Intelligent Tuning of a Kalman Filter forINS/GPS Navigation Applications

摘要

The demand for low-cost portable civil navigationsystems has been growing over the last several years. TheGlobal Positioning System has been the backbone of mostcurrent navigation systems, but its usefulness in harshenvironments, such as downtown urban areas or underheavily treed terrain is limited due to signal blockages.To help bridge these signal gaps inertial navigation systems have been commonly used. An integratedINS/GPS system can provide a continuous navigationsolution regardless of the environment.For civil applications the use of MEMS sensors areneeded due to cost, size and regulatory restrictions ofhigher grade inertial units. The Kalman Filter hastraditionally been used to optimally weight the GPS andINS measurements, but when using MEMS grade sensorsthe a priori tuning parameters given by the designer ormanufacturer are not always optimal. In these cases, theposition errors during loss of the GPS signals accumulatefaster than the ideally tuned case. To aid in the on-linetuning process, a reinforcement learning algorithm wasused to tune the Kalman filter parameters as navigationdata was collected.Tuning any Kalman filter is a difficult task and is oftendone before navigation with the aid of the filter designer.This process often entails much iteration using humanexpertise and is in no way guaranteed to result in optimalparameters. Reinforcement learning is an intelligent andadaptable solution to this problem which uses acombination of dynamic programming and trial and errorexploration to develop a set of parameters that improve asnavigation data is collected by the user.In comparison to a manual tuning approach it was foundthat using reinforcement learning led to similar estimatesof the tuning parameter values for a single integratedsystem. This is encouraging since the reinforcementlearning was done online without any need of interventionby the designer. Furthermore, when the results of tuningone sensor were applied to a second integrated sensor, thetuning converged almost immediately and improved thepositioning accuracy of the system by several meters incomparison to using the values generated from tuning the first sensor.