Dual Neural Network Approach for Virtual Sensor at Indoor Positioning System

Individuals, with mental or physical disabilities, need that others know their localization within an indoor environment in order to receive adequate healthcare. This paper presents an indoor positioning system based on a received signal strength indicator (RSSI) sensor network, where positions are determined by an artificial neural network (ANN) from the received signals. This work investigates the effect of using the past and present data from the other sensors to estimate one missing signal, using a second ANN, and using it as a virtual sensor in the main ANN. For the study, a database was built in a typical residential environment with one transmitter and four receivers. The research studies the effect on the performance caused by the failure of one sensor showing the gains of using virtual signals, as well as a comparison of this virtual data with the measured data. The ANNs are trained with the cross-validation method to avoid overfitting. The selected number of neurons in the inner layer, for each case, was the complexity capable of presenting at least the same performance of an oversized ANN, which was also trained without overfitting. The system developed achieved a considerable efficiency, being able to reproduce the position of the individual with less than 0.36 m of average error when all four receivers were working properly. However, this average error can increase to 0.52–0.91 m when a receiver is at failure, depending on which one fails. Nevertheless, the use of the proposed virtual sensor can diminish about 0.2 m of average error in case of failure. Therefore, the use of virtual data proved to be a feature capable of improving positioning when a sensor fails, in relation to the alternative of performing this positioning without this sensor nor its corresponding virtual signal.