Desenvolvimento de modelos de predição de falhas em turbogeradores por meio da análise multivariada de séries temporais.

Abstract

This report presents the documentation of a study on multivariate time series classification models. As a case study, a failure prediction model in turbogenerators of an isolated generation system was developed. For this, operational data and monitoring records from one of the system’s turbogenerators were used. This data went through a pre-processing stage, in which the data was cleaned and the model’s most relevant attributes were selected. Throughout the work, a multivariate time series prediction model based on VAR (Vector AutoRegression) was implemented to constitute an artificial database in order to contribute with unbalance issues in the failures detection. The fault classifier was developed using a machine learning model based on a special kind of recurrent neural networks called LSTM (Long Short Term Memory). The tests carried out so far indicate that the implemented techniques present satisfactory results in the detection of turbogenerator failures, contributing to the asset management process.