Pattern-Recognition-Based Dual-Point Fiber Temperature Sensor Using a Reliable Synthetic Database

We propose to use the nonlinearity in wavelength sweeping of a distributed feedback (DFB) diode laser to generate a reliable synthetic database implemented in a pattern recognition-based fiber sensor. First, the experimentally extracted wavelength-sweeping nonlinearity permits obtaining the time-varying tuning rate along the sweeping period. Second, this tuning rate is used to simulate a two-point interferometric sensing system under a complete set of temperature variations to generate a reliable synthetic database used to train a pattern recognition algorithm. Third, the trained algorithm is successfully implemented for correctly identifying the experimental sensing signals of a dual-point temperature sensor. In fiber sensor systems employing machine learning (ML) algorithms, a huge amount of experimental data is required to ensure accurate pattern classification, which becomes a challenging task. The methodology for generating a reliable synthetic database provides time and resource savings in the lab, without compromising the accuracy of the results. A standard DFB diode laser, wavelength tuned over a few tens of pico-meters, is used as an optical source, and a p-i-n photodetector is used as an optical detector. A description of the wavelength-nonlinearity extraction approach, a mathematical model of the interferometric fiber sensor, and experimental results confirming the effectiveness of the proposed sensing system are reported. Also, classification results using a database generated without including the nonlinearity effect are presented to highlight the importance of considering this nonlinearity.