Optimization for a Neural Network Based on Input-Vectors Correlation and Its Application to a Truck Scale

Abstract

It is important and valuable to improve the generalization ability of a neural network (NN) when there is a lack of training samples. This paper presents an approach of optimizing a NN based on input-vectors correlation (IVCNN) to improve the NN's ability in the case of lacking samples, and then uses this method to compensate for the weighing errors of a truck scale. First, we analyze the truck scale's weighing principle and the spatial correlation of the truck scale's input signals, and then use this correlation to construct the constraint conditions and the performance index for training an NN. Finally, we give the detailed training algorithm of IVCNN. In addition, we prove the IVCNN's convergence and analyze its anti-interference performance, convergent speed, and computational complexity. The experimental results demonstrate the effectiveness of IVCNN by comparison with an NN based on the data induction method (DINN, i.e., an NN trained only by data samples, not prior knowledge) and supported vector regression.