Estimation of the artificial neural network uncertainty used for measurand reconstruction in a sampling transducer

The use of artificial neural networks in the field of measurement is mainly because of the need to meet ever increasing demands on the speed of obtaining the measurement results and increase their accuracy. A neural network, which is one of the components of a measurement signal processing chain, represents a specific type of a transducer, and therefore there is a need to determine the uncertainty of the results obtained by the network. Determination of this uncertainty is necessary for the purpose of comparing the metrological properties of neural networks of different structures, learned using different algorithms, as well as for comparing the classical and neural data processing methods that allow obtaining sufficiently accurate measurement results. This study presents a method for uncertainty estimation which is based on the histogram of neural network output result errors obtained in the testing process. The method allows determining the width of the uncertainty interval of final results for a given level of confidence. The theoretical considerations are complemented by simulation results of the measuring chain which uses neural networks for reconstruction of the input signal of a non-linear sensor.