Abstract
To address environmental sound event recognition in scenarios subject to electromagnetic interference, this paper proposes a reliable classification system for low signal-to-noise ratio (SNR) conditions. The system integrates a fiber-optic extrinsic Fabry–Perot interferometer (EFPI) acoustic sensor equipped with a graphene-based diaphragm and an improved residual neural network. Environmental sound signals are collected by the sensor and obtained via a phase demodulation algorithm. Next, an SE-ResNet18 model is constructed by introducing a squeeze-and-excitation network(SE-Net) into a standard ResNet structure. Multiple data augmentation strategies, including SpecAugment and time-domain segmentation, are adopted to alleviate the limited training data issue. Low-SNR datasets ( − 5 to 15 dB) are synthesized using the ESC-10 and NoiseX-92 datasets for model training and evaluation. Experimental results show the proposed system achieves an average classification accuracy of 82.50% under SNR conditions ranging from−5 dBto15dB. This reflects an improvement of 2.50 and 6.25 percentage points compared to models without the SE-Net module and data augmentation, respectively. The developed system demonstrates superior performance in low-SNR environments, providing an effective technical approach for sound event monitoring in electromagnetically sensitive scenarios with considerable practical potential.