Banca de DEFESA: LEONARDO MENDES SOUSA XAVIER

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : LEONARDO MENDES SOUSA XAVIER
DATE: 29/02/2024
TIME: 10:00
LOCAL: Depto. de Engenharia Elétrica / Galpão - Sala 03
TITLE:

MACHINE LEARNING APPLICATIONS IN DETECTING ANOMALIES IN WIND SYSTEMS.

KEY WORDS:

Machine learning. Fault detection. Predictive analysis. Wind energy.

PAGES: 100
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

In this work, a methodology focused on identifying anomalies in wind turbines is presented, aiming to anticipate temperature variations in critical machine components. The elements subjected to analysis include the gearbox bearing and the drive-end bearing of the generator. The study is based on modeling and applying machine learning algorithms with the goal of predicting the temperature in these components. The employed regression algorithms encompass linear multiple regression, extreme gradient boosting, and a recurrent neural network called long short-term memory. For modeling these techniques, data from the supervision system of three wind turbines in a Brazilian wind farm consisting of twelve machines are utilized. Machines were chosen with distinct temperature behaviors to assess potential variations in the performance of learning models in the face of diverse thermal behaviors. The machine data undergo a preprocessing stage to identify outliers from the normal operation of wind turbines. Subsequently, the data is divided into specific sets for algorithm application. In the case of the extreme gradient boosting model, a Bayesian optimization technique was employed to find optimal parameters that suit the proposed dataset. The results of the regression algorithms are analyzed in terms of performance metrics, and comparisons between actual and predicted temperatures are conducted within defined control limits, aiming to identify anomalies in the temperature of the studied elements. Finally, the models applied to the three machines are compared for each analyzed component. The main advantages of this model include its ability to provide excellent results for complex prediction problems, low financial implementation costs, and high adaptability for implementation in other machines.

COMMITTEE MEMBERS:
Externo ao Programa - 1061963 - JEYDSON LOPES DA SILVA - nullInterno - 1237647 - JOSE FILHO DA COSTA CASTRO
Presidente - 1543575 - PEDRO ANDRE CARVALHO ROSAS