基于卡尔曼滤波器的GPS近实时定位时钟估计毕业论文外文翻译.doc

外文原文Kalman-filter-based GPS clock estimation for near real-time positioningAndre´ Hauschild . Oliver MontenbruckAbstract： In this article, an algorithm for clock offset estimation of the GPS satellites is presented. The algorithm is based on a Kalman-filter and processes undifferenced code and carrier-phase measurements of a global tracking network. The clock offset and drift of the satellite clocks are estimated along with tracking station clock offsets, troposphericzenith path delay and carrier-phase ambiguities. The article provides a brief overview of already existing nearreal-time and real-time clock products. The filter algorithm and data processing scheme is presented. Finally, theaccuracy of the orbit and clock product is assessed with aprecise orbit determination of the MetOp satellite andcompared to results gained with other real-time products.Keyword :Clock estimation Precise orbit determination Real-time Kalman filterIntroductionA growing number of near real-time precise point positioning (PPP) applications raise the need for precise GPS orbit and clock products with short latency. One of these applications is the precise orbit determination (POD) of remote-sensing satellites, which is to be performed shortly after a ground station pass. The observations of the satellites GPS receiver are available immediately after the download to the ground station. For processing these data,the user requires precise orbit and clock data for thecomplete GPS constellation. The rubidium and cesium atomic standards of the GPS satellites are subject to clock noise and frequency variations, which can originate from a variety of effects and are hard to forecast. Predictions of clock offset and drift, which are provided for example in the predicted part of the ultra-rapid orbits provided by IGS or the broadcast ephemerides, will deviate quickly from the true values by several decimeters or even meters. Thus,these orbit/clock-products become unusable for PPP applications, where a carrier-phase based positioning accuracy down to centimeter level is desired. The solution to this problem is the use of clock offsets, which have been estimated from GPS measurements originating from a network of sensor stations. Currently, only a limited set of providers for precise (near-) real-time orbit/clock-products is available. Among them are three of the IGS Analysis Centers: JPL (Bar-Sever et al. 2003), NRCan and ESA (Pe´rez et al. 2006). The JPL products are transmitted to the user with a latency of about 5 s and can be accessed in various ways, for example, internet data streams and satellite broadcast. The real-time orbit and clock product generation at ESA is currently under development and not publicly available. For our article, however, near real-time orbit and clock products dedicated for the support of the MetOp-Mission have been used. A batch algorithm has been used to generate these products by processing a 2-day data arc for the satellite orbits. The corresponding clocks are computed from shorter data arcs of 23 min including an overlap of 8 min to the previous batch (Zandbergen et al.2006). The real-time orbits and clocks from NRCan are based on data from a global real-time station network. The products are not publicly available.The Astronomical Institute of University Berne (AIUB) has also computed near-real-time clock and orbit products for the test period used in this article. AIBU generates orbit- and clock-data by post-processing of short 100-min batches of GPS observations (Bock et al. 2008).A real-time system for clock estimation is currently under development at the German Space Operations Center of DLR. The generated orbit/clock-products will be used to support orbit determination of low-earth-orbit satellites (LEO satellites) for up-coming space missions, which require near real-time orbit determination accuracies downto 810 cm. The software is based on a Kalman-filter, which processes undifferenced code and carrier phase observations from a worldwide network of GPS stations. The filter uses the orbit information from the predicted part of the latest ultra-rapid IGS products and estimates clock offsets and drifts for the complete GPS constellation. In this article, the complete filter algorithm including the preprocessingof the raw measurements is introduced. The orbit and clock products computed with the filter algorithm are used for a POD with real GPS measurements from the GNSS Receiver for Atmospheric Sounding (GRAS) onboard the MetOp-A satellite. The same analysis has also been performed with the IGS ultra-rapid, JPL, ESA and AIUB products and the results are compared and discussed.Filter algorithmThe clock-estimation algorithm is based on a Kalman-filter,which can be used as a conventional Kalman-filter as well as a forward-/backward-filter with smoother. The filterprocesses ionosphere-free linear data combinations of code and carrier phase measurements on the L1- and L2-frequency.The filter state includes the satellite clock error and the clock drift for the complete constellation of 32 satellites.The state vector additionally comprises the receiver clock offset, a differential tropospheric zenith delay as well as the float carrier phase ambiguities of all satellites in view of each station. The station positions are extracted from recent IGS Sinex-files (IGS 2008) and held fixed in the filter. The current GPS constellation has 32 active satellites and typical tracking network size for the filter is about 20 stations. Assuming that each station tracks on average 10 GPS satellites leads to a total number of about 300 elements in the state vector.Some of the state vector elements require further explanation: the estimated receiver clock offsets for the tracking stations do not represent the offset of the real receiver clocks, since the observation data has been preprocessed before being used in the filter. The pseudo range observations are used together with the a priori orbits and known station position to compute a coarse estimation of the receivers clock error. All observations and the measurement epoch are then corrected by the estimated clockoffset. This pre-processing reduces large clock jumps in the order of milliseconds to less then a microsecond and is beneficial for two reasons: first, the process noise for the receiver clocks can be reduced by several orders of magnitude,as ground station clock jumps do not have to be compensated for. It has been found that this procedure improved the filter stability during measurement updates.Second, elimination during pre-processing eases the filter implementation in later filter steps, as no further measures are necessary for a consistent handling of the ground station clocks. In addition, individual process noise settings for each ground station are avoided, which would need to be maintained in case of changes in the station setup. The differential tropospheric zenith delay shall also be explained in further detail here. The model of the ionosphere-free code and carrier phase observables already includes corrections for the tropospheric delay using a model of the standard atmosphere, which will be introduced later in this section. The true tropospheric delay will differ from the values provided by the empirical model,since the actual local weather conditions deviate from the model parameters. To compensate these deviations, a differential zenith path delay is estimated for each station,which is then mapped into a differential tropospheric slant delay, using an elevation dependent mapping function. Thecarrier phase ambiguities in the filter state are estimated as float values and are not fixed.In order to be able to perform the Kalman-filter time update, the state vector must be predicted towards the next update epoch using a system model. For this algorithm,the GPS satellite clocks are predicted linearly in time. The clock drift and all other state parameters are assumed to be constant. Of course, the satellite clock drift is not strictly constant but it undergoes slow variations.These variations are due to the characteristics of the individual satellite clocks and are driven by hardly predictable effects like thermal variations onboard the GPS satellites. Furthermore, the ground station clock offset and the differential tropospheric delay are subject to variations.In order to compensate the deviations of the system model from the truth, process noise is introduced on these elements of the state vector. Without process noise, the covariance of the state vector would decrease over time and as a result, the weight of the measurements during the filter update decreases, which leads to divergence of the filter.Figure 1 depicts a flowchart of the complete filter algorithm. At the beginning, the forward filter is initialized.The coarse values from the IGS ultra-rapid product are used as a priori values for the satellite clock offset and drift. All other elements of the state vector are set to zero.Additionally the process noise for the filter state and the measurement noise are set during this step.The selection of the process noise and measurement noise determines whether the filter adds more weight to the propagated state based on the system model or to the actual measurements. That is, if the process noise is low compared to the measurement noise, the filter will rely more on the system model and will only gradually correct the filter state during the measurement update. Meaningful settings for the noise of the observables can easily be found from an assessment of the measurement precision. In our case, the carrier phase observables have been assigned a measurement noise of 2 cm.The process noise of the state vector elements is in general more difficult to determine. For simplicity, it is assumed to result from an integrated white noise process,which means that the process noise increases linear in time. It is denoted qi for the filter state element i and is characterized with the standard deviation and time constant s. The process noise matrix has diagonal structure and the elements of the main diagonal are found from qi =i2t/i.The time difference t denotes the time between the consecutive epochs.For the process noise settings of the satellite clock states, no distinction is made between the individual clock types. Instead, the process noise settings are the same for all GPS satellites. The clock offsets have a process noise with a standard deviation of 3 cm and a time-constant of 600 s. The clock drift process noise has a standard deviation of 0.0005 m/s (&10-12 s/s) over 900 s. Though these simplified assumptions do not strictly reflect the selected two-state clock-model, they are favored for a real-time capable process compared to more elaborated models.Using clock models with characteristic process noise settings depending on the satellite block type (Senior et al.2008) or even on the individual satellite clocks (Hutsell1996) adds more complexity, since changes of the used onboard frequency standard or untypical clock behavior must be detected to adjust the process noise settings.Otherwise, the benefits of the model would not be fully exploited. Adapting the settings in real-time from recent data significantly increases the computational load and has therefore not been attempted. However, a rigor assessment of the benefits different clock models deems helpful to identify possible enhancements.The differential zenith path delay of the ground stations are assumed to vary only marginally over time. Consequently,only a small amount of process noise with a standard deviation of 2 mm over 1 h is assigned. On the contrary, the ground station clock offset will exhibit noiselike behavior with deviations in the order of tens of meters due to the clock-jump elimination procedure mentioned previously. Therefore, the comparably large process noise has been chosen to compensate for these deviations. The ambiguities of the carrier phase measurements are assumed to be constant parameters and therefore no process noise is introduced. Upon filter initialization, the initial covariance matrix is set up as a diagonal matrix with the square of the initial standard deviation on the main diagonal. Table 1 provides an overview of the filter settings.In the next step, the filter state is propagated towards the first epoch where measurements are available. During preprocessing in the following step, the ground station clock jumps are eliminated from the data as previously explained. Additionally, the observables are screened for missing data and satellites, which have dropped below an elevation cutoff angle of 10. The core part of the data screening is an integrity monitoring which is performed on the pseudo range and the carrier phase measurements in order to detect and remove outliers. During this monitoring,the orbits and clocks of the predicted IGU product are used together with the known station position to compute the residuals of the ionosphere-free observations for each satellite. Since the position is known, only the station clock offset, which is common for all measurements, must be computed and removed from the residuals. If the RMS of the pseudo ranges exceeds a predefined threshold, the residuals are recursively recomputed with a single satellite excluded at a time. The combination, which yields the lowest residual, identifies the satellite with the outlier in the pseudo range measurement. This satellite is excluded from the filter at this epoch. If the residual-threshold is still exceeded, the procedure of recursively excluding satellite is repeated until the threshold is met or the number of valid satellites drops to two. In the latter case, all remaining satellites are rejected as well, since the monitoring procedure cannot further be performed. A similar approach has been chosen for the monitoring and screening of the carrier phase measurements, but instead time differences of the carrier phases between the current and the previous epoch are used, in order to avoid the complication of estimating ambiguities at this step. With this monitoring procedure,measurement outliers and cycle slips can be detected and the associated satellites are excluded from the measurement update.Afterwards, the ambiguities in the state vector are examined. If satellites have dropped below the elevation limit of the filter or are no longer tracked, their ambiguities are deleted and the space in the filter state is freed. If satellites are newly acquired, their ambiguities are initialized using code-carrier differences to provide their initial values. In addition, ambiguities of satellites, which have been rejected during the data screening, are removed from the filter and initialized again as soon as valid measurements for the satellite are available.Prior to the measurement update the filter applies a cloc