Técnicas de tolerância a falhas em perceptron multicamadas baseado em FPGA - estudo de caso: salve todas

Abstract

The concept of fault tolerance can be understood as the ability of a system to maintain its correct operation even after the occurrence of a failure (Avizienis et al. 2004). This area of study emerged in the 1950s, aimed at dealing with failures in military and aerospace equipment operating in hostile and/or remote environments, and since then it has proven to be a prominent field of study, especially with the popularization of the use of computers and embedded systems. And that’s the research field of this work: the application of fault tolerance techniques in an Artificial Neural Network with Multilayer Perceptron (MLP) architecture embedded in an FPGA. The MLP is part of a system aimed at women’s safety that works on the identification of possible risk situations for users. The system has vital signs, sudden movements and geolocation sensors that provide information about the user’s current situation to the MLP Network responsible for analyzing these data. Since the MLP Network plays a critical role in identifying risk situations, it is necessary to apply techniques to increase the system’s reliability, aiming at greater safety for the user. Therefore, this work analyzes the gains and impacts of applying three fault tolerance techniques combined in embedded MLP. The techniques used include: refining the weights and biases of neurons in the network’s processing layers; changes in the MLP architecture, involving the removal of hidden neurons that are less sensitive to failures and the duplication of hidden neurons that are more sensitive to failures (a technique known as Augmentation); and the Triple Modular Redundancy of the neurons in the input and output layers of the network. The results obtained with the application of the three mentioned techniques contributed to significant gains in the overall reliability of the system. The advantages of applying the techniques combined stand out, thus maximizing improvements in Reliability for the system. Furthermore, it also draws attention to the advantages of applying the techniques for Refining the Weights and Biases of the MLP Network, Removal of Hidden Neurons that are less sensitive to failures, since these techniques do not add additional costs to the project and, in the case of the Removal technique, also brings improvements in processing and system latency.