MACHINE LEARNING FOR INVERSE CUBIC SENSOR CALIBRATION: A REGULARIZED BAYESIAN APPROACH

Abstract

Sensors often exhibit nonlinear responses that can be approximated by cubic polynomial transformations-balancing complexity and accuracy. The inverse problem of recovering the physical input from the sensor output is critical in medical diagnoses, control systems and environmental monitoring. The recovery is challenging due to ambiguity in reconstructing the input, error propagation and parameter estimation in the presence of noise or limited data points. The paper proposes a regularized Bayesian framework to ensure stable, unique solutions using machine learning (ML) algorithms to select the physically meaningful root with high accuracy and low computational cost. The ML algorithms achieve over 95% accuracy in root selection with XG Boost offering a 15% inference time than exhaustive root evaluation.