Multi-satellite intelligent access method, system and device based on AI prediction

Abstract

The application relates to the technical field of satellite intelligent access, and discloses a multi-satellite intelligent access method, system and equipment based on AI prediction, which comprises the following steps: collecting satellite orbit data and user terminal position data of a multi-satellite system, and performing satellite position prediction through an orbit prediction neural network to obtain satellite predicted positions and communication window information; each intelligent agent perceives communication environment parameters and constructs an environment perception vector; a multi-agent strategy network is driven based on the satellite predicted positions and the environment perception vector to generate a first access strategy; each intelligent agent performs resource conflict processing according to the first access strategy to obtain a second access strategy; the application adopts a multi-agent architecture to replace traditional centralized control, each intelligent agent has independent decision-making capability and realizes cooperation through a distributed negotiation protocol, single-point failure problems are avoided, robustness and scalability are improved, and resource competition and access conflicts are effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of satellite intelligent access technology, and in particular to a multi-satellite intelligent access method, system and device based on AI prediction. Background Technology

[0002] Traditional multi-satellite access methods employ preset access rules and static resource allocation strategies. When faced with complex and ever-changing space environments and user demands, these methods often exhibit slow response times and poor adaptability. In application scenarios with extremely high requirements for connection reliability, such as emergency communications, maritime operations, and communications in remote areas, traditional methods struggle to meet the demands for rapid and accurate satellite access, leading to significant prediction errors and affecting the accurate timing of access. Summary of the Invention

[0003] This invention provides a multi-satellite intelligent access method, system, and device based on AI prediction. This invention adopts a multi-agent architecture to replace the traditional centralized control. Each agent has independent decision-making capabilities and achieves collaboration through a distributed negotiation protocol, avoiding single point of failure, improving robustness and scalability, and effectively reducing resource contention and access conflicts.

[0004] The first aspect of this invention provides a multi-satellite intelligent access method based on AI prediction, the multi-satellite intelligent access method based on AI prediction comprising: Collect satellite orbit data and user terminal location data from multiple satellite systems, and use an orbit prediction neural network to predict satellite positions to obtain predicted satellite positions and communication window information; Each intelligent agent perceives communication environment parameters and constructs an environment perception vector; Based on the satellite predicted location and the environmental perception vector, a multi-agent policy network is driven to generate a first access policy; Each agent performs resource conflict resolution according to the first access strategy to obtain the second access strategy.

[0005] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the step of collecting satellite orbit data and user terminal location data from a multi-satellite system, and predicting satellite positions using an orbit prediction neural network to obtain predicted satellite positions and communication window information, includes: Each intelligent agent acquires the latitude and longitude coordinates, UTC timestamps, and historical trajectory sequences of each satellite in the multi-satellite system in real time, and uses the latitude and longitude coordinates, the UTC timestamps, and the historical trajectory sequences as raw orbit data; The original orbit data is subjected to outlier detection and time-series prediction interpolation to obtain satellite orbit data; Each intelligent agent simultaneously collects the geographic coordinates, motion vector, and altitude of the corresponding user terminal to obtain the user terminal location data; The satellite orbit data and the user terminal location data are input into the orbit prediction neural network to predict the satellite position, thereby obtaining the predicted satellite position and communication window information.

[0006] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of inputting the satellite orbit data and the user terminal location data into an orbit prediction neural network to predict the satellite position and obtain the predicted satellite position and communication window information includes: The satellite orbit data and the user terminal location data are input into the embedding layer of the orbit prediction neural network for feature extraction and feature combination to obtain the input feature vector; The input feature vector is input into the bidirectional LSTM layer of the orbit prediction neural network for forward and backward temporal modeling to obtain the context feature vector; The context feature vector is input into the position prediction branch network of the orbit prediction neural network for calculation to obtain the satellite's predicted position. The context feature vector is input into the communication window evaluation branch network of the orbit prediction neural network to calculate geometric visibility and obtain communication window information.

[0007] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of inputting the context feature vector into the communication window evaluation branch network of the orbit prediction neural network to perform geometric visibility calculation and obtain communication window information includes: The context feature vector is input into the communication window to evaluate the geometric parameter decoding layer of the branch network for feature decoding to obtain the spatial geometric parameters; The spherical geometry calculation module of the branch network, evaluated through the communication window, calculates the angular parameters of each satellite relative to the user terminal based on the spatial geometry parameters. The angle parameter is input into the Earth occlusion judgment module of the communication window evaluation branch network to perform visibility evaluation and obtain the visibility flag vector; The signal propagation loss calculation module of the branch network is evaluated through the communication window, and the communication quality is evaluated based on the visibility flag vector to obtain the communication window information.

[0008] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the process of each intelligent agent perceiving communication environment parameters and constructing an environment perception vector includes: Each intelligent agent synchronously collects the received signal strength RSSI, carrier-to-noise ratio, Doppler frequency shift, link delay, and bit error rate of the communication link to obtain communication environment parameters; The communication environment parameters are cleaned to obtain an environment parameter sequence; The environmental parameter sequence is input into a one-dimensional convolutional neural network for feature fusion to obtain a fused feature vector; Based on the fused feature vector, a perception quality assessment index is calculated, and the perception quality assessment index is used as a weighting coefficient to adjust the fused feature vector to obtain an environmental perception vector.

[0009] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, the step of generating a first access strategy based on the satellite predicted position and the environmental perception vector-driven multi-agent policy network includes: The satellite predicted position, the environmental perception vector, the current queue status information, and the historical behavior records are spliced ​​and combined to form the standardized state vector of each intelligent agent. The standardized state vector is input into the Actor network corresponding to each agent in the multi-agent policy network for independent decision calculation, and the policy feature vector of each agent is obtained. The original action vectors of each agent are obtained by generating multi-task decisions based on the policy feature vectors through the Critic network corresponding to each agent in the multi-agent policy network. Gaussian exploration noise is added to the original action vector and boundary restrictions are applied through an action constraint mechanism to obtain the first access strategy.

[0010] In conjunction with the first aspect, in a sixth implementation of the first aspect of the present invention, each intelligent agent performs resource conflict processing according to the first access strategy to obtain a second access strategy, including: Each agent extracts the target satellite number, time slot allocation, frequency band selection, and power level based on the first access strategy to form an intent vector and broadcasts it to neighboring agents through a distributed communication protocol. At the same time, it receives the intent vectors of neighboring agents and generates a global intent vector set. The resource overlap rate between agents is calculated based on the global intent vector set, and the policy difference is measured by Euclidean distance to obtain the agent cooperation relationship matrix. Based on the aforementioned agent collaboration relationship matrix, conflict detection and priority arbitration are performed to obtain a resource allocation scheme. Based on the resource allocation scheme, the satellite selection, frequency configuration, and power settings of each agent are readjusted and the constraints are verified to obtain the second access strategy.

[0011] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the AI-based prediction-based multi-satellite intelligent access method further includes: Each agent performs satellite access operations according to the second access strategy and monitors the communication link establishment status during the access process in real time to obtain performance indicators; The current state vector, the second access strategy, the performance index, the next state vector, and the termination flag are used to form a state transition tuple and stored in the experience replay buffer. At the same time, the experience sample set is calculated based on the absolute value of the time difference error. Randomly sample from the empirical sample set and calculate the loss functions of each agent's Actor network and Critic network to obtain network parameter gradient information; Based on the network parameter gradient information, the Actor network and Critic network are updated using cosine annealing learning rate scheduling and adaptive exploration noise attenuation mechanism to obtain the updated multi-agent policy network.

[0012] A second aspect of the present invention provides an AI-predictive multi-satellite intelligent access system, the AI-predictive multi-satellite intelligent access system comprising: The acquisition module is used to collect satellite orbit data and user terminal location data from a multi-satellite system, and to predict satellite positions through an orbit prediction neural network to obtain satellite predicted positions and communication window information. A module is built to enable each agent to perceive communication environment parameters and construct environment perception vectors. The generation module is used to drive a multi-agent policy network based on the satellite's predicted position and the environmental perception vector to generate a first access policy; The resource conflict handling module is used by each agent to handle resource conflicts according to the first access strategy to obtain the second access strategy.

[0013] A third aspect of the present invention provides an electronic device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the electronic device to execute the above-described AI-predictive multi-satellite intelligent access method.

[0014] Compared with existing technologies, this invention has the following advantages: It replaces traditional centralized control with a multi-agent architecture, where each agent has independent decision-making capabilities and collaborates through a distributed negotiation protocol, avoiding single-point-of-failure problems and improving robustness and scalability. Simultaneously, through a collaborative metric function and conflict avoidance mechanism, it effectively reduces resource contention and access conflicts. The policy network based on the MADDPG architecture can handle continuous action spaces, generating decisions such as satellite selection, frequency allocation, and power control. Compared to traditional discrete action selection methods, it has stronger decision-making flexibility and adaptability. The Actor-Critic structure organically combines policy optimization and value assessment. It deeply integrates environmental perception information with communication decisions, guiding policy generation through perception quality assessment indicators, achieving a shift from separate processing to integrated collaboration, improving decision accuracy and real-time performance. Simultaneously, it extracts the temporal correlation of perception features through a one-dimensional convolutional neural network. The orbit prediction neural network employing bidirectional LSTM and a multi-head attention mechanism can adaptively learn the impact of complex perturbation factors on satellite orbits, achieving higher prediction accuracy compared to traditional Kepler orbit prediction methods. Through the parallel design of the position prediction branch and the communication window evaluation branch, it simultaneously obtains position and visibility information. Through experience replay buffers and priority sampling mechanisms, the system can continuously learn from historical experience and optimize strategies. Employing cosine annealing learning rate scheduling and adaptive exploration noise attenuation mechanisms, it achieves a dynamic balance between exploration and utilization, enabling autonomous adjustments based on environmental changes. Through intent vector broadcasting and priority arbitration mechanisms, it achieves effective negotiation and resource conflict resolution among agents. The negotiation process makes comprehensive decisions based on multiple dimensions such as perception quality, business urgency, and historical success rate, ensuring fairness and efficiency in resource allocation. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] The structures, proportions, sizes, etc., shown in the accompanying drawings of this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed in the specification, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0017] Figure 1This is a flowchart illustrating the AI-based intelligent access method for multiple satellites provided in an embodiment of the present invention. Figure 2 This is a schematic block diagram of the structure of the AI-predictive multi-satellite intelligent access system provided in the embodiments of the present invention; Figure 3 This is a schematic block diagram of the structure of the electronic device provided in the embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0020] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0021] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items, and all possible combinations, and includes such combinations. See also Figure 1 One embodiment of the AI-predictive-based multi-satellite intelligent access method of the present invention includes: Step 100: Collect satellite orbit data and user terminal location data from the multi-satellite system, and predict satellite positions using an orbit prediction neural network to obtain predicted satellite positions and communication window information; Specifically, multiple intelligent agent nodes deployed in the network acquire in real time the latitude and longitude coordinates, unified coordination time (UTC) timestamps, and historical trajectory sequences for each satellite in the current multi-satellite system. This spatiotemporal vector set, composed of these three types of data, serves as the input set for the raw orbital data. Outlier detection is performed on the raw orbital data, employing a sliding window mechanism with statistical bias constraints combined with the 3σ criterion to identify outliers in the orbital coordinates. After detecting missing segments or discontinuous time points, temporal prediction interpolation is performed using a Long Short-Term Memory (LSTM) network to generate a complete and numerically reliable orbital data sequence. Simultaneously, each intelligent agent collects the three-dimensional geographic location information of the user equipment at its corresponding ground user terminal, including two-dimensional planar geographic coordinates, a motion vector consisting of the direction and velocity of motion, and the altitude corresponding to the current location, constructing the user terminal location data. After concatenating and combining satellite orbit data with user terminal location data, the data is input into the orbit prediction neural network. The orbit prediction neural network adopts a multi-layer structure and integrates LSTM and multi-head attention mechanisms. It can learn the potential higher-order perturbation relationships in orbital motion and output the predicted positions of each satellite in the Earth coordinate system in multiple future time slices. Combined with user terminal location information, Earth curvature model and line-of-sight occlusion judgment logic, the communication window information between the agent and each candidate satellite is derived, specifically including the start time of visibility, the end time of visibility, and the coverage range of the accessible faith angle.

[0022] Step 200: Each agent senses the communication environment parameters and constructs an environment perception vector; Specifically, agents deployed at various access nodes or terminals synchronously collect key physical layer parameters of the current communication link state through local sensing modules. These parameters include Received Signal Strength Indication (RSSI), Carrier-to-Noise Power Ratio (C / N0), Doppler frequency shift, end-to-end link transmission delay, and the measured Bit Error Rate (BER), forming the original set of communication environment parameters. Data cleaning processing is performed on these parameters, including outlier removal based on statistical distribution, short-term smoothing using a sliding window approach, and time-series interpolation repair of discontinuous sampling segments. The output is a consistent and time-aligned sequence of environment parameters. This sequence is then input into a predefined one-dimensional convolutional neural network. The network uses a configuration with a kernel size of 3, a stride of 1, and 64 channels, supplemented by batch normalization layers to enhance numerical stability. The convolutional operation extracts the temporal variation features of the link state using a sliding window approach and outputs a high-dimensional fused feature vector, representing the comprehensive communication environment state perceived by the agent at the current moment. The perceived quality assessment index is calculated based on the fused feature vector. The perceived quality assessment index is constructed by combining the exponential penalty function of link delay, the linear influence term of bit error rate, and the sigmoid mapping function of carrier-to-noise ratio. The perceived quality assessment index numerically reflects the availability of the overall communication environment. The perceived quality assessment index is used as a weighting coefficient to perform element-wise weighted adjustment on each dimension of the fused feature vector to generate an environment perception vector with communication availability offset perception capability.

[0023] Step 300: Based on satellite-predicted location and environmental perception vectors, drive a multi-agent policy network to generate the first access policy; Specifically, a standardized state vector is constructed within each agent node. This standardized state vector is composed of multiple components, including future time-series satellite prediction position information generated by an orbit prediction neural network, an environmental perception vector output by a communication environment perception module, current communication scheduling queue status information, and historical action sequence features constructed based on behavior records within several past time windows. After combination, mean-variance normalization is applied to ensure that the parameters of each dimension remain consistent on the numerical scale. The standardized state vector is then input into a multi-agent policy network based on deep deterministic policy gradients. Each agent corresponds to an independently configured Actor network in the network. The Actor network uses a multilayer perceptron structure as its backbone, employs a nonlinear activation function for layer-by-layer feature transformation, and outputs the agent's policy feature vector, representing its decision bias in the current state. The Critic network corresponding to each agent is invoked. The Critic network receives the current global state and the policy feature vectors of all agents as joint input, and outputs a continuous raw action vector through residual connections and value function regression modules. This raw action vector encodes the agent's prediction of the optimal access behavior at the current moment, including target satellite selection, frequency band search priority, power control allocation factor, and handover policy parameters. To ensure appropriate exploratory nature during policy generation and avoid getting trapped in local optima, a Gaussian exploration noise term with zero mean and dynamic variance control is introduced into each raw action vector. During the action output stage, the action constraint module performs boundary constraints and physical feasibility checks on the raw action vectors, ensuring that the sum of power factors equals a preset upper limit and the handover threshold fluctuates within a defined range. The perturbed and constrained action output is then used as the first access policy vector for the current agent.

[0024] Step 400: Each agent performs resource conflict resolution according to the first access strategy to obtain the second access strategy.

[0025] Specifically, after receiving the first access policy generated by itself based on the multi-agent policy network, each agent extracts four key control information items from the first access policy vector: target satellite number, allocated communication time slot, selected frequency band type, and corresponding transmission power level. These four decision parameters are then encoded into a uniformly structured intent vector. This intent vector is broadcast in real-time to other neighboring agents through a distributed communication protocol interface deployed in the system. Simultaneously, it receives all intent vectors sent from neighboring nodes, constructing a global intent vector set locally. Based on this global intent set, potential conflicts in communication resource usage between different agents are quantitatively analyzed. The policy differences between any two agents are measured by calculating the resource overlap rate of their selected target satellites, frequency bands, and time slots, as well as their Euclidean distance in the access action space. This generates a symmetric agent cooperation relationship matrix, where each item reflects both the intensity of resource competition and the similarity of behavior among agents. The collaborative relationship matrix is ​​analyzed to identify conflict areas and trigger a conflict resolution mechanism. A priority arbitration algorithm is used to allocate ownership of resource items with high conflict concentration. The arbitration mechanism ranks agents based on a weighted priority index composed of their current perceived quality score (QoS), the urgency of service requests, and historical access success rates. Agents with higher priority receive priority in resource allocation. The original first access strategy of each agent is modified according to the resource allocation scheme, including target satellite reselection, frequency configuration adjustment, scaling or upper limit pruning of transmit power factors, etc. The action constraint verification module is then invoked to perform physical boundary legality checks and strategy consistency verification on the modified access parameters, outputting a second access strategy that conforms to the conflict coordination rules and resource feasibility conditions.

[0026] In one specific embodiment, the process of performing step 100 may specifically include the following steps: Each intelligent agent acquires the latitude and longitude coordinates, UTC timestamps, and historical trajectory sequences of each satellite in the multi-satellite system in real time, and uses the latitude and longitude coordinates, UTC timestamps, and historical trajectory sequences as raw orbital data; Outlier detection and time-series prediction interpolation are performed on the raw orbit data to obtain satellite orbit data; Each intelligent agent simultaneously collects the geographic coordinates, motion vector, and altitude of the corresponding user terminal to obtain the user terminal location data; Satellite orbit data and user terminal location data are input into the orbit prediction neural network to predict satellite position, thereby obtaining the predicted satellite position and communication window information.

[0027] Specifically, in the multi-agent architecture, each agent is given an independent data acquisition module, enabling it to maintain a continuous link with the multi-satellite system. A synchronization clock module is also configured to obtain coordinated time information. During actual operation, each agent acquires the current latitude and longitude coordinates from each on-orbit satellite in the system at fixed time intervals. This includes the real-time projection coordinates of the satellite's sub-satellite points relative to the Earth's surface and the corresponding UTC timestamps. Simultaneously, it extracts and maintains the historical trajectory sequence for each satellite. This historical trajectory sequence consists of latitude and longitude trajectory points from multiple consecutive moments, reflecting the temporal continuity and trajectory change patterns of the satellite's on-orbit operation. These three types of data together constitute the original orbital data input set. Considering the existence of data gaps, abrupt changes, and jumps in actual orbital data due to factors such as inter-satellite link interruptions, observational obstruction, synchronization drift, or equipment errors, an outlier detection operation is performed. A sliding window mechanism based on a statistical threshold of three standard deviations is used to identify data points whose rate of position change exceeds the prediction range of the orbital physical model within a short time span, and these points are marked as anomalies. For time fragments or gaps formed after removing anomalies, an interpolation method based on time series prediction is introduced. An LSTM structure is used to construct a time series prediction completion module, which predicts the coordinate values ​​corresponding to the missing time points based on historical trajectories as input, generating satellite orbital data. Simultaneously, each agent on the ground synchronously collects the three-dimensional position status data of the corresponding user terminals within its service range, including: the terminal's current geographic coordinates (longitude and latitude), representing its two-dimensional position distribution on the Earth's surface; the terminal's motion vector, including velocity and orientation angle; and the altitude information of the terminal's location, used to correct for the impact of Earth's curvature on communication visibility, forming a user terminal location data set. After concatenating satellite orbit data with user terminal location data, the data is input into the orbit prediction neural network. The neural network structure employs a deep neural network model consisting of an input layer, an LSTM layer, a multi-head attention mechanism layer, a fully connected mapping layer, and an output layer. The input layer accepts a 128-dimensional vector, including the current satellite latitude and longitude, UTC timestamp, historical trajectory encoding, and relevant user terminal location status information. The LSTM layer captures the temporal variation patterns of the orbit. The attention mechanism learns the importance of the orbital state at different times in influencing the future trajectory and performs weighted fusion to improve prediction accuracy. Subsequently, the fully connected layer performs feature dimensionality reduction and nonlinear mapping, and the output layer outputs the predicted satellite position data for multiple future prediction time slices. Based on the visibility determination algorithm between the user terminal's coordinates and the predicted orbit points, spatial visibility is determined by considering factors such as Earth radius, terminal elevation angle threshold, and trajectory curvature. This derives the communication window information between each agent and the corresponding candidate satellite, including the continuous time interval from the start of the satellite's visibility to the end of obstruction, the elevation angle change trend corresponding to the window, and link quality fluctuation estimates.

[0028] In one specific embodiment, the process of inputting satellite orbit data and user terminal location data into an orbit prediction neural network to predict satellite position and obtain predicted satellite position and communication window information can specifically include the following steps: Satellite orbit data and user terminal location data are input into the embedding layer of the orbit prediction neural network for feature extraction and feature combination to obtain the input feature vector; The input feature vector is input into the bidirectional LSTM layer of the trajectory prediction neural network to perform forward and backward temporal modeling, thereby obtaining the context feature vector; The context feature vector is input into the position prediction branch of the orbit prediction neural network for calculation to obtain the satellite's predicted position. The context feature vector is input into the communication window evaluation branch of the orbit prediction neural network to calculate geometric visibility and obtain the communication window information.

[0029] Specifically, a neural network structure for orbit prediction with multi-branch output capability is constructed, and an embedding layer module for spatiotemporal information encoding is introduced at the model input to uniformly represent multimodal input features. Satellite orbit data and user terminal location data are structurally aligned. Satellite orbit data includes multiple continuous temporal features such as latitude and longitude coordinates, UTC timestamps, trajectory change vectors, and velocity direction information, while user terminal location data includes static and dynamic attributes such as 3D geographic coordinates, motion vectors, altitude, and terrain constraint identifiers. These two types of features are concatenated to form a composite input vector, which is then input into the embedding layer. The embedding layer consists of multiple linear transformation modules and nonlinear activation functions, capable of mapping the original high-dimensional heterogeneous features to a unified low-dimensional vector space. Simultaneously, a positional encoding method is used to introduce temporal continuity labels, ensuring the recognizability of the time series context. After initial feature fusion and compression, a standardized input feature vector is output. The input feature vector is passed to a bidirectional LSTM layer of the neural network for deep temporal modeling. The LSTM layer consists of two independent recurrent units, forward and backward. The forward unit processes trajectory information sequentially to capture historical dependencies, while the backward unit scans the time series in reverse to identify future state constraints. The hidden state vectors output by both units are concatenated to form a context feature vector, representing the orbital context representation at the current time point, combining past trajectories and future trends, and possessing the ability to model long-term dependencies and perceive short-term disturbances. The context feature vector is then input into two functional branches of the orbit prediction neural network to complete position prediction and communication window evaluation. The position prediction branch consists of a fully connected mapping layer and a position decoder, which performs a nonlinear projection operation on the context feature vector to decode the potential orbital change trend into a sequence of future predicted trajectory points in the Earth coordinate system. The output format is the longitude, latitude, and altitude information of the satellite in several consecutive time slices, used to reconstruct the satellite's trajectory in three-dimensional space. The communication window evaluation branch network is used for spatial visibility determination and link reachability assessment. The communication window evaluation branch network performs geometric relationship parsing on the context feature vector, mapping the context feature vector into a relative spatial angle vector and a motion rate tensor. It uses the geometric visibility model to calculate the angle change between the user terminal's location and the predicted satellite trajectory, and combines geographical constraints such as Earth curvature, obstruction height, and minimum elevation angle threshold to determine whether there is a continuous communicable path. When the judgment conditions are met, it marks the corresponding time period as a valid communication window and outputs communication window information, including the visibility start and end time, window duration, reachable angle range, and link stability index.

[0030] In one specific embodiment, the process of inputting the context feature vector into the communication window evaluation branch network of the orbit prediction neural network to perform geometric visibility calculation and obtain communication window information can specifically include the following steps: The context feature vector is input into the communication window to evaluate the geometric parameters of the branch network. The decoding layer performs feature decoding to obtain the spatial geometric parameters. The spherical geometry calculation module of the branch network is evaluated through the communication window to calculate the angle parameters of each satellite relative to the user terminal based on the spatial geometry parameters. The angle parameter is input into the communication window to evaluate the Earth occlusion judgment module of the branch network for visibility assessment, and a visibility flag vector is obtained. The signal propagation loss calculation module of the branch network is evaluated through the communication window, and the communication quality is evaluated based on the visibility flag vector to obtain the communication window information.

[0031] Specifically, the current orbital state and user position state are back-interpreted based on contextual feature vectors. The contextual feature vectors are input into the geometric parameter decoding layer of the communication window evaluation branch network. This decoding layer consists of multiple fully connected mapping layers and nonlinear activation functions, and combines learnable parameters trained in the network to perform back-projection and decoding operations on the feature vectors. This restores the abstract orbital state representation to structured spatial geometric parameters, including the three-dimensional position vectors of the satellite and user terminal in the geocentric coordinate system, the unit vector of the satellite's orbital direction, the geographic orientation and relative altitude of the user terminal, and the instantaneous velocity and acceleration vectors of the satellite relative to the terminal. These spatial geometric parameters are then input into the spherical geometry calculation module of the communication window evaluation branch network. This module models and calculates the relative angle information between the satellite and user terminal, such as the viewing angle, azimuth angle, and elevation angle, based on spherical trigonometric relationships. The satellite position vector and the user position vector are normalized to unit vectors, and the angle between them is calculated by vector dot product to obtain the center line-of-sight angle between the satellite and the user. Then, the elevation angle parameter is calculated by combining the user terminal's orientation vector and the satellite's relative displacement direction vector. The Earth spherical model is introduced to correct the azimuth deviation, resulting in a three-dimensional angle parameter set, including the center angle, elevation angle, azimuth angle, polar angle, and projection angle, to describe whether the communication link has the basic conditions for geographical visibility. The three-dimensional angle parameters are input into the Earth occlusion judgment module for visibility determination. The Earth occlusion judgment module performs logical judgment based on spherical geometric constraints and the elevation angle threshold model. If the elevation angle value is greater than the preset minimum passable belief angle threshold (between 5° and 10°), and the path line represented by the center angle does not cross the Earth's surface, then the user terminal is determined to have line-of-sight access to the satellite at the current moment, and the visibility flag is output as "1"; otherwise, if the path is blocked by the curvature of the Earth, or the elevation angle is below the antenna's field of view, the visibility flag is output as "0". The binary result vector composed of the visibility status of all satellites and terminals is the visibility flag vector. The visibility marker vector is input into the signal propagation loss calculation module, which calculates the link quality index based on the effective line-of-sight term in the marker vector. The link distance for each visible satellite is solved, and the basic propagation attenuation is calculated using a free-space path loss model based on the three-dimensional spatial distance between the satellite and the terminal in the geocentric coordinate system. Angle correction loss is calculated based on the elevation angle parameter between the terminal and the satellite, combined with a surface obstruction probability model and an atmospheric refraction attenuation model. An additional channel attenuation term is introduced by the meteorological disturbance coefficients (such as cloud cover, water vapor content, etc.) that may exist in the region during the current time period. All loss terms are then weighted and summed to convert the result into an estimated average link signal-to-noise ratio (SNR) per unit time. The SNR estimate is then combined with the defined communication window validity judgment rule: when the SNR is greater than a certain minimum communicable threshold (e.g., 10 dB) and the line-of-sight condition is met, the corresponding time period is marked as an effective communication window.Through the calculation process, the communication window information between each satellite and the user terminal within a specific time period is output, including the start and end time of the window, the duration of the communication window, the link signal-to-noise ratio change curve within the window, the communication interruption prediction point, and the window stability level score.

[0032] In one specific embodiment, the process of performing step 200 may specifically include the following steps: Each intelligent agent synchronously collects the received signal strength RSSI, carrier-to-noise ratio, Doppler frequency shift, link delay, and bit error rate of the communication link to obtain communication environment parameters; The communication environment parameters are cleaned to obtain an environmental parameter sequence; The environmental parameter sequence is input into a one-dimensional convolutional neural network for feature fusion to obtain a fused feature vector; The perception quality assessment index is calculated based on the fused feature vector, and the perception quality assessment index is used as a weighting coefficient to adjust the fused feature vector to obtain the environmental perception vector.

[0033] Specifically, each agent is equipped with an independent communication environment perception module, which continuously monitors the connected satellite communication link and periodically collects multiple physical layer parameters at high frequency, including received signal strength RSSI, carrier noise ratio (C / N0), Doppler frequency shift, link delay, and bit error rate. These parameters reflect the power of signal propagation, link signal-to-noise ratio, frequency offset, round-trip transmission efficiency, and bit-level transmission reliability, respectively. They constitute a five-tuple of communication environment, characterizing the time-varying and dynamic fluctuation characteristics of the communication link, and are recorded in a structured manner in a local buffer queue in real time. Considering the measurement errors caused by uncertainties such as sudden interference, obstruction, numerical jitter, and observational anomalies in actual communication links, a data cleaning process is performed on the raw communication parameter data after acquisition. A sliding window is constructed based on the historical statistical characteristic values ​​of each parameter, and its mean and standard deviation are calculated. Outliers are removed using a three-standard-deviation criterion. For local abrupt changes, a moving median filter is used for smoothing and repair. For missing or discontinuous data segments, interpolation and regression methods are used for completion. Linear interpolation is used for stationary parameters such as RSSI and C / N0, while a time series prediction interpolation algorithm based on LSTM is used for fluctuating parameters such as Doppler shift and bit error rate to enhance the fitting ability of dynamic trends. A communication environment parameter sequence is constructed and standardized into a unified data format for neural network input. A multidimensional environmental parameter sequence is input into a one-dimensional convolutional neural network for feature fusion processing. The neural network structure consists of multiple stacked Conv1D convolutional layers and ReLU nonlinear activation units. The kernel size is set to 3, the stride to 1, and the number of channels to 64. A sliding window method is used to extract the local correlations between five types of communication parameters in the time dimension. The ability of the model to express the coexistence of short-term disturbances and long-term trends is enhanced by stacking layers. At the same time, a batch normalization module is placed after each convolutional layer to improve numerical stability and accelerate convergence efficiency, outputting a set of fused feature vectors of fixed length. The fused feature vectors are then used to calculate a perceptual quality assessment index to help identify the availability level of the current link in the access strategy. The perceptual quality assessment index comprehensively considers the joint effects of link latency, bit error rate, and carrier-to-noise ratio. The normalized average link delay is applied in an exponentially decaying form as a penalty factor. The bit error rate is then mapped to a negative coefficient in a first-order linear manner. The Sigmoid function is used to map the C / N0 value to the [0,1] interval to introduce nonlinear dynamic enhancement characteristics. The product of the three results forms the perceived quality score (QoS) value, with a numerical range limited to [0,1] to reflect the overall quality level of the communication link. The QoS value is used as the weight coefficient of the fused feature vector, and each dimension of the vector is adjusted element-wise. This weakens the interference of unstable features on the subsequent policy network under low-quality link conditions, while enhancing the driving effect of perceived features on policy generation under high-quality conditions, thus constructing an environment-aware vector with communication quality sensitivity.

[0034] In one specific embodiment, the process of performing step 300 may specifically include the following steps: The satellite predicted position, environmental perception vector, current queue status information, and historical behavior records are spliced ​​and combined to form the standardized state vector of each intelligent agent. The standardized state vector is input into the Actor network corresponding to each agent in the multi-agent policy network to perform independent decision calculations, thereby obtaining the policy feature vector of each agent. The original action vectors of each agent are obtained by generating multi-task decisions based on the policy feature vectors through the Critic network corresponding to each agent in the multi-agent policy network. Gaussian exploration noise is added to the original action vector and boundary restrictions are applied through an action constraint mechanism to obtain the first access strategy.

[0035] Specifically, a standardized state vector construction module is established in the multi-agent system architecture. This module is designed for each agent with independent perception and decision-making capabilities. In each policy update cycle, it extracts real-time state information from four core data sources and performs format concatenation. It acquires the position data of the candidate satellites corresponding to the current agent in the prediction time slice, output by the orbit prediction neural network. This data includes satellite latitude and longitude coordinates, altitude, prediction residuals, and confidence scores, and optionally incorporates linear extrapolation parameters of future orbit trends for trajectory prediction. Simultaneously, it acquires environmental perception vectors from the perception module after cleaning and convolutional feature fusion. These vectors integrate communication link metrics such as RSSI, C / N0, Doppler shift, link delay, and bit error rate. It also acquires statistical data on the communication queue status from the resource scheduling module, including the number of waiting tasks, the distribution of remaining task duration, the histogram of task priority distribution, and historical congestion records. Finally, it acquires historical behavior records, maintaining the policy action sequences taken by each agent in the past several time slots and their corresponding reward feedback to capture behavioral inertia and learning bias trends. After mapping the four types of data to numerical vector form, a concatenation operation is performed, and the data is normalized using mean and standard deviation to form a standardized state vector with consistent dimensions and stable values. The standardized state vectors are then input into the Actor network corresponding to each agent. The Actor network serves as the policy generation branch, and its internal structure consists of three fully connected neural networks. ReLU activation and LayerNorm layers are used to improve training stability. The output of the last layer is compressed to the [-1,1] interval using the tanh activation function. The vector dimension is set to 12 dimensions, representing satellite selection probability distribution, frequency band search priority weight, power control scaling factor, and switching threshold offset, etc., to constitute the current agent's preferred access behavior intention. The policy feature vectors are then input into each Critic network. The Critic network structure receives the current state vectors of all agents and the policy vectors output by the corresponding Actor networks, concatenates them as joint input, and performs value function estimation through a multi-task evaluation channel. With the support of a deep residual structure, it outputs the value estimate Q-value of the current joint state-action pair. Simultaneously, along the Actor path, policy features are transformed into action vectors in an operable form. These action vectors maintain the same dimensionality as the policy output, but each term represents an executable control value for the corresponding parameter, such as a satellite index, frequency band sorting index, normalized power factor, and channel switching threshold in dB. A Gaussian exploration noise mechanism is introduced after the Actor outputs the action vectors. A normally distributed noise vector with zero mean and dynamic standard deviation is added to the action output, forming perturbed candidate solutions. Throughout the training cycle, exponential decay or cosine annealing mechanisms are used to gradually reduce the noise variance, achieving a smooth transition from exploration to convergence.To ensure that the action output meets the requirements of physical feasibility and system resource constraints, after the action vector is generated, the action constraint module performs item-by-item boundary pruning and normalization verification operations. This includes performing softmax normalization on the satellite selection probability vector to ensure that its sum is 1, pruning the power factor to between 0 and 1 and ensuring that the total allocated power does not exceed the equipment limit, limiting the handover threshold to the range of [-3,3]dB and introducing a smoothing regularization term to suppress large and frequent handover behavior. The output continuous action vector after disturbance and constraint verification is used as the first access strategy and submitted to the execution module for actual scheduling decision.

[0036] In one specific embodiment, the process of performing step 400 may specifically include the following steps: Each agent extracts the target satellite number, time slot allocation, frequency band selection, and power level based on the first access strategy to form an intent vector and broadcasts it to neighboring agents through a distributed communication protocol. At the same time, it receives the intent vectors of neighboring agents and generates a global intent vector set. The resource overlap rate between agents is calculated based on the global intent vector set, and the policy difference is measured by Euclidean distance to obtain the agent cooperation relationship matrix. Conflict detection and priority arbitration are performed based on the agent cooperative relationship matrix to obtain a resource allocation scheme; Based on the resource allocation scheme, the satellite selection, frequency configuration, and power settings of each agent are readjusted and the constraints are verified to obtain the second access strategy.

[0037] Specifically, a structurally unified policy decoding module is set up in the policy execution phase of the multi-agent system. This module operates on each agent, extracting four resource control parameters upon receiving its own generated first access policy: the currently selected target satellite number, the allocated communication time slot number, the frequency band category identifier, and the transmission power level. These four control elements are then concatenated in a fixed order into a structured intent vector. The intent vector represents the agent's communication access intent and resource allocation plan at the current moment using a hybrid encoding method of integer index and floating-point power value. The intent vector is broadcast via a lightweight distributed communication protocol deployed in the system, with its propagation range limited to the set of physically nearby communicable agents. During the broadcast period, the agent synchronously listens for and receives intent vector data packets sent by surrounding agents. After reception, the intent vector data packets sent by surrounding agents are combined with the local intent vector to form a global intent vector set. Conflict assessment is performed based on the global intent vector set. The resource usage overlap between each pair of agents in the set is calculated and analyzed. A resource overlap rate calculation method is used to evaluate the degree of intersection of each pair of agents in the target satellite number, communication time slot number, and frequency band selection fields. If two agents select the same target satellite and have the same frequency band field, and their allocated time slots overlap, they are identified as candidate pairs for resource conflict. The overlap rate is defined as the ratio of the intersection length to the union length of these three items, with a value range of [0,1]. Simultaneously, each agent's first access policy vector is considered a vector sample in the point set. The Euclidean distance is used to measure the numerical difference between its policy vector and those of other agents, assessing the similarity or deviation of policy behaviors. The resource overlap rate and policy Euclidean distance are integrated to construct a symmetric agent cooperation relationship matrix. The main diagonal elements represent the agent's own cooperation score, and the off-diagonal elements represent the cooperation conflict measure between any two agents, quantifying the degree of mutual coupling of all agents in the system regarding communication resource usage. Based on the agent collaboration matrix, a conflict detection and arbitration mechanism is initiated. Agents with a median value higher than a preset conflict threshold are screened one by one, and priority allocation is performed on the most significant conflict groups. The arbitration process adopts a weighted ranking rule, and the priority of each agent is determined by three factors: QoS score, reflecting the quality level of its current communication link; business urgency score, calculated based on the time limit and priority of tasks pending in the queue; and historical success rate, recording the number of times and success rate of successful access by the agent in the past scheduling cycles. The three scores are combined according to weighted coefficients (e.g., QoS 0.5, urgency 0.3, historical record 0.2) to form a total priority score. All conflicting agents are ranked from highest to lowest, with higher-ranked agents retaining their original first access strategy configuration, while lower-ranked agents readjust their access configuration based on resource availability.After resource arbitration is completed, a resource allocation plan is generated based on the adjustment results and fed back to each agent to update access policy parameters. For agents involved in reconfiguration, their satellite selection field is replaced with the one with the longest communication window or the best link quality among the currently available satellites. The frequency configuration field is reallocated to a non-conflicting frequency band, and the power setting is adjusted within a safe range according to the remaining link budget to be no lower than the minimum communication threshold. After the above adjustments are completed, each field is again subject to boundary verification through an action constraint mechanism to ensure parameter legality, including that the power factor does not exceed the physical transmission limit, the time slot number does not exceed the boundary, and the frequency band selection conforms to the equipment's support capabilities. Once all verifications are passed, the second access policy after conflict coordination is formed.

[0038] In one specific embodiment, the multi-satellite intelligent access method based on AI prediction further includes the following steps: Each agent performs satellite access operations according to the second access strategy and monitors the communication link establishment status in real time during the access process to obtain performance indicators; The current state vector, the second access strategy, the performance index, the next state vector, and the termination flag are used to form a state transition tuple and stored in the experience replay buffer. At the same time, the experience sample set is calculated based on the absolute value of the time difference error. Randomly sample from the empirical sample set and calculate the loss functions of each agent's Actor network and Critic network to obtain the network parameter gradient information; Based on the network parameter gradient information, the Actor network and Critic network are updated using cosine annealing learning rate scheduling and adaptive exploration noise attenuation mechanism to obtain the updated multi-agent policy network.

[0039] Specifically, each agent executes the satellite access operation procedure based on the target satellite number, communication frequency band selection, time slot allocation, and power control factor parameters included in the second access strategy. This involves sending resource requests and access link establishment signaling to the designated satellite and monitoring the access status in real time during the establishment process at the physical and link layers. This includes dynamic indicators such as link establishment delay, response ACK feedback, access success flag, retransmission count, and signal-to-noise ratio changes. The access evaluation module performs statistical reduction on the access process to obtain multiple performance indicators in dimensions such as access success rate, average establishment delay, link stability score, and effective throughput. After each round of decision execution, the agent's original state vector (i.e., standardized perception input), the currently adopted second access strategy (i.e., execution action vector), the performance indicators generated during the access process (as reward or loss signals), the next state vector after access completion (generated by the perception module at the next moment), and a Boolean flag indicating whether the termination state has been entered are encapsulated into a state transition tuple, stored in a locally maintained experience replay buffer, and marked with the current timestamp. The time difference (TD) error between the current agent's predicted state-action value function Q(s,a) by the Critic network and the predicted Q(s',a') by the target network is calculated. The absolute value of this TD error is then added with a small positive bias term to form a priority index for that sample. This index is used to construct a priority empirical sample set with the TD error magnitude as the sampling weight, and normalization is performed. During each round of policy network training, several state transition tuples are randomly sampled from the empirical sample set according to a priority-weighted distribution to form a training batch. These batches are then input into the Actor and Critic networks respectively to calculate their loss functions. The loss of the Critic network is defined based on the mean squared error as the squared deviation between the predicted Q value and the target Q value, while the loss of the Actor network is solved based on the gradient direction of the negative dominance function in the policy gradient method. This solution is then backpropagated to the network parameters to form parameter update gradient information. To improve training stability and convergence efficiency, a cosine annealing learning rate scheduling mechanism is introduced during the optimization process. This involves setting an initial maximum and minimum learning rate, and periodically adjusting the learning rate cosinely according to the current training steps. This ensures a high update rate in the early stages of training for thorough exploration, while gradually reducing the learning rate in the later stages to stabilize the convergence direction. Simultaneously, an adaptive exploration noise attenuation mechanism is deployed, dynamically adjusting the exploration noise variance based on the increasing training steps. This maintains high noise levels early on to enhance exploration capabilities, while gradually attenuating the noise to a minimum during the later convergence phase to prevent policy oscillations. The gradient and scheduling information mentioned above are used to update the parameter weights in the Actor and Critic networks corresponding to each agent, forming the updated multi-agent policy network structure.

[0040] The AI-based prediction-based multi-satellite intelligent access method in the embodiments of the present invention has been described above. The AI-based prediction-based multi-satellite intelligent access system in the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 2 One embodiment of the AI-predictive multi-satellite intelligent access system of the present invention includes: The acquisition module 11 is used to acquire satellite orbit data and user terminal location data of the multi-satellite system, and to predict satellite positions through an orbit prediction neural network to obtain satellite predicted positions and communication window information. Module 12 is used for each agent to perceive communication environment parameters and construct environment perception vectors; Generation module 13 is used to drive a multi-agent policy network based on satellite-predicted position and environmental perception vectors to generate a first access policy; The resource conflict handling module 14 is used by each intelligent agent to handle resource conflicts according to the first access strategy and obtain the second access strategy.

[0041] Through the collaborative efforts of the aforementioned components, a multi-agent architecture replaces traditional centralized control. Each agent possesses independent decision-making capabilities and collaborates through a distributed negotiation protocol, avoiding single points of failure and improving robustness and scalability. Simultaneously, resource contention and access conflicts are effectively reduced through collaborative metric functions and conflict avoidance mechanisms. The policy network based on the MADDPG architecture can handle continuous action spaces, generating decisions such as satellite selection, frequency allocation, and power control. Compared to traditional discrete action selection methods, it exhibits greater decision-making flexibility and adaptability. The Actor-Critic structure organically combines policy optimization and value assessment. Deeply integrating environmental perception information with communication decisions, and guiding policy generation through perception quality assessment indicators, it achieves a shift from separate processing to integrated collaboration, improving decision accuracy and real-time performance. Furthermore, a one-dimensional convolutional neural network extracts the temporal correlation of perception features. The orbit prediction neural network employing bidirectional LSTM and multi-head attention mechanisms can adaptively learn the impact of complex perturbation factors on satellite orbits, achieving higher prediction accuracy compared to traditional Kepler orbit prediction methods. Through the parallel design of the position prediction branch and the communication window evaluation branch, it simultaneously obtains position and visibility information. Through experience replay buffers and priority sampling mechanisms, the system can continuously learn from historical experience and optimize strategies. Employing cosine annealing learning rate scheduling and adaptive exploration noise attenuation mechanisms, it achieves a dynamic balance between exploration and utilization, enabling autonomous adjustments based on environmental changes. Through intent vector broadcasting and priority arbitration mechanisms, it achieves effective negotiation and resource conflict resolution among agents. The negotiation process makes comprehensive decisions based on multiple dimensions such as perception quality, business urgency, and historical success rate, ensuring fairness and efficiency in resource allocation.

[0042] Please see Figure 3 , Figure 3 The present invention provides a schematic block diagram of the structure of an electronic device 300. The electronic device 300 includes a processor 301 and a memory 302, which are connected through a system bus 303. The memory 302 may include a non-volatile storage medium and internal memory.

[0043] The non-volatile storage medium can store a computer program. The computer program includes program instructions that, when executed by the processor 301, cause the processor 301 to perform any of the aforementioned AI-predictive multi-satellite intelligent access methods.

[0044] The processor 301 provides computing and control capabilities to support the operation of the entire electronic device 300.

[0045] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor 301, the processor 301 can execute any of the above-mentioned AI-predictive multi-satellite intelligent access methods.

[0046] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the electronic device 300 involved in the present invention. The specific electronic device 300 may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0047] It should be understood that processor 301 can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0048] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device 300 described above can be referred to the corresponding process of the aforementioned AI-based prediction-based multi-satellite intelligent access method, and will not be repeated here.

[0049] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0050] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0051] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention.

Claims

1. A multi-satellite intelligent access method based on AI prediction, characterized in that, include: Collect satellite orbit data and user terminal location data from multiple satellite systems, and use an orbit prediction neural network to predict satellite positions to obtain predicted satellite positions and communication window information; Each intelligent agent perceives communication environment parameters and constructs an environment perception vector; Based on the satellite predicted location and the environmental perception vector, a multi-agent policy network is driven to generate a first access policy; Each agent performs resource conflict resolution according to the first access strategy to obtain the second access strategy.

2. The AI-based intelligent multi-satellite access method according to claim 1, characterized in that, The process involves collecting satellite orbit data from a multi-satellite system and user terminal location data, and then using an orbit prediction neural network to predict satellite positions to obtain predicted satellite positions and communication window information, including: Each intelligent agent acquires the latitude and longitude coordinates, UTC timestamps, and historical trajectory sequences of each satellite in the multi-satellite system in real time, and uses the latitude and longitude coordinates, the UTC timestamps, and the historical trajectory sequences as raw orbit data; The original orbit data is subjected to outlier detection and time-series prediction interpolation to obtain satellite orbit data; Each intelligent agent simultaneously collects the geographic coordinates, motion vector, and altitude of the corresponding user terminal to obtain the user terminal location data; The satellite orbit data and the user terminal location data are input into the orbit prediction neural network to predict the satellite position, thereby obtaining the predicted satellite position and communication window information.

3. The AI-based intelligent multi-satellite access method according to claim 2, characterized in that, The step of inputting the satellite orbit data and the user terminal location data into the orbit prediction neural network to predict the satellite position and obtain the predicted satellite position and communication window information includes: The satellite orbit data and the user terminal location data are input into the embedding layer of the orbit prediction neural network for feature extraction and feature combination to obtain the input feature vector; The input feature vector is input into the bidirectional LSTM layer of the orbit prediction neural network for forward and backward temporal modeling to obtain the context feature vector; The context feature vector is input into the position prediction branch network of the orbit prediction neural network for calculation to obtain the satellite's predicted position. The context feature vector is input into the communication window evaluation branch network of the orbit prediction neural network to calculate geometric visibility and obtain communication window information.

4. The AI-based intelligent multi-satellite access method according to claim 3, characterized in that, The step of inputting the context feature vector into the communication window evaluation branch network of the orbit prediction neural network to perform geometric visibility calculation and obtain communication window information includes: The context feature vector is input into the communication window to evaluate the geometric parameter decoding layer of the branch network for feature decoding to obtain the spatial geometric parameters; The spherical geometry calculation module of the branch network, evaluated through the communication window, calculates the angular parameters of each satellite relative to the user terminal based on the spatial geometry parameters. The angle parameter is input into the Earth occlusion judgment module of the communication window evaluation branch network to perform visibility evaluation and obtain the visibility flag vector; The signal propagation loss calculation module of the branch network is evaluated through the communication window, and the communication quality is evaluated based on the visibility flag vector to obtain the communication window information.

5. The AI-based intelligent multi-satellite access method according to claim 1, characterized in that, Each intelligent agent perceives communication environment parameters and constructs an environment perception vector, including: Each intelligent agent synchronously collects the received signal strength RSSI, carrier-to-noise ratio, Doppler frequency shift, link delay, and bit error rate of the communication link to obtain communication environment parameters; The communication environment parameters are cleaned to obtain an environment parameter sequence; The environmental parameter sequence is input into a one-dimensional convolutional neural network for feature fusion to obtain a fused feature vector; Based on the fused feature vector, a perception quality assessment index is calculated, and the perception quality assessment index is used as a weighting coefficient to adjust the fused feature vector to obtain an environmental perception vector.

6. The AI-based intelligent multi-satellite access method according to claim 1, characterized in that, The first access policy is generated by driving a multi-agent policy network based on the satellite predicted position and the environmental perception vector, including: The satellite predicted position, the environmental perception vector, the current queue status information, and the historical behavior records are spliced ​​and combined to form the standardized state vector of each intelligent agent. The standardized state vector is input into the Actor network corresponding to each agent in the multi-agent policy network for independent decision calculation, and the policy feature vector of each agent is obtained. The original action vectors of each agent are obtained by generating multi-task decisions based on the policy feature vectors through the Critic network corresponding to each agent in the multi-agent policy network. Gaussian exploration noise is added to the original action vector and boundary restrictions are applied through an action constraint mechanism to obtain the first access strategy.

7. The AI-based intelligent multi-satellite access method according to claim 1, characterized in that, Each intelligent agent performs resource conflict handling according to the first access strategy to obtain a second access strategy, including: Each agent extracts the target satellite number, time slot allocation, frequency band selection, and power level based on the first access strategy to form an intent vector and broadcasts it to neighboring agents through a distributed communication protocol. At the same time, it receives the intent vectors of neighboring agents and generates a global intent vector set. The resource overlap rate between agents is calculated based on the global intent vector set, and the policy difference is measured by Euclidean distance to obtain the agent cooperation relationship matrix. Based on the aforementioned agent collaboration relationship matrix, conflict detection and priority arbitration are performed to obtain a resource allocation scheme. Based on the resource allocation scheme, the satellite selection, frequency configuration, and power settings of each agent are readjusted and the constraints are verified to obtain the second access strategy.

8. The AI-based intelligent multi-satellite access method according to claim 1, characterized in that, The AI-based prediction-based multi-satellite intelligent access method also includes: Each agent performs satellite access operations according to the second access strategy and monitors the communication link establishment status during the access process in real time to obtain performance indicators; The current state vector, the second access strategy, the performance index, the next state vector, and the termination flag are used to form a state transition tuple and stored in the experience replay buffer. At the same time, the experience sample set is calculated based on the absolute value of the time difference error. Randomly sample from the empirical sample set and calculate the loss functions of each agent's Actor network and Critic network to obtain network parameter gradient information; Based on the network parameter gradient information, the Actor network and Critic network are updated using cosine annealing learning rate scheduling and adaptive exploration noise attenuation mechanism to obtain the updated multi-agent policy network.

9. A multi-satellite intelligent access system based on AI prediction, characterized in that, The method for performing the AI-predictive multi-satellite intelligent access method as described in any one of claims 1-8 includes: The acquisition module is used to collect satellite orbit data and user terminal location data from a multi-satellite system, and to predict satellite positions through an orbit prediction neural network to obtain satellite predicted positions and communication window information. A module is built to enable each agent to perceive communication environment parameters and construct environment perception vectors. The generation module is used to drive a multi-agent policy network based on the satellite's predicted position and the environmental perception vector to generate a first access policy; The resource conflict handling module is used by each agent to handle resource conflicts according to the first access strategy to obtain the second access strategy.

10. An electronic device, characterized in that, The electronic device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the electronic device to execute the AI-predictive multi-satellite intelligent access method as described in any one of claims 1-8.

Images