ODE System Identification of a Dynamic Weight Acquisition Process Using Feedforward Neural Networks

Abstract

For detailed modelling of the dynamic weight acquisition process for combination scales, very high effort is required to obtain white box models, since the mechatronic system is highly dependent on fluctuating properties. They are related to the products to be weighed such as piece weight, target weight, the type of impact to the weighing bucket, to the load cell, and superimposed external mechanical and electric disturbances. Therefore this work develops a black box approach to model these measurement systems. The black box model consists of a system of ordinary differential equations which are described as feedforward neural network. The network is trained by 5,897 load cell signal measurements of the impact pulse of plastic anchors with different target weights. The ODE system of the weight acquisition is described by a black box model with a mean error of the static weight value of 0.28 % and a standard deviation of the error of the static weight value of 0.93%. Analyses show that the loss is more scattered for heavier target weight than for lighter target weight. External disturbances could not be generated in the test due to the missing input signal.