Machine learning for sensing with a multimode exposed core fiber specklegram sensor.

Area of Science:

• Photonics and Optical Sensing
• Artificial Intelligence in Scientific Instrumentation
• Fiber Optic Sensor Technology

Background:

• Traditional fiber specklegram sensors (FSSs) rely on statistical analysis, which can be susceptible to noise and limited in dynamic range.
• Existing methods for analyzing fiber specklegrams face challenges in robustness and precision for various sensing applications.

Purpose of the Study:

• To demonstrate the efficacy of deep learning (DL) techniques in improving the analysis of fiber specklegrams for enhanced sensing capabilities.
• To compare the performance of deep neural networks (DNNs) against traditional correlation methods in fiber optic sensing.

Main Methods:

• Utilized two distinct deep neural networks: a convolutional neural network (CNN) and a multi-layer perceptron (MLP).
• Employed an exposed-core multimode fiber for data acquisition in sensing experiments.
• Trained DNNs to be resilient against noise, including random specklegram translations.

Main Results:

• Deep learning models, specifically CNNs and MLPs, demonstrated superior performance in analyzing fiber specklegrams compared to traditional correlation techniques.
• Successful application of DL for accurate air temperature and water immersion length measurements using FSSs.
• Validated the capability of DNNs to generalize and perform robustly even in the presence of noise.

Conclusions:

• Deep learning offers a significant advancement over traditional statistical methods for fiber specklegram sensor data analysis.
• The proposed DNN-based approach enhances the accuracy, robustness, and dynamic range of FSSs for practical sensing applications.
• This study highlights the potential of AI to revolutionize fiber optic sensing technologies.