Separation of distorted overlapping spectra in FBG sensor networks using self-supervised contrastive learning

Fiber Bragg grating (FBG) sensor networks encounter spectral overlap and distortion under non-uniform physical fields (e.g., residual strain monitoring in composite structures), limiting multiplexing capability and sensing accuracy. This study proposes a contrastive spectrum separation model (CSSM) as a self-supervised learning framework for separating distorted overlapping spectra in FBG sensor networks. The framework employs a dual-encoder architecture with parallel convolutional neural networks, enabling direct feature extraction from distorted overlapping spectra without extensive labeled training data. CSSM demonstrates superior robustness under varying degrees of spectral overlap and noise conditions (e.g., 54.5% SNR improvement at 15 dB). Simulation results demonstrate CSSM's capability to achieve wavelength detection accuracy of 1.6388 pm with a structural similarity index of 0.9074 for separated distorted spectra. Experimental results with strain gradients reaching -650 µε/mm verify the model's practical effectiveness, reducing strain measurement error from ±37.2 µε to approximately 1.35 µε. The framework enhances FBG networks' multiplexing capability and accuracy in complex operational scenarios, advancing structural health monitoring applications.