Architecture design and implementation for sensing equipment faults from ultra-weak acoustic emission signals

In acoustic emission (AE) detection, the weakness of the acoustic source signal, the interference from background noise, and the attenuation during signal propagation result in the sensor-received signal being completely submerged by noise, severely impacting downstream fault identification and anomaly analysis. This study focuses on the fault detection scenario of AE signals characterized by a low signal-to-noise ratio (SNR). It uses low-loss noise reduction and Mel-frequency spectrum transformation for preprocessing, then extracts features with an optimized stacked auto-encoder combined with CNN (OSAE-CNN) innovatively. These features are input into an SVM for fault classification. The proposed method significantly improves fault identification accuracy to 91.38% for signals with an SNR of −20 dB, a 30% increase over the previous method. The research findings can provide technical support for the fault monitoring and safe operation of electromechanical equipment and also offer empirical references for ultra-weak signal processing in various domains.