A general method for sensor linearization based on neural networks

We propose a general method for linearizing the response of an arbitrary sensor. The procedure consists of a simple artificial neural network that compensates the nonlinear characteristic of the sensor. The neural network (a multilayer perceptron) is trained with input-output data: the (nonlinear) output of the sensor is used as input data, and the difference between sensor outputs and the desired linear responses are the target values. In this paper, an NTC sensor is used as application example of the procedure, and several practical results are provided. As this method is particularly suitable for embedded systems based on simple, low-resolution microcontrollers, its implementation on this kind of system is studied and analyzed.