Measurement accuracy enhancement with multi-event detection using the deep learning approach in Raman distributed temperature sensors

In this work, we present a novel deep learning framework for multi-event detection with enhanced measurement accuracy from the measured data of a Raman Optical Time Domain Reflectometer (Raman-OTDR). We demonstrate the utility of a deep learning-based approach by comparing the results from three popular neural networks, i.e. vanilla recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU). Before feeding the experimentally obtained data to the neural network, we sanitize our data through a correlation filtering operation to suppress outlier noise spikes. Based on experiments with Raman-OTDR traces consisting of single temperature event, we show that the GRU is able to provide better performance compared to RNN and LSTM models. Specifically, a bidirectional-GRU (bi-GRU) architecture is found to outperform other architectures owing to its use of data from both previous as well as later time steps. Although this feature is similar to that used recently in one dimension convolutional neural network (1D-CNN), the bi-GRU is found to be more effective in providing enhanced measurement accuracy while maintaining good spatial resolution. We also propose and demonstrate a threshold-based algorithm for accurate and fast estimation of multiple events. We demonstrate a 4x improvement in the spatial resolution compared to post-processing using conventional total variational denoising (TVD) filters, while the temperature accuracy is maintained within ± 0.5 ^oC of the set temperature.