Machine learning based generic demodulation framework in distributed temperature sensing

Distributed temperature sensing is a rapidly developing branch within the realm of optical fiber sensing. To establish a generic framework for temperature demodulation, it is essential to delineate the mapping relationship between the acquired signals and the distinctive features of the measurement system. Accordingly, a meticulous analysis of the interplay between the signals and the system’s features is conducted, wherein machine learning techniques are employed to capture intricate mapping relationships. To enhance the network’s capability in extracting relevant features, the loss function employed for network training incorporates additional physical information that is inherently embedded within the temperature demodulation process. To evaluate the performance and generalizability of various demodulation techniques, the temperature sensing system is employed comprising a light source with varying intensity levels and an internal thermostat. Three datasets are created by capturing scattered light intensities and reference temperatures from various locations within different signal intensity. The effectiveness of three demodulation approaches, namely fixed-parameter demodulation, direct prediction demodulation, and physical constrained prediction demodulation, is assessed and compared in terms of their generalizability across the datasets. The results obtained indicate that physical constrained prediction demodulation exhibits a significant improvement in temperature demodulation accuracy across diverse datasets, while concurrently demonstrating a robust feature extraction capability as a measurement system.