Detecção de anomalias nas respostas de vibração de rotores usando técnicas de inteligência artificial

Abstract

Rotating machinery plays an essential role in industrial sectors such as oil and gas, power generation and mining, and its failures may result in unplanned shutdowns, significant economic losses and safety risks. Vibration analysis stands out as one of the main condition monitoring techniques, enabling the early identification of mechanical problems and supporting predictive maintenance strategies. The increasing adoption of continuous monitoring systems has led to large volumes of data, imposing limitations on traditional approaches based exclusively on manual analysis. This dissertation proposes and evaluates Machine Learning methodologies for anomaly detection in rotor vibration responses, with emphasis on unsupervised strategies aligned with the industrial reality of scarce labeled fault data. The work integrates and extends the Edge Analytics framework, developed by the Structural Mechanics Laboratory of the Federal University of Uberlândia (LMEst - UFU), which combines simple statistical techniques and deep learning models for anomaly detection. In addition to the original models of the Edge Analytics architecture—Gates, Autoencoder, DeepAnT, and Histogram — new models based on Multilayer Perceptron (MLP) and Long Short-Term Memory (LSTM) are developed and evaluated. Furthermore, the combination of these models with Change Point Detection (CPD) techniques is investigated in order to enhance sensitivity in identifying subtle and sustained changes in vibration responses. The validation of the proposed methodologies is conducted through five benchmarks, involving numerical and experimental data from different rotating systems, including rotors supported by magnetic bearings, an industrial exhaust fan and bearing degradation tests. Performance is assessed using classical classification metrics, with emphasis on the F1 score due to the strong class imbalance. The results indicate that there is no universally superior model, however LSTMand MLP-based models stand out in faults with progressive evolution, while the combination with CPD contributes to the detection of subtle changes. Thus, the study provides technical support for industrial applications in predictive maintenance.