Sagnac Loop Based Sensing System for Intrusion Localization Using Machine Learning

Among all optical sensing techniques, the distributed Sagnac loop (SI) sensor has the advantage of being simple to implement with low cost. Most of the proposed techniques for using SI exploit the frequency null method for event localization. However, such a technique suffers from the low spectrum signal power, complicating event localization under environmental noise. In this work, event localization using time-domain instead of frequency null signals is achieved using machine learning (ML), which is increasingly being exploited in many science fields, including sensing applications. First, a training dataset that includes 200 events is generated over a 50 km effective sensing fiber. These time-domain signals are considered as features for training the ML algorithm. Then, the random forest (RF) ML algorithm is used to develop a model for event location prediction. The results show the capability of ML in predicting the event's location with 55 m mean absolute error (MAE). Further, the percentage of test realizations with prediction error > 200 m is 0.7%. The sensing signal bandwidth is investigated, showing better performance results for sensing signals of larger bandwidths. Finally, the proposed model is validated experimentally. The results showed good accuracy with MAE < 100 m.