Garuda - Garba Rujukan Digital

Studi Komparasi Terhadap Kapabilitas Generalisasi dari Jaringan Saraf Tiruan Berbasis Incremental Projection Learning Murfi, Hendri; Kusumoputro, Benyamin
Jurnal Teknik Elektro Vol 1, No 2 (2001): SEPTEMBER 2001
Publisher : Institute of Research and Community Outreach

One of the essences of supervised learning in neural network is generalization capability. It is an ability to give an accurate result for data that are not learned in learning process. One of supervised learning method that theoretically guarantees the optimal generalization capability is incremental projection learning. This paper will describe an experimental evaluation of generalization capability of the incremental projection learning in neural networks%2C called projection generalizing neural networks%2C for solving function approximation problem. Then%2C Make comparison with other general used neural networks%2C i.e. back propagation networks and radial basis function networks. Base on our experiment%2C projection generalizing neural networks doesn%5C%27t always give better generalization capability than the two other neural networks. It gives better generalization capability when the number of learning data is small enough or the noise variance of learning data is large enough. Otherwise%2C it does not always give better generalization capability. Even though%2C In case the number of learning data is big enough and the noise variance of learning data is small enough%2C projection generalizing neural networks gives worse generalization capability than back propagation networks Abstract in Bahasa Indonesia : Salah satu hal yang penting dari suatu metode pembelajaran pada jaringan saraf tiruan adalah kapabilitas generalisasi. Yaitu kemampuan untuk memberikan hasil yang akurat terhadap data yang tidak diajarkan pada tahap pembelajaran. Salah satu metode pembelajaran yang memberikan jaminan secara teori diperolehnya kapabilitas generalisasi yang optimal adalah projection learning. Pada tulisan ini kami akan melakukan evaluasi eksperimental terhadap kapabilitas generalisasi dari jaringan saraf tiruan berbasis projection learning yang bersifat incremental%2C yang disebut projection generalizing neural networks%2C untuk memecahkan masalah aproksimasi fungsi. Kemudian melakukan studi komparasi dengan jaringan saraf tiruan yang sudah umum digunakan%2C yaitu back propagation networks dan radial basis functions networks. Berdasarkan hasil uji coba komputasi yang kami lakukan%2C projection generalizing neural networks tidak selalu memberikan kapabilitas generalisasi yang lebih baik. projection generalizing neural networks memberikan kapabilitas generalisasi yang lebih baik ketika jumlah data pembelajaran cukup kecil atau variansi noise dari data pembelajaran cukup besar. Selain dari dua kondisi tersebut%2C projection generalizing neural networks tidak selalu memberikan kapabilitas generalisasi yang lebih baik. Bahkan%2C untuk kondisi dimana jumlah data pembelajaran cukup besar dan variansi noise cukup kecil%2C projection generalizing neural networks memberikan kapabilitas generalisasi yang lebih buruk dari back propagation networks. supervised+learning%2C+incremental+projection+learning%2C+generalization+capability%2C+artificial+neural+networks%2C+function+approximation+problem

Sensorless-BLDC motor speed control with ensemble Kalman filter and neural network Rif'an, Muhammad; Yusivar, Feri; Kusumoputro, Benyamin
Journal of Mechatronics, Electrical Power and Vehicular Technology Vol 10, No 1 (2019)
Publisher : Research Centre for Electrical Power and Mechatronics, Indonesian Istitutes of Sciences

The use of sensorless technology at BLDC is mainly to improve operational reliability and play a role for wider use of BLDC motors in the future. This research aims to predict load changes and to improve the accuracy of estimation results of sensorless-BLDC. In this paper, a new filtering algorithm is proposed for sensorless brushless DC motor based on Ensemble Kalman filter (EnKF) and neural network. The proposed EnKF algorithm is used to estimate speed and rotor position, while neural network is used to estimate the disturbance by simulation. The proposed algorithm requires only the terminal voltage and the current of three phases for estimated speed and disturbance. A model of non-linear systems is carried out for simulation. Variations in disturbances such as external mechanical loads are given for testing the performance of the proposed algorithm. The experimental results show that the proposed algorithm has sufficient control with error speed of 3 % in a disturbance of 50 % of the rated-torque. Simulation results show that the speed can be tracked and adjusted accordingly either by disturbances or the presence of disturbances.

Sensorless-BLDC motor speed control with ensemble Kalman filter and neural network Muhammad Rif'an; Feri Yusivar; Benyamin Kusumoputro
Journal of Mechatronics, Electrical Power and Vehicular Technology Vol 10, No 1 (2019)
Publisher : National Research and Innovation Agency

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.14203/j.mev.2019.v10.1-6

The use of sensorless technology at BLDC is mainly to improve operational reliability and play a role for wider use of BLDC motors in the future. This research aims to predict load changes and to improve the accuracy of estimation results of sensorless-BLDC. In this paper, a new filtering algorithm is proposed for sensorless brushless DC motor based on Ensemble Kalman filter (EnKF) and neural network. The proposed EnKF algorithm is used to estimate speed and rotor position, while neural network is used to estimate the disturbance by simulation. The proposed algorithm requires only the terminal voltage and the current of three phases for estimated speed and disturbance. A model of non-linear systems is carried out for simulation. Variations in disturbances such as external mechanical loads are given for testing the performance of the proposed algorithm. The experimental results show that the proposed algorithm has sufficient control with error speed of 3 % in a disturbance of 50 % of the rated-torque. Simulation results show that the speed can be tracked and adjusted accordingly either by disturbances or the presence of disturbances.

PERLUASAN METODE MFCC 1D KE 2D SEBAGAI ESKTRAKSI CIRI PADA SISTEM IDENTIFIKASI PEMBICARA MENGGUNAKAN HIDDEN MARKOV MODEL (HMM) Buono, Agus; Jatmiko, Wisnu; Kusumoputro, Benyamin
Makara Journal of Science Vol. 13, No. 1
Publisher : UI Scholars Hub

Show Abstract | Download Original | Original Source | Check in Google Scholar

The Extention of MFCC Technique from 1D to 2D as Feature Extractor for Speaker Identification System Using HMM. In this paper, we introduce an extension of Mel-Frequency Cepstrum Coefficients (1D-MFCC) methodology to bispectrum data, referred to as 2D-MFCC, for feature extraction. 2D-MFCC is based on 2D bispectrum data rather than 1D spectrum vector yielded by Fourier transform, so the filter in 1D-MFCC must be extend to 2D filter and using 2D cosine transform to get the mel-cepstrum coefficients from the filtered bispectrum values. Based on 2D-MFCC, we develop a speaker recognition system with Hidden Markov Model (HMM) as classifier. The experimental results show that the recognition rate is around 88%, 92% and 99% for 20, 40 and 60 data training, respectively

PENDETEKSIAN JENIS DAN KELAS AROMA DENGAN MENGGUNAKAN METODE ONE-VS-ONE DAN METODE ONE-VS-REST Rustam, Zuherman; Kusumoputro, Benyamin; Widjaja, Belawati
Makara Journal of Science Vol. 7, No. 3
Publisher : UI Scholars Hub

Show Abstract | Download Original | Original Source | Check in Google Scholar

Aroma classification using one-vs-one and one-vs-rest methods. Computational Intelligence used in pattern classification problem can be divided into two different parts, one based on Neural Network and the other based on Statistical Learning. The Statistical Learning discovered by Vapnik on 70-est decade. For the pattern classification, Vapnik developed hyperplane optimal separation, which is known as Support Vector Machines Method (SVM). In the beginning, SVM was designed only to solve binary classification problem, where data existing are classified into two classes. To classify data whose consist of more than two classes, the SVM method can not directly be used. There are several methods can be used to solve SVM multiclasses classification problem, they are One-vs-One Method and One-vs-Rest Method. Both of this methods are the extension of SVM binary classification, they will be discussed in this article so that we can see their performance in aroma classification process. Data of aroma used in this experiment is consisted of three classes of aroma, each of them has six classes. The division of this class is based on alcohol concentration mixed into each of those aromas. For example, for aroma A, there are six kinds of aroma A with different alcohol concentration: 0%, 15%, 25%, 30%, 45% and 75%. The performance of these methods is measured based on their ability to recognize and classify aroma, precisely and match with the right class or variety of data existed.

SISTEM PENGENALAN WAJAH 3-D MENGGUNAKAN PENAMBAHAN GARIS CIRI PADA METODE PERHITUNGAN JARAK TERPENDEK DALAM RUANG EIGEN Lina, Lina; Kusumoputro, Benyamin
Makara Journal of Science Vol. 7, No. 1
Publisher : UI Scholars Hub

Show Abstract | Download Original | Original Source | Check in Google Scholar

3-D Face Recognition System using Additional Feature Lines in Nearest Feature Line Method in Eigenspace Representation. In this paper, the authors propose a new method in 3-D face recognition system using additional feature lines in Nearest Feature Line method, called the Modified Nearest Feature Line method. The additional feature lines can be acquired by projecting each feature point to other feature lines in the same class without increasing the number of feature points. With these additional lines, the system will have the ability to capture more variations of face images, so it can increase the recognition rate of the system. The authors also propose KL-TSubspace1 and KL-TSubspace2 as methods in transforming the 3-D face images from its spatial domain to their eigenspace domain. The experiments use the 3-D human faces of Indonesian people in various expressions and positions. Then, the system is applied to recognize unknown face images with different viewpoints. Experimental results shown that the system using KL-TSubspace2 and Modified Nearest Feature Line method can have the highest recognition rate of 99.17%.

PENGEMBANGAN SISTEM PENCIUMAN ELEKTRONIK DENGAN 16 BUAH SENSOR KUARSA DAN ALGORITMA NEURAL PROPAGASI BALIK UNTUK PENGENALAN AROMA CAMPURAN Kusumoputro, Benyamin; Jatmiko, Wisnu
Makara Journal of Science Vol. 6, No. 3
Publisher : UI Scholars Hub

Show Abstract | Download Original | Original Source | Check in Google Scholar

An artificial odor recognition system is developed for discriminating odors. This artificial system consisted of 16 quartz resonator crystals as the sensor array, a frequency modulator and a frequency counter for each sensor that are connected directly to a microcomputer. We have already shown that the artificial odor recognition system with 4 sensors is high enough to discriminate simple odor correctly, however, when it was used to discriminate compound odors, the recognition capability of this system is dropped significantly to be about 40%. Results of experiments show that the developed artificial system with 16 sensors could discriminate compound aroma based on 6 gradient of alcohol concentrations with high recognition rate of 89.9% for non batch processing system, and 82.4% for batch processing of the classes of odors

PENGEMBANGAN SISTEM PENCIUMAN ELEKTRONIK DENGAN 16 BUAH SENSOR KUARSA DAN ALGORITMA NEURAL PROPAGASI BALIK UNTUK PENGENALAN AROMA CAMPURAN Kusumoputro, Benyamin; Jatmiko, Wisnu
Makara Journal of Science Vol. 6, No. 3
Publisher : UI Scholars Hub

Show Abstract | Download Original | Original Source | Check in Google Scholar

An artificial odor recognition system is developed for discriminating odors. This artificial system consisted of 16 quartz resonator crystals as the sensor array, a frequency modulator and a frequency counter for each sensor that are connected directly to a microcomputer. We have already shown that the artificial odor recognition system with 4 sensors is high enough to discriminate simple odor correctly, however, when it was used to discriminate compound odors, the recognition capability of this system is dropped significantly to be about 40%. Results of experiments show that the developed artificial system with 16 sensors could discriminate compound aroma based on 6 gradient of alcohol concentrations with high recognition rate of 89.9% for non batch processing system, and 82.4% for batch processing of the classes of odors

Comparative Analysis of LSTM and Bi-LSTM Models for Earthquake Occurrence Prediction in Tokai-Japan Region Hamdi, Azhari Haris Al; Nugroho, Hapsoro Agung; Kusumoputro, Benyamin
International Journal of Electrical, Computer, and Biomedical Engineering Vol. 2 No. 4 (2024)
Publisher : Universitas Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.62146/ijecbe.v2i4.87

This study compares the performance of Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) models in predicting earthquake occurrences in the Tokai region, using data from the United States Geological Survey (USGS) dataset. Given the importance of accurate earthquake prediction, particularly in high-risk regions, this research focuses on assessing the effectiveness of each model in identifying occurrence and non-occurrence events. Both models were tuned to optimize sensitivity and specificity through adjustments in sequence length, learning rate, and additional hyperparameters, with results evaluated using metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the curve (AUC). Findings reveal that while both models achieved high sensitivity, the LSTM model demonstrated superior specificity and AUC, indicating a more balanced performance in distinguishing between earthquake occurrences and non-occurrences. The results show that LSTM outperforms Bi-LSTM in terms its classification metrics. LSTM achieved an accuracy of 76%, compared to 55% for Bi-LSTM. For the AUC metric, LSTM scored 66%, while Bi-LSTM scored 67%.

Komparasi Metode Optimasi Adam dan SGD dalam Skema Direct Inverse Control untuk Sistem Kendali Data Sikap dan Ketinggian Quadcopter HAQQI, MUHAMMAD SABILA; KUSUMOPUTRO, BENYAMIN
ELKOMIKA: Jurnal Teknik Energi Elektrik, Teknik Telekomunikasi, & Teknik Elektronika Vol 10, No 2: Published April 2022
Publisher : Institut Teknologi Nasional, Bandung

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.26760/elkomika.v10i2.458

ABSTRAKTerdapat banyak variable nonlinear dalam sistem kendali untuk quadcopter sehingga cukup rumit untuk mengatur dinamika penerbangan wahana ini. Untuk mengatasi masalah tersebut akan dikembangkan suatu skema sistem kendali Direct Inverse Control menggunakan Deep Learning berbasis Artificial Neural Network (ANN). Sistem yang dikembangkan akan mengendalikan data sikap dan ketinggian quadcopter. Pada artikel ini akan dibandingkan kinerja dari dua metode optimasi untuk Mean Squared Error pada simulasi, yaitu Adaptive Moment Estimation dan Stochastic Gradient Descent. Hasil menunjukkan metode Adaptive Moment Estimation mampu memberikan nilai Mean Squared Error yang lebih kecil dibandingkan metode Stochastic Gradient Descent untuk semua data sikap dan ketinggian yang dikendalikan dengan nilai 0.0069 untuk roll rate, 0.0057 untuk pitch rate, 0.0062 untuk yaw rate, dan 0.0042 untuk data ketinggian.Kata kunci: Deep Learning, Artificial Neural Network, Adam, SGD, MSEÂ ABSTRACTThere are many nonlinear variables in the control system for the quadcopter so it is quite complicated to regulate the flight dynamics of this vehicle. To overcome this problem, a Direct Inverse Control control system scheme using Deep Learning based on Artificial Neural Network (ANN) will be developed. The system developed will control the attitude and altitude data of the quadcopter. In this article, we will compare the performance of two optimization methods for Mean Squared Error in simulation, namely Adaptive Moment Estimation and Stochastic Gradient Descent. The results show that the Adaptive Moment Estimation method is able to provide a smaller Mean Squared Error value than the Stochastic Gradient Descent method for all attitude and altitude data controlled with values of 0.0069 for roll rate, 0.0057 for pitch rate, 0.0062 for yaw rate, and 0.0042 for altitude data. Keywords: Deep Learning, Artificial Neural Network, Adam, SGD, MSE

Title

Found 12 Documents
Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title Search

Found 12 Documents Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title

Found 12 Documents
Search