Practical Approaches To Accuracy Enhancement For MEMS-based Inertial Navigation Aided By GNSS

The advent of high-performance and low-cost MEMS inertial sensors drives the miniaturization,cost-reduction and mass-production of once expensive and bulky inertial navigation system(INS),providing affordable GNSS/SINS integrated system for small unmanned aerial vehicles(UAV).As the core system to provide real-time,accurate and reliable navigation results(i.e.,time,position,velocity and attitude),GNSS/MEMS-SINS integrated navigation system paves the way for UAV's challenging missions,such as autonomous flight and automatic landing.However,traditional GNSS/MEMS-SINS integration suffers from performance degradation due to low-cost MEMS sensors and GNSS outages,thus weakening the autonomous navigation ability of UAV.This dissertation focuses on accurate GNSS/MEMS-SINS integrated navigation system and algorithm,aiming at improving the navigation accuracy limited by traditional design.The main researches include GNSS/MEMS-SINS loosely-coupled stochastic error modeling method,time-differenced carrier phase(TDCP)aided GNSS/MEMSSINS tightly-coupled navigation method,neural network(NN)aided inertial navigation method,and MEMS-based in-flight initial alignment method.The prototype system has been designed and produced to conduct vehicle experiments for demonstration and performance evaluation.The work and contributions are the following:(1)Optimal fusion entails accurate structures and parameters of both system noises and measurement noises for GNSS/MEMS-SINS integration.This dissertation proposes a new Allan variance-based stochastic error modeling method for GNSS/MEMS-SINS loosely-coupled optimal state estimation.Allan variance analysis technique is used to identify and model the stochastic errors in MEMS sensors,and Chi-square test is introduced to detect the possible GNSS outliers based on normalized innovation sequences.In the meantime,the adverse effects of GNSS outliers is restrained by means of covariance matching.The measurement noise parameters are estimated accurately according to the frequency division caused by band-pass characteristics derived from the calculation of Allan variance.The field test demonstrates that the proposed method significantly improves the traditional GNSS/MEMS-SINS integrated navigation accuracy.Specifically,compared with traditional RTK/MEMS-SINS loose integration,the proposed method undergoes an obvious increase in overall position and attitude accuracy by 47% and 15%,from 0.089 m and 0.138° to 0.047 m and 0.117°.When it comes to SPP/MEMS-SINS integration,the proposed method witnesses a significant improvement in overall position,velocity and attitude accuracy by 18%,56% and 19%,from 5.1852 m,0.1465m/s and 0.2131° to 4.2497 m,0.0645m/s and 0.1735°.(2)The adoption of carrier phase observations is essential for GNSS precise positioning,while it suffers from either high-dimensional ambiguity resolution problem or baseline length restriction.To accomplish real-time and wide-area navigation,timedifferenced carrier phase(TDCP)measurement is proposed to improve the traditional GNSS/MEMS-SINS tight integration without the need of ambiguity resolution and reference station.Two measurement models are derived from two aspects.On the one hand,from the perspective of velocity integration,TDCP measurements can be expressed as velocity integration between two successive GNSS epochs.The measurement model is established to relate TDCP observation error with system error state,thus forming a tightly-coupled(TC)model.On the other hand,from the perspective of position difference,the position difference error derived from TDCP observations and geometry is used to relate with system error state,thus forming a quasi-tightly-coupled(QTC)model.The field test results show that the QTC scheme can achieve the navigation accuracy of 2.4866 m,0.0533m/s and 0.1599°(1?),while the TC approach can reach the accuracy of 2.6424 m,0.0465m/s and 0.0990°(1?)with much more smooth position error.Apparently,TC scheme outperforms QTC in accuracy.However,QTC has advantage over TC method in real-time realization,because its measurement dimension(=3)is always much smaller than that of TC method(=m×n,n represents the number of available GNSS satellites,and m represents the number of GNSS frequency).In addition,MEMSSINS-aided detection for small GNSS single-frequency cycle-slip is also analyzed and its feasibility is demonstrated initially.(3)A new neural network(NN)aided Kalman filter is proposed to overcome the performance degradation caused by frequently occurred GNSS outages in urban areas.When GNSS signals are available,the NN module is trained to establish an empirical mapping model between inertial measurements and GNSS position difference.Once GNSS outage happens,the well-trained NN module will be used to predict the GNSS position difference according to current inertial measurements.Adaptively robust Kalman filter is also devised to estimate and correct SINS errors and resist the adverse effects of possible outlying NN predictions.The semi-physical simulation test indicates that the proposed approach can reduce the maximum position and velocity error by 21.33% and 19.35% with slightly improved attitude accuracy,when GNSS outage length is set as 40 s.This result basically demonstrates the potentials of NN module in improving the navigation accuracy during relatively long GNSS outages.(4)Airborne low-cost MEMS-SINS encounters large initial attitude uncertainty when UAV is flying,leading to nonlinear INS model errors.To solve this problem,a new in-flight initial alignment algorithm is proposed.The in-flight coarse alignment directly uses both GNSS velocity and MEMS acceleration to conduct double-vector attitude determination.The flight test indicates that the coarse attitude accuracy(1?)can reach 15°(yaw),7°(pitch)and 9°(roll)for small fixed-wing UAV in balanced flight,satisfying the small misalignment angle condition for the following fine alignment phase.As for the in-flight fine alignment method,Cubature transformation is adopted to handle the model nonlinearity caused by improper coarse attitude,and state augmentation technique is introduced to capture meaningful odd-order moment information and reduce the adverse impacts of non-additive noise in inertial measurements.The flight test results indicate the proposed alignment algorithm can complete the initial alignment more quickly and accurately compared with the conventional EKF/UKF-based in-motion alignment approaches,especially when the initial attitude errors are large.As a unified in-flight alignment,it can guarantee accurate and reliable alignment in situations of either large or small initial attitude errors without model changes for small UAV applications.(5)The performance evaluation of navigation system calls for high-accuracy position and attitude reference.As a result,smoothing algorithm is indispensable in this situation.This dissertation proposes a reduced backward RTS smoother for reference determination of GNSS/MEMS-SINS integrated navigation.The results are also analyzed by means of RMS(root mean square)and Allan variance to evaluate the absolute and relative accuracy.The field test results indicate that the proposed smoother can further improve the position,velocity and attitude accuracy by about 40%,12% and 20%,compared with traditional Kalman filtering.It also points out that RTS smoother not only significantly improves the mid-term and long-term navigation accuracy,but also improves the short-term accuracy,but fails to decrease the effects of high-frequency noises.The reduced-RTS smoother can be used as a standard tool for GNSS/INS navigation reference determination.(6)A prototype GNSS/MEMS-SINS integrated navigation system is established.The core navigation algorithm is integrated within embedded computer,including ARTK algorithm,in-flight initial alignment algorithm,the loosely integrated navigation algorithm,the tightly integrated navigation algorithm,etc.Several land vehicle experiments have been conducted to evaluate the system and algorithm performance.The “blind pilot” navigation experiment,meaning that the car driver operates the vehicle only with the real-time navigation visual output from prototype system during the experiment,has also been conducted successfully,which justifies the accuracy and real-time ability.