• Proceedings of the IEEK Conference
• /
• 2002.06d
• /
• pp.191-194
• /
• 2002

The final purpose of this paper is the enhancement of speech recognition rate under the matched telephone environment between training data and test data. To analyze the effect by the distortion of the changing telephone channel on every call, MFCC is used as the feature parameter and CMN, RTCN, and RASTA are used as channel compensation techniques. For each case, the variation of feature parameters of all phones is analyzed. And, we find recognition rates according to each compensation method using the continuous HMM recognizer, and examine the relationship between variation and recognition rate.

• 제어로봇시스템학회:학술대회논문집
• /
• 1998.10a
• /
• pp.215-220
• /
• 1998

In order to optimize fuzzy model, we use the optimal algorithm with a hybrid type in the identification of premise parameters and standard least square method in the identification of consequence parameters of a fuzzy model. The hybrid optimal identification algorithm is carried out using a genetic algorithm and improved complex method. Also, the performance index with weighting factor is proposed to achieve a balance between the insults of performance for the training and testing data. Several numerical examples are used to evaluate the performance of the proposed model.

• Journal of the Korean Institute of Telematics and Electronics C
• /
• v.34C no.6
• /
• pp.79-88
• /
• 1997

New technologies, systems, and services in telecommunication have increased the need for an efficient and robust control mechanism to protect switching systems from overload. To achieve proper control, it is necessary to find a set of parameters that can describe the system. However, it is difficult to find types of data that can form a suitable basis for control. In this paper, we categorize the load status of a switching system into three classes (i.e., normal state, pre-overload state, and overload state) and formulate the overload detection as a classification problem. We find the relationships between the load classes and a set of monitored switching system parameters by applying a fuzzy rule-generation method. The rules are automatically generated from training data. Simulation results involving a switching system is given.

• Speech Sciences
• /
• v.7 no.4
• /
• pp.11-26
• /
• 2000

In this paper, we propose the HMM with the MIN module. Because initial and re-estimated variance vectors are important elements for performance in HMM recognition systems, we propose a method which compensates for the mismatched statistical feature of training and test data. The MIN module function is a differentiable function similar to the sigmoid function. Unlike a continuous density function, it does not include variance vectors of the data set. The proposed hybrid HMM/MIN module is a unified network in which the observation probability in the HMM is replaced by the MIN module neural network. The parameters in the unified network are re-estimated by the gradient descent method for the Maximum Likelihood (ML) criterion. In estimating parameters, the variance vector is not estimated because there is no variance element in the MIN module function. The experiment was performed to compare the performance of the proposed HMM and the conventional HMM. The experiment measured an isolated number for speaker independent recognition.

• Speech Sciences
• /
• v.11 no.4
• /
• pp.75-88
• /
• 2004

In this paper, we propose a new approach to sequential linear regression adaptation of continuous density hidden Markov models (CDHMMs) based on transformation space model (TSM). The proposed TSM which characterizes the a priori knowledge of the training speakers associated with maximum likelihood linear regression (MLLR) matrix parameters is effectively described in terms of the latent variable models. The TSM provides various sources of information such as the correlation information, the prior distribution, and the prior knowledge of the regression parameters that are very useful for rapid adaptation. The quasi-Bayes (QB) estimation algorithm is formulated to incrementally update the hyperparameters of the TSM and regression matrices simultaneously. Experimental results showed that the proposed TSM approach is better than that of the conventional quasi-Bayes linear regression (QBLR) algorithm for a small amount of adaptation data.

• Journal of the Korea Institute of Military Science and Technology
• /
• v.13 no.3
• /
• pp.461-469
• /
• 2010

In this paper, we have proposed an improved fusion method of detection features which can enhance the detection probability under the given false alarm rate in the prescreening stage of SAR ATR(Synthetic Aperture Radar Automatic Target Recognition) system. Since the detection features have the positive correlation, the detection performance can be improved if the joint probability distribution of detection features is considered in the fusion process. The detection region is designed as a simple piecewise linear function which can be represented by few parameters. The parameters for the detection region can be derived by training the sample SAR images to maximize the detection probability with the given false alarm rate. Simulation result shows that the detection performance of the proposed method is improved for all combinations of detection features.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.22 no.3
• /
• pp.481-492
• /
• 1997

This paper presents an algorithm for recognizing surface mount device(SMD) IC pattern based on the error back propoagation(EBP) neural network and discrete cosine transform(DCT). In this approach, we chose such parameters as frequency, angle, translation and amplitude for the shape informantion of SMD IC, which are calculated from the coefficient matrix of DCT. These feature parameters are normalized and then used for the input vector of neural network which is capable of adapting the surroundings such as variation of illumination, arrangement of objects and translation. Learning of EBP neural network is carried out until maximum error of the output layer is less then 0.020 and consequently, after the learning of forty thousand times, the maximum error have got to this value. Experimental results show that the rate of recognition is 100% in case of the random pattern taken at a similar circumstance as well as normalized training pattern. It also show that proposed method is not only relatively relatively simple compare with the traditional space domain method in extracting the feature parameter but also able to re recognize the pattern's class, position, and existence.

• Proceedings of the KIEE Conference
• /
• 1993.07b
• /
• pp.1004-1006
• /
• 1993

A new method for the optimal design of a single-sided linear induction motor(SLIM) is presented. The method utilizes the neural networks and finite element method for optimizing the design parameters of SLIM. The finite element analysis is used to produce a variety of neural networks training data and the neural networks is used for optimizing the design parameters by sequential unconstrained minimization technique(SUMT). As a result, it is known that the novel method is very efficient and accurate as an optimization technique.

• Journal of the Semiconductor & Display Technology
• /
• v.19 no.4
• /
• pp.126-129
• /
• 2020

We introduce a lifetime assessment technique using deep learning algorithm with complex electrical parameters such as resistivity, permittivity, impedance parameters as integrated indicators for predicting the degradation of the organic molecules. The evaluation system consists of fully automated in-situ measurement system and multiple layer perceptron learning system with five hidden layers and 1011 perceptra in each layer. Prediction accuracies are calculated and compared depending on the physical feature, learning hyperparameters. 62.5% of full time-series data are used for training and its prediction accuracy is estimated as r-square value of 0.99. Remaining 37.5% of the data are used for testing with prediction accuracy of 0.95. With k-fold cross-validation, the stability to the instantaneous changes in the measured data is also improved.

• Journal of the Semiconductor & Display Technology
• /
• v.22 no.4
• /
• pp.131-135
• /
• 2023

Neural network-based time series models called time series neural networks (TSNNs) are trained by the error backpropagation algorithm and used to predict process shifts of parameters such as gas flow, RF power, and chamber pressure in reactive ion etching (RIE). The training data consists of process conditions, as well as principal components (PCs) of optical emission spectroscopy (OES) data collected in-situ. Data are generated during the etching of benzocyclobutene (BCB) in a SF₆/O₂ plasma. Combinations of baseline and faulty responses for each process parameter are simulated, and a moving average of TSNN predictions successfully identifies process shifts in the recipe parameters for various degrees of faults.