Objective The fiber Bragg grating sensing system, with its advantages of electromagnetic interference resistance, lightweight, and high-temperature resistance, has been widely used in structural health monitoring, aerospace, biomedical, and marine engineering fields. However, due to external factors, FBG reflected spectrum data often suffer from Gaussian noise, bad points, and baseline interference, leading to significant errors in central wavelength calculations. Although data filtering on the host computer software has good noise reduction effects, it is difficult to meet the real-time demodulation requirements in industrial applications. Moreover, the commonly used algorithms on the lower computer are limited by the measurement range, which cannot be applied to spectrum filtering of multiple abnormal reflections, and have low signal-to-noise ratio and poor key information extraction capabilities, as well as poor robustness.

Methods To address the limitations of lower machine filtering algorithms, a novel FPGA-accelerated processing method for FBG reflection spectra, grounded in sliding correlation filtering, is proposed. The characteristics and origins of FBG abnormal reflection spectra are scrutinized, and the standard deviation of the second derivative of the Gaussian function is meticulously determined, predicated on the half-width of the FBG peak. The filter template length is calculated using sensor sampling compensation, and the negative values of the function are employed as the filter template weights. The spectrum data is correlated with the filter template through a sliding data window. A designed sliding correlation filtering processing module is crafted, and a devised circular queue strategy is employed to design storage memory. This crafted algorithm is implemented on the FPGA.

Results and Discussions To assess the efficacy of the sliding correlation filtering algorithm in denoising various FBG abnormal reflection spectra, the algorithm was applied to process simulated signals containing varying levels of Gaussian noise, bad points, and baselines. The results were juxtaposed against those obtained from four other commonly employed lower machine filtering algorithms. The findings revealed that the sliding correlation filtering exhibited commendable key data extraction capabilities and robustness. Its suppression effect on Gaussian noise and baselines was markedly superior to that of other algorithms (Fig.4-5, Fig.8-9), with a maximum SNR improvement of 28.23 dB relative to other algorithms. The suppression effect of sliding correlation filtering on aberrant points was second only to that of median filtering and median averaging filtering (Fig.6-7), with a maximum SNR difference of no more than 3.14 dB. The mean deviation and standard deviation of the calculated central wavelength are the smallest, measuring 1.712 pm and 2.996 pm respectively. An experimental setup was established to collect the original FBG reflection spectra. The experiment (Fig. 14) underscored the algorithm's adeptness in rectifying FBG reflection spectra containing various abnormal conditions, while maintaining a low FPGA resource utilization rate (Tab.2). Remarkably, processing one spectral data point took a mere 5.09 μs.

Conclusions This paper introduces a novel approach for processing FBG reflected spectra using FPGA, leveraging the sliding correlation filtering technique. The method ingeniously employs the negative of the second derivative of the Gaussian function as the weight of the filtering template, and meticulously performs correlation calculations between the reflected spectrum data and the filtering template through a sliding data window. The accelerated design of this algorithm is meticulously implemented on the FPGA platform. Experimental results demonstrate the efficacy of the proposed method in effectively correcting various abnormal FBG reflection spectra while preserving their fundamental data characteristics. It achieves a maximum SNR improvement of up to 28.23 dB compared to alternative algorithms, with the highest precision observed in central wavelength demodulation. The mean deviation and standard deviation of the central wavelength are reported as 1.712 pm and 2.996 pm, respectively. Additionally, the processing time for one data point is only 5.09 μs to process one data point. This research holds significant implications for engineering applications.