Correcting inertia sensors of a navigation systems in spite to perform a successfull neural network approach

A challenging open question regards the combined use of low cost GPS systems together with low cost Inertia navigation systems in vehicle navigation tools. Apart from the principle of working, The two systems are characterized by di erent features. The rst provides (when suitably operating) optimal precision for long distance and observation time, the second one is very e cient only within a short observation time as being a ected by the so called random walking e ect. In many recent papers the NN have been used to increase the accuracy of the Inertia Measurement Systems together with GPS signals. However, the direct use of the cinematic variables in the network provided results whose reliability are restricted to limited classes of events (or trip). A more interesting approach is the possibility to use Neural Network system to try to increase the reliability of the inertia measurements while being able to foresee the origin of sensor mistakes. In this paper attention is focused on the capability to get a successful training of a neural network. As a matter of fact, the optimal choice of the input parameters of a NN is fundamental to achieve a general training, that is to say not focused on a particular set of observations. The more these parameters are directly connected to the physical source of errors, the more the network will be capable of a general (not specialised) use.