Uso de técnicas de aprendizado de máquina para previsão de falhas em turbogeradores

Abstract

In the industry, anomaly detection capability under the operating conditions is of great interest. If identi ed well in advance, maintenance interventions can be planned on demand, which determines a condition based maintenance. With the increase of the amount of data acquired for supervision and of the computational power for processing, the development of machine learning techniques can aid in the detection of operating conditions that indicate maintenance needs. In this dissertation, such techniques are applied to allow the identi cation of failures in turbo generators. Methodologies are presented for the treatment of the operation databases, for the selection of variables and for the identi cation of characteristics that represent the operational cases properly. Classi ers are designed with this data and compared to each other to evaluate the e ectiveness of these methods.