The architecture of a neural network is inspired by the structure of the biological nervous system. In contrast to traditional technical systems with a powerful central processing unit a neural network is composed of a complex linkage of a large number of simple processing units, the so-called neurons. Neural networks differ in terms of the transmission behavior of a neuron or network node (the so-called activation function), as well as the arrangement and linking of the various network nodes. The most common neural networks for function approximation are the multi-layer Perzeptron (MLP) and the radial basis functions network (RBF). Both are strictly forward networks, that means the flow of information is arranged through the layers of the network from the inputs to the outputs without processing any information of following or the same layer. As activation functions, different approaches are used, see table.

Typically these are for the RBF networks radialsymmetric functions with a single extremum and monotonous sloping outer flanks. Often Gaussian functions are used as activation functions for RBF networks, which are then overlaid for representing the corresponding target function. The most important parameters are the location of the nodes, the validity radius and the weights of the transitions between the layers, which are optimized in the training of the network such that the function is approximated as well as possible.

An MLP network consists of an input, one or more hidden and an output layer. That means, the input values are distributed by the input layer to the hidden layer, processed there and hand over to the following layer parties. The output layer combines the values of the last hidden layer and processes them to the outputs of the network. The activation function in the neurons is often linear in the input and output layer, while the functions of the hidden layers may vary. Often, these are hyperbolic tangent functions. The neurons of the individual layers are fully linked. There is just forward linkages (feed forward net). There are no layers skipped.

The architecture of a neural network is inspired by the structure of the biological nervous system. In contrast to traditional technical systems with a powerful central processing unit a neural network is composed of a complex linkage of a large number of simple processing units, the so-called neurons. Neural networks differ in terms of the transmission behavior of a neuron or network node (the so-called activation function), as well as the arrangement and linking of the various network nodes. The most common neural networks for function approximation are the multi-layer Perzeptron (MLP) and the radial basis functions network (RBF). Both are strictly forward networks, that means the flow of information is arranged through the layers of the network from the inputs to the outputs without processing any information of following or the same layer. As activation functions, different approaches are used, see table.

Typically these are for the RBF networks radialsymmetric functions with a single extremum and monotonous sloping outer flanks. Often Gaussian functions are used as activation functions for RBF networks, which are then overlaid for representing the corresponding target function. The most important parameters are the location of the nodes, the validity radius and the weights of the transitions between the layers, which are optimized in the training of the network such that the function is approximated as well as possible.

An MLP network consists of an input, one or more hidden and an output layer. That means, the input values are distributed by the input layer to the hidden layer, processed there and hand over to the following layer parties. The output layer combines the values of the last hidden layer and processes them to the outputs of the network. The activation function in the neurons is often linear in the input and output layer, while the functions of the hidden layers may vary. Often, these are hyperbolic tangent functions. The neurons of the individual layers are fully linked. There is just forward linkages (feed forward net). There are no layers skipped.

The modelling and identification with the aid of artificial neural networks was used successfully at IRT as part of the process control of the laser beam welding. For this purpose, an MLP network was implemented, which additionally is extended by analytical process knowledge linked.

The identified process model served as a basis for a Model based Predictive Controller.

The modelling and identification with the aid of artificial neural networks was used successfully at IRT as part of the process control of the laser beam welding. For this purpose, an MLP network was implemented, which additionally is extended by analytical process knowledge linked.

The identified process model served as a basis for a Model based Predictive Controller.