed for signal processing in this research adopts a MLP (Multi-Layer Perceptron) architecture, which analyzes the in-phase and quadrature signals collected within a specified time window, using two distinct input layers. The training of the neural network utilizes CW radar signals and reference heart rates obtained from the ECG. In the experimental evaluation, networks trained on different datasets were compared, and their performance was assessed based on loss and frequency accuracy. The proposed methodology exhibits substantial potential for achieving precise vital signals through non-contact measurements, effectively mitigating the limitations of existing methodologies.

• 한국수자원학회:학술대회논문집
• /
• 한국수자원학회 2023년도 학술발표회
• /
• pp.37-37
• /
• 2023

There is currently a high-accuracy modern forecasting method that uses machine learning algorithms or artificial neural network models to forecast river water levels or flowrate. As a result, this study aims to develop a mathematical model based on artificial neural networks to effectively forecast river water levels upstream of Tranh Culvert in North Vietnam's Bac Hung Hai irrigation system. The mathematical model was thoroughly studied and evaluated by using hydrological data from six gauge stations over a period of twenty-two years between 2000 and 2022. Furthermore, the results of the developed model were also compared to those of the long-short-term memory neural networks model. This study performs four predictions, with a forecast time ranging from 6 to 24 hours and a time step of 6 hours. To validate and test the model's performance, the Nash-Sutcliffe efficiency coefficient (NSE), mean absolute error, and root mean squared error were calculated. During the testing phase, the NSE of the model varies from 0.981 to 0.879, corresponding to forecast cases from one to four time steps ahead. The forecast results from the model are very reasonable, indicating that the model performed excellently. Therefore, the proposed model can be used to forecast water levels in North Vietnam's irrigation system or rivers impacted by tides.

• Journal of Electrical Engineering and Technology
• /
• 제5권4호
• /
• pp.666-671
• /
• 2010

Multi-layer perceptron neural networks have been constructed to classify four types of defects in steam generator tubes. Three features are extracted from the signals of the eddy current testing method. These include maximum impedance, phase angle at the point of maximum impedance, and an angle between the point of maximum impedance and the point of half the maximum impedance. Two hundred sets of these features are used for training and assessing the networks. Two approaches are involved to train the networks and to classify the defect type. One is the conjugate gradient method and the other is the Broydon-Fletcher-Goldfarb-Shanno method which is recognized as the most popular algorithm of quasi-Newton methods. It is found from the computation results that the training time of the Broydon-Fletcher-Goldfarb-Shanno method is much faster than that of the conjugate gradient method in most cases. On the other hand, no significant difference of the classification performance between the two methods is observed.