Uso e aprendizado de máquina visando maior eficiência energética de aplicações com microcontroladores embarcados.

Abstract

Embedded systems can be easily found in many domains of technology, from mobile phones to satellites, from toys to industrial plants. As the name suggests, these systems are embedded in more complex ones. And for this reason, many concerns arise from their strict requirements of small size and low weight. When developing embedded systems, these requirements are translated into constraints on memory, processing performance, cost, and low energy consumption. In the last decades, industry has joined forces with academic research for developing new methods to increase energy efficiency of embedded systems, from hardware to software, and to improve usability, cost, autonomy and reliability altogether. The goal this project was to reduce the amount of transmitted data by applying machine learning techniques for processing data on the edge. This strategy should increase the autonomy of low power wireless applications. This work presents a study case as a guide. A review of the technologies and of the developed activities in the last months is presented. Among other topics, the process of gathering, preprocessing and creating a database; defining, implementing and training a neural network on CPU; the conversion from the CPU model to a model adapted to MCU architecture; and the proposal and implementation of a solution adapted to low power in a commercial MCU (NXP QN9080)