Machine learning-based methods for temperature estimation of fiber Bragg grating sensors

Fiber Bragg Grating sensors are a technology useful for measuring strain and temperature in harsh or tight environments, or places with no access to power supplies, where electronic sensors are difficult to use. Their robustness to electromagnetic interference makes them attractive for monitoring high voltage direct current transmission systems, which helps to stabilise the ”global” power grid. Data processing of these sensors deals with converting shifts in the reflected spectrum into changes in the environment. Analysing the state-of-the-art data processing algorithms, it was found that they have issues getting accurate estimates for non-symmetric spectra. This occurs when the spectrum is sampled with low resolution, as is the case for spectrometers. Another concern is that these algorithms process each sensor sequentially, which adds up to a long processing time for sensor arrays with hundreds of sensors. This thesis addresses these problems by introducing machine learning into the data processing. This has potential to solve the non-symmetric spectra issue because the spectrum shape should be unimportant for consistent responses.
This allows for better utilization of spectrometers, reducing the sensing setup’s hardware complexity and cost. Furthermore, the simultaneous processing of several sensors will be examined to reduce the processing time for large arrays. Two machine learning algorithms, the Gaussian process regression and the backpropagation neural network, are explored with simulated data to judge whether machine learning algorithms are useful. The performance is compared with the conventional algorithms, centroid, Gaussian fit, and cross-correlation. Then the results are validated using experimental data. The data is collected from three different setups to show the models’ versatility. Finally, experimental data is estimated using a model trained with simulated data to reduce the number of experimental measurements and increase the unbiased data used to train the models. The results demonstrate that it is possible to estimate the temperature of a single sensor based on its spectrum with an accuracy of around 1 ◦C, which was better
than the conventional algorithms. Furthermore, the temperatures for 19 sensors were simultaneously estimated with an accuracy of 1.5 ◦C and a processing time of 1.74 ms/sensor for spectrometer data, creating the basis for a real-time viable method.