Using principles of machine learning, we propose a six-component(6C) seismic waves polarization vector identification method based on multi-layer fully connected(MFC) neural network. Firstly, each 5 000 polarization vector data sets for five wave types and noise are obtained by using the mathematical model of polarization vectors of 6C wave types under a series of simulation parameters. 5 000 of them are randomly selected as test sets and the others as training sets. We make a comprehensive comparison for identification performance between MFC neural network and support vector machine(SVM) model. The results show that the MFC neural network model is significantly better than the SVM model both in identifying five(SH and Love waves are treated one type) and six polarization vector types, with the average recognition rate of 99.786% and 87.940%, respectively.

In order to further test the tracking indicators of microwave fluctuation impending earthquake, based on the data from real-time monitoring and tracking system of microwave fluctuation impending earthquake in Lanzhou, we analyze the waveform data recorded by stations within 150 km of earthquakes of magnitude 5.0 and above in Qinghai region from 2020 to 2022. We believe that almost all stations within a radius of 150 km from the epicenters can record impending microwave motion events. The stations that achieve the index of microwave fluctuation impending earthquake anomalies are not the closest to the epicenter. This may be related to regional tectonic stress, strike of seismogenic fault and process of earthquake rupture. Moreover, the larger the magnitude of the main earthquake, the greater the likelihood of the occurrence of the index of microwave fluctuation impending earthquake anomalies.

We construct a three-dimensional viscoelastic finite element model of the eastern boundary fault zone of the Sichuan-Yunnan block. We calculate long-term average tectonic stress loading rate of the main faults with GNSS observation data as a constraint. According to the historical earthquake rupture models, we calculate the co- and post-seismic Coulomb stress changes in Xianshuihe fault zone, Longmenshan fault zone and so on. Then, we obtain the total stress evolution of the main fault zone in the study area. The results show that the tectonic stress loading rate on Xianshuihe fault zone, the southern section of Daliangshan fault zone, and Litang fault zone are high, which can reach 1.0 to 1.5 kPa/a. In addition, the tectonic stress accumulation in the 2022 Luding M_S6.8 earthquake rupture area is about 0.177 MPa during the seismic recurrence cycle, and the historical events produced 0.07 MPa stress loading. The two stress accumulations together led to the Luding earthquake’s occurrence. In addition, the total stress accumulation on the Anninghe fault zone seismic gap area since the occurrence of the previous event(about 487 years ago) is about 0.3 MPa. The last rupture events occurred about 1 000 years ago in the southern section of Daliangshan fault zone and Litang fault zone, and the total stress accumulation in the seismic recurrence cycle is more than 1.0 MPa, which may indicate that these areas have greater seismic activity potential.

We analyze the characteristics and changes of system time offset parameters of each GNSS, and compare and evaluate the broadcast system time offsets according to the data published by BIPM and the time and frequency bulletin, respectively. The results show that there is a bias of more than 10 ns in the system time offsets of GPS/GLONASS/Galileo broadcasted by BDS. The accuracy of system time offsets between GLONASS and GPS, as well as Galileo and GPS, obtained indirectly through BDS, are better than that of the same parameters broadcasted by GLONASS and Galileo.

Based on the third-order analytic solution of orbit of circular restricted three-body problem, we analyze the orbital evolution character under different initial conditions. Based on the theory of error propagation, we examine the influence of cislunar ephemerides error on the circular collinear libration points orbits at different initial values and initial errors. The results show  that:1)The orbit error is in the same order of magnitude as the cislunar ephemerides error, and its peak value is about 2.5 times the cislunar ephemerides error when the mean value of the cislunar ephemerides is not zero, and the dimensionless period of orbit error is about half of that of the orbit. 2)There are great differences in the influence of cislunar ephemerides error at different positions of the same orbit. The error peak points of plane Lyapunov orbit, Halo orbit and vertical Lyapunov orbit are at or near the maximum and minimum values in X direction, at the maximum and minimum values in Z direction and z=0, respectively.

The sparrow search algorithm(SSA) is used to optimize the variational mode decomposition(VMD), then combining wavelet decomposition(WD), we propose a GNSS coordinate time series noise reduction method(IVMD-WD). This article conducts experiments using simulated signals and measured data from 10 reference stations.The results show that compared with empirical mode decomposition(EMD), ensemble empirical mode decomposition(EEMD) and WD, the IVMD-WD method has the best noise reduction effect, it can more effectively eliminate noise in the GNSS coordinate time series.

Taking the domestic Feiteng CPU as an example, we discuss the efficiency optimization methods of precise orbit determination for navigation satellites based on domestic ARM architecture CPU. Firstly, based on the precise orbit determination process for navigation satellites, we identify the reduction of clock errors and the inverse computation of normal equations as the main time-consuming steps. Secondly, we use multi-threaded and OpenBlas to optimize the two steps separately. The results show that the optimization significantly improves computational efficiency. In terms of clock error reduction, when solving for 100 stations and 32 navigation satellites, the original average time consuming of each epoch is 1.105 s, which is reduced to 0.188 s after optimization. In terms of the inverse computation of normal equations, the original average time consuming is 2 264 s, which is reduced to only 78 s after optimization.

The real-time observation data transmitted by BNC(BKG NTRIP client) is used to evaluate and analyze the positioning performance of GPS and QZSS in the Asia Pacific region from the number of visible satellites, position dilution of precision(PDOP), multipath effect, signal-to-noise ratio, real-time relative positioning accuracy, etc. The results show that after combining QZSS with GPS system, the number of visible satellites increases, the geometric configuration of satellites is better, and the availability and reliability of positioning accuracy improves. The variation pattern of the multipath effect of QZSS satellites is consistent with that of GPS satellites, and the multipath error of each frequency point of QZSS satellites is mostly smaller than that of GPS satellites. The signal-to-noise ratio and altitude angle variation trend of each frequency point of GPS and QZSS satellites are basically the same. In the N, E, and U directions, the real-time relative positioning accuracy of the GPS/QZSS combination is improved compared to GPS.

Based on least squares method, wavelet denoising, wavelet neural network and BP neural network, we construct nine kinds of prediction models of Earth rotation parameters. The results show that the prediction of Earth rotation parameters based on BP neural network is not effective, and the RMSE is greater than 1.5 mas. The least squares and wavelet neural network combined model has the best prediction effect, and the RMSE is less than 1.3 mas. For the prediction of length of day, the combined model of least squares and wavelet neural network is not good, and the RMSE is greater than 0.18 ms. The RMSE of wavelet neural network prediction model is the smallest, which is 0.07 ms. The results show that the wavelet neural network prediction model has the best prediction effect on the length of day.

The grey forecasting models for deformation monitoring are divided into three types: the traditional GM(1,1) model and its improved models, the non-homogeneous grey model, and the GM(1,1) power model and its improved models. We adopt the Origin fitting functions Exp2PModl, Exponential, and SRichards2 as the alternative methods of the three types of grey forecasting models. Based on the theoretical research and example validation, we compare and analyze the three types of grey forecasting models and their alternative methods. The results show that there are significant differences among the three types of grey forecasting models in terms of the fitting function, the presence or absence of limit values, the suitability for equidistant or non-equidistant modeling, and the scope of applicability; it is necessary to select a suitable grey forecasting model according to the characteristics of deformation monitoring data. Compared with the three types of grey forecasting models, the Origin fitting function is more advantageous in parameter solving and requirements on modeling data; it can obtain comparable or even higher fitting or prediction accuracy, which can be completely used in deformation monitoring instead of grey forecasting models, except for special optimization objectives that need to be realized by programming.

We take Xi’an-Xianyang region as the study area and use Sentinel-1A satellite data covering the study area from June 2015 to January 2023. Using time-series InSAR as the main technical means, combined with other data, we obtain information such as the distribution characteristics and deformation velocity of land subsidence in Xi’an-Xianyang area, and analyze the laws of land subsidence occurrence and evolution. The results show that the results obtained by SBAS-InSAR and StaMPS-InSAR are basically consistent. The ground deformation areas in Xi’an are mostly central and southern, and the main factor of deformation is groundwater exploitation. The distribution and trend are influenced by ground fissures and faults. Most areas of Xianyang are in a stable state, with only some areas experiencing significant deformation, which is influenced by various factors such as groundwater exploitation and urban construction.

Aiming at the inconsistency of deformation characteristics of two large landslides in Guxue town, Deyong county, obtained from Sentinel-1A data of ascending and descending orbits, we study the applicability of InSAR technology for landslide monitoring in this area by using R index, sensitivity, and coherence parameters. The results show that the descending data is more suitable for the identification and monitoring of the two large landslides than the ascending data and can explain the reason of the inconsistency of the landslide results. We undertake time series deformation analysis and multi-dimensional deformation feature decomposition for the two landslides. The results show that the maximum cumulative deformation values of the two landslides from 2019 to 2020 are about 300 mm and 230 mm, respectively, and more attention should be paid to them.

By establishing the stress-driven post-seismic afterslip model, this paper numerically simulates the post-seismic afterslip effect of 2021 M_W7.4 Maduo earthquake in Qinghai province. It then fits the GPS post-seismic deformation time-series data 160 days after the earthquake. The results show that afterslip is mainly distributed on the down-dip region and both sides of the coseismic rupture. The maximum slip is 1.24 m, and the cumulative released seismic moment is about 1.08×10¹⁹ Nm, which is equivalent to a M_W6.62 earthquake. Afterslip at the depth of 5 to 15 km accounts for about 80% of the total afterslip. The optimal fitting model shows that the reference slip rate parameters V₀ at the upper and lower regions of the fault are 6.5 m/a and 0.15 m/a, respectively, with the boundary depth of 13 km, which indicates that the difference of frictional properties at the Maduo earthquake afterslip interface is significant in the depth.

This study employs the GAMIT/GLOBK software to process data from 83 GNSS stations spanning the period from January of 2009 to June of 2019. The study obtains the velocities of the GNSS stations within the ITRF2014 framework and constructs a three-dimensional velocity field model of Guangxi using the Kriging interpolation method. By combining observations from 136 GNSS stations in Guangxi and its surrounding areas, we calculate the regional strain field using the grid distance weighting method and analyze its characteristics. The results indicate that the average horizontal velocity of the Guangxi region within the ITRF2014 framework is 34.95 mm/a with the predominant direction of N106.3°E. The vertical motion exhibits regional slow uplift and subsidence. The northern area, located north of 23.5°N, experiences compressional deformation characterized by dominant compressive strain in the SN and NW directions. In contrast, the southern region displays extensional deformation with dominant tensile strain oriented in a west-to-east fan-shaped pattern. The shear strain field demonstrates higher values in the peripheral regions and lower values in the central transitional zone. Over the past decade, earthquakes in the region primarily occur in the transition zone from high to low shear strain, showcasing spatial and temporal migration as well as back-and-forth movement along the fault zone.

By employing the Bayesian three-cornered hat method, we integrate multiple GRACE/GRACE-FO reconstruction models and improve the precision of the reconstruction models using deep learning methods. This approach provides valuable guidance for accurately filling the data gaps in GRACE/GRACE-FO missions. The experimental results show that the combined model exhibits the lowest uncertainty in global terrestrial water storage changes and maintains better agreement with the reference model in most basins. When compared to the optimal individual model, the combined model demonstrates superior performance in semi-arid, semi-humid, and humid regions, with a significant 15% reduction in overall uncertainty and an approximate 5% enhancement in Nash-Sutcliffe efficiency for both global and basin-scale terrestrial water storage changes.

To address the problem of massive data computation in the process of gravity field calculation, we jointly complete a parallel acceleration algorithm for gravity field calculation based on the least squares method using MPI(massage passing interface) and CUDA(compute unified device architecture). MPI is used to complete the task allocation of complex processes and achieve global-level parallel acceleration. We write a parallel acceleration program for large-scale matrix multiplication based on CUDA and adapted for different types of matrices. At the same time, we use MPI to further subdivide the calculation process of the law matrix to compress the peak memory value within the sub-process. We complete the 30 and 120 order gravity field simulation calculation tasks on a single machine. The results show that when inverting the 30 order gravity field, the acceleration ratio can reach 180; when inverting the 120 order gravity field, parallel computing only takes 2 hours for a single iteration, while it cannot be calculated in serial mode.

We combine improved complete empirical mode decomposition with adaptive noise (ICEEMDAN) and distributive entropy(DistEn) to propose a method to suppress the random noise of the extensometer signal without customizing the parameters and with good denoising effect. Firstly, the signal of the extensometer is processed by ICEEMDAN, and several intrinsic mode functions (IMF) are obtained.Then the distributive entropy value of each IMF component is calculated, and according to the magnitude of different distributive entropy values and the degree of chaos of the characterized component signals, each IMF is targeted to be traded off. Finally, linear reconstruction is performed. Designing simulated signal denoising experiments and SS-Y extensometer signal denoising experiments, the results show that the scaler signal reconstruction based on the ICEEMDAN-DistEn denoising model has significantly better reduction and denoising effect than several denoising models such as CEEMDAN-DistEn, wavelet denoising and Kalman filtering.

A CCFE model is used to detect regional seismic signals(epicenter distance ＜150 km) in a single waveform signal and identify P wave and S wave phase. Continuous waveform data recorded at Foziling station in Huoshan area, Anhui province from June to August 2017 are used for seismic detection and phase identification. A total of 164 earthquakes were detected, about 2.16 times more than the earthquake catalog. The missing earthquakes significantly improved the integrity of the seismic catalog in the M_L－1.7 to ML0.0 range. The difference between the time of P wave and S wave phases detected in the cataloged earthquake and the cataloged results are about 0.03 seconds. Compared with the more common deep learning models CRED, EQT and GPD, CCFE model has a higher success rate of identifying seismic events and phase locations for earthquakes with smaller magnitude and some special cases, such as two seismic waveforms that are close to each other, especially when a larger waveform is preceded by a much smaller one.