Design parameter optimization method for sensor device development considering recognition error correcting cost

Abstract

Machine learning has come to be widely used in the new development of sensor devices. By developing sensors, training data can be modified to be suitable for higher discrimination accuracy. Studies of training data optimization methods to improve discrimination accuracy have been proposed, but they have no consideration of the sensor development cost. In this letter, we propose a cost-conscious training data optimization method. To handle training data optimization, we take distribution characteristics of the training data as design parameters. Derive the design parameters modification ways to cancel prediction errors of machine learning model. Development cost of design parameters modification ways are evaluated by using predetermined development cost condition, and lowest cost way is selected as the design parameter optimization way. Design parameter optimization process is investigated in several condition of the sensor development cost. Design parameter optimization process demonstrates optimization results referring to each sensor development cost condition.