Name: WILLIAN PACHECO SILVA
Publication date: 23/06/2026
Examining board:
Name |
Role |
|---|---|
| CLAUNIR PAVAN | Examinador Externo |
| HELDER ROBERTO DE OLIVEIRA ROCHA | Presidente |
| MARIA JOSE PONTES | Coorientador |
| MARIANA LYRA SILVEIRA | Examinador Interno |
Summary: This dissertation develops and evaluates a liquid-level estimation system in water-oil biphasic environments, with the objective of performing the sensor fusion of FBG optical transducers with dynamic deep learning architectures. The methodology combines amplitude and wavelength measurements from pressure- and temperature-sensitive sensors to mitigate cross-sensitivity and temperature interference in liquid-level measurement. This strategy makes it possible to map the operational limit of FBG technology based on the interaction between the physical sensor configuration and the computational architecture. In an experimental dataset with approximately 8,300 samples, the MLP, CNN, and LNN neural network architectures were applied, and their results were compared after optimization with Optuna. Next, the obtained results were subjected to robustness tests under Gaussian noise and colored 1/f noise at different signal-to-noise ratios (SNR). Under controlled laboratory conditions, the LNN showed greater robustness than the static architectures in degraded scenarios, with lower noise growth, observed from the RMSE value of the obtained results, and greater R² stability; even so, the very high R² values in the noise-free scenario must be read with caution, as they reflect a favorable experimental condition. In parallel, the metaheuristic optimization algorithm MOGWO was applied, which made it possible to identify Pareto-optimal configurations for the geometric parameters of the sensor system, such as thickness, diameter, and positioning of the sensors on the fiber, that is, solutions in which any improvement in one parameter implies loss of performance in another. These results should be interpreted as a design guideline, and not as direct experimental validation of the optimized physical configuration of the sensors. The results indicate that the integration of continuous-time neural networks, sensor fusion, and metaheuristic optimization is promising for fluid monitoring in dynamic and noisy environments.
