Artificial intelligence for optical fiber sensors

• Main Author: Li, Baocheng
• Other Authors: Wong Liang Jie
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/164277
• Institution: Nanyang Technological University

Optical fiber sensor technology has been studied intensively in the last few decades, it has great applications in structural health monitoring, civil engineering, biochemistry, and medical devices. Optical fiber sensor can be classified intro different categories based on number of sensing point namely single point sensing, quasi-distributed sensing, and fully distributed sensing. In this thesis, we focus on single point sensing and AI techniques for quasi-distributed sensing. We investigated a single-point optical fiber sensor for temperature sensing using a twin-core photonic crystal fiber (TCPCF). The proposed TCPCF has a hybrid light-guiding mechanism where one TCPCF core guides light in the total internal reflection mechanism and the other guides light in the photonic bandgap mechanism. The liquid crystal 5 CB is selectively infiltrated into one of the inner rings of the TCPCF air holes. The 5CB has a higher refractive index than the silica core; therefore, the photonic bandgap mechanism guides the light. The theoretical analysis showed that the photonic bandgap core modes, total internal reflection core mode, and 5CB channel modes could simultaneously propagate and interfere with each other in the proposed TCPCF. The experiment results showed a good agreement with the theoretical studies. The proposed TCPCF was tested experimentally for temperature response and achieved high-temperature sensitivity of 4.91 nm/°C for the nematic phase of 5CB, and −3.68 nm/°C for the isotropic phase of 5CB, respectively. The proposed TCPCF with a hybrid light-guiding mechanism could have significant applications in optical fiber communication, optical fiber sensing, and optical laser systems. We demonstrated a novel way of fabricating a long period grating fiber for refractive index sensing. The multicore fiber (MCF) is mechanically twisted with a CO2 laser system to form a helical long period grating (HLPG). The HLPG has unique advantages compared to conventional long period grating, such as high mechanical strength, low fabrication cost, and ease of fabricating. In addition, the proposed HLPG made from MCF has a refractive index response up to 1077 nm/RIU and an extensive free spectral range. Challenges arise when the signal spectra from the quasi-distributed fiber Bragg grating sensor network are highly overlapped. The overlapped spectra are hard to separate, resulting in a high signal demodulation error. The conventional peak detection method cannot be applied to the overlapped signal. The non-machine learning algorithms like particle swarm optimizers, genetic algorithms, and differential evolution algorithms are proposed to deal with the overlapped issue. However, these methods suffered from a long computation time which could not realize real-time monitoring of sensor networks. We proposed a multi-peak detection machine learning model based on a dilated convolutional neural network (CNN) to overcome the overlapping signal problem and enhance the demodulation speed. The proposed dilated CNN model has enhanced the sensor network multiplexing capability and achieved a low root-mean-square (RMS) error of less than 0.05 pm in overlapped signal demodulation for two-FBGs pairs. Furthermore, the demodulation time for each pair of FBGs is only 15 ms. The proposed dilated CNN model is robust again noise, and signal demodulation RMS error is less than 0.47 pm even though the signal-to-noise ratio is 15 dB. Our findings have realized a real-time detection system for the quasi-distributed FBG sensor network and proved that neural network algorithms are promising in signal demodulation problems. Besides, the signal resolution of reflected spectrum is directly related to the network's sensitivity, and the physical equipment like spectrometer often limits the resolution. We propose the dilated U-Net with residual connection model to perform signal demodulation and enhance the resolution of the overlapping spectrum from the FBG sensor network. The noisy overlapped signal with low resolution went through a signal demodulation process to separate the signal, followed by a resolution enhancement process to increase the network sensitivity. The model has demonstrated high capability to enhance the network multiplexity, demodulation efficiency, and resolution enhancement. Furthermore, the model is trained using noisy spectra with different levels of signal-to-noise ratio, and thus the proposed model is robust toward the noise. Our model allows the FBG sensor network to bypass the equipment limitations to increase its sensitivity without any modifications to existing equipment. Different degrees of resolution enhancement have been demonstrated, and the error is remained low for a 100 times of resolution enhancement.