PRINCIPAL COMPONENT ANALYSIS FOR SUPPORT VECTOR MACHINE ALGORITHM-BASED MODULATION CLASSIFIER MODEL OPTIMIZATION

• Main Author: Damario Lukito, William
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/64890
• Institution: Institut Teknologi Bandung

ABSTRACT PRINCIPAL COMPONENT ANALYSIS FOR SUPPORT VECTOR MACHINE ALGORITHM-BASED MODULATION CLASSIFIER MODEL OPTIMIZATION By William Damario Lukito NIM: 23221018 (Master’s Program in Electrical Engineering) Automatic modulation classification (AMC) is one of the noteworthy and challenging parts in the development of intelligent radio systems. An AMC model should be accurate, having consistent accuracies over various channel conditions, computationally efficient, and able to classify as many modulation types as possible. Various problems and challenges appeared related to AMC cases, especially in complex interclass and intraclass modulation classification. Various methods and techniques have been also presented by previous researchers. However, those methods were well-performed just for interclass or intraclass only. As the technology develops, the classification technique using machine learning has been popularly used to solve diverse complex cases. Based on several references, support vector machine (SVM) is one of the machine learning algorithms that could obtain a great performance. Nevertheless, the challenge of machine learning applications is the determination of input predictors. Increasing the number of predictors increases computational complexity, even though it should make the model more exact. In this Thesis, 21 features based on signal spectral, Wavelet transform, higher-order statistics, and cyclostationary analysis are listed on the features pool. Those features will be selected for AWGN and Rayleigh fading channels. Moreover, in order to solve the complexity issues, dimension reduction techniques such as principal component analysis (PCA) could be used. The PCA works by transforming a set of classification features into a set of principal components (PC). The transformed PCs should have the same dimension as the number of used classification features. As a consequence, we need to perform an evaluation regarding the performance effects of the decreasing number of PCs over the obtained AMC models. In this Thesis, the modulations are limited to 6 types that represent interclass and intraclass, which are BPSK, QPSK, 8-PSK, 16-QAM, BFSK, and 8-PAM. The modulated signals are in the radio frequency (RF) band. The quadratic kernel iii SVM is used as the machine learning algorithm. Training will be performed twice, without PCA (non-PCA) and with PCA. After training, we found that the decreasing number of PCs decreased the average accuracy, and increased the elapsed training time. However, no linear nor reciprocal relation is found between the number of PCs and the prediction speed. Nonetheless, with a certain fixed number of PCs, PCA application should upturn the average accuracy, hasten the elapsed training time, or accelerate the prediction speed. Non-PCA and PCA training resulted in a set of AMC models for AWGN and Rayleigh fading channels. The optimum model for AWGN channel is SVM PCA with 8 out of 11 PCs, which obtained 97.10% average accuracy, with 6.7986 seconds of elapsed training time and about 64000 observation/second of prediction speed. Meanwhile, the optimum model for Rayleigh fading channel is SVM PCA with 8 out of 8 PCs, which obtained 75.20% average accuracy, with 296.22 seconds of elapsed training time and about 9300 observation/second of prediction speed. Keywords: PCA, AMC, SVM.