Spectral Demodulation of Mixed-Linewidth FBG Sensor Networks Using Cloud-Based Deep Learning for Land Monitoring

Abstract

Fiber Bragg grating (FBG) sensing systems face significant challenges in resolving overlapping spectral signatures when multiple sensors operate within limited wavelength ranges, severely limiting sensor density and network scalability. This study introduces a novel Transformer-based neural network architecture that effectively resolves spectral overlap in both uniform and mixed-linewidth FBG sensor arrays, operating under bidirectional drift. The system uniquely combines dual-linewidth configurations with reflection and transmission mode fusion to enhance demodulation accuracy and sensing capacity. By integrating cloud computing, the model enables scalable deployment and near-real-time inference even in large-scale monitoring environments. The proposed approach supports self-healing functionality through dynamic switching between spectral modes during fiber breaks and enhances resilience against spectral congestion. Comprehensive evaluation across twelve drift scenarios demonstrates exceptional demodulation performance under severe spectral overlap conditions that challenge conventional peak-finding algorithms. This breakthrough establishes a new paradigm for high-density, distributed FBG sensing networks applicable to land monitoring, soil stability assessment, groundwater detection, maritime surveillance, and smart agriculture.

Keywords: agriculture; deep learning; fiber Bragg grating; land monitoring; maritime sensing; optical sensors; sensor networks; spectral analysis; urban infrastructure.

Figure 1

Conceptual diagram of the proposed dual-linewidth FBG sensing network.

Figure 2

Reflection spectra of three FBG sets with different linewidths: (a) narrow-band gratings with 0.2 nm FWHM; (b) medium-band gratings with 0.5 nm FWHM; (c) wide-band gratings with 0.8 nm FWHM.

Figure 3

Spectral responses of the uniform-linewidth three-FBG array (Case 1) under symmetrical bidirectional drift applied to FBG01 and FBG03, with FBG02 fixed: (a) reflection spectrum at ±1 pm drift; (b) transmission spectrum at ±1 pm drift; (c) reflection spectrum at ±5 pm drift; (d) transmission spectrum at ±5 pm drift; (e) reflection spectrum at ±10 pm drift; (f) transmission spectrum at ±10 pm drift. Each trace represents a specific incremental drift step, and the color gradient visualizes the stepwise spectral changes from deep purple to bright yellow. A zoomed-in view is included in each subfigure to better visualize the overall stepwise spectral shifts under each drift condition.

Figure 4

Spectral responses of the mixed-linewidth six-FBG array (Case 2) under symmetrical bidirectional drift applied to FBG11 and FBG16, with FBG12–FBG15 fixed: (a) reflection spectrum at ±1 pm drift; (b) transmission spectrum at ±1 pm drift; (c) reflection spectrum at ±5 pm drift; (d) transmission spectrum at ±5 pm drift; (e) reflection spectrum at ±10 pm drift; (f) transmission spectrum at ±10 pm drift. Each trace corresponds to an incremental drift step, with the color gradient showing the progression from deep purple to bright yellow. A zoomed-in view is included in each subfigure to better visualize the overall stepwise spectral shifts under each drift condition.

Figure 5

Performance of the Transformer demodulator on the six drift scenarios of Case 2 (mixed-linewidth array): (a) training and validation loss versus epoch; (b) final demodulation accuracy for the same six scenarios from Case 2.

Figure 6

Comprehensive error analysis of the Transformer demodulator across all twelve drift scenarios (Cases 1 & 2): (a) Mean squared error (MSE); (b) root-mean-squared error (RMSE); (c) mean absolute error (MAE); (d) mean percentage error (MPE).

Figure 7

Computational cost and dataset size for Transformer training and inference across the twelve drift scenarios (Cases 1 and 2): (a) GPU training time per scenario; (b) inference (testing) time per scenario; (c) mean training time aggregated by case (uniform versus mixed linewidth array); (d) number of synthetic training spectra generated for each scenario.

Figure 8

Transformer-predicted peak wavelengths for the strained gratings in Case 2 (mixed-linewidth array) under symmetrical bidirectional drift applied to FBG11 and FBG16: (a) reflection spectrum at ±1 pm drift with predicted peaks of FBG11 (+Δλ) and FBG16 (−Δλ); (b) transmission spectrum at ±1 pm drift showing corresponding notch detections; (c) reflection spectrum at ±5 pm drift with peak localization amid partial overlap; (d) transmission spectrum at ±5 pm drift confirming dual-peak detection; (e) reflection spectrum at ±10 pm drift with isolated peaks from overlapping gratings; (f) transmission spectrum at ±10 pm drift showing successful peak identification within the broad composite pass-band.