Enhancing fault and failure prediction in redundancy models: a novel approach using generalized stochastic petri networks and spiking neural networks

• Ano de defesa: 2024
• Autor(a) principal: Abreu, Rute Souza de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: Spiking neural networks; Dependability; Fault prediction; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/58340

Failure prediction plays an important role across several sectors such as industry, technology, medical sector, among others. This task can help in the reducing of equipment maintenance costs, prevention of accidents and disasters, and improvement of system dependability since it can increase availability by reducing system downtime. This work presents a methodology for fault prediction in redundancy models designed using the formality of Generalized Stochastic Petri Networks. The approach comprehends the steps of modeling and simulation of systems with active and passive redundancy under different fault scenarios, such as non-perfect switches, standby failures, and common cause failures, as well as fault datasets generation and the implementation of a machine learning model for performing the fault prediction. For forecasting, this research utilizes Spiking Neural Networks (SNNs), which have been recognized as the third generation of Artificial Neural Networks. Just like typical artificial neural networks, SNNs draw inspiration from the biological dynamics of the brain, incorporating the interconnected topology of neurons into their architecture. However, while conventional neural networks rely on error minimization by weight adjustment, SNNs aim to replicate the learning process by simulating neuron behavior by taking into account elements of the biological process such as synapse, energy accumulation, electric impulse firing, and refractory periods between emissions. Due to the ability to capture temporal aspects from data, SNNs are vastly used in problems with time dynamics. Additionally, literature has shown these networks to be task and energy-efficient serving as a low-cost alternative compared to conventional ANNs.