Learning unknown nonlinearities using a discrete observer in combination with neural networks

In this paper a method is proposed of how to enable a time discrete observer to learn online the nonlinearities of the observed system by using neural networks. The method presented is especially designed for industrial applications. The authors' method employs a new aspect of how to use a time discrete observer in combination with neural networks fed by a special learning rule, which does not need exactly known system parameters. First, the observer structure and its extension for the use with not exactly known system parameters are derived. Afterwards, the practical use is demonstrated by means of an example which shows the learning of the nonlinear characteristics of friction observed at the feed drive of a milling machine.

[1]  D. Schroder,et al.  Neural-net based observers for sensorless drives , 1994, Proceedings of IECON'94 - 20th Annual Conference of IEEE Industrial Electronics.

[2]  D. Luenberger Observing the State of a Linear System , 1964, IEEE Transactions on Military Electronics.

[3]  R. E. Eckmiller,et al.  ‘Partition of Unity’ RBF Networks are Universal Function Approximators , 1994 .

[4]  Anuradha M. Annaswamy,et al.  Stable Adaptive Systems , 1989 .

[5]  Donald F. Specht,et al.  A general regression neural network , 1991, IEEE Trans. Neural Networks.

[6]  Weiping Li,et al.  Applied Nonlinear Control , 1991 .