Design of Adaptive Decision Correntropy Induced Metric Functional Link Artificial Neural Network Based Channel Equalizer

• Main Authors: Yi-Hao Wang, 王翊豪
• Other Authors: Wan-De Weng
• Format: Others
• Language: zh-TW
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/24124363412585076414

碩士 === 國立雲林科技大學 === 電機工程系 === 104 === The goal of this thesis is to design a channel equalizer to compensate signal distortion during transmission. The channel equalizer is composed of functional link artificial neural network (FLANN) structure and correntropy induced metric (CIM) algorithm. Traditionally, mean square error (MSE) has been widely used as the cost function for updating weights in equalizers. This thesis introduces a new FLANN based equalizer that utilizes CIM instead. One of the most important factors that influence equalizer performance is the weight updating formulae. Conventional MSE algorithms directly apply the output error to the weight updating calculation. CIM algorithm, however, uses output error to correct the sign of weight updated values. So it is likely to obtain more accurate corrections and achieve faster convergence. For example, we see from one of our simulation results that FLANNCIM achieves stabilization after 800 iterations of training, but FLANNMSE requires 1200 iterations to attain similar results. It can also be seen that FLANNCIM exhibits lower steady state error. The purpose of expansion of basis functions is to expand the received interference signals into a higher order dimensionality. And then from the expanded terms the FLANN structure can extract information with higher correlation to restore original signals. Trigonometric polynomials are used as the basis functions of FLANNCIM. In this thesis the equalizer performance is compared with other basis functions, such as Chebyshev polynomial, Legendre polynomial, etc. We found that Chebyshev polynomial FLANN (CFLANN) and Legendre polynomial FLANN (LFLANN) both suffer larger output deviation and slower rates of convergence. The performance degradations of CFLANN and LFLANN become more severe when inter-symbol interference is getting higher. Bit error rate (BER) performance and convergence rate of FLANNCIM, however, only slightly change under the same situation. Simulation results demonstrate that FLANNCIM has better performance and is relatively robust to channel variations. We conclude that FLANNCIM structures can potentially be applied to realistic channel equalizers.