Two steps hybrid calibration algorithm of support vector regression and K-nearest neighbors

Abstract

Errors in the measurements of pipeline nondestructive tools may lead to faulty decisions causing economic and environmental loss, or system failures. Calibration models are an effective tool that is used to enhance the quality of corrosion measurements affected by inline inspection tools sizing accuracy. Parametric calibration models are limited to datasets with Gaussian behavior. On the other hand, non-parametric calibration models can overcome the normality limitation, however, they provide only a local or general estimation. This paper presents a new hybrid calibration model that is based on two steps K nearest neighbor interpolation and support vector regression. The suggested hybrid model uses both general and local estimation behaviors for the calibration process, hence resulting in a better prediction. The hybrid algorithm was evaluated using a dataset of pipeline corrosion measurements collected by a Magnetic Flux Leakage (MFL) sensor (with an error margin of ±20% of the true values), and an Ultrasonic (UT) device (with an error margin of ±4%). The suggested approach resulted in reducing the errors in MFL corrosion measurements to be only ±6.82% instead of the original ±20%.