Neural Estimator of Sound Speed in Air for Ultrasonic Sensing in Manufacturing Environments

Abstract

The influence of air characteristics on the sound speed is well known, and this can affect the accuracy and performance of ultrasonic (US) sensors. The supposition of constant sound speed in air is true only for constant environment properties, like indoor environments. Even then, the speed of sound can have different constant values for different sets of environment properties, generated by their spatial or temporal variations, as in manufacturing environments. In addition, for specific manufacturing or outdoor environments, the air characteristics can have large absolute values with wide variations, which can not be neglected. In this case, the sound speed estimation is very important. In this paper, a model with uncertainties of sound speed in air as function of air properties is presented. Also, a neural estimator of sound speed is proposed using feed-forward neural networks, based on multi-variable model with uncertainties of sound speed in air. Several neural networks are analyzed using different topology, neuron characteristics, and training data sets. The selection was made based on the relative error and the mean square error of the neural network output, compared with the sound speed model. The predictor is working well even if both the training and testing sets are affected by important noise, with estimation in the noise range.