INS Error Processing And Filtering Model Refining For Integrated Navigation Based On Time-frequency Analysis

This dissertation mainly focuses on the theories and algorithms of INS error processing andfiltering model refining for integrated navigation based on time-frequency analysis. Since INSerror accumulates and the filtering model of integrated system is difficult to determine,time-frequency characteristic analysis method is used to analyze and de-noise INS error, as wellas to refine priori filtering model, in addition, several issues concerning colored noise such as thedetection problem are researched. The main works and contributions are summarized as follows:1. Several basic time-frequency analysis methods are summarized and their relationship isgiven; power spectral characteristics of several major error sources in inertial sensors areanalyzed theoretically. By time-frequency analysis of actual data, the characteristics andmanifestations of the error sources are showed and time-frequency characteristics in static anddynamic condition are compared.2. After analyzing the fundamental principle of wavelet threshold de-noising, it is showedthat wavelet threshold de-noising has many limitations under certain conditions, especially inprocessing inertial sensor errors. The amplitude of useful signal in static inertial data is low, Theuse of wavelet threshold de-noising goes against the principle of threshold de-noising; waveletthreshold de-noising assumes that the noise is white, however, the signal output of the inertialsensor contains a lot of colored noise component, when the effect of wavelet thresholdde-noising is limited.3. For static data, EMD compulsive de-noising method proposed in this thesis first disposesIMFs of exceptional noise by2sigma criterion and then the number of IMFs of high frequencynoise is determined by correlation coefficient. The de-noising process is finally done byreconstructing the other IMFs. EMD compulsive de-noising overcomes the disadvantages thatwavelet threshold denoising bears under static condition.4. For dynamic data, EMD threshold de-noising method proposed here utilizes fractionalGaussian noise as the model of inertial sensor errors, taking colored noise which istime-correlated into consideration. The model parameter estimation method by power spectraldensity is given. Noise variance in IMFs is derived and noise thresholds of IMFs are estimatedthrough the obtained variance. EMD threshold de-noising is effective on reducing sensor errorssince it estimates thresholds of IMFs in conjunction with proper noise model.5. Appropriate prior information is of great importance to both classical Kalman filteringand adaptive filtering. From the viewpoint of stochastic model in the frequency domain like PSD,time-frequency analysis is used to refine prior filtering model of integrated navigation. This method reduces the subjectivity and empiricism in determining filtering model and avoids thecomplicated tuning in utilizing Kalman filtering.6. Taking the case of adjacent-correlated state noise as an example, the rigorous filteringsolution under the condition of colored noise is derived by orthogonal projection theory. Thissolution is an extension of classic Kalman Filtering and classic Kalman Filtering is a special caseof this solution.7. Two methods are proposed here to do real-time detection for the time-correlationcharacteristic of noise: the graphic method and the hypothesis test method. The graphic methodis more intuitionistic while the hypothesis test method has rigorous theory and completelysatisfies real-time demand. When the detection methods are applied, if noise is time-correlated,filtering algorithms under the condition of colored noise are employed to enhance accuracy, andif noise is not correlated, the classic Kalman Filtering will be used to maintain efficiency.