COMPARATIVE ANALYSIS OF CLASSICAL AND QUANTUM AUTOENCODERS IN THE DETECTION OF ANOMALIES IN INDUSTRIAL ACOUSTIC SIGNALS UNDER LOW SIGNAL-TO-NOISE RATIO

Fault detection in noisy industrial environments is a critical challenge for predictive maintenance. This work investigates the use of Autoencoders for anomaly detection in acoustic signals from the MIMII dataset, focusing on the most challenging scenario: pump under -6 dB, where noise exceeds the signal. Three approaches were evaluated: Autoencoder (AE), Variational Autoencoder (VAE), and Quantum Autoencoder (QAE), all trained using only normal data and assessed through reconstruction error. Results show that AE and VAE achieve higher overall accuracy (0.91 and 0.92), but low recall for the fault class (0.30 and 0.37). In contrast, the QAE achieved a recall of 0.75 and F1-score of 0.72 for faults, with an accuracy of 0.90. These findings indicate that QAE is more effective in detecting anomalies under highly noisy conditions, showing strong potential for critical industrial applications.