LOCALLY WEIGHTED REGRESSION IN SENSOR CALIBRATION TECHNIQUES
Date
2003-11
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
ORCID
Type
Degree Level
Masters
Abstract
Recently, the technology of sensor arrays combined with a neural network system applied to multi-component analysis has been used in many industrial areas. For some difficult or inaccessible environments, such as the agricultural environment, nuclear reactors, underwater, and space environments, it is desirable that sensors are calibrated online for more reliable measurements.
In this thesis, research into a new technique of determining the individual component concentration in a mixture is described. Locally Weighted Regression (LWR), a memory-based statistical learning algorithm, combined with sensor fusion and a simulated, partially selective sensor array were used to identify individual concentrations in a mixture. The performance of the LWR technique was compared with the performance of the neural network technique. This thesis also presents an online sensor calibration scheme for an un-calibrated replacement sensor using the LWR technique.
The validity of the proposed scheme was evaluated through computer simulations and compared with artificial neural network results. The results show that the LWR technique is able to accurately determine the individual component concentrations in a mixture. Compared with the neural network technique, the LWR algorithm is more flexible and time saving for real-time measurements. This research has also shown that the calibration of a replacement sensor could be accomplished
online using the outputs of the other sensors in the sensor array as a reference for the calibration process.
Description
Keywords
Citation
Degree
Master of Science (M.Sc.)
Department
Electrical Engineering