Distributed sensing is a key requirement for the health and condition monitoring of several capital-intensive structures such as aerospace vehicles, bridges, dams, tunnels, oil/gas pipelines, and electric power distribution networks. Specifically, distributed temperature sensing is quite useful in some of the above applications for which Raman scattering in optical fibers is a well known solution. Distributed sensing can be carried out by marrying the Raman scattering with optical time domain reflectometry (OTDR), a technique popularly known as Distributed Anti-Stokes Raman Thermometry (DART). The fundamental limitation in DART is the trade-off between spatial resolution and temperature uncertainty, which in turn depends on signal to noise ratio (SNR) of the measurement. Previously, several approaches have been pursued to enhance the performance of such systems through the sensing fiber, the detection mechanism, and the post-processing techniques to improve the SNR.

In this work, we have developed a theoretical model of the DART system to study its performance metrics with respect to the individual component specifications and the post-processing technique. In particular, we have studied the use of Total Variation Denoising (TVD) to improve the SNR of the measured signal without compromising the spatial resolution. We have also explored the use of a loop scheme to address the calibration issues with the conventional approach based on the measurement of the ratio of anti-Stokes to Stokes intensity or the ratio of anti-Stokes to Rayleigh intensity to estimate the temperature map. Such a loop-based DART scheme for been demonstrated for real-time power line monitoring in overhead power transmission (OPGW/OPC) cables. Temperature experienced by the optical fiber embedded in the power cable is estimated through a simple heat transfer model and is experimentally validated using DART measurements. Further, we have explored the ability of recurrent neural network (RNN) to improve the performance of the DART system. Based on systematic studies, we find that the bi-directional Gated Recurrent Unit (bi-GRU) is best suited for achieving low temperature uncertainty (±0.5 ˚C) and 4-fold improvement in spatial resolution compared with conventional filtering. We show that the proposed approach could help to improve the performance of other distributed sensing systems as well.