Impact of noise on machine learning models for simultaneously predicting water and oil levels using optical fiber sensors

We address the problem of simultaneously estimating the water-oil interface and total levels in three-phase separator tanks by applying three machine learning methods: Multilayer Perceptron (MLP), Kolmogorov-Arnold Networks (KAN), and Random Forest (RF). Data was collected from Fiber Bragg Grating-based optical sensors and processed to suppress outliers using the Local Outlier Factor algorithm. Hyperparameters for each model were optimized using Grid Search, and their performance was compared. The trained models were also exposed to scenarios with different levels of noise, and performance was evaluated. The results suggest that KAN performs effectively in predicting liquid levels, achieving a Root Mean Square Error of less than 3 mm and a Mean Absolute Percentage Error below 0.3% in scenarios without noise. Both MLP and KAN exhibit similar accuracy when the noise level is up to 1%. However, the MLP model outperforms KAN in higher noisy scenarios. On the other hand, the RF model shows the least effectiveness overall in noisy environments, though it does maintain a relatively stable maximum error across different noise levels. Therefore, we demonstrate that KAN has advantages in low-noise scenarios compared to conventional MLP and RF models. Conversely, MLP is more effective under higher noise conditions. These findings can aid in the research and development of monitoring systems.