Abstract:
The calculation of a resistive sensor is considered as the load of an unstable communication line using a neural network. In the corresponding approximation or regression problem, the feedforward neural network is trained using training data and additional control data. Such data are calculated from a mathematical model of the communication line in the form of a resistive two-port with some type of change step (regular or irregular) of the load and line parameters. The training data is traditionally divided into training, validation and test sets. It was established that in the training epochs, the neural network reveals this internal pattern (inherent in the used step of change) in these three sets. Therefore, when training the network and then applying the additional control data, small errors are obtained. But for the additional control data with a different type of step, the errors appear. The use of mixed training data by combining data with diverse type of change step eliminates said internal regularity and the neural network shows the capability to generalization and small errors by presented numerical experiments. To quantify the quality of the trained network, a special index is introduced as a repeatability of the specified relative error in percent for multiple retraining.