Título: Comparative Study of Time –Frequency Representations for Anomaly Detection in Industrial Equipament
Autores: Pedro Henrique Meira de Araújo, Rodrigo de Paula Monteiro, Rafael Bérgamo Barreto de Holanda, Carmelo José Albanez Bastos Fil ho & Mariela
Resumo: In Industry 4.0, companies focus on predictive maintenance to reduce expenses and limit asset depreciation. We actively investigate anomaly detection as a way to identify potential equipment failures. Given the critical role of rotating machines in various industrial applications, our study examines how images generated from the signals of rotating machine components can aid in modeling a classification network for detecting anomalies. We utilized the MIMII dataset and implemented classification techniques using Spectrogram, Mel Spectrogram, and MFCCs methods derived from both anomalous and normal signals. We employed the generated images in a Convolutional Neural Network to distinguish between anomalous and normal instances. The Mel spectrogram clearly outshined the other methods, particularly when faced with fluctuations in signal noise. At an SNR of 6 dB, it achieved remarkable accuracy and recall scores of 0.99 and 0.98, respectively. Even when the conditions worsened to -6 dB, the Mel spectrogram still delivered impressive results, boasting an accuracy of 0.87 and a recall of 0.90, which underscores its remarkable stability and performance under challenging conditions. In comparison, the traditional Spectrogram recorded both accuracy and recall at 0.85 in the same -6 dB scenario, while the MFCC Spectrogram lagged behind significantly, with scores of only 0.74 for accuracy and 0.66 for recall. These findings highlight the Mel spectrogram’s superior effectiveness in tackling noisy environments. This work sets the stage for developing a framework that enhances machine learning techniques for detecting anomalies in industrial machinery.
Palavras-chave: industry 4.0; predictive maintenance; rotating machines; machine learning; anomaly detection.
Páginas: 7
Código DOI: 10.21528/CBIC2025-1191054
Artigo em PDF: CBIC_2025_paper1191054.pdf
Arquivo BibTeX:
CBIC_2025_1191054.bib