Measurement uncertainty of contact and non-contact techniques on condition monitoring of complex industrial components

The paper discusses the effect of changing the sensor's configuration and the data processing procedure on the results of Neural Networks (NN) algorithms. They are used for both classification and prediction in Condition Monitoring (CM) applications of real cutting machines of industrial automated production lines. In particular, a comparison between non-contact techniques (laser displacement sensors) and contact ones (tri-axial accelerometers) is carried out. The robustness of results is studied, depending on the type of sensor and of the feature for monitoring, on the position and orientation of the sensors themselves. The physical meaning of choosing is taken into account throughout the procedure, in order to define optimized practical configurations based on the uncertainty evaluation. Experimental results show the possibility of validating a configuration of good compromise based on a mixing of different types of sensors, for the benefit of the monitoring strategy itself and of the consequent preventive and condition-based maintenance actions.