A multicore fiber platform for distributed temperature sensing enhanced by machine learning algorithms

Machine learning algorithms have attracted much interest for their efficiency in processing signals measured by photonic sensors. In this study, we propose a multicore fiber platform for distributed temperature sensors enhanced by machine learning algorithms. Our experimental setup involves densely inscribed FBGs in the central core of a multicore fiber, whereas sparsely located FBGs in the peripheral cores serve as reference temperature sensors. The goal of the machine learning algorithm is to predict the positions of the individual FBG reflectance peaks from the raw reflectance spectrum of the dense FBG array. We evaluated the performance of Long Short-Term Memory (LSTM) neural network and CatBoost algorithm for measuring the temperature distribution. We have shown that both algorithms maintain high precision in predicting the temperature distribution even in cases where different reflectance peaks overlap. Our findings highlight the significant impact of temperature–time dynamics during the training process, which can greatly increase accuracy. In our study, the CatBoost algorithm outperformed the LSTM model when a variety of temporal dynamics of temperature change were used in the training dataset. The LSTM model demonstrated greater generalization in learning sensor responses, performing better on an unseen dataset with pronounced seasonality. Our results demonstrate the potential to enhance the wavelength-division multiplexing capabilities of distributed FBG sensors, even with a limited spectral bandwidth of the optical interrogator, using machine learning algorithms. This can improve spatial resolution and extend the sensing range of distributed FBG sensors.