Machine learning-based long-period fiber grating demodulation: a promising tool for cost-effective in-field implementation

Abstract

Long-period fiber gratings (LPGs) are versatile optical fiber devices, and one of their uses is the manufacture of optical sensors. These sensors have gained significant attention in recent years due to their unique detection capabilities and ease of manufacturing. However, their widespread adoption has been hindered by the complexity and cost of traditional interrogation methods, which often rely on bulky and expensive equipment. This thesis presents a novel approach to LPG interrogation that leverages a sparse optical filter bank and machine learning algorithms to achieve cost-effective and reliable measurements. The proposed approach is investigated through three distinct methods, each building upon the previous one, culminating in a fully connected neural network based on self-attention. This approach enables the use of FBGs with arbitrary positions in a multi-sensor configuration, enhancing system cost-effectiveness by sensor multiplexing and incorporating multi-point sensing capabilities. The results demonstrate the effectiveness of the proposed approach by demonstrating its high accuracy, robustness to noise, and ability to generalize to a wide range of LPG sensors and FBG array configurations. The use of synthetic data for training overcomes the limitations of acquiring large datasets, allowing for the development of more robust and complex models. To assess the effectiveness of the proposed technique, a refractive index sensor was built and interrogated by the proposed system with an error less than 0.2%. This thesis should contribute significantly to the field of LPG sensor interrogation, offering a promising solution for various applications, mainly in structural health monitoring, environmental sensing, and industrial process control. The proposed approach paves the way for further advances in LPG sensing technology and its widespread adoption in the field.