Auto-calibração e linearização de sensores utilizando técnicas deinteligência computacional

Abstract

In this work, a general system for self-calibration and linearization of sensors is proposed, involving hardware and software solutions. The hardware is responsible for the data acquisition and signal conditioning. The software solutions employs radial basis function neural networks for the linearization of the input-output relationship of a temperature sensor (thermocouple), trained with multiobjective least square method. By varying the width of the radial basis function, di erent Pareto sets are obtained. A decision-making strategy over the nondominated set is proposed based on linear regression, in order to choose the network with best structure for the problem. The advantage of multiobjective learning in the context of sensor linearization is providing an adequate network for the problem that also presents low structural complexity, reducing the hardware implementation cost. Four experiments are presented for the validaton of the multiobjective learning algorithm. The three rst experiments involve temperature sensors in di erents ranges of operation. The fourth experiment involves a capacitive balance, in which the voltage is related to the variation of the capacitance. The results show that the proposed methodology is suitable for the linearization problem, allowing the selection of a neural network structure with a low implementation cost.