Distributed fiber-optic sensing technology is crucial for monitoring various key parameters such as strain, temperature fluctuations, and vibrations. Among these technologies, Brillouin optical time-domain reflectometry (BOTDR) systems have gained prominence in fields like civil engineering, transportation, electric power, and oil and gas pipelines because of their distinct advantages. These advantages include the use of single-ended fiber-optic cables, ease of deployment, operational efficiency, and real-time continuous measurements, making them highly suitable for long-term monitoring in diverse environments. However, in practical applications, BOTDR systems often encounter various noise sources, including environmental and operational interference, which can severely compromise the accuracy of the demodulated Brillouin frequency shift (BFS) measurements. Because the system relies on detecting and demodulating weak, spontaneous Brillouin-scattering signals propagating along the sensing fiber, it is inherently susceptible to background noise interference. Therefore, effective noise mitigation and signal-to-noise ratio (SNR) enhancement are essential for improving BOTDR system performance and reliability in practical applications.

We propose a denoising algorithm that inverse free sparse Bayesian learning (IFSBL) with K-singular value decomposition (KSVD) (IFSBL-KSVD). The proposed algorithm effectively combines the adaptive sparse representation capabilities of IFSBL with the efficient iterative optimization and reconstruction advantages of KSVD. First, the IFSBL utilizes a Bayesian framework to adaptively select sparse components from the spectral-demodulated signals in a BOTDR system. This approach enables the algorithm to effectively retain the primary signal features associated with the BFS while simultaneously suppressing background noise, thereby accurately isolating the BFS signals from noise interference. Subsequently, KSVD optimizes the dictionary atoms through iterative updates based on the sparse representation provided by IFSBL. This iterative process enhances the dictionary’s ability to accurately capture and represent the characteristic signal frequencies, thus enhancing the overall noise suppression effect. By leveraging the dictionary-updating mechanism inherent in sparse coding, the IFSBL-KSVD algorithm achieves high noise-suppression capabilities while maintaining high efficiency and accuracy in BFS extraction from the BOTDR system. Thus, the proposed algorithm significantly enhances the signal quality and accuracy of frequency shift extraction, leading to better overall performance in real-world applications.

Temperature detection experiments are conducted using engineering-grade temperature-sensing fiber-optic cables, and the experimental data are analyzed and processed using the proposed algorithm. The results show that the IFSBL-KSVD algorithm significantly reduces random noise in the three-dimensional Brillouin gain spectrum (BGS) and its top view, improving data smoothness and making heating segments more distinguishable (Fig. 5). After applying the proposed algorithm, the fluctuations in individual BGS curves are notably suppressed, and their smoothness is greatly enhanced. After noise reduction, the central frequency of the BGS curve is determined more accurately, thereby aiding in the precise quantification of frequency shift variations by the BOTDR system (Fig. 6). The proposed algorithm achieves substantial noise reduction, with the SNR improving to 38.46 dB, which is 6.90 dB and 4.21 dB higher than the SNR of the original data and the data provided by the discrete cosine transform (DCT) algorithm, respectively. In the non-temperature-variable region, the fluctuation range decreases to 1.35 MHz, while that in the temperature-variable region decreases to 0.88 MHz; these reductions correspond to an improvement in temperature measurement accuracy to ±0.83 ℃ and ±0.40 ℃, respectively, with an additional runtime increase of only 26.38 s (Fig. 7). By adjusting the pulse width (50‒100 ns), the proposed algorithm enhances the SNR by 6.32 dB‒6.90 dB (Fig. 8). Furthermore, the algorithm effectively reduces the BFS standard deviation in the temperature-variable region to below 2 MHz across all pulse widths. This improvement in SNR and measurement accuracy is achieved without compromising the spatial resolution, highlighting the algorithm’s robustness and flexibility (Table 1).

An innovative denoising algorithm based on IFSBL and dictionary learning (i.e., KSVD) is proposed to improve the detection performance of BOTDR systems in engineering applications. The proposed algorithm exploits the adaptive sparse representation capability of IFSBL and the efficient iterative optimization feature of KSVD in dictionary learning, thereby enabling effective noise reduction for BOTDR spectral signals. Experimental results from temperature change detection using engineering-grade temperature-sensing fiber-optic cables demonstrate that the proposed algorithm significantly improves the SNR and increases the BFS detection accuracy of the BOTDR system by 2.16 MHz. Despite its improved performance, the proposed algorithm requires only an additional runtime of 26.38 s, demonstrating its high computational efficiency. These findings confirm that the proposed algorithm excels in both noise reduction and computational efficiency, significantly enhancing the overall performance of low-cost BOTDR systems. The algorithm’s high efficiency and robustness strongly support the widespread adoption of BODTR technology in complex and challenging engineering environments.