Artificial Intelligence-Based Multi-Target Trajectory Prediction Method, System, Equipment, and Medium for Space-Based Communication, Navigation, and Remote Sensing Data
By extracting cross-modal correlation features through distributed acquisition and multimodal fusion models, and combining hybrid prediction and consortium blockchain evidence storage, the problem of multi-source data correlation in space-based communication, navigation and remote sensing systems has been solved, achieving high-precision and high-reliability trajectory prediction, and adapting to the complex environment of aerospace scenarios.
Patent Information
- Application Number
- CN202611629005.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies in space-based communication, navigation, and remote sensing integrated systems struggle to integrate the inherent correlations of multi-source data, leading to severe error accumulation in trajectory prediction under complex perturbation forces. This results in failure to meet the accuracy and reliability requirements of long-term time-series predictions and hinders support for space situation assessment and avoidance decision-making.
By employing distributed collection of multi-source heterogeneous data, a multimodal fusion model with embedded spatiotemporal attention mechanism is used to extract cross-modal correlation features, generating a continuous and uninterrupted historical trajectory sequence. A hybrid prediction model is then used for multi-scale trajectory prediction, and a consortium blockchain with Byzantine fault-tolerant consensus mechanism is used for evidence storage, thereby realizing real-time collision warning and trend analysis under a cloud-edge collaborative architecture.
It improves the accuracy and reliability of trajectory prediction, reduces error accumulation, provides highly reliable quantitative support, and provides reliable data support for space situation assessment and avoidance decision-making, adapting to the high dynamic and high-risk requirements of aerospace scenarios.
Smart Images

Figure CN123115740A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aerospace information technology, specifically to a method, system, device, and medium for multi-target trajectory prediction based on artificial intelligence-based space-based communication, navigation, and remote sensing data. Background Technology
[0002] The space-based integrated communication, navigation, and remote sensing system is a core infrastructure for maintaining aerospace security. It is widely used in scenarios such as low-Earth orbit satellite constellation management, space debris monitoring, and spacecraft collision early warning. The operating environment is characterized by high risk, high dynamism, and strong adversarial features. Therefore, there are rigid requirements for the long-term accuracy, physical reliability, and collaborative response capabilities of target trajectory prediction. While the industry has gradually adopted data-driven approaches to assist trajectory analysis and prediction with the development of artificial intelligence and big data technologies, existing technical solutions still have key shortcomings and are difficult to adapt to the comprehensive control requirements of complex space environments. Specifically: Most existing technologies rely on single-modal data for analysis. Even when multi-source data is introduced, features are extracted independently from each type of data and then simply superimposed. These technologies fail to analyze the inherent relationships between different types of data in light of the physical laws governing the motion of spatial targets, nor do they connect the dynamic evolution patterns of various data over time. This makes it difficult for existing solutions to accurately reflect the motion state of targets under complex perturbation forces, and trajectory predictions are prone to significant deviations. This leads to severe accumulation of long-term prediction errors, making it difficult to provide reliable support for subsequent space situation assessment and avoidance decisions.
[0003] To address these issues, the present invention provides a method, system, device, and medium for multi-target trajectory prediction based on artificial intelligence-based space-based communication, navigation, and remote sensing data, thereby resolving the aforementioned technical problems. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, this invention provides a method, system, device, and medium for multi-target trajectory prediction based on artificial intelligence-based space-based communication, navigation, and remote sensing data, in order to solve the problems in existing technologies.
[0005] One embodiment of the present invention provides a multi-target trajectory prediction method based on artificial intelligence-based space-based communication, navigation, and remote sensing data, comprising the following steps: Distributed collection of multi-source heterogeneous data and spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference. The multi-source heterogeneous data includes space-based communication signal data, satellite navigation and positioning data and space-based space remote sensing image data. A multimodal fusion model with embedded spatiotemporal attention mechanism is constructed, and cross-modal correlation features including communication signal features, navigation and positioning features and spatial remote sensing image features are extracted based on the multimodal fusion model; Based on the cross-modal association features, multi-space target detection and association tracking are performed to generate a continuous and uninterrupted historical trajectory sequence. The historical trajectory sequence is input into a pre-trained hybrid prediction model to calculate the long-range temporal dependency and local temporal evolution features of the historical trajectory sequence. Based on the long-range temporal dependency and local temporal evolution features, the model outputs multi-scale trajectory prediction results of multiple spatial targets within a preset time range in the future, and quantifies the prediction uncertainty to generate confidence intervals. The model parameters, multi-scale trajectory prediction results, and confidence intervals of the multi-source heterogeneous data, multi-modal fusion model, and hybrid prediction model are written into a consortium blockchain using a Byzantine fault-tolerant consensus mechanism for storage, and real-time collision warning at the edge and trend analysis in the cloud are completed through a preset cloud-edge collaborative architecture.
[0006] This application also relates to a multi-target trajectory prediction system based on artificial intelligence-based space-based communication, navigation, and remote sensing data, including: The data acquisition module is used to collect multi-source heterogeneous data in a distributed manner and perform spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference. The multi-source heterogeneous data includes space-based communication signal data, satellite navigation and positioning data, and space-based space remote sensing image data. The feature extraction module is used to construct a multimodal fusion model with an embedded spatiotemporal attention mechanism, and extract cross-modal correlation features, including communication signal features, navigation and positioning features, and spatial remote sensing image features, based on the multimodal fusion model. The sequence generation module is used to perform multi-spatial target detection and correlation tracking based on the cross-modal correlation features, and generate a continuous and uninterrupted historical trajectory sequence. The hybrid prediction module is used to input the historical trajectory sequence into a pre-trained hybrid prediction model, calculate the long-range temporal dependency and local temporal evolution features of the historical trajectory sequence, output the multi-scale trajectory prediction results of multiple spatial targets within a preset time range based on the long-range temporal dependency and local temporal evolution features, and quantify the prediction uncertainty to generate confidence intervals. The evidence storage module is used to write the multi-source heterogeneous data, model parameters of the multi-modal fusion model and the hybrid prediction model, multi-scale trajectory prediction results and confidence intervals into the consortium blockchain with Byzantine fault-tolerant consensus mechanism for evidence storage, and to complete real-time collision warning at the edge and trend analysis in the cloud through a preset cloud-edge collaborative architecture.
[0007] This application also relates to a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method for multi-target trajectory prediction of space-based communication, navigation, and remote sensing data based on artificial intelligence.
[0008] This application also relates to a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for multi-target trajectory prediction of space-based communication, navigation, and remote sensing data based on artificial intelligence.
[0009] The multi-target trajectory prediction method, system, device, and medium based on artificial intelligence for space-based communication, navigation, and remote sensing data provided in the above embodiments have the following beneficial effects: 1. This invention employs distributed synchronous acquisition of three types of heterogeneous observation data: space-based communication signals, satellite navigation and positioning, and space-based remote sensing images, and completes standardized spatiotemporal registration. Based on a multimodal fusion model with an embedded spatiotemporal attention mechanism, it extracts cross-modal correlation features. It abandons the shallow fusion method of simply superimposing multi-source features in existing technologies, utilizes spatial attention to mine the spatial physical correlation between targets, and temporal attention to capture the temporal evolution law of observation data. Combined with the dynamic physical law of space targets, it achieves deep coupling of multi-source information, which can restore the true motion state of targets under complex perturbation environment and reduce the basic error of trajectory analysis from the source.
[0010] 2. This invention performs multi-target detection and inter-frame correlation matching based on deeply fused cross-modal correlation features, generating continuous and uninterrupted complete historical trajectories, avoiding the accumulation of initial errors caused by target loss and trajectory gaps; it calculates the long-distance temporal dependence and local temporal evolution features of the historical trajectory sequence through a hybrid prediction model, and outputs multi-scale trajectory prediction results of multiple targets within a preset time range in the future, and quantifies the prediction uncertainty to output confidence intervals, solving the problem of error accumulation in long-term trajectory prediction, improving the accuracy and physical reliability of prediction results, and providing highly reliable quantitative support for spatial situation assessment and avoidance decision-making.
[0011] 3. This invention utilizes a consortium blockchain with a Byzantine fault-tolerant consensus mechanism to uniformly store original observation data, model parameters, and trajectory prediction results, ensuring that the data throughout the entire process is tamper-proof and traceable. Combined with a cloud-edge collaborative architecture, it enables low-latency real-time collision warning at the edge and comprehensive situational assessment across the cloud, adapting to the highly dynamic, high-risk, and highly collaborative business requirements of aerospace scenarios, and comprehensively improving the collaborative response capabilities of aerospace safety management and control systems. Attached Figure Description
[0012] Figure 1 A flowchart illustrating a multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence, provided in an embodiment of the present invention; Figure 2 This is a schematic block diagram of a computer device provided in an embodiment of the present invention. Detailed Implementation
[0013] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.
[0014] Reference Figure 1 One embodiment of the present invention provides a multi-target trajectory prediction method based on artificial intelligence for space-based communication, navigation, and remote sensing data, comprising the following steps: S10. Distributed collection of multi-source heterogeneous data and spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference. The multi-source heterogeneous data includes space-based communication signal data, satellite navigation and positioning data and space-based space remote sensing image data. S20. Construct a multimodal fusion model with embedded spatiotemporal attention mechanism, and extract cross-modal correlation features including communication signal features, navigation and positioning features and spatial remote sensing image features based on the multimodal fusion model; S30. Based on the cross-modal correlation features, perform multi-space target detection and correlation tracking to generate a continuous and uninterrupted historical trajectory sequence; S40. Input the historical trajectory sequence into the pre-trained hybrid prediction model, calculate the long-distance temporal dependency and local temporal evolution characteristics of the historical trajectory sequence, output the multi-scale trajectory prediction results of multiple spatial targets within a preset time range based on the long-distance temporal dependency and local temporal evolution characteristics, and quantify the prediction uncertainty to generate a confidence interval. S50. The model parameters, multi-scale trajectory prediction results and confidence intervals of the multi-source heterogeneous data, multi-modal fusion model and hybrid prediction model are written into the consortium blockchain with Byzantine fault-tolerant consensus mechanism for storage, and real-time collision warning at the edge and trend analysis in the cloud are completed through the preset cloud-edge collaborative architecture.
[0015] In this embodiment, step S10 is used to complete the distributed global acquisition and spatiotemporal reference unified preprocessing of multi-source heterogeneous observation data in the space-based communication, navigation, and remote sensing scenario. This provides standardized basic data input for subsequent cross-modal feature fusion and multi-space target trajectory prediction. Distributed acquisition involves parallel and synchronous acquisition of multiple types of observation data based on a network architecture composed of space-based observation nodes and ground receiving stations in different orbits, ensuring observation coverage and data timeliness. The multi-source heterogeneous data includes three categories: space-based communication signal data, satellite navigation and positioning data, and space-based space remote sensing image data. Space-based communication signal data consists of the time-series amplitude and frequency data of space target communication scattering signals and telemetry signals captured by spaceborne or ground-based telemetry and control stations. Satellite navigation and positioning data consists of the space target position, velocity, and timing sequence data output by the navigation constellation. Space-based space remote sensing image data consists of space-based space remote sensing image frame sequences containing space targets, captured by optical / synthetic aperture radar space-based satellites carrying space surveillance payloads in the orbital airspace. The three types of data are analyzed from the perspectives of signal physical attributes, The three dimensions of motion state and geometric shape characterize the features of spatial targets; spatiotemporal alignment refers to the registration and calibration of timestamps and spatial coordinate systems of different data sources to ensure that the time nodes and spatial positions of the same spatial target in different modal data correspond one-to-one; standardized preprocessing refers to the regularization operations such as noise removal, format normalization, and outlier filtering of the raw data; the preset spatiotemporal reference is a pre-defined unified spatial coordinate system and time reference, with the time reference adopting UTC Coordinated Universal Time; the spatial coordinate system includes two sets used in conjunction: the CGCS2000 geocentric geofixed coordinate system is used for spatiotemporal alignment of multi-source heterogeneous observation data and registration of remote sensing imaging; when carrying out long-term orbital dynamics extrapolation and trajectory prediction calculation of spatial targets, the conversion is made to the J2000 (ICRF) geocentric inertial coordinate system to complete the orbit extrapolation calculation; after the orbit prediction is completed, the prediction results are converted back to the CGCS2000 geocentric geofixed coordinate system through a coordinate transformation model with precession-nutation, Earth rotation, and polar motion correction for subsequent output, collision warning, and consortium blockchain evidence storage.
[0016] Step S20 is used to mine the inherent spatiotemporal correlation of the three types of observation data based on the preprocessed multi-source heterogeneous data through a multimodal fusion model with embedded spatiotemporal attention mechanism, and extract deeply coupled cross-modal correlation features to provide high-dimensional fusion feature support for subsequent multi-space target detection and trajectory prediction. Among them, the multimodal fusion model with embedded spatiotemporal attention mechanism is a feature fusion network integrating spatial attention, temporal attention and cross-attention three-layer structure, which can adaptively focus on key spatial locations, key time nodes and cross-modal strong correlation information. Communication signal features are features extracted from space-based communication signal data that characterize the frequency and amplitude variation of space target signals. Navigation and positioning features are features extracted from satellite navigation and positioning data that characterize the position and velocity motion state of space targets. Space remote sensing image features are features extracted from space-based space remote sensing image data that characterize the geometric shape and texture attributes of space targets. Cross-modal correlation features are comprehensive features obtained by fusing communication signal features, navigation and positioning features and space remote sensing image features. They also contain the signal attributes, motion state and geometric shape information of space targets, and characterize the spatial correlation and temporal evolution coupling relationship between the three types of features.
[0017] Step S30 is used to perform spatial target detection and inter-frame correlation tracking based on the cross-modal correlation features after deep fusion, eliminate false interference spatial targets, fill in the gaps in trajectory observation, and generate a continuous and uninterrupted historical trajectory sequence, providing complete and temporally consistent motion state input for trajectory prediction. Among them, multi-spatial target detection is the process of identifying and locating all candidate spatial targets from space-based remote sensing observation data. Combining cross-modal correlation features can improve the accuracy of identifying weak and occluded spatial targets. Correlation tracking is the process of matching the identity and continuing the state of the same spatial target in continuous observation frames, ensuring that the observation results of the same spatial target at different times correspond to a unique trajectory identifier. The historical trajectory sequence is a set of data on the position and motion state of the same spatial target at multiple times, arranged in chronological order after correlation tracking. Continuous and uninterrupted means that there are no identity jumps in the trajectory or gaps caused by long-term observation loss.
[0018] Step S40 is used to input the historical trajectory sequence generated by the associated tracking into the pre-trained hybrid prediction model, analyze the global temporal patterns and local state changes of the spatial target motion, output the multi-scale trajectory prediction results of multiple spatial targets within a preset time period in the future, and quantify the fluctuation range of the prediction error and generate the corresponding confidence interval, providing highly reliable quantitative support for subsequent spatial situation assessment and collision avoidance decision-making. The pre-trained hybrid prediction model is a dedicated trajectory prediction network that integrates temporal encoding, physical constraint decoding, and uncertainty quantification capabilities, balancing long-term prediction accuracy and physical rationality. Long-distance temporal dependencies represent the overall trend and periodicity of the spatial target motion over a long period in the historical trajectory sequence. Global temporal correlation features, such as periodic patterns, reflect the long-term inertia and evolutionary laws of space target motion; local temporal evolution features are the instantaneous changes and local disturbances of space target motion state within a short time window, reflecting the short-term fluctuations and maneuvering characteristics of space target motion; the preset time range is the prediction duration interval pre-set according to business needs, which can be adapted to different early warning cycles and analysis scenarios; multi-scale trajectory prediction results are trajectory prediction outputs with different time granularities and different spatial precision levels, covering various business needs such as real-time early warning and long-term situation analysis; the confidence interval is the trajectory fluctuation boundary at the corresponding confidence level obtained based on the quantification of prediction uncertainty, used to intuitively characterize the credibility of the prediction results.
[0019] Step S50 securely and reliably stores the raw observation data, model assets, and prediction results of the entire space-based communication, navigation, and remote sensing trajectory prediction process. Simultaneously, a layered, collaborative cloud-edge architecture enables differentiated business responses, balancing low-latency real-time early warning at the edge with comprehensive situational analysis capabilities in the cloud. This provides traceable data and highly efficient collaborative support for space-based target control and collision avoidance. The core data written to the consortium blockchain for storage includes four categories: raw multi-source heterogeneous observation data, full model parameters of multi-modal fusion models and hybrid prediction models, multi-scale trajectory prediction results, and corresponding confidence intervals. This covers the entire chain from data input, model carrier, to output results, meeting the data traceability requirements for subsequent result review, model reproduction, and accountability auditing. The consortium blockchain, employing a Byzantine fault-tolerant consensus mechanism, is a permissioned distributed ledger network jointly established by relevant space-based control participants. The Byzantine fault-tolerant consensus mechanism ensures data consistency across the entire network even when some nodes exhibit abnormal or malicious behavior, guaranteeing that the stored data is tamper-proof and fully traceable, adapting to the high-security and high-reliability storage requirements of space-based scenarios. The pre-defined cloud-edge collaborative architecture adopts a three-tiered deployment architecture of "central cloud - regional edge station - spaceborne edge unit": the top layer is the global central cloud node, the middle layer is the ground regional edge computing station, and the bottom layer is the spaceborne lightweight edge unit carried on the observation satellite. The three-tiered nodes are interconnected through high-speed space-to-ground links and ground private networks. The cloud-based unified deployment of nodes includes a complete version of the multimodal fusion model and a complete hybrid prediction model, focusing on the long-term trajectory trend assessment and judgment of multi-space targets across the entire domain, global model training and iterative optimization, comprehensive situational analysis of track risks across the entire domain, and full data on-chain archiving and management. The edge terminals are divided into two types of deployment carriers: ground-based edge stations and on-board computing units. After the complete model training is completed in the cloud, lightweight prediction sub-models are obtained through model pruning, quantization, and distillation lightweight processing. These are then distributed to each edge terminal for local deployment, forming a lightweight prediction capability. The edge terminal does not need a complete large model; it only runs a lightweight inference module, focusing on completing real-time collision warnings for near-space targets based on real-time observation data and lightweight prediction results. Real-time collision warning execution process at the edge: The edge receives multi-source real-time observation segments synchronously collected by space-based systems in real time, calls the local lightweight prediction sub-model to output short-scale instantaneous trajectories, and has built-in trajectory proximity comparison rules and orbit safety distance threshold discrimination logic. Once the intersection distance between the orbits of two space targets is lower than the safety threshold, a high-risk near space target warning is immediately output locally without the need to send back to the cloud for calculation, ensuring millisecond-level low-latency warnings. The cloud-based trend analysis periodically receives short-term trajectory fragments, local early warning events, and lightweight inference intermediate results from various edge terminals. Based on full historical observation data and a global, complete multi-scale prediction model, it performs cross-regional, long-term, full-domain spatial target orbital status simulation and long-term collision risk assessment. A hierarchical data synchronization mechanism is set up between the cloud and the edge: the edge caches short-term real-time observations and short-term prediction results locally, and only periodically compresses and uploads alarm events, key trajectory segments, and model inference features to the cloud; after the cloud completes the global model iteration update, it distributes the lightweight sub-model incremental package to all edge terminals to complete the model synchronization update. The two achieve hierarchical data synchronization and business collaboration through high-speed satellite-to-ground communication links and encrypted terrestrial private networks.
[0020] In one feasible embodiment, it is assumed that: there are 3 space debris and 2 operational satellites in orbit in a certain near-Earth orbit. The ground-based aerospace control system needs to continuously monitor the trajectory of multiple space targets in the orbit and provide collision risk warnings. The space targets move at high relative speeds and the observations are affected by ionospheric disturbances, resulting in local data deviations. It is necessary to achieve high-precision trajectory prediction and low-latency warning response based on the fusion of multi-source data from communication, navigation, and remote sensing. Step S10: A distributed acquisition network consisting of two space-based optical remote sensing satellites in sun-synchronous orbit, the BeiDou navigation constellation, and three ground tracking stations synchronously acquires communication scattering signal data, navigation reflection positioning data, and space-based optical remote sensing image data of space targets. Using UTC (Coordinated Universal Time) as the reference, the timestamps of the three types of data are unified to a 1-second granularity. All position coordinates are uniformly transformed to the CGCS2000 coordinate system. Ionospheric scintillation noise is removed from the communication signals, gross errors are eliminated from the navigation data, and radiometric correction and geometric registration are performed on the remote sensing images. All data are unified to a preset spatiotemporal reference. In this example, the observation data alignment uses the CGCS2000 coordinate system. When performing subsequent long-term trajectory prediction and extrapolation, the data is converted to the J2000 (ICRF) geocentric inertial coordinate system for orbital dynamics calculations. After the prediction is completed, the data is converted back to the CGCS2000 coordinate system for output.
[0021] Step S20: Construct a multimodal fusion model with embedded spatiotemporal attention mechanism. Input the three types of preprocessed data into the model and extract communication signal features, navigation and positioning features and spatial remote sensing image features respectively. Through the spatiotemporal attention mechanism of the model, mine the spatial position correlation and trajectory temporal evolution correlation between multiple spatial targets and fuse them to obtain cross-modal correlation features that simultaneously contain signal attributes, motion state and geometric morphology information.
[0022] Step S30: Based on cross-modal correlation features, perform multi-space target detection on space-based remote sensing images, remove star background clutter and false space targets, and confirm 5 valid monitored space targets; use correlation matching algorithm to perform identity association and status update of space targets in continuous observation frames, fill in the missing trajectory segments in local observations, and generate a continuous and uninterrupted historical trajectory sequence of 5 space targets over the past 30 minutes.
[0023] Step S40: Input the continuous historical trajectory sequence into the pre-trained hybrid prediction model. The model analyzes the long-distance temporal dependence and local temporal evolution characteristics of the spatial target motion, and outputs the multi-spatial target trajectory prediction results at three time scales: 30 minutes, 1 hour, and 2 hours. At the same time, it quantifies the prediction uncertainty and generates the trajectory fluctuation confidence interval corresponding to a 95% confidence level.
[0024] Step S50: The original multi-source heterogeneous data, the full parameters of the multi-modal fusion model and the hybrid prediction model, the multi-scale trajectory prediction results and confidence intervals are encapsulated into standard data blocks and written into the space-air control alliance chain with Byzantine fault-tolerant consensus mechanism to complete distributed storage; through the preset three-level cloud-edge collaborative architecture of "central cloud - regional edge station - spaceborne edge unit", the ground regional edge station deploys the lightweight prediction sub-model distributed to the cloud, compares the proximity distance of space targets in real time based on the local short-term prediction results, and quickly outputs real-time warnings to the local area for space targets with collision risks higher than the threshold; the cloud control center deploys the complete prediction model, receives the trajectory and alarm data returned by each edge, conducts long-term collision trend analysis of the entire orbital plane, and regularly updates the lightweight model and distributes it to each edge node.
[0025] It should be noted that the above scenarios and the number of spatial targets, time parameters, prediction scale, confidence level, etc., are all illustrative examples, used only to assist those skilled in the art in understanding the overall logic of the solution, and do not constitute any limitation on the present invention. In practical applications, relevant parameters and thresholds can be flexibly adjusted according to the monitoring orbital altitude, spatial target type, observation equipment configuration, early warning requirements, etc.
[0026] In one embodiment, step S10 specifically includes: S11. Distributed acquisition includes multi-source heterogeneous data such as space-based communication signal data, satellite navigation and positioning data, and space-based remote sensing image data; S12. Extract the Doppler frequency shift extrema and signal attenuation characteristics from space-based communication signal data, and use them as communication physical tags to characterize the radial velocity characteristics and signal transmission loss characteristics of space targets relative to the observation station. S13. Detect cycle slips from satellite navigation and positioning data and extract ionospheric delay corrections as navigation physical tags characterizing the integrity features and space atmospheric delay features of satellite navigation and positioning data; S14. Extract the size, attitude angle, and illumination and shadow features of space targets from space-based remote sensing image data, and use them as remote sensing physical tags to characterize the geometric and physical properties and illumination environment features of space targets. S15. Based on the communication physical tag, navigation physical tag and remote sensing physical tag, perform physical consistency verification on multi-source heterogeneous data, remove abnormal data with tag conflicts, and unify them to a preset spatiotemporal reference.
[0027] In this embodiment, as described in steps S11-S15 above, specifically: Step S11 is used to perform distributed network acquisition of multi-source heterogeneous data from space-based communication, navigation, and remote sensing, providing raw observation data input for subsequent physical tag extraction and data consistency verification. Distributed acquisition is achieved through a collaborative acquisition network composed of multi-orbit space-based observation nodes and multi-site ground receiving nodes. Space-based observation nodes include space-based space remote sensing satellites at different orbital altitudes and navigation constellation-borne sensing payloads carrying space surveillance payloads. Ground receiving nodes include ground tracking and control stations and signal receiving stations distributed across different geographical regions. All types of nodes initiate acquisition in parallel and synchronously to ensure data time synchronization and spatial coverage. The acquired space-based communication signal data consists of the time-series amplitude and frequency sampling number of space target communication scattering signals and tracking and control response signals captured by ground tracking and control stations or spaceborne communication payloads. According to reports, satellite navigation and positioning data consists of three-dimensional position, velocity, and high-precision timing sequence data of space targets output by the navigation constellation reflection signal receiving system or onboard GNSS receiver. Space-based remote sensing image data consists of continuous space-based remote sensing image frame sequences containing space targets, captured by optical / synthetic aperture radar space-based satellites carrying space surveillance payloads at preset revisit periods, facing the orbital airspace. For example, in the near-Earth orbit space debris monitoring scenario, a collaborative acquisition network is formed by two sun-synchronous orbit space-based optical remote sensing satellites, the BeiDou navigation constellation ground reflection receiving system, and three ground tracking and control stations distributed at different longitudes to simultaneously acquire three types of raw observation data: communication scattering signal data of space target debris, navigation reflection positioning calculation data, and space-based optical remote sensing image data.
[0028] Step S12 is used to extract signal features with clear physical meaning from the original space-based communication signal data, construct communication physical labels, and provide a physical dimension basis for subsequent data physical consistency verification and feature fusion weight constraints. Among these features, the Doppler frequency shift extreme value is the peak value of the offset of the received signal frequency relative to the transmitted signal frequency when there is relative motion between the space target and the observation station. It is extracted through Fast Fourier Transform spectral analysis and directly characterizes the radial velocity and direction of motion of the space target relative to the observation station. The signal attenuation feature is the degree of amplitude attenuation of the signal after propagation through a space path. It is extracted through a space link loss estimation algorithm and characterizes transmission loss characteristics such as ionospheric loss and space obstruction during signal transmission. Physical tags are quantifiable identifiers extracted from communication signals that have a clear physical correspondence. They correspond to the radial velocity characteristics and signal transmission loss characteristics of space targets, respectively, and can be used to verify the physical consistency of the motion state of space targets in different modal data. For example, in the scenario of monitoring space debris in low Earth orbit, the maximum value of the signal frequency shift is extracted from the time-series data of the space target scattering communication signal collected from the ground telemetry and control station by fast Fourier transform spectrum analysis as the extreme value of Doppler frequency shift. The radial velocity of the space target relative to the observation station is calculated to be 7.2 km / s. At the same time, the attenuation amplitude of the signal is extracted as the signal attenuation characteristic by the space link loss estimation algorithm. The two together constitute the communication physical tag of the space target.
[0029] Step S13 is used to extract physical features related to positioning quality and space environment from the original satellite navigation and positioning data, and construct navigation physical labels to provide a physical dimension basis for the spatial alignment constraints of subsequent data physical consistency verification and feature fusion. Among them, cycle slip is the phase integer jump phenomenon caused by signal loss in satellite navigation carrier phase observation values, which is detected by the phase difference change rate method. The cycle slip detection result directly characterizes the integrity and observation reliability of satellite navigation and positioning data. Ionospheric delay correction is a quantitative correction value for the propagation path delay of navigation signals caused by ionospheric effects, which is calculated by the Klobuchar ionospheric model and characterizes the signal propagation path. The space atmospheric delay characteristics; navigation physical tags are quantifiable identifiers extracted from the navigation and positioning dimension that have a clear physical correspondence. They correspond to the integrity characteristics of the space target's position state and the space atmospheric delay characteristics, respectively, and can be used to verify the physical consistency of the space target's position state in different modal data. For example, in the near-Earth orbit space debris monitoring scenario, in the space target positioning time series data collected from the Beidou navigation reflection receiving system, cycle slips are detected and the data integrity state is marked by the phase difference change rate method. At the same time, the ionospheric delay correction amount of 2.3m is extracted by solving the Klobuchar ionospheric model. The two together constitute the navigation physical tag of the space target.
[0030] Step S14 is used to extract physical features related to the geometric shape of space targets and the observation environment from the original space-based remote sensing image data, construct remote sensing physical labels, and provide physical dimension basis for subsequent data physical consistency verification and geometric matching constraints for feature fusion. Among them, the space target size is the geometric scale parameter of the space target entity obtained from the remote sensing image through edge detection and contour fitting, representing the spatial size attribute of the space target; the attitude angle is the deflection angle of the space target's body coordinate system relative to the observation coordinate system, obtained through the spatial target contour and imaging geometric relationship, representing the spatial orientation and attitude state of the space target; the illumination and shadow features are the shape, size, and position features of the shadow formed by the space target under illumination conditions, extracted by combining the solar altitude angle at the observation time and the geometric contour of the space target, representing the illumination ring at the observation time. The correspondence between the environment and the geometric shape of the space target; remote sensing physical tags are quantifiable identifiers with clear physical correspondences extracted from space-based remote sensing images. They correspond to the geometric and physical attributes and illumination environment characteristics of the space target, respectively, and can be used to verify the physical consistency between the geometric shape and spatial position of the space target in different modal data. For example, in the near-Earth orbit space debris monitoring scenario, from the monitoring area image taken by the sun-synchronous orbit space-based optical remote sensing satellite, the major axis dimension of the space target debris is 1.2m and the minor axis dimension is 0.8m, which is obtained by edge detection and contour fitting. Combined with the imaging geometric relationship, the attitude angles include a roll angle of 15° and a pitch angle of 8°. At the same time, the length and orientation of the space target's shadow are extracted as illumination shadow features by combining the solar altitude angle at the observation time. The three together constitute the remote sensing physical tag of the space target.
[0031] Step S15 is used to perform physical dimension consistency verification on multi-source heterogeneous observation data based on the communication physical tags, navigation physical tags, and remote sensing physical tags extracted in the previous steps, remove abnormal data with physical conflicts, and complete the final unified alignment of multi-source data to a preset spatiotemporal reference. The physical consistency verification is a process of matching and verifying the physical laws of multimodal observation results of the same spatial target under the same spatiotemporal reference by constructing a multi-tag physical constraint threshold matrix based on the physical meaning of each of the three types of tags. This covers verification dimensions such as motion state consistency, position coordinate consistency, and geometric attribute consistency, ensuring that the spatial target state reflected by different modal data conforms to the objective physical laws of spatial target motion and signal propagation. Tag conflict refers to a significant deviation from physical laws in the spatial target state reflected by different modal physical tags, such as a mismatch between the navigation positioning result and the radial velocity calculated by Doppler frequency shift, or a discrepancy between the spatial target position calculated by space-based remote sensing imagery and the navigation positioning coordinates. In cases of deviation exceeding the threshold, any deviation in any dimension exceeding the preset allowable threshold is considered a label conflict, and the corresponding data becomes abnormal data that needs to be removed. Unifying to a preset spatiotemporal reference means that the verified data is finally precisely aligned in spatiotemporal coordinates to form a valid observation dataset that conforms to the unified reference specification. For example, in the scenario of monitoring space debris in near-Earth orbit, for three types of observation data of the same space target, the radial velocity calculated by the communication physical tag, the positioning integrity and position deviation reflected by the navigation physical tag, and the geometric position of the space target corresponding to the remote sensing physical tag are cross-validated based on the multi-label physical constraint threshold matrix. If the navigation data at a certain observation time shows a position jump of hundreds of meters due to cycle slip, and there is a significant deviation from the motion state calculated by the communication Doppler and the position of the space target calculated by the space-based space remote sensing image, it is judged as abnormal data with a label conflict and is removed. The remaining verified data are uniformly aligned to the CGCS2000 coordinate system and the UTC Coordinated Universal Time reference.
[0032] In one embodiment, the multimodal fusion model includes a spatial attention sublayer, a temporal attention sublayer, and a cross-attention sublayer. Step S20 specifically includes: S21. Based on the communication physical tag, navigation physical tag and remote sensing physical tag, generate the confidence weights corresponding to each multi-source heterogeneous data after physical consistency verification, and input the confidence weights as suppression coefficients into the multimodal fusion model. S22. Using a multimodal fusion model, feature extraction is performed on the multi-source heterogeneous data after physical consistency verification. Communication signal features are extracted from space-based communication signal data, navigation and positioning features are extracted from satellite navigation and positioning data, and space remote sensing image features are extracted from space-based space remote sensing image data. S23. In the spatial attention sub-layer of the multimodal fusion model, based on navigation and positioning features and with the ionospheric delay correction in the navigation physical label as the alignment reference, the inhibition coefficient is used to suppress the spatial attention weights, eliminate inconsistent associations caused by cycle slips that lead to position jumps, and characterize the spatial position associations between multiple spatial targets. S24. In the temporal attention sub-layer of the multimodal fusion model, based on the characteristics of the communication signal and the extreme value of the Doppler frequency shift in the communication physical label, the temporal attention weight is suppressed by the suppression coefficient to characterize the temporal evolution correlation of the trajectory. S25. In the cross-attention sublayer of the multimodal fusion model, spatial remote sensing image features are used as the geometric reference. Spatial location association and temporal evolution association are introduced as structural constraints into the cross-attention calculation. Spatial target size and attitude angle in remote sensing physical labels are used as the geometric matching reference. Suppression coefficients of communication side and navigation side are applied to weight the cross-attention weights to suppress them. Cross-modal interaction between communication signal features and navigation positioning features in the spatial remote sensing image observation space is established, and cross-modal association features are obtained by fusion.
[0033] In this embodiment, the multimodal fusion model includes a spatial attention sublayer, a temporal attention sublayer, and a cross-attention sublayer, forming a hierarchical and progressive feature fusion processing architecture. This provides structural support for the accurate extraction of cross-modal correlation features. The spatial attention sublayer is the feature computation unit in the multimodal fusion model responsible for mining the spatial distribution correlations of multiple spatial targets and filtering effective spatial feature information. It generates a spatial attention weight map through an adaptive spatial weight allocation algorithm, which can adaptively adjust the spatial attention weights for different spatial targets and regions, filtering out abnormal correlation interference in the spatial dimension. The temporal attention sublayer is responsible for capturing the temporal patterns of spatial target motion and characterizing the dynamic evolution characteristics of trajectories. The feature computation unit generates a temporal attention weight sequence through a dynamic weighting algorithm based on temporal weights. This optimizes the temporal attention weights at different observation times, strengthens effective temporal correlations, and eliminates instantaneous anomalies. The cross-attention sublayer is the core fusion computation unit in the multimodal fusion model responsible for realizing the interactive coupling of multimodal features. It constructs a cross-modal correlation matrix through a scaling dot product attention computation algorithm, which can establish cross-domain mapping and correlation matching between different modal features, and complete the deep coupling of the three types of features. The three sublayers work together in sequence to complete feature optimization and fusion step by step from the spatial dimension, temporal dimension, and cross-modal interaction dimension, ensuring that the final output cross-modal correlation features have spatial accuracy, temporal continuity, and modal coupling.
[0034] As described in steps S21~S25 above, specifically: Step S21 is used to quantify and generate the credibility weights corresponding to each multi-source heterogeneous data after physical consistency verification based on communication physical tags, navigation physical tags, and remote sensing physical tags. The credibility weights are then mapped to suppression coefficients and input into the multimodal fusion model, providing a physical quantification constraint for the adaptive suppression optimization of weights in each attention sublayer. The credibility weights are single-modal observation data quality scores quantified based on the data integrity, environmental disturbance level, and physical parameter rationality reflected by the three types of physical tags. These scores are calculated using a multi-dimensional tag quantization mapping algorithm, and their values are positively correlated with data observation quality, intuitively reflecting the credibility level of each modality's observation data. The suppression coefficients are model weight adjustment parameters obtained by normalizing the credibility weights and mapping them through linear mapping. The algorithm converts weights into coefficients, with the numerical values positively correlated with the confidence weights. Modal data with local anomalies or strong disturbances correspond to lower suppression coefficients, which are used to adaptively weaken the constraints on the attention weights of each attention sublayer of the model. For example, in the near-Earth orbit space debris monitoring scenario, for three types of observation data that have passed the physical consistency verification, the confidence weights of the three types of data are generated by a multi-dimensional label quantization mapping algorithm, which combines the signal attenuation degree of communication physical tags, the cycle slip state and 2.3m ionospheric delay correction amount of navigation physical tags, and the illumination and shadow distortion degree of remote sensing physical tags. Then, the confidence weights of the three types of data are obtained by linear mapping transformation, and the exclusive suppression coefficients of the corresponding modes are obtained. All suppression coefficients are synchronously input into the multi-modal fusion model for the adjustment of the attention weights of each sublayer.
[0035] Step S22 utilizes the dedicated feature encoding branch of the multimodal fusion model to independently extract single-modal features from the multi-source heterogeneous data after physical consistency verification, sequentially outputting communication signal features, navigation and positioning features, and space remote sensing image features. This provides standardized feature inputs with unified dimensions for weight optimization and cross-modal fusion of each attention sublayer. Specifically, the communication signal features are core time-domain and frequency-domain features extracted from space-based communication signal time-series data through a one-dimensional convolutional coding network, representing the dynamic changes in frequency and amplitude of space target signals. The navigation and positioning features are continuous position and velocity motion features extracted from satellite navigation and positioning time-series data through a time-series cyclic coding network, representing the spatial target's position and velocity. The motion state and changing trend of space targets; spatial remote sensing image features are the geometric, texture and spatial location features of space targets extracted from space-based space remote sensing image data through a two-dimensional convolutional coding network, representing the geometric shape and spatial distribution attributes of space targets; for example, in the near-Earth orbit space debris monitoring scenario, the multimodal fusion model extracts features from communication scattering signal data, BeiDou navigation and positioning time series data and space-based space optical remote sensing image data that have passed physical consistency verification through corresponding dedicated coding networks, and outputs the communication signal features, navigation and positioning features and space remote sensing image features of the corresponding space targets in sequence. The three types of features are unified in dimension and have complete physical information, and can be directly connected to the subsequent attention sublayer for processing.
[0036] Step S23 is used in the spatial attention sublayer of the multimodal fusion model to adaptively suppress spatial attention weights using a suppression coefficient. This suppresses inconsistent spatial relationships caused by cycle slips, eliminating them and accurately characterizing the true spatial positional relationships between multiple spatial targets. Specifically, the spatial attention sublayer uses navigation and positioning features as the core benchmark for matching the spatial positions of multiple spatial targets, ensuring a unified coordinate benchmark for the spatial relationships of all spatial targets. It uses the ionospheric delay correction in the navigation physical labels as the spatial alignment accuracy benchmark, quantifying the spatial error level of the positioning data and providing an accuracy reference for weight suppression. This is achieved through adaptive suppression of spatial weights. The algorithm adjusts the original spatial attention weights by combining corresponding suppression coefficients, and weakens the abnormal spatial association weights corresponding to position jumps caused by navigation data cycle slips. It eliminates inconsistent associations that violate the laws of space physics and finally outputs optimized spatial position association features of multiple spatial targets. For example, in the near-Earth orbit space debris monitoring scenario, within the spatial attention sub-layer, the navigation and positioning features of spatial targets are used as the spatial matching benchmark, and the 2.3m ionospheric delay correction is used as the alignment accuracy benchmark. Through the spatial weight adaptive suppression algorithm, the corresponding spatial attention weights with cycle slips and position jumps are suppressed using the suppression coefficients on the navigation side, eliminating misaligned abnormal spatial associations and characterizing the true spatial position associations between multiple space debris targets.
[0037] Step S24 is used in the temporal attention sublayer of the multimodal fusion model to adaptively suppress and optimize the temporal attention weights using the suppression coefficient, taking the communication signal features as the unified temporal benchmark and the Doppler frequency shift extreme value in the communication physical label as the temporal evolution accuracy benchmark, thereby eliminating abnormal perturbation correlations in the temporal dimension and characterizing the continuous and stable temporal evolution correlation of the spatial target trajectory. The temporal attention sublayer uses high-temporal-resolution communication signal features as the core benchmark for temporal motion analysis of space targets, ensuring the continuity and stability of temporal correlation analysis. It uses the Doppler frequency shift extreme value in the communication physical tag as the criterion for judging the radial motion rate and temporal changes of space targets, constraining the physical scale of trajectory temporal evolution. Through a temporal weight dynamic suppression algorithm, combined with the suppression coefficient corresponding to the communication mode, the original temporal attention weights are adaptively adjusted to weaken invalid temporal correlation weights caused by abnormal signal transmission loss and instantaneous noise disturbances, retaining valid temporal correlation information that conforms to the true motion law of space targets, and accurately characterizing the dynamic evolution relationship of space target trajectories at different times. For example, in the near-Earth orbit space debris monitoring scenario, within the temporal attention sublayer, the communication signal features of space targets are used as the temporal matching benchmark, and the Doppler frequency shift extreme value corresponding to a radial velocity of 7.2 km / s is used as the temporal evolution benchmark. The temporal weight dynamic suppression algorithm uses the communication-side suppression coefficient to suppress and filter abnormal temporal weights, effectively eliminating erroneous correlations caused by temporal disturbances, and characterizing the continuous temporal evolution correlation of space debris target trajectories over time.
[0038] Step S25 is used to construct a unified geometric benchmark based on spatial remote sensing image features in the cross-attention sublayer of the multimodal fusion model, introduce spatial location association and temporal evolution association as structural constraints, and complete the weighted suppression optimization of cross-attention weights by combining the spatial target geometric parameters of remote sensing physical labels and multimodal suppression coefficients. This realizes cross-modal interactive fusion of communication and navigation features in the spatial remote sensing image observation space, generating highly consistent and strongly coupled cross-modal association features. Specifically, the cross-attention sublayer uses spatial remote sensing image features as a globally unified geometric matching benchmark, establishes a standardized geometric matching scale based on the spatial target size and attitude angle in the remote sensing physical labels, and constrains the geometric consistency of cross-modal feature mapping. The spatial location association of multiple spatial targets output by the spatial attention sublayer and the trajectory temporal evolution association output by the temporal attention sublayer are embedded into the cross-attention matrix calculation process as spatiotemporal structural constraints to ensure the physical rationality of the fused features. Through the cross-modal weighted suppression algorithm, the communication side and navigation side are respectively optimized. The original cross-attention weights are subjected to the corresponding modal suppression coefficients to adaptively weaken the erroneous cross-modal association weights caused by geometric misalignment, temporal mismatch, and spatial anomalies. After weight optimization, the projection and bidirectional interactive coupling of communication signal features and navigation positioning features to the spatial remote sensing image observation space are realized. The unique information of the three types of single-modal features is fused to obtain complete cross-modal association features that take into account geometric shape, motion state, signal attributes, and spatiotemporal patterns. For example, in the near-Earth orbit space debris monitoring scenario, the size parameters of the space target (1.2m×0.8m), attitude parameters (15° roll angle and 8° pitch angle) are used as the geometric matching benchmark in the cross-attention sub-layer. The optimized spatial position association and temporal evolution association are introduced as structural constraints. The cross-attention weights are optimized by applying the communication and navigation dual-modal suppression coefficients through the cross-modal weighted suppression algorithm, filtering modal mismatch associations, establishing cross-domain interaction of communication and navigation features in the spatial remote sensing image observation space, and fusing to obtain the cross-modal association features of the space debris target.
[0039] In one embodiment, step S25 specifically includes: S251. Construct a cross-attention query vector using the spatial remote sensing image features, and construct a key vector and a value vector by concatenating the communication signal features and navigation positioning features. S252. Calculate the similarity matrix between the query vector and the key vector, and add the spatial location association as a location bias term and the temporal evolution association as a time decay term to the similarity matrix to obtain the structurally constrained similarity matrix. S253. Based on the spatial target size and attitude angle in the remote sensing physical label, calculate the geometric matching score matrix, and multiply the geometric matching score matrix element by element with the similarity matrix after structural constraints to obtain the geometrically weighted similarity matrix. S254. Using the credibility weights generated by the communication physical tags and navigation physical tags as suppression coefficients, the geometrically weighted similarity matrix is masked and scaled to obtain attention weights. The value vector is then weighted and aggregated using the attention weights to output cross-modal association features.
[0040] In this embodiment, as described in steps S251-S254 above, specifically: Step S251 is used to complete the initial vector construction for cross-attention calculation. Different modal features are assigned differentially to achieve modal division of labor among query, key, and value vectors. The query vector is the feature vector responsible for retrieving matching criteria in the cross-attention calculation. It uses spatial remote sensing image features as the sole source of construction and performs a dimensional uniform transformation on the spatial remote sensing image features using a feature dimension mapping algorithm to generate a dimensionally regular and semantically focused geometric reference query vector. The key and value vectors are the feature vectors responsible for providing the feature information to be matched in the cross-attention calculation. Communication signal features and navigation positioning features are concatenated and fused according to feature dimensions. A feature concatenation and fusion algorithm completes the dual-modal feature... The features are sequentially concatenated to construct key and value vectors of a unified dimension, realizing the aggregation of information from motion and signal modal features. This step uses a differentiated vector construction method, fixing remote sensing geometric features as the retrieval benchmark and communication and navigation features as the objects to be matched, ensuring the uniqueness of the benchmark for cross-modal interaction. For example, in the near-Earth orbit space debris monitoring scenario, the preprocessed spatial remote sensing image features of the space target are converted into cross-attention query vectors through a dimension mapping algorithm. At the same time, the communication signal features and navigation and positioning features of the corresponding space target are sequentially concatenated through a feature concatenation and fusion algorithm to generate key and value vectors for matching, completing the pre-vector initialization for cross-attention calculation.
[0041] Step S252 is used to complete the similarity association calculation between the query vector and the key vector, and introduces spatiotemporal prior constraints to modify and optimize the original similarity matrix, resulting in a constrained similarity matrix that combines spatial rationality and temporal continuity. The similarity matrix is a two-dimensional association matrix used to quantify the matching degree between the query vector and the key vector, calculated using a scaled dot product similarity algorithm. The values of each element in the matrix correspond to the initial matching degree between different modal features. The positional bias term is a spatial correction parameter generated by the spatial positional association quantization output from the preceding spatial attention sublayer, used to characterize the true spatial distribution relationship of multiple spatial targets and correct false matches caused by spatial misalignment. The temporal decay term is a temporal correction term generated by the temporal evolution association quantization output from the preceding temporal attention sublayer. The parameters are used to characterize the temporal evolution of spatial target trajectories, weakening invalid matches with large temporal spans and inconsistent motion patterns. Through a matrix element superposition algorithm, the position bias term and the time decay term are superimposed element by element to the original similarity matrix, incorporating the spatiotemporal structural prior constraints, eliminating abnormal matching associations in the spatiotemporal dimension, and obtaining a structurally constrained similarity matrix. For example, in the near-Earth orbit space debris monitoring scenario, the initial similarity matrix between the remote sensing feature query vector and the communication and navigation feature key vectors is calculated by the scaling dot product similarity algorithm. Then, the quantized spatial position association is used as the position bias term and the temporal evolution association is used as the time decay term for element-by-element superposition correction, discarding false association information with spatiotemporal mismatch, and generating a structurally constrained similarity matrix with spatiotemporal physical rationality.
[0042] Step S253 is used to complete the cross-modal geometric matching quantization based on the geometric parameters of the remote sensing physical tags. The similarity matrix after structural constraints is then weighted and corrected using the geometric matching score matrix to achieve the implantation of geometric physical constraints. The geometric matching score matrix is a two-dimensional quantized scoring matrix generated by a geometric similarity matching algorithm based on two core geometric parameters—spatial target size and attitude angle—from the remote sensing physical tags. It characterizes the degree of matching between the spatial target state corresponding to communication and navigation features and the actual geometric shape of the remote sensing object. Element-wise matrix multiplication is achieved using the Hadamard product algorithm, where the score of each element in the geometric matching score matrix is weighted and applied to the similarity matrix after structural constraints. The corresponding elements of the matrix are adaptively reduced for cross-modal associations with low geometric matching degree and attitude deviation exceeding the threshold, while effective associations with high geometric fit are retained and strengthened. Finally, spatiotemporal constraints and geometric constraints are integrated to obtain a geometrically weighted similarity matrix that takes into account both spatiotemporal rationality and geometric authenticity. For example, in the near-Earth orbit space debris monitoring scenario, based on the size parameters of the space target (1.2m×0.8m), attitude parameters (15° roll angle and 8° pitch angle), a global geometric matching score matrix is calculated through a geometric similarity matching algorithm. Then, an element-wise Hadamard product operation is performed between this matrix and the similarity matrix after structural constraints to reduce the weight of false associations with geometric mismatch and generate a geometrically weighted optimized similarity matrix.
[0043] Step S254 uses the suppression coefficients obtained by transforming the credibility weights corresponding to the communication and navigation modes to perform masking constraints and numerical scaling on the geometrically weighted similarity matrix to solve for the attention weights. Then, it performs value vector weighted aggregation based on the attention weights, and finally outputs cross-modal association features that fuse multi-dimensional physical constraints. Among them, the masking and scaling processing is implemented through the modality suppression masking and scaling algorithm. The masking operation is used to shield invalid association elements corresponding to modes with extremely low credibility based on the suppression coefficients. The scaling operation is used to adaptively scale down the matching scores of each element in the matrix based on the suppression coefficients, weakening the association weights corresponding to modes with defects such as signal loss and positioning jumps. The attention weights are standardized two-dimensional weight moments obtained after masking, scaling and normalization. The matrix elements represent the degree of reliable matching between the remote sensing query vector and the communication and navigation key-value vectors. Weighted aggregation is accomplished through a feature weighted summation algorithm, which uses optimized attention weights to weightedly superimpose and fuse the original communication and navigation features corresponding to the key-value vectors, aggregating deep features of a unified dimension. For example, in the near-Earth orbit space debris monitoring scenario, the two types of modal suppression coefficients obtained by converting communication physical labels and navigation physical labels are retrieved. The modal suppression mask scaling algorithm is used to mask and scale the similarity matrix after geometric weighting, and normalize it to obtain reliable attention weights. Then, the attention weights are used to perform weighted summation aggregation on the spliced communication and navigation value vectors, and finally output the complete cross-modal association features of the space debris target.
[0044] In one embodiment, step S30 specifically includes: S31. Spatial target detection is performed on the spatial remote sensing image features to obtain a candidate detection set. The target response intensity of cross-modal correlation features in the spatial remote sensing image observation space is used as the judgment criterion. The spatial target size and attitude angle in the remote sensing physical label are used as geometric constraints to eliminate interference targets in the candidate detection set, confirm valid spatial targets, and obtain an initial candidate spatial target set. S32. Based on the Doppler frequency shift extremum in the communication physical tag and the ionospheric delay correction amount in the navigation physical tag, apply physical motion constraints to the initial candidate spatial target set and predict the motion trend. S33. Input the cross-modal association features into a preset association matching algorithm, integrate the motion trend prediction results and the cycle slip detection results in the navigation physical label, perform association matching on spatial targets between consecutive frames, perform state update on the successfully matched trajectory, perform trajectory initialization on the unmatched candidate spatial targets, and generate a continuous and uninterrupted historical trajectory sequence.
[0045] In this embodiment, as described in steps S31-S33 above, specifically: Step S31 is used to perform initial target detection based on the optimized spatial remote sensing image features, and to perform interference screening by combining cross-modal association feature response mechanism and remote sensing physical geometric constraints, accurately eliminating false interference spatial targets, screening out real and effective spatial targets, and constructing an initial candidate spatial target set. The candidate detection set is the initial set of spatial targets extracted by the target detection algorithm through full-domain target retrieval of spatial remote sensing image features, including multiple types of candidate objects such as real spatial targets, noisy false spatial targets, and background interference spatial targets. The target response intensity is the spatial target association confidence score output after the cross-modal association features are mapped to the spatial remote sensing image observation space, quantifying the credibility of the cross-modal matching between the image region and the real spatial target. The spatial target size and attitude angle are inherent geometric constraint parameters of the remote sensing physical label, and are determined by geometric thresholding. The screening algorithm verifies the shape and attitude compliance of candidate spatial targets. Through a dual screening mechanism combining target response intensity threshold judgment and geometric constraints, it filters out interfering spatial targets with abnormal size, attitude distortion, and weak cross-modal response, retaining real and valid spatial targets that conform to physical characteristics, thus constructing an accurate and low-noise initial candidate spatial target set. For example, in the near-Earth orbit space debris monitoring scenario, the optimized space remote sensing image features are first processed by the target detection algorithm to output a global candidate spatial target detection set. Then, the target response intensity based on cross-modal correlation features is used for initial screening. Combined with the attitude constraints of 1.2m×0.8m spatial target size, 15° roll angle and 8° pitch angle, a second verification is performed to remove background noise, debris pseudo-spatial targets and other interfering objects, and real spatial debris targets are selected to form the initial candidate spatial target set.
[0046] Step S32 is used to apply objective physical motion constraints to the initial candidate space target set based on the Doppler frequency shift extrema in the communication physical tag and the ionospheric delay correction in the navigation physical tag, thereby completing the high-precision motion trend prediction of space targets. The Doppler frequency shift extrema are core parameters in the communication physical tag that characterize the radial velocity and direction of motion of space targets, accurately reflecting their instantaneous speed and dynamic characteristics. The ionospheric delay correction is a correction parameter in the navigation physical tag that characterizes atmospheric propagation errors, compensating for spatial observation biases in navigation and positioning data and improving the accuracy of motion parameter calculation. The physical motion constraint prediction algorithm is a short-time trajectory prediction algorithm that integrates the kinematic laws of space targets. Specifically, it is implemented through radial velocity motion modeling combined with spatial error compensation iterative calculation. The core position prediction formula adopts a uniform motion error correction model, as shown in the following formula:
[0047] in, Predict the position of a spatial target at the next moment; The current observed position of the space target; The radial velocity of the space target is obtained by inversion from the extreme value of Doppler frequency shift; The time interval between adjacent frames; This is the spatial position error compensation value corresponding to the ionospheric delay correction.
[0048] This algorithm uses the radial velocity of the space target calculated from the Doppler frequency shift extremum as the motion state input and the ionospheric delay correction as the observation error compensation term. It iteratively corrects the position offset error of the space target, constrains the motion amplitude and direction of the space target to conform to the physical characteristics of the motion of the space target in near-Earth orbit, and eliminates abnormal motion inference results that violate the motion law of the space target, so as to obtain the short-term motion trend prediction results of the space target that fits the actual orbit characteristics. For example, in the near-Earth orbit space debris monitoring scenario, combined with the 7.2 km / s radial velocity Doppler frequency shift extremum corresponding to the communication physical tag and the 2.3 m ionospheric delay correction of the navigation physical tag, the algorithm applies orbital motion physical constraints to the initial candidate space target set through the above radial velocity modeling and error compensation iterative calculation method, corrects the motion offset caused by the observation error, and predicts the short-term motion trend and position change law of each candidate space target.
[0049] Step S33 is used to input the optimized cross-modal association features into a preset association matching algorithm, fuse the motion trend prediction results with the cycle slip detection results carried by the navigation physical tag to perform continuous inter-frame spatial target association pairing, complete trajectory state updates and new trajectory initialization in different scenarios, and output a complete historical trajectory sequence without breaks in time sequence; among them, the cross-modal association features are fused features obtained through multi-layer attention constraint optimization, used to provide the basis for judging the similarity of inter-frame spatial target features; the motion trend prediction results are the spatial target's next moment position prior derived from the physical motion constraint model, used to narrow the inter-frame spatial target matching retrieval range and reduce cross-frame mismatch phenomena; the cycle slip detection results are the positioning anomaly markers output by the navigation physical tag, used to identify navigation observation jump distortion data, dynamically reduce the weight of abnormal positioning data in the matching calculation, and avoid matching errors caused by positioning jumps; the preset association matching algorithm is a multi-dimensional cost joint matching algorithm, which calculates the intra-frame candidate spatial target matching cost by comprehensively considering feature similarity, position prediction deviation, and positioning anomaly penalty term, and selects the minimum cost combination as the minimum cost combination. The optimal matching relationship is achieved by employing a temporal state iteration update mechanism to synchronously refresh the trajectory position and velocity motion parameters for successfully matched spatial target trajectories across frames, maintaining temporal continuity of the trajectory. For new candidate spatial targets without matching objects or spatial targets that reappear after brief occlusion, an independent trajectory initialization process is executed to create a dedicated trajectory branch. By coordinating the two sets of processing logics—matching update and trajectory initialization—the defects of missing, mismatched, and broken trajectories of spatial targets are eliminated, resulting in a continuous and uninterrupted complete historical trajectory sequence. For example, in the near-Earth orbit space debris monitoring scenario, the cross-modal association features output by multimodal fusion are input into a multi-dimensional cost joint matching algorithm. The matching search interval is limited by combining the short-term motion trend prediction results of each debris target. At the same time, the jump distortion positioning data is masked based on the cycle slip detection results of the navigation physical label, completing the accurate association matching of debris targets in adjacent frames. The motion state and coordinate information of continuously matched debris trajectories are updated in real time, and the trajectory initialization is completed for newly appearing candidate debris targets in the image. Finally, a long-term uninterrupted historical trajectory sequence of multiple spatial targets is generated.
[0050] In one embodiment, the construction and training of the pre-trained hybrid prediction model in step S40 specifically includes: S401. Construct a hybrid prediction model, which includes a time-series coding module, a physical constraint decoding module, and an uncertainty prediction module connected in sequence. S402. Configure the temporal coding module to extract long-range temporal dependencies and local temporal evolution features from historical trajectory sequences, and fuse the two into a trajectory coding representation; configure the physical constraint decoding module to embed the trajectory coding representation, the Doppler frequency shift extrema in the communication physical tag, and the ionospheric delay correction amount in the navigation physical tag as kinematic constraints into the prediction process, and decode to generate multi-scale trajectory prediction results; configure the uncertainty prediction module to output the confidence intervals corresponding to each scale based on the trajectory coding representation and the multi-scale trajectory prediction results. S403. Obtain a sample dataset containing historical trajectory sequence samples and the corresponding future real trajectories of the samples, and divide the sample dataset into a training set and a validation set; S404. Using the training set divided by the historical trajectory sequence samples as input and the corresponding future real trajectory as the supervision label, construct a multi-task joint loss function based on physical consistency loss and uncertainty calibration loss. S405. Based on the multi-task joint loss function, the hybrid prediction model is iteratively trained using the training set until the loss on the validation set meets the preset convergence condition, thus obtaining the trained hybrid prediction model.
[0051] In this embodiment, as described in steps S401-S405 above, specifically: Step S401 is used to construct the overall network architecture of the hybrid prediction model for orbit extrapolation of space targets, and to determine the serial connection relationship of the three core functional modules of the model, providing a basic network carrier for temporal feature extraction, physical constraint decoding, and prediction confidence evaluation. Specifically, the temporal encoding module uses the Transformer-LSTM hybrid temporal algorithm as the feature extraction unit responsible for mining trajectory temporal information, used to analyze the internal variation patterns of long-term trajectories; the physical constraint decoding module uses a multi-layer fully connected decoding algorithm with kinematic physical embedding as the prediction extrapolation unit embedding prior orbital motion, used to combine measured physical labels to constrain trajectory prediction output; and the uncertainty prediction module uses a Gaussian variance estimation algorithm as the error quantization evaluation unit, used to calculate the reliability of prediction results at different scales. The three modules are connected in series in the order of temporal coding module, physical constraint decoding module, and uncertainty prediction module. The output features of the previous module are completely input into the next module as the input for operation, realizing the operation process of temporal feature extraction, physical constraint prediction, and confidence interval solution. For example, in the scenario of near-Earth orbit space debris monitoring, a three-layer series hybrid prediction model is built. The front end is set with a Transformer-LSTM-based temporal coding module to read the historical trajectory of long-term debris. The middle end is connected to a multi-layer fully connected physical constraint decoding module with embedded kinematic equations to introduce Doppler and ionospheric physical parameter constraints for the prediction process. The back end is cascaded with an uncertainty prediction module based on Gaussian variance estimation to output the confidence interval of orbit prediction at each scale, thus completing the basic architecture of the hybrid prediction model.
[0052] Step S402 is used to complete the functional parameters and input / output logic configuration of the three core modules of the hybrid prediction model. It defines the complete computational flow for feature extraction, constraint embedding, and uncertainty quantification for each module, unifies the data interaction rules of the entire model, and ensures the orderly linkage of temporal features, physical constraints, and confidence assessment processes. Specifically, the temporal encoding module performs feature extraction based on the Transformer-LSTM hybrid temporal algorithm, mining long-distance temporal dependencies and short-interval local temporal evolution features from the input continuous historical trajectory sequence. Then, it merges and normalizes the two types of temporal features through a feature fusion and concatenation algorithm, generating a trajectory encoding representation with unified dimensions that considers both global temporal patterns and local motion details. This trajectory encoding representation is then transmitted to the downstream physical constraint decoding module. The physical constraint decoding module performs trajectory deduction based on a multi-layer fully connected decoding algorithm with kinematic physical embedding. It simultaneously receives three types of inputs: the trajectory encoding representation output from the upstream, the Doppler frequency shift extrema extracted from the communication physical tag, and the ionospheric delay correction amount extracted from the navigation physical tag. It uses the Doppler radial velocity and ionospheric error compensation amount as the trajectory... Kinematic constraints are embedded in the decoding computation link, and multi-scale trajectory prediction results of different duration dimensions such as short-term, medium-term and long-term are generated through hierarchical deduction. The trajectory encoding representation and all multi-scale trajectory prediction results are simultaneously sent to the uncertainty prediction module. The uncertainty prediction module performs error quantization calculation based on the Gaussian variance estimation algorithm. Using the trajectory encoding representation and trajectory prediction results of each scale as joint input, it fits the position deviation distribution under each prediction duration, solves and outputs the confidence interval matching the trajectory prediction results of each scale, and quantifies the reliable fluctuation range of orbit predictions at different durations. For example, in the near-Earth orbit space debris monitoring scenario, the Transformer-LSTM temporal encoding module is configured to extract global temporal correlation and local orbit change features from the long-term historical trajectory of debris and fuse them into trajectory encoding. The multi-layer fully connected physical decoding module is configured to introduce Doppler frequency shift extrema and ionospheric delay correction as orbital motion constraints, and generate short-term and medium-term multi-scale debris orbit prediction results through deduction. Then, the Gaussian variance estimation uncertainty module is configured to combine the encoding features and multi-scale prediction values to output the orbit confidence intervals corresponding to each prediction duration.
[0053] Step S403 is used to complete the collection, construction, and normalization of the trajectory sample dataset required for model training. This provides standardized and independent sample data support for subsequent multi-task loss function constraints, model iterative training, and accuracy verification, ensuring the standardization and convergence effectiveness of the model training process. The sample dataset is a specially constructed collection of orbital trajectory samples, corely containing several historical trajectory sequence samples of space targets and future real trajectory labels corresponding to each historical trajectory sample. It comprehensively covers trajectory evolution samples under different orbital states and perturbation environments, ensuring the scenario completeness and data diversity of the sample data. The historical trajectory sequence samples are continuous, uninterrupted long-time-series trajectory data generated through multimodal fusion and inter-frame correlation matching, carrying the complete spatiotemporal motion evolution laws of space targets. The future real trajectory is the actual orbital motion of space targets obtained through synchronous observation. The dynamic trajectory serves as the ground truth supervision for model training. A hierarchical random partitioning algorithm is used to randomly sample and partition the entire dataset. Following a fixed ratio (e.g., 7:3, 8:2), all samples are divided into a training set for iterative updates of model parameters and a validation set for real-time verification of the model's generalization ability and convergence. This strictly ensures that the two datasets have no overlap and consistent distribution, avoiding overfitting and validation failures caused by sample overlap. For example, in near-Earth orbit space debris monitoring scenarios, continuous historical trajectory sequences of space debris and corresponding real future observation trajectories are collected in batches under multiple time periods and perturbation environments. A complete trajectory sample dataset is constructed, and then the dataset is split using a hierarchical random partitioning algorithm to obtain a training set for model parameter training and a validation set for accuracy verification. This provides standardized sample support for subsequent iterative optimization of the hybrid prediction model.
[0054] Step S404 is used to construct a multi-task joint loss function for model training based on the partitioned training set samples and ground truth labels. This function simultaneously optimizes trajectory prediction accuracy and uncertainty quantification reliability through a dual-loss collaborative constraint mechanism, achieving bidirectional correction of the physical compliance of model prediction results and the reliability of error assessment. Specifically, historical trajectory sequence samples from the training set serve as input feature data for the model's forward inference, driving the hybrid prediction model to complete the entire process of temporal encoding, physical decoding, and uncertainty prediction. Future real trajectories serve as global supervision ground truth labels, used to quantify the deviation between the model's prediction output and the actual orbital motion laws. The physical consistency loss is the fundamental loss term constraining the trajectory prediction results to conform to the physical laws of orbital kinematics. It is used to correct trajectory prediction deviations and avoid unreasonable prediction results where velocity and displacement evolution violate the motion characteristics of the spatial target. The formula for calculating the physical consistency loss is:
[0055] in, This represents a loss of physical consistency. This represents the total number of samples in the training set. For the first Multi-scale prediction of location vectors for each sample; For the first Each sample corresponds to a real trajectory position vector; The radial velocity vector is calculated from the extreme values of the Doppler frequency shift. This refers to the inter-frame time interval. This is the ionospheric delay correction vector; this loss constrains the error between the predicted value and the true value through the orbital motion equations, filtering out prediction results that violate the physical rules of space motion.
[0056] The uncertainty calibration loss is a specific loss term adapted to the confidence interval output. It is used to calibrate the confidence interval distribution corresponding to the prediction results at each scale, correct the quantization bias of overconfidence and underconfidence in the model, and ensure that the uncertainty assessment results are true and effective. The formula for calculating the uncertainty calibration loss is as follows:
[0057] in, Calibration loss for uncertainty; For the first The variance of the confidence interval for each sample corresponding to the trajectory prediction scale; this term balances the prediction error and variance estimation, achieving adaptive calibration of the confidence interval.
[0058] A multi-task weighted fusion algorithm is used to weight and fuse the physical consistency loss and the uncertainty calibration loss, constructing a joint loss function that takes into account prediction accuracy, physical compliance, and uncertainty credibility. The formula for the multi-task joint loss function is as follows:
[0059] in, Total loss due to multi-task collaboration; The weighting coefficient for physical consistency loss. To calibrate the loss weighting coefficients for uncertainty, These are preset hyperparameters; The total loss is used to achieve simultaneous optimization and mutual constraint correction of dual tasks. For example, in the scenario of monitoring space debris in near-Earth orbit, a batch of historical trajectory training set samples of debris are input into the hybrid prediction model to obtain multi-scale predicted trajectories and confidence interval variances. The corresponding real observed future trajectories are used as supervision ground values. The loss of physical consistency of orbital motion and the loss of confidence calibration are respectively substituted into the formula to solve the loss function of orbital motion physical consistency and the loss function of confidence calibration. Then, the joint total loss function of multi-task is obtained by weighted fusion of hyperparameters, which simultaneously constrains the physical rationality and uncertainty quantification accuracy of debris trajectory prediction.
[0060] Step S405 uses the constructed multi-task joint loss function as the basis for gradient optimization, performs multiple rounds of parameter iteration updates for the hybrid prediction model based on the partitioned training set, and introduces validation set loss to monitor the model's convergence status in real time. The training process terminates when the preset convergence criteria are met, outputting a hybrid prediction model with stable trajectory prediction and uncertainty quantification capabilities. The iterative training employs an adaptive gradient descent optimization algorithm to update parameters. Each round of training sequentially completes the entire chain of trajectory sample forward inference, total loss calculation, gradient backpropagation, and iterative correction of network weight parameters. In each iteration, validation set samples are simultaneously input into the current network to calculate the validation total loss. The multi-task joint loss function is the core evaluation metric for model iterative optimization, synchronously constraining the parameter update directions of the three modules: temporal encoding, physical constraint decoding, and uncertainty prediction, while balancing the physical compliance of trajectory prediction with the accuracy of confidence interval calibration. The preset convergence criteria are used to terminate the training process. The convergence condition is a pre-configured training termination rule, including two types of judgment logic: the total loss of the validation set decreases below a set threshold for multiple consecutive rounds, and the number of iterations reaches the maximum limit. Training can be terminated when either condition is triggered, avoiding the problems of model overfitting and unnecessary iterations consuming computing power. The validation set does not participate in weight updates throughout the process, but only serves as an independent dataset to objectively evaluate the model's generalization performance, ensuring that the convergence judgment result is unbiased and reliable. For example, in the scenario of near-Earth orbit space debris monitoring, the debris trajectory training set is input into the hybrid prediction model in batches, and adaptive gradient descent iterative training is performed based on the joint total loss of multiple tasks. After each round of training, the validation loss is calculated using an independent debris trajectory validation set, and the trend of loss decrease is continuously monitored. When the change of the total loss of the validation set is below the threshold of 0.001 for 10 consecutive rounds, the convergence condition is met and the training is terminated. Finally, the trained hybrid prediction model with optimized parameters is output and can be directly used for multi-scale orbit prediction.
[0061] In one embodiment, step S40 specifically includes: S41. Input the historical trajectory sequence into the trained hybrid prediction model, use the temporal coding module to extract the long-distance temporal dependency and local temporal evolution features in the historical trajectory sequence, and concatenate the long-distance temporal dependency and local temporal evolution features with the cross-modal association features to form a trajectory coding representation that includes spatiotemporal semantics and physical attributes. S42. The trajectory encoding representation is decoded based on the physical constraint decoding module. During the decoding process, the allowable domain of the predicted velocity is limited by the extreme value of the Doppler frequency shift, and the correction boundary of the predicted position is limited by the ionospheric delay correction amount. The deviation of the excess part is calculated as velocity residual and position residual, and the velocity residual and position residual are embedded as regularization terms in the decoding process to suppress non-physical interpretation. The decoding generates multi-scale trajectory prediction results of multiple spatial targets within a future preset time range. S43. The uncertainty prediction module receives the trajectory encoding representation and multi-scale trajectory prediction results, calculates the prediction uncertainty through a preset uncertainty quantization method, and outputs the confidence intervals corresponding to the trajectories at each scale.
[0062] In this embodiment, as described in steps S41-S43 above, specifically: Step S41 involves inputting the preprocessed continuous historical trajectory sequence into the training convergent hybrid prediction model. The temporal coding module performs two-layer mining of trajectory temporal features and combines this with the cross-modal correlation features obtained in the previous step for cross-domain feature splicing and fusion. This constructs a high-dimensional trajectory coding representation that incorporates temporal evolution patterns, spatial geometric characteristics, and modal physical properties, providing pre-feature input for subsequent physical constraint decoding and uncertainty quantification prediction. The temporal coding module uses the Transformer-LSTM hybrid temporal algorithm, specifically designed for two-layer feature extraction of long-term trajectory data. It captures both long-distance temporal dependencies within the global trajectory sequence and local temporal evolution features between single frames, comprehensively covering the long-term evolution patterns and short-term motion details of spatial target orbits. The cross-modal correlation features are fused features obtained through multi-modal fusion, spatiotemporal geometric constraints, and modal weight suppression optimization, carrying remote sensing geometry and communication information... The system integrates multidimensional physical attributes and spatial correlation information of the signal and navigation positioning. Feature splicing and fusion are achieved through a dimension alignment splicing algorithm, which merges the channel dimensions and normalizes the features of the two-dimensional temporal features and cross-modal correlation features to compensate for the lack of physical information and insufficient semantic correlation of single trajectory features. The final fused trajectory encoding representation includes the spatiotemporal motion semantics, geometric morphological attributes, and signal positioning physical features of the spatial target, and has strong physical prior and high representation accuracy. For example, in the near-Earth orbit space debris monitoring scenario, the continuous and uninterrupted historical trajectory sequence of space debris is input into the trained hybrid prediction model. The long-period orbit temporal change law and short-term local motion features of the debris trajectory are extracted through the temporal coding module. Then, the model is dimensionally aligned and spliced with the optimized cross-modal correlation features of the debris to generate an integrated trajectory encoding representation that integrates multi-source physical information and spatiotemporal information, providing a high-precision feature base for subsequent multi-scale orbit prediction.
[0063] Step S42 is used to perform deep decoding and deduction on the high-dimensional trajectory encoding representation output by S41 through the physical constraint decoding module. Communication and navigation dual-modal physical parameters are introduced into the decoding operation link to implement double-layer boundary constraints. A residual regularization mechanism is used to suppress non-physical outputs that violate spatial motion laws. The deduction yields standardized trajectory prediction results for multiple spatial targets and multiple time periods, ensuring the motion compliance and position accuracy of the predicted trajectory. The physical constraint decoding module employs a multi-layer fully connected decoding algorithm with kinematic physics embedding, using the trajectory encoding representation that integrates spatiotemporal semantics and physical attributes as the core input for multi-dimensional decoding. The trajectory feature decoding mapping is performed; the Doppler frequency shift extremum is the velocity boundary constraint parameter corresponding to the communication physical tag, used to define the legal allowable domain of the space target's orbital motion velocity, and to truncate and control abnormal velocity components exceeding the normal orbital motion rate during the decoding process; the ionospheric delay correction is the position correction constraint parameter corresponding to the navigation physical tag, used to limit the error correction boundary of the space target's predicted position, and to constrain the maximum allowable deviation range of the orbital predicted position; the velocity residual and position residual are the deviation quantization values after physical boundary verification, corresponding to the deviation components of the predicted velocity exceeding the velocity allowable domain and the predicted position exceeding the position allowable domain, respectively. The system sets a correction boundary for the deviation component; embeds two types of residuals as physical regularization terms into the decoding loss and feature mapping process to penalize and correct non-physical and unreasonable prediction components in real time, adaptively suppressing distorted prediction results that do not conform to the laws of orbital kinematics; through a dual correction mechanism of boundary constraints and residual regularization, the trajectory feature decoding and deduction are completed layer by layer. The long-term orbit deduction calculation is completed in the J2000 (ICRF) geocentric inertial coordinate system. The predicted orbit state obtained from the deduction is then transformed to the CGCS2000 geocentric Earth-fixed coordinate system for output. Finally, the system outputs multiple spatial targets at different preset times in the future. Accurate, compliant, and multi-scale trajectory prediction results at various scales; for example, in the scenario of monitoring space debris in near-Earth orbit, the physical constraint decoding module performs dimension-by-dimensional decoding and deduction of the trajectory encoding representation of debris targets, limits the legal range of debris orbital motion velocity based on the extreme value of Doppler frequency shift, locks the prediction position error correction boundary based on the ionospheric delay correction amount, calculates the excess velocity residual and position residual in real time and embeds them into the decoding operation in a regularized form, filters out non-physical prediction solutions with excessive velocity distortion and position offset, and finally outputs stable multi-scale orbital trajectory prediction results of debris targets within different preset time ranges such as short-term, medium-term and long-term.
[0064] Step S43 is used to call the uncertainty prediction module to receive the trajectory encoding representation and multi-scale trajectory prediction results output from the upstream. It employs standardized uncertainty quantification methods to quantify and evaluate the trajectory prediction errors at various time spans, outputting confidence intervals for matching prediction deviation distributions at different scales. This quantifies the reliability of trajectory prediction results across different time spans, providing error references for trajectory situation assessment and risk assessment. Specifically, the uncertainty prediction module uses a Gaussian variance estimation algorithm for quantization, simultaneously reading two types of input data: trajectory encoding representations with both spatiotemporal semantics and physical attributes, and multi-scale trajectory prediction results for multiple spatial targets. The preset uncertainty quantification method is an error estimation algorithm based on Gaussian distribution variance fitting. Based on the spatial target motion characteristics carried by the trajectory encoding and the predicted coordinate deviations at each scale, it fits the data to obtain... The module calculates the variance parameters of the position error for each prediction duration. The confidence interval is generated by converting the variance parameters obtained from the fitting and is used to intuitively represent the normal fluctuation range of the trajectory prediction results at the corresponding scale. The larger the variance value, the higher the uncertainty of the trajectory prediction for that period. The module independently completes the variance fitting and interval solution for each type of prediction scale in the short and medium-to-long term to achieve scale-specific uncertainty calibration. For example, in the near-Earth orbit space debris monitoring scenario, the uncertainty prediction module reads the trajectory encoding representation after the debris target is fused and the short-term and medium-to-long-term multi-scale trajectory prediction results. It solves the position error variance for each prediction period by using the uncertainty quantification method of Gaussian variance fitting, and outputs the confidence interval for each type of debris trajectory at each time scale, intuitively distinguishing the credible fluctuation range of the trajectory prediction results for different extrapolation durations.
[0065] In one embodiment, an AI-based multi-target trajectory prediction system for space-based communication, navigation, and remote sensing data is provided. This AI-based multi-target trajectory prediction system for space-based communication, navigation, and remote sensing data corresponds to the AI-based multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data described in the previous embodiment. The AI-based multi-target trajectory prediction system for space-based communication, navigation, and remote sensing data includes: The data acquisition module is used to collect multi-source heterogeneous data in a distributed manner and perform spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference. The multi-source heterogeneous data includes space-based communication signal data, satellite navigation and positioning data, and space-based space remote sensing image data. The feature extraction module is used to construct a multimodal fusion model with an embedded spatiotemporal attention mechanism, and extract cross-modal correlation features, including communication signal features, navigation and positioning features, and spatial remote sensing image features, based on the multimodal fusion model. The sequence generation module is used to perform multi-spatial target detection and correlation tracking based on the cross-modal correlation features, and generate a continuous and uninterrupted historical trajectory sequence. The hybrid prediction module is used to input the historical trajectory sequence into a pre-trained hybrid prediction model, calculate the long-range temporal dependency and local temporal evolution features of the historical trajectory sequence, output the multi-scale trajectory prediction results of multiple spatial targets within a preset time range based on the long-range temporal dependency and local temporal evolution features, and quantify the prediction uncertainty to generate confidence intervals. The evidence storage module is used to write the multi-source heterogeneous data, model parameters of the multi-modal fusion model and the hybrid prediction model, multi-scale trajectory prediction results and confidence intervals into the consortium blockchain with Byzantine fault-tolerant consensus mechanism for evidence storage, and to complete real-time collision warning at the edge and trend analysis in the cloud through a preset cloud-edge collaborative architecture.
[0066] Specific limitations regarding the multi-target trajectory prediction system for space-based communication, navigation, and remote sensing data based on artificial intelligence can be found in the limitations of the multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence described above, and will not be repeated here. Each module in the aforementioned multi-target trajectory prediction system for space-based communication, navigation, and remote sensing data based on artificial intelligence can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.
[0067] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows. Figure 2 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database is used for data storage, data processing, and data analysis. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-target trajectory prediction method based on artificial intelligence for space-based communication, navigation, and remote sensing data.
[0068] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a multi-target trajectory prediction method based on artificial intelligence-based space-based communication, navigation, and remote sensing data.
[0069] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a multi-target trajectory prediction method based on artificial intelligence-based space-based communication, navigation, and remote sensing data.
[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0071] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.
[0072] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.
Claims
1. A multi-target trajectory prediction method based on artificial intelligence for space-based communication, navigation, and remote sensing data, characterized in that, Includes the following steps: Distributed collection of multi-source heterogeneous data and spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference. The multi-source heterogeneous data includes space-based communication signal data, satellite navigation and positioning data and space-based space remote sensing image data. A multimodal fusion model with embedded spatiotemporal attention mechanism is constructed, and cross-modal correlation features including communication signal features, navigation and positioning features and spatial remote sensing image features are extracted based on the multimodal fusion model; Based on the cross-modal association features, multi-space target detection and association tracking are performed to generate a continuous and uninterrupted historical trajectory sequence. The historical trajectory sequence is input into a pre-trained hybrid prediction model to calculate the long-range temporal dependency and local temporal evolution features of the historical trajectory sequence. Based on the long-range temporal dependency and local temporal evolution features, the model outputs multi-scale trajectory prediction results of multiple spatial targets within a preset time range in the future, and quantifies the prediction uncertainty to generate confidence intervals. The model parameters, multi-scale trajectory prediction results, and confidence intervals of the multi-source heterogeneous data, multi-modal fusion model, and hybrid prediction model are written into a consortium blockchain using a Byzantine fault-tolerant consensus mechanism for storage, and real-time collision warning at the edge and trend analysis in the cloud are completed through a preset cloud-edge collaborative architecture.
2. The multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in claim 1, characterized in that, The steps of distributively collecting multi-source heterogeneous data and performing spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference specifically include: Distributed acquisition includes multi-source heterogeneous data such as space-based communication signal data, satellite navigation and positioning data, and space-based remote sensing image data; The Doppler frequency shift extrema and signal attenuation characteristics are extracted from space-based communication signal data and used as communication physical labels to characterize the radial velocity characteristics and signal transmission loss characteristics of space targets relative to the observation station. Cycle slips are detected from satellite navigation and positioning data, and ionospheric delay corrections are extracted as navigation physical labels that characterize the integrity of satellite navigation and positioning data and the space atmospheric delay characteristics. Extract space target size, attitude angle, and illumination and shadow features from space-based remote sensing image data to serve as remote sensing physical labels characterizing the geometric and physical properties and illumination environment features of space targets; Based on the aforementioned communication physical tags, navigation physical tags, and remote sensing physical tags, physical consistency verification is performed on multi-source heterogeneous data to eliminate abnormal data with tag conflicts, so as to unify them to a preset spatiotemporal reference.
3. The multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in claim 2, characterized in that, The multimodal fusion model includes a spatial attention sublayer, a temporal attention sublayer, and a cross-attention sublayer. The step of extracting cross-modal correlation features, including communication signal features, navigation and positioning features, and spatial remote sensing image features, based on the multimodal fusion model specifically includes: Based on the communication physical tags, navigation physical tags, and remote sensing physical tags, credibility weights corresponding to each multi-source heterogeneous data after physical consistency verification are generated, and the credibility weights are used as suppression coefficients to input into the multimodal fusion model. The multimodal fusion model was used to extract features from the multi-source heterogeneous data after physical consistency verification. Communication signal features were extracted from space-based communication signal data, navigation and positioning features were extracted from satellite navigation and positioning data, and space remote sensing image features were extracted from space-based space remote sensing image data. In the spatial attention sublayer of the multimodal fusion model, navigation and positioning features are used as the benchmark, and the ionospheric delay correction in the navigation physical label is used as the alignment benchmark. The suppression coefficient is used to suppress the spatial attention weights, eliminate the inconsistent associations caused by cycle slips and position jumps, and characterize the spatial position associations between multiple spatial targets. In the temporal attention sublayer of the multimodal fusion model, based on the characteristics of communication signals and the extreme value of Doppler frequency shift in the communication physical label, the suppression coefficient is used to suppress the temporal attention weights and characterize the temporal evolution correlation of the trajectory. In the cross-attention sublayer of the multimodal fusion model, spatial remote sensing image features are used as the geometric reference. Spatial location association and temporal evolution association are introduced as structural constraints into the cross-attention calculation. Spatial target size and attitude angle in remote sensing physical labels are used as the geometric matching reference. Suppression coefficients on the communication side and the navigation side are applied to weight the cross-attention weights to suppress them. Cross-modal interaction between communication signal features and navigation positioning features in the spatial remote sensing image observation space is established, and cross-modal association features are obtained by fusion.
4. The multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in claim 3, characterized in that, The steps of using spatial remote sensing image features as a geometric reference, incorporating spatial location association and temporal evolution association as structural constraints into cross-attention calculation, and using the spatial target size and attitude angle in remote sensing physical labels as a geometric matching reference, applying suppression coefficients from the communication side and the navigation side to weight and suppress the cross-attention weights, establishing cross-modal interaction between communication signal features and navigation positioning features in the spatial remote sensing image observation space, and fusing them to obtain cross-modal association features, specifically include: A query vector for cross-attention is constructed using the spatial remote sensing image features, and a key vector and a value vector are constructed by concatenating the communication signal features and navigation positioning features. Calculate the similarity matrix between the query vector and the key vector, and add the spatial location association as a location bias term and the temporal evolution association as a time decay term to the similarity matrix to obtain the structurally constrained similarity matrix; Based on the spatial target size and attitude angle in the remote sensing physical label, the geometric matching score matrix is calculated, and the geometric matching score matrix is multiplied element-wise with the similarity matrix after structural constraints to obtain the geometrically weighted similarity matrix. The credibility weights generated by the communication physical tags and navigation physical tags are used as suppression coefficients to mask and scale the geometrically weighted similarity matrix to obtain attention weights. The attention weights are then used to weight and aggregate the value vectors to output cross-modal association features.
5. The multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in claim 2, characterized in that, The step of performing multi-spatial target detection and correlation tracking based on the cross-modal correlation features to generate a continuous and uninterrupted historical trajectory sequence specifically includes: The spatial remote sensing image features are used to detect spatial targets to obtain a candidate detection set. The target response intensity of cross-modal correlation features in the spatial remote sensing image observation space is used as the judgment criterion. The spatial target size and attitude angle in the remote sensing physical label are used as geometric constraints to eliminate interfering targets in the candidate detection set, confirm the valid spatial targets, and obtain the initial candidate spatial target set. Based on the Doppler frequency shift extremum in the communication physical tag and the ionospheric delay correction in the navigation physical tag, physical motion constraints are applied to the initial candidate spatial target set to predict motion trends. The cross-modal association features are input into a preset association matching algorithm, which integrates motion trend prediction results and cycle slip detection results in navigation physical tags to perform association matching on spatial targets between consecutive frames. The algorithm updates the status of successfully matched trajectories and initializes the trajectories of unmatched candidate spatial targets, thereby generating a continuous and uninterrupted historical trajectory sequence.
6. The multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in claim 1, characterized in that, The construction and training of the pre-trained hybrid prediction model specifically includes: A hybrid prediction model is constructed, which includes a time-series coding module, a physical constraint decoding module, and an uncertainty prediction module connected in sequence. The temporal coding module is configured to extract long-range temporal dependencies and local temporal evolution features from historical trajectory sequences, and fuse the two into a trajectory coding representation; the physical constraint decoding module is configured to embed the trajectory coding representation, the Doppler frequency shift extrema in the communication physical tag, and the ionospheric delay correction amount in the navigation physical tag as kinematic constraints into the prediction process, and decode to generate multi-scale trajectory prediction results; the uncertainty prediction module is configured to output the confidence intervals corresponding to each scale based on the trajectory coding representation and the multi-scale trajectory prediction results. Obtain a sample dataset containing historical trajectory sequence samples and their corresponding future real trajectories, and divide the sample dataset into a training set and a validation set; Using the training set divided from the historical trajectory sequence samples as input and their corresponding future real trajectories as supervision labels, a multi-task joint loss function is constructed based on physical consistency loss and uncertainty calibration loss. Based on the multi-task joint loss function, the hybrid prediction model is iteratively trained using the training set until the loss on the validation set meets the preset convergence condition, thus obtaining the trained hybrid prediction model.
7. The multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in claim 6, characterized in that, The steps of inputting the historical trajectory sequence into a pre-trained hybrid prediction model, calculating the long-range temporal dependencies and local temporal evolution features of the historical trajectory sequence, outputting multi-scale trajectory prediction results for multiple spatial targets within a preset future time range based on the long-range temporal dependencies and local temporal evolution features, and quantifying prediction uncertainty to generate confidence intervals, specifically include: The historical trajectory sequence is input into the trained hybrid prediction model. The temporal coding module is used to extract long-distance temporal dependencies and local temporal evolution features from the historical trajectory sequence. The long-distance temporal dependencies, local temporal evolution features and cross-modal association features are concatenated and fused into a trajectory coding representation that includes spatiotemporal semantics and physical attributes. The trajectory encoding representation is decoded based on the physical constraint decoding module. During the decoding process, the allowable domain of the predicted velocity is limited by the extreme value of the Doppler frequency shift, and the correction boundary of the predicted position is limited by the ionospheric delay correction amount. The deviation of the excess part is calculated as velocity residual and position residual, and the velocity residual and position residual are embedded as regularization terms in the decoding process to suppress non-physical interpretation. The decoding generates multi-scale trajectory prediction results of multiple spatial targets within a future preset time range. The uncertainty prediction module receives the trajectory encoding representation and multi-scale trajectory prediction results, calculates the prediction uncertainty through a preset uncertainty quantification method, and outputs the confidence intervals corresponding to the trajectories at each scale.
8. A multi-target trajectory prediction system for space-based communication, navigation, and remote sensing data based on artificial intelligence, used to implement the steps of the multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to collect multi-source heterogeneous data in a distributed manner and perform spatiotemporal alignment and standardization preprocessing to unify it to a preset spatiotemporal reference. The multi-source heterogeneous data includes space-based communication signal data, satellite navigation and positioning data, and space-based space remote sensing image data. The feature extraction module is used to construct a multimodal fusion model with an embedded spatiotemporal attention mechanism, and extract cross-modal correlation features, including communication signal features, navigation and positioning features, and spatial remote sensing image features, based on the multimodal fusion model. The sequence generation module is used to perform multi-spatial target detection and correlation tracking based on the cross-modal correlation features, and generate a continuous and uninterrupted historical trajectory sequence. The hybrid prediction module is used to input the historical trajectory sequence into a pre-trained hybrid prediction model, calculate the long-range temporal dependency and local temporal evolution features of the historical trajectory sequence, output the multi-scale trajectory prediction results of multiple spatial targets within a preset time range based on the long-range temporal dependency and local temporal evolution features, and quantify the prediction uncertainty to generate confidence intervals. The evidence storage module is used to write the multi-source heterogeneous data, model parameters of the multi-modal fusion model and the hybrid prediction model, multi-scale trajectory prediction results and confidence intervals into the consortium blockchain with Byzantine fault-tolerant consensus mechanism for evidence storage, and to complete real-time collision warning at the edge and trend analysis in the cloud through a preset cloud-edge collaborative architecture.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in any one of claims 1-7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-target trajectory prediction method for space-based communication, navigation, and remote sensing data based on artificial intelligence as described in any one of claims 1-7.
