A Systematic Review of Deep Learning Approaches for Fiber Bragg Grating Sensor Data Interpretation in Subsea Multi-Parameter Monitoring

Fiber Bragg Grating (FBG) sensors have recently been identified as one of the most important technologies to provide subsea monitoring owing to high sensitivity, Electromagnetic Interference (EMI) immunity and the ability to measure a number of parameters at once. Although FBG-based sensing solutions are very promising, there are numerous challenges associated with spectrum interpretation in subsea conditions. Particularly, temperature, salinity and pressure parameters are cross-sensitive, with spectral distortions (polarisation effects, birefringence, wavelength effects and noise) causing complex nonlinear dependencies that are intractable to traditional signal processing methods. Recent development in the field of deep learning demonstrates remarkable achievements in the modelling of complicated spectral patterns to do multi-parameter estimation. Various kinds of neural architecture, such as convolutional, recurrent, and attention-based neural networks are employed to learn the local and global interactions among the spectral signals. In addition, new learning methods like self-supervised learning, federated learning, and physics-aware modelling are designed to alleviate the issues such as data scarcity, domain variability, and deployment-related challenges. Nonetheless, the current deep learning approaches have several critical weaknesses, such as the lack of modelling cross-sensitivity effects, the lack of good generalisation when operating under realistic settings, poor resistance to uncertainties, and real-time and edge-deployment issues. This paper presents a review of our work on the use of deep learning to interpret FBG spectral data in detail. In particular, we derive a theoretical framework to explain the problem, create a taxonomy to analyse the state-of-the-art approaches, and find out significant gaps in research and future directions.