Neural networks are applied to the identification of non-linear structural dynamic systems. Two complementary problems inspired from customer surveys are successively considered. Each of them calls for a different neural approach. First, the mass of the system is identified based on acceleration recordings. Statistical experiments are carried out to simultaneously characterize optimal pre-processing of the accelerations and optimal neural network models. It is found that key features for mass identification are the fourth statistical moment and the normalized power spectral density of the acceleration. Second, two architectures of recurrent neural networks, an autoregressive and a state-space model, are derived and tested for dynamic simulations, showing higher robustness of the autoregressive form. Discussion is first based on a non-linear two-degree-of-freedom problem. Neural identification is then used to calculate the load from seven acceleration measurements on a car. Eighty three per cent of network estimations show below 5% error.