Adaptive IIR diffusion networks using hybrid FIR-IIR combination.

• Ano de defesa: 2021
• Autor(a) principal: Coelho, Rodrigo Midea
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Adaptive networks; Filters combination; FIR; Identificação de sistemas; IIR adaptive filters; Output error; Padé approximant; Processamento de sinais adaptativos; Processamento digital de sinais
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/3/3142/tde-29042022-083345/

Application fields such as Internet of Things, wearables and 5G are typically battery and cost constrained, demanding algorithms with improved performance at reduced computational cost. Infinite impulse response (IIR) adaptive filters (AFs), are known by their ability to model an unknown system in a compact manner, but their typically low convergence rates may pose a major obstacle to their widespread use when identifying rational plants. In this work, two new iterative methods based on Padé approximants mappings (PAM) are introduced to be used in hybrid finite impulse response (FIR)-IIR adaptive filter combination to accelerate convergence while preserving steady-state performance. An iterative linear system solver is employed to leverage the FIR AF by performing the PAM at all iterations, in contrast to previous approaches that performs the mappings either once or in a cyclic manner. The proposed methods spread the computational cost and makes them suitable for hard real time applications. Simulations show the novel mappings efficiency by outperforming previous cyclic mapping approaches and even modern metaheuristic algorithms. After addressing the standalone IIR AF applications, a distributed estimation implementation for adaptive networks (ANs) of the hybrid FIR-IIR structure is introduced. Simulations show that the proposed IIR-ANs are able to outperform traditional FIR-ANs when estimating auto-regressive moving-average (ARMA) processes by reducing from 70% to 90% the total quantity of multiplications, making it suitable for IoT and sensor networks (SNs) applications.