In order to non-invasively track a capsule endoscopy in the gastrointestinal tract, a telemetric localization method using alternating magnetic fields was presented. Focusing on the method, a Bayesian-regularization neural network based on the Levenberg-Marquart algorithm was investigated to reduce system errors. Firstly, the neural network structure for localization calibration was designed. Then, both Bayesian-regularization and Levenberg-Marquart algorithms were used to train the neural network to limit an over-fitting. Using an experimental platform for localization, both the calibration table for training the network and the validation table for verifying the calibration quality were established,and the location data were calibrated by the trained neural network. The calibration experiment shows that the proposed neural network can be trained well enough to efficiently compensate the errors in electromagnetic localizing system. The mean errors of X, Y, Z, α, β respectively have been reduced to 8.7 mm,10.1 mm,7.3 mm,0.086 rad and 0.081 rad after calibration. Comparing with the standard Back-Propagation(BP) algorithm, the Bayesian-regularization neural network based on Levenberg-Marquart algorithm has better performance in the generalization capability and convergence precision.信息科学