Softcomputing Techniques on Embedded Systems for Industrial Engineering and Information Technology

Abstract

This dissertation documents a research on e cient Soft Computing tech- niques on Embedded Systems. The study begins with an analysis of the numerical techniques that constitutes the state of the art of practical Soft Computing, identifying in the Arti cial Neural Networks and the Non-linear Optimization Algorithms the most interesting techniques to develop in em- bedded environment. A comprehensive foundation of the theoretical aspects involving these topics, and a documented review of the techniques used in practical engineering scenarios is described in the rst part of the disserta- tion. The topics investigated involves: Advanced Arti cial Neural Network Architectures, Enhanced Generalization Sizing Techniques, First and Second Order Training Algorithms, Non-Linear Optimization problems, Evolution- ary and Swarm Intelligence Algorithms, Local Search Techniques and Hybrid Optimization Algorithms. The problem of embedded implementation of Arti cial Neural Networks, and the optimization algorithms involved for their training, is described in the second part of the dissertation. First, an overview of the most interesting embedded devices available for this techniques implementation is presented. Following, an analysis on the performance, accuracy and memory footprint 1 for neural techniques is presented. Two aspects where considered in the anal- ysis: the numerical approximations of the non-linear activation function , and the algorithmic e ciency of the linear operations. The topics investigated involved: Numerical Approximation of the Activation Function, Microcon- troller Optimization of the Resources, FPGA based Hardware Accelerators for Floating-Point Operations and Vector-Matrix Multiplication. The nal part of this dissertation describes the scenarios where the devel- oped techniques were used. The applications were divided in speci c elds. The rst regards the control of a Photovoltaic System operating-point for the extraction of the maximum power, referred as Maximum Power Point Tracking. The second describes the assessment of the Solar Irradiance from indirect measurements on PV devices, and the prediction of its future value using Recurrent Neural Networks. The third is inherent to an Inverse Biome- chanical Problem relative to the estimation of the muscle load of an cyclist using kinematic and dynamic measurements from an instrumental pedal.