Diagnóstico de falhas em processos industriais baseado em dados usando resíduos de redes neurais recorrentes

Abstract

This work proposes a data-driven approach for fault detection and diagnosis in industrial processes using residuals generated by a Long Short-Term Memory (LSTM) network. The LSTM model is trained with normal operating data in order to learn the nominal dynamics of the system and estimate the monitored process variables. The residuals are computed as the difference between the measured and predicted values, representing deviations from the expected system behavior. The generated residuals are used as input features for a supervised Random Forest classifier, which is responsible for identifying the operating condition of the process. The proposed methodology is validated using the Sim3Tanks benchmark, considering both normal operating scenarios and fault scenarios. The obtained results indicate that the combination of LSTM, residual generation and Random Forest classification is capable of distinguishing normal and faulty conditions, achieving good performance in the evaluated classification and detection metrics.