Event Recognition in Distributed Optical Fiber Sensing Systems Using a Fourier-Enhanced Deep Learning Framework

Distributed optical fiber sensing (DOFS) systems have gained significant attention for their ability to monitor and detect various events through vibration signals. However, real-world environments are often complex and noisy, which poses significant challenges to accurate event recognition. In this article, we propose a novel deep learning framework to address these issues by integrating a Fourier transform-based time–frequency adaptive denoising (TFAD) module and a multiscale feature extraction (MSFE) network. The TFAD module transforms vibration signals from the time domain to the frequency domain, leveraging the powerful learning capabilities of deep learning to distinguish between noise components and the relevant vibration signal components. This allows for the filtering of frequency components that interfere with event recognition. Additionally, the time-series reconstructor is used to rebuild any missing information from the filtered signal, thereby improving the signal quality. The MSFE module employs fast Fourier convolution (FFC) with a global receptive field, combining it with standard convolution and incorporating frequency attention (FA) to enable lightweight and efficient extraction as well as fusion of both global and local features. Extensive experiments are conducted on a private distributed fiber sensing dataset and several public datasets. Results show that the proposed method achieves state-of-the-art performance while maintaining high efficiency, making it well-suited for edge deployment in real-world scenarios.