Over the past decades, the national demand for structural health monitoring of large infrastructures such as bridges and oil and gas pipelines has gradually increased. Based on the scattering effect in optical fibers, researchers have proposed a distributed optical fiber sensing (DOFS) system. It is not only very sensitive to external parameters such as temperature change, strain, and vibration, but also has the advantages of long-distance multipoint monitoring, low cost, corrosion resistance, radiation resistance, and large bandwidth, which makes it an important technological tool for structural health monitoring of large-scale infrastructures. DOFS is mainly categorized based on scattering mechanisms, which are Rayleigh scattering, Brillouin scattering, and Raman scattering. Compared with other DOFS, DOFS based on Brillouin scattering has high temperature and strain sensitivity, thus providing accurate measurements. In addition, it is also capable of long-distance distributed monitoring of external strains and temperature changes with high spatial resolution, which has attracted the attention of a large number of researchers and has been widely used.

However, with the increase in sensing distance in DOFS, the decrease in signal-to-noise ratio (SNR) will lead to an increase in measurement uncertainty. In addition, massive data will be generated in the process of long-distance continuous measurement, and the required measurement time will increase correspondingly. How to process massive data intelligently, quickly, and accurately to further improve system performance and obtain more accurate physical parameters is the biggest problem facing the development of the system. Currently, the development of system hardware technology is particularly insufficient in the face of massive data processing, which creates an opportunity for advanced signal processing and analysis using digital signal processing (DSP) technology, which can effectively obtain effective information from the massive data generated by the system. In the past few years, the development of powerful computer processors has laid the foundation for the development of advanced DSP technologies, and recent advances in big data and cloud technologies have provided tools for efficient storage and massive data processing. With the development, DSP technology has the advantages of smaller back-end processing time overhead and no increase in system hardware complexity.

We review the DSP techniques used for data processing in Brillouin-DOFS in recent years and focus on the applications of image and video denoising technology and machine learning information extraction and recognition technology in it, so as to provide a reference for future research of DSP technology in Brillouin-DOFS.

The multi-dimensional (time, frequency, and position) domain of Brillouin signals contains redundancy and structural similarity. However, none of the earlier denoising methods have utilized the feature. Thus, the researchers have introduced the image-video denoising technique to reduce the noise of the sensing signals. At first, some traditional image and video denoising algorithms are summarized, and the principles of the algorithms, as well as the performance of denoising effects are generally introduced. It also shows that the optimization of algorithm parameters and the transformation of 3D BGS can enhance the denoising performance. However, the traditional algorithms still affect the spatial resolution and measurement reliability. With the research and development of machine learning, neural networks have also been used for denoising Brillouin signals by the powerful nonlinear fitting ability. Neural networks have many architectures such as artificial neural networks (ANNs), convolutional neural networks (CNNs), and generative adversarial networks (GANs), and all of them are capable of fast and high-fidelity denoising.

Machine learning has a strong ability to fit complex nonlinear functions, which is very suitable for solving regression and classification problems. In addition, the machine learning algorithm is extremely short in time, showing its potential in information extraction. First, the application of traditional machine learning algorithms to the direct extraction of temperature or frequency is presented. These algorithms demonstrate much higher extraction speed than traditional fitting algorithms and have stronger robustness. With the increase in computing power, the neural network can be well-trained by simulating a large number of Brillouin gain spectra in different situations. In addition, by constructing the dataset in different cases, the corresponding purpose can be realized, such as solving the frequency extraction error caused by non-local effects. Finally, some studies on the performance evaluation of neural network model extraction and the integration of neural networks with other techniques are also presented.

DSP technology can process massive data intelligently, quickly, and accurately, so as to further improve the performance of the system. Firstly, the concept of image and video denoising makes use of the repeated structures of information in the multi-dimensional domain of Brillouin signal. Then, a variety of traditional denoising algorithms and machine learning methods have been applied. Secondly, since traditional fitting methods are time-consuming, machine learning techniques are also introduced into Brillouin-DOFS. It can directly learn the nonlinear mapping between input and output so that information such as frequency, temperature, or strain can be accurately and quickly extracted from BGS. In the future, in addition to developing more advanced techniques to achieve longer, more accurate, and faster sensing systems, how to better evaluate and interpret machine learning algorithms is also the focus of research. It is believed that Brillouin-DOFS based on DSP technology will play an increasingly important role in infrastructure, aerospace, energy transportation, and other fields.