Predição de tempo de vida restante em sistemas integrados digitais considerando condições ambientais dinâmicas

• Ano de defesa: 2017
• Autor(a) principal: Pedro Fausto Rodrigues Leite Junior
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Minas Gerais; Brasil; ENG - DEPARTAMENTO DE ENGENHARIA ELÉTRICA; Programa de Pós-Graduação em Engenharia Elétrica; UFMG
• Palavras-chave em Português: Tempo de vida restante; Confiabilidade; Envelhecimento de circuitos integrados; Predição de falhas; Engenharia elétrica; Confiabilidade (Engenharia); Circuitos integrados digitais; Localização de falhas (Engenharia) -
• Link de acesso: http://hdl.handle.net/1843/51446

The study of integrated circuits reliability has become of sudden importance to the understanding, detection and correction of their failures. Compreending how or under which conditions this aging becomes harmful to a system enable decisions to mitigate or prevent these conditions. This work establishes a flow for analysis and simulation of integrated systems that allows us to understand the aging of it in different conditions. In addition, it allows us to extract and analyze data that are relevant to predict their failure and also serve as input to verification, evaluation and fault-tolerance systems using machine learning techniques. The developed methodology allows the integration of offline and online data collection techniques to update estimation methods, as well as allowing new ones to be added. This work uses three different methods to predict the Mean Time To Failure and the Remaining Useful Lifetime for test circuits. The MTTF is estimated for each of them using a Generalized Linear Model (specifically a Partial Least Squares Regression), Euclidean Distance and Pearson’s Correlation as prediction methods. Our results indicate that the representation of the operating conditions of the systems through dynamic profiles is more realistic than the representation through a operation profile that does not vary in time, and more precise. Additionally, the MTTF prediction was approximately 90% for Partial Least Squares Regression and Euclidean Distance models.