Correction method for wireless electromagnetic localization of microcapsule devices

Abstract

In the development of a telemetry localization system for wireless tracking of a microcapsule medical device based on alternating current magnetic fields, further improvements are required in terms of localization accuracy and reductions in systematic error. A new correction method is proposed based on an improved neural network algorithm for wireless localization. Based on the wireless localization model and its prototype, a single neural network with five input and five output neurons was designed for correction. Because the position and attitude angle are defined on different domains, both the input and output variables were normalized to improve network convergence. To prevent overfitting, the Levenberg-Marquardt Bayesian regularization algorithm was used as an effective learning algorithm for the neural network. Through experimental testing, the tracked and true locations were obtained, and the effects of neural network correction on improving localization accuracy were assessed. The experiments demonstrated reductions in localization errors when using the improved neural network correction algorithm. After correction, average errors of the X, Y, Z, α, and β components reduced to 8.1 mm, 9.3 mm, 7.2 mm, 0.075 rad, and 0.071 rad, respectively. Compared with the basic back propagation algorithm, the Levenberg-Marquardt Bayesian regularization algorithm effectively improves the generalizability and convergence accuracy of neural networks in wireless localization correction. In addition, this method provides a feasible solution for improving the accuracy when wirelessly tracking a microcapsule device.