Neural Networks as Soft Sensors: a Comparison in a Real World Application.

Physical atmosphere parameters, as temperature or humidity, can be indirectly estimated on the surface of a monument by means of soft sensors based on neural networks, if an ambient air monitoring station works in the neighborhood of the monument itself. Since the soft sensors work as virtual instruments, the accuracy of such measurements has to be analyzed and validated from statistical and metrological points of view. The paper compares different typologies of neural networks, which can be used as soft sensors in a complex real world application: a non invasive monitoring of the conservation state of old monuments. In this context, several designed connessionistic systems, based on radial basis function, Elman networks and support vector regression, are trained and tested on rich sets. An original statistical procedure, based on specific estimators and the substitution error, is applied for comparison. It is also able to characterize the performances of the soft sensors in the physical domain of interest.