Analysis of time–frequency signal representations for anomaly detection in industrial environments
Abstract
In the context of Industry 4.0, sound-based anomaly detection has emerged as a relevant strategy to support predictive maintenance and reduce unplanned downtime in industrial environments. This study investigates the impact of different time–frequency representations on acoustic anomaly detection in industrial machines operating under noisy conditions. Using the MIMII dataset, experiments were conducted on two machine elements (pump and slider) under three signal-to-noise ratio (SNR) levels: 6 dB, 0 dB, and −6 dB. Three feature representations, spectrogram, MelSpectrogram, and Mel-frequency cepstral coefficients (MFCC), were generated and used as inputs to a convolutional neural network for binary classification between normal and anomalous states. In addition to standard evaluation metrics such as accuracy, precision, and recall, Kullback–Leibler divergence maps were employed to analyze the separability between acoustic patterns. The results indicate that MelSpectrogram provides more stable and consistent performance across different SNR levels, especially under severe noise conditions, while MFCC proves to be more sensitive to noise variations. Statistical tests based on the Wilcoxon method confirm the superior robustness of MelSpectrogram compared to the other representations, highlighting its potential for future applications in industrial acoustic anomaly detection.