Previsão de falhas em componentes aeronáuticos: um estudo comparativo entre redes neurais artificiais

Abstract

This study addresses predictive maintenance to predict aircraft engine failures, employing neural network methods such as Recurrent Neural Network (RNN), Bidirectional Recurrent Neural Networks (BRNN), Long short-term memory (LSTM), Extreme Gradient Boosting (XGBoost) and Gated Recurrent Unit (GRU). The work is justified by the crucial need to keep engines running excellent conditions to ensure passenger safety and reduce maintenance expenses. The fundamental objective is to develop models that, By delving into cycle history and sensory data, they can predict whether a motor will fail within a specific cycle, thus making a comparison between methods and evaluating which is best for that application. The methodology covers importing libraries, loading and pre-processing the dataset, generating of the target variable for classification, data normalization and exploratory data analysis (EDA). Several RNN models were built and evaluated, including a simple RNN with one and 25 features, as well as a bidirectional model with 25 features. Subsequently, the LSTM, XGBOOST and GRU models were implemented. The results show that the more complex RNN models, with all 25 features, achieved better accuracy and evaluation metrics compared to the simpler models. The conclusion highlights the better accuracy of the LSTMs model for this application, being more suitable in aircraft maintenance to avoid costly repairs or unnecessary replacements, significantly contributing to the safety and efficiency of maintenance operations.