AE-IRMamba: Low-Complexity Inverted Residual Mamba for Identification of Piezoelectric Ceramic and Optical Fiber Acoustic Emission Sensors Signals

The deep learning networks currently used for the identification and classification of acoustic emission (AE) signals collected by sensors are highly complex, which makes networks unsuitable for portable devices due to resource limitations. Therefore, we propose an ultralow-complexity network architecture named AE-IRMamba for AE signals collected by optical fiber AE sensors and traditional ceramic AE sensors. We are the first to introduce the state-space models (SSMs) into the field of AE signal identification and classification. Specifically, from the inverted residual network to extract features, adjustable parallel mamba was used to establish long-range spatial dependencies with linear computational complexity, we have achieved a holistic observation of global features spanning temporal and spatial. The proposed model has only 25.8k parameters, and the classification accuracy on two different sensor datasets is 99.9% and 93.94%, respectively. Compared to commonly used network architectures, the model complexity is reduced by 99.9%, while still maintaining state-of-the-art recognition accuracy. The proposed network architecture offers a viable solution for the use of resource-constrained portable devices.