Título: Fault Detection in Industrial Pumps Using Cyclic Frequencies with Autoencoder
Autores: Daniela Maia da Silva, Luiz Felipe Queiroz Silveira
Resumo: Industrial machines play a fundamental role in the productive sector, being responsible for maintaining efficiency and operational continuity across various industrial segments. Early fault detection in such equipment, through machine learning techniques, is essential to increase operational reliability, reduce corrective maintenance costs, and avoid unplanned downtimes. In this work, we propose an approach for anomaly detection in industrial pumps based on acoustic signal analysis, using Spectral Correlation Density (SCD) combined with unsupervised learning models. SCD is employed to capture cyclostationary spectral patterns present in the signals, allowing the extraction of maximum-value vectors per cyclic frequency (α-profile). Subsequently, the Random Forest algorithm is used to select the most relevant cyclic frequencies, resulting in a lower-dimensional yet highly discriminative representation. The extracted descriptors are used as input to autoencoders trained exclusively on normal data, with anomaly detection performed based on reconstruction error. The methodology was evaluated using real data from the MIMII Dataset, covering sections 00, 02, and 04, and showed consistent performance, with AUCs above 0.90 and an average F1-score of 0.80, outperforming previous spectrogram-based approaches.
Palavras-chave: Fault detection; Industrial pumps; Cyclostationary signals; Spectral correlation density; Machine learning.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1191777
Artigo em PDF: CBIC_2025_paper1191777.pdf
Arquivo BibTeX:
CBIC_2025_1191777.bib