Banca de DEFESA: DANIELA MAIA DA SILVA

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : DANIELA MAIA DA SILVA
DATE: 31/03/2026
TIME: 09:00
LOCAL: Sala virtual: https://meet.google.com/aqc-qfpr-zku
TITLE:

Fault Detection in Industrial Machines Using Cyclostationary Acoustic Descriptors and Semi-Supervised Learning

KEY WORDS:

Fault detection, Industrial pumps, Spectral correlation density, Machine learning.

PAGES: 82
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

Early fault detection in industrial machines is essential for predictive maintenance strategies, reducing operational costs, and mitigating risks. Among non-invasive approaches, acoustic signal analysis stands out because it enables equipment monitoring without physical contact; however, noise and the spectral complexity of these signals hinder the application of conventional techniques. This study proposes and evaluates an acoustic-based fault detection methodology that combines cyclostationary analysis through Cyclic Spectral Density (SCD) with machine learning techniques. First, SCD is used to extract alpha-profiles capable of highlighting cyclic modulations associated with mechanical behavior. Next, the Random Forest algorithm is employed to reduce dimensionality and automatically select the most informative $\alpha$ coefficients. The selected descriptors are then fed into a semi-supervised autoencoder, trained only on normal samples, which detects anomalies based on reconstruction error and a threshold calibrated from nominal behavior. The methodology was validated on acoustic data from the MIMII dataset, covering different machine types and SNR conditions. The results indicate that performance varies significantly depending on machine type and the signal section analyzed: Valve and Pump showed greater stability and robustness across multiple noise scenarios, whereas Fans were more sensitive, with limitations under severe SNR conditions. Overall, the proposed approach shows practical potential for predictive maintenance by integrating cyclostationary feature extraction, automatic feature selection, and reconstruction-based detection, while also making explicit the factors (machine type, section, and noise) that condition diagnostic reliability.

 

COMMITTEE MEMBERS:
Presidente - 1543191 - LUIZ FELIPE DE QUEIROZ SILVEIRA
Interno - 2579664 - ALLAN DE MEDEIROS MARTINS
Externo ao Programa - 1640260 - JOSE JOSEMAR DE OLIVEIRA JUNIOR - UFRNExterno à Instituição - FLÁVIO BEZERRA COSTA