Failure prognosis methods and offline performance evaluation

• Autor(a) principal: Bruno Paes Leão
• Data de Publicação: 2011
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do ITA
• Texto Completo: http://www.bd.bibl.ita.br/tde_busca/arquivo.php?codArquivo=1970
• topic: Monitoramento da saúde de sistemas; Análise de falhas; Detecção de falhas; Vida útil; Estimação de vida útil restante; Filtros de Kalman; Filtros de rastreamento; Algoritmos; Engenharia eletrônica

The capability of predicting failure events of systems and components can provide benefits in equipment operation and maintenance. For this reason, the subject of failure prognosis is gaining greater attention from academia and industry over the last years. This work presents novel contributions related to the development and performance evaluation of failure prognosis solutions. One important step in failure prognosis is the estimation of the health state of the monitored equipment and its trend. Here, Sigma-Point Kalman Filter (SPKF) algorithms are employed for this purpose and their performance is compared to Particle Filter (PF) algorithms which are commonly cited in literature for this kind of application. Once the health state and its trend are estimated, in order to proceed with the failure prognosis, it is necessary to use this information to predict the remaining useful life (RUL) of the equipment. The RUL estimate is commonly yielded on the form of a probability distribution. A novel method, based on the Unscented Transform (UT), is presented and evaluated for this purpose. Results indicate that this approach may provide benefits when compared to the usual Monte Carlo based method. Finally, after a failure prognosis solution is developed, it is necessary to adequately evaluate its performance. This work also comprises the proposition of a novel method for such assessment, based on the use of the Probability Integral Transform (PIT). Such new method provides a measure of how adequately the proposed RUL probability distributions fit the available set of ground truth validation data. Also, additional proposed features make it possible to take into consideration the impact of the size of the validation data set into the uncertainty of the resulting metrics.