Título: Low Cost GPS/INS Navigation Systems with Error Compensation by Artificial Neural Networks
Autores: Marques Filho, Edmundo Alberto; Rios Neto, Atair; Kuga, Helio Koiti
Resumo: This paper addresses the use of artificial neural networks (ANN) to compensate for errors in inertial measurement units (IMU) of global positioning system (GPS) aided inertial navigation systems (INS). The GPS technology dominates, nowadays, the positioning and navigation (POS/NAV) market, and alternative POS/NAV systems are only needed because GPS does not work in all environments, or can not provide reliable solutions during some time interval. There are different solutions to fulfill information during GPS blockage and integrated inertial sensors systems with GPS are frequently used. However, low cost inertial sensors have the disadvantage of accumulating continuous errors in great extension, leading to poor system performance. In this context, ANN is applied to provide better NAV/POS solutions during the lack of information in GPS outages. This work introduces a review of the main used concepts and techniques, an approach to define input-output ANN signals based on a reduced set of inertial navigation equations, and the ANN prediction and training operation modes. It also presents a training algorithm, based on adaptive Kalman filtering approach and proposes a method for ANN training with the characteristic of alternating the training patterns from batch mode, with a constant data set size, to sequential mode, by filtering individual pattern-by-pattern of training data, which gives to the method some real time training capacity. Finally, numerical simulation results are assessed from urban vehicular positioning application, with data acquired from an MEMS IMU Crossbow CD400_200 and an Ashtech Z12 GPS receiver. The proposed methods were tested with different land vehicle dynamic situations and the position errors, computed in prediction mode or simulated GPS outage, were assessed. When compared to a conventional INS/GPS system, integrated by a Kalman filter and operating without GPS updates, the ANN position errors have lower magnitudes. These results indicate that ANN was more capable to learn the vehicle’s kinematics, for a certain time interval, than the modeling presented by the conventional navigation system.
Palavras-chave: GPS aided inertial navigation; low cost navigation; positioning and navigation systems; artificial neural networks; adaptive Kalman filter
Páginas: 15
Código DOI: 10.21528/lmln-vol10-no4-art3
Artigo em PDF: vol10-no4-art3.pdf
Arquivo BibTex: vol10-no4-art3.bib