The durability and reliability of a structural component are usually estimated on the basis of its design loading spectrum, which is a combination of the loading spectra that correspond to various operating conditions. However, if some of the loading spectra that should be used to build the design loading spectrum were not measured, because of cost and time limitations, they must be determined in another way. In order to predict the missing loading spectra a relationship between the operating conditions and the corresponding loading spectra must be known. We showed previously that it is possible to estimate this relationship using a localised basis function neural network. However, with the localised basis function neural network only discrete loading spectra can be predicted and an extrapolation outside the domain of the known operating conditions is impossible. To avoid these problems we have developed a new approach for predicting the loading spectra. It is based on a combined multilayer perceptron neural network. This new approach enables the prediction of smoothed loading spectra and does not impose any restrictions on their extrapolation. In this paper we present the theoretical background to this new approach and then apply it to examples of simulated and measured load states.