Unmanned aerial vehicle cluster cross-domain control and data calibration system based on satellite communication cooperation

Abstract

The application discloses a satellite communication cooperation-based unmanned aerial vehicle cluster cross-domain control and data calibration system and belongs to the technical field of unmanned aerial vehicle control; the system comprises a multi-constellation heterogeneous fusion communication module, a dynamic cross-domain autonomous domain control module, an endogenous benchmark dynamic data calibration module and an endogenous security protection module; the system adopts a dynamic cross-domain autonomous domain architecture and does not depend on any single satellite; the system is free of a global control center and realizes distributed self-organization; the system realizes centimeter-level data calibration in any area without a ground benchmark station or an anchor node through an endogenous space-time benchmark; and the application solves the defects in communication, control, data calibration and security protection in the prior art.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) control technology, specifically relating to a cross-domain control and data calibration system for UAV swarms based on satellite communication collaboration. Background Technology

[0002] Cross-domain control and data calibration of UAV swarms based on satellite communication is a core technology of an integrated air-space-ground unmanned system. It uses satellite communication networks as a global link to integrate multiple UAVs scattered in different geographical areas and operational domains into an organic whole, realizing unified command, collaborative perception and precise operation. It also uses high-precision spatiotemporal references provided by satellites to systematically calibrate and fuse the status data and perception data of all nodes, ultimately forming a swarm efficiency of "1+1>2".

[0003] Currently, cross-domain UAV control and data calibration technologies based on satellite communication still have at least the following technical shortcomings: 1) Satellite communication relies on a single constellation, which poses single points of failure and geographical risks. The separation of communication, navigation and computing leads to resource waste and weak anti-interference capabilities. 2) Cross-domain control adopts a centralized architecture, which has poor scalability, high response latency, insufficient cross-domain collaboration, and excessive reliance on human intervention in the loop; 3) Data calibration relies entirely on external benchmarks; accuracy drops sharply in areas without benchmarks; calibration is not real-time; and heterogeneous data fusion is difficult. 4) Security protection relies on traditional cryptography, has weak anti-spoofing capabilities, and lacks an effective distributed trust mechanism. Summary of the Invention

[0004] This invention provides a cross-domain control and data calibration system for UAV swarms based on satellite communication collaboration. It is based on a heterogeneous communication architecture of multi-constellation dynamic aggregation and high-altitude pseudo-satellite network supplementation, without relying on any single satellite. It adopts a dynamic cross-domain autonomous domain architecture, without a global control center, and realizes distributed self-organization. Through the swarm's endogenous spatiotemporal reference, it can achieve centimeter-level data calibration in any region without the need for ground reference stations or anchor nodes. This invention solves the defects of existing technologies in terms of communication, control, data calibration, and security protection.

[0005] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution: A satellite communication-based cross-domain control and data calibration system for unmanned aerial vehicle (UAV) swarms includes: The multi-constellation heterogeneous converged communication module includes: The multi-constellation dynamic aggregation and slicing unit is used to monitor the channel quality, bandwidth, latency, and anti-interference capability of available satellites worldwide in real time. According to actual mission requirements, the communication link is divided into low-latency and high-reliability control slices, high-bandwidth sensing slices, and high-precision calibration slices. Each slice uses links from 2 to 3 different constellations for aggregation and transmission at the same time. If a single link fails, it will not affect the completion of the overall communication transmission. The communication signal-derived navigation unit utilizes the time difference of arrival, angle of arrival, and Doppler frequency shift of the communication signal for navigation and positioning. It does not require a dedicated navigation signal, and can achieve meter-level positioning using only the communication signal even when the navigation signal is completely interfered with. After the communication signal is fused with the satellite navigation signal, the positioning accuracy is improved to the centimeter level. The satellite-machine distributed edge computing unit dynamically distributes computing tasks among the satellite's on-board processor, the UAV's edge processor, and the ground server; simple data calibration and decision-making tasks are completed on the UAV's local edge processor, while complex fusion tasks are completed on the satellite; it is configured to only transmit the final results and key data back to the ground, reducing the satellite communication bandwidth requirements. The dynamic cross-domain autonomous system control module uses an improved 3D DBSCAN algorithm to automatically divide platforms with task similarity >85% and geographical distance <100km into an autonomous system. An autonomous system has a minimum of 3 nodes and a maximum of 50 nodes. If there are more than 50 nodes, it will automatically split into multiple subdomains. An improved Practical Byzantine Fault Tolerance (PBFT) algorithm is used for domain leader election. Different autonomous systems exchange model parameters through domain leader nodes to achieve global collaboration. The cluster-endogenous benchmark dynamic data calibration module selects at least three nodes as benchmark nodes, and other nodes perform relative observations with the benchmark nodes. After fusing the observation data, a unified endogenous spatiotemporal benchmark for the entire cluster is generated. The sensors have their own built-in data acquisition components, and the parameters of the sensors are corrected in real time using a pre-trained Transformer model. Cross-validation is performed between sensors of the same type to achieve sensor calibration. For sensor data with different domain differences, a multimodal fusion model based on an attention mechanism is used to transform the data to the unified endogenous spatiotemporal benchmark. The overall observation redundancy of the cluster is used to eliminate random and systematic errors of individual sensors. The intrinsic security protection module uses a Physically Unclonable Function (PUF) to assign a unique PUF identification fingerprint to the UAV; it maintains a local blockchain for each autonomous region, which uses an improved RAFT consensus algorithm to prevent operation record data from being tampered with; it adopts a triple anti-interference technology combining frequency hopping, spread spectrum, and beamforming, and uses consistency verification of multi-constellation observation data to detect navigation signal spoofing attacks.

[0006] Preferably, the method for dividing the autonomous region is as follows: 1) Input parameter preprocessing and feature vectorization: All input parameters are converted into computable numerical vectors. One-hot encoding is used for task type vectors, and complex tasks can be represented as multi-hot vectors. Geographic coordinates are converted to geocentric rectangular coordinates (X,Y,Z) to avoid errors in spherical distance calculation. Time windows are represented as [start time, end time], and the time overlap between two tasks is calculated. For platform type weights: the adaptability (0~1) of different platforms for performing different tasks is predefined, such as the weight of UAV performing reconnaissance tasks = 0.9, and the weight of unmanned vehicle performing reconnaissance tasks = 0.6. The remaining energy is normalized to the [0,1] space. 2) 3D task similarity calculation: Define a comprehensive distance metric function; the smaller the distance, the higher the similarity. in, For task type distance, use cosine distance to calculate the difference between two task vectors; For geographic distance, the spherical distance is calculated using the Haversine formula and normalized to [0,1]. The time distance is calculated as 1 minus the time overlap. When the times do not overlap at all, D_time = 1. , , The sum of the three is 1, and the default setting is... , , ; Similarity conversion formula: similarity S = 1 - D. When S > 0.85, the two nodes belong to the same potential autonomous region. 3) Improved 3D DBSCAN clustering algorithm: Initialize all nodes to an unvisited state, and randomly select an unvisited node P; Calculate the ε-neighborhood of P: the set of all nodes with a combined distance D < ε from P and a similarity S > 0.85; if the neighborhood size is ≥ 3, then P is a core node, and a new autonomous region C is created; add P and all nodes in its neighborhood to C and mark them as visited; recursively process each node in C: if the node is a core node, add the unvisited nodes in its neighborhood to C; repeat the above operation until all nodes have been visited. For isolated nodes, those with a similarity greater than 0.7 are merged into the nearest autonomous system, and the remaining isolated nodes are processed separately. 4) Dynamic partitioning and merging of autonomous regions: When the number of nodes in an autonomous region exceeds 50, partitioning is performed. The partitioning mechanism is as follows: Extract the task feature vectors of all nodes within the autonomous system; use the K-means algorithm to divide the nodes into K subsets, where K = ceil, calculated as the number of nodes / 40; reserve 10 nodes for expansion space; re-execute 3D DBSCAN on each subset to ensure that the number of nodes in each subdomain is between 3 and 50; elect a new domain master node for each subdomain. The merging mechanism is as follows: when the average similarity S between two adjacent autonomous regions is greater than 0.9 and the total number of nodes is less than 40, they are automatically merged into one autonomous region. 5) Incremental update mechanism in dynamic scenarios When a new node is added, only the average similarity between the node and all existing autonomous regions is calculated, and the node is added to the autonomous region with the highest similarity and less than 45 nodes. When a node leaves, check the size of its autonomous system. If it is less than 3, merge the remaining nodes into other similar autonomous systems. When task parameters change, only the similarity of the affected nodes is recalculated, and their respective autonomous regions are adjusted.

[0007] Preferably, the domain leader election method is as follows: 1) Initial elections immediately after the autonomous region is established 1.1) All nodes calculate their own and their neighbors' real-time reputation values ​​according to the formula: Reputation score R = 0.4 * Communication quality + 0.3 * Remaining energy + 0.2 * Computing power + 0.1 * Historical performance Each node maintains a globally sorted list of reputation values, which is updated synchronously via the Gossip protocol; the node with the highest reputation value automatically becomes the initial candidate node, without waiting for triggering conditions. 1.2) Nomination Phase: The initial candidate node broadcasts a Nominate message containing its own ID, current highest reputation value, and view number v=0; other nodes, upon receiving the nomination message, perform three verifications: a) verify the validity of the message signature; b) verify that the candidate node's reputation value is indeed the highest known at present; c) verify that the view number matches the local view number; if the verification passes, an Acknowledge message is sent to the candidate node; otherwise, the malicious behavior is ignored and recorded. 1.3) Confirmation Phase: Candidate nodes collect confirmation messages. When they receive valid confirmations from at least 2f+1 different nodes (f=⌊(n-1) / 3⌋), the candidate nodes use the BLS algorithm to aggregate all confirmation signatures, generate an Announce message, and broadcast it to the entire network. After receiving the Announce message, all nodes verify the aggregated signatures and update their local domain master information and view number. The election is complete, and the new domain master begins sending heartbeat messages. 2) Normal operation and heartbeat mechanism The domain owner sends a heartbeat message every 100ms, containing a summary of the current cluster status; normal nodes set a heartbeat timeout of 300ms, and if no heartbeat is received within the timeout period, the domain owner is considered to be faulty; the domain owner attaches the latest node list and reputation ranking to the heartbeat, and all nodes are updated synchronously; when the reputation value of a node changes by more than 10%, it actively broadcasts an update message, triggering a network-wide synchronization; 3) View Change If any node does not receive a heartbeat from the domain master within 300ms, a view change is immediately triggered. The node updates its local view number from v to v+1. The node sends a ViewChange message to the top-ranked node based on its latest reputation value. When the top-ranked node receives at least 2f+1 ViewChange messages, it becomes a new candidate node. The new candidate node initiates a new round of elections, and the process is exactly the same as the initial election. If the election is not completed within 1 second, the view number is automatically incremented by 1, and the next node with the highest reputation value initiates the election. 4) Handling the addition or departure of dynamic nodes When a new node joins, it sends a join request to the current domain owner. After verifying its identity, the domain owner adds it to the node list and broadcasts the updated node list and the fault tolerance threshold f. When a node leaves normally, it sends a Leave message, the domain owner updates the node list and the f value, and broadcasts the notification to the entire network. When a node leaves abnormally, the domain owner marks it as a faulty node through heartbeat timeout detection, updates the node list and f value, and broadcasts a notification to the entire network. Each time the node list changes, all nodes automatically recalculate f = ⌊(n-1) / 3⌋.

[0008] Preferably, model parameters are exchanged between different autonomous regions to achieve global collaborative multi-agent decision-making using federated reinforcement learning and attention mechanisms. Each autonomous region trains its model locally, uploading only the model parameters and not the original data. The specific implementation method is as follows: 1) Preliminary Preparation and System Initialization: After each autonomous system is established and a domain master node is elected through the improved PBFT algorithm, the domain master node randomly initializes a set of global domain model parameters. The domain master distributes these parameters to all UAV nodes within its domain via satellite communication links to ensure that all nodes have completely consistent initial models. The domain master nodes of all autonomous systems establish point-to-point communication connections to form a global collaborative network. Each domain master node maintains a list of basic information for all other domains, including domain ID, task type, geographical range, and current model version number. 2) Local training loop within the autonomous domain: At the beginning of each training cycle, the domain master node sends the latest domain model parameters to all drones in the domain. After the drones load the new parameters, they replace the old local model. While performing actual tasks, all drones interact with the environment using the current model; each drone independently collects its own status, actions, rewards, and next status data to form a local experience trajectory; all raw experience data is stored only locally on the drone. Once each drone has collected enough empirical data, after a default of 100 steps, the gradient of the model parameters is calculated locally using the Proximal Policy Optimization (PPO) algorithm. The calculation process is entirely completed on the drone's edge processor, generating only gradient vectors and not retaining the original empirical data. All drones encrypt the calculated gradient vectors and upload them to the domain master node. After collecting the gradients of all drones in the domain, the domain master node uses the FedAvg algorithm to perform weighted aggregation, with the weights proportional to the number of experience samples collected by each drone. The domain master node updates the domain model parameters using the aggregated global gradient. After the update is complete, the next training cycle begins, and the above steps are repeated; each autonomous region performs local training independently without interfering with others. 3) Cross-Autonomous Domain Attention Fusion Process: After an autonomous domain completes 5 rounds of local training, its domain master node broadcasts the parameters of the current domain model to all other domain master nodes in the global collaborative network; at the same time, the domain master also receives model parameters sent by all other domains. Each domain master node calculates the cosine similarity between each other based on the task feature vectors of its own domain and other domains, including task type, geographical environment, target characteristics, time window, etc. The higher the similarity, the more relevant the tasks of the two domains are, and the more valuable the experience of the other is. The calculated inter-domain similarity values ​​are input into the softmax function to generate an attention weight vector; the larger the weight value, the higher the proportion of the model in the corresponding domain during fusion; the base weight of the model in this domain is fixed at 0.5, and all other domains share the remaining 0.5 weight; Each domain master node performs a weighted sum of the model parameters from all other domains based on the generated attention weights, and then fuses this sum with the model parameters from its own domain; the fusion formula is: New model parameters = 0.5 × parameters of this domain + Σ(attention weights i × parameters of other domains i) The domain master node distributes the merged new model parameters to all drones in the domain, replacing the original domain model; after the drones load the new model, they continue local training and task execution. 4) Dynamic scene adaptive processing: When the number of nodes in an autonomous system exceeds 50, it is automatically split into two subdomains. The domain master nodes of the two new subdomains inherit the latest model parameters of the original autonomous system as their respective initial domain models. Then, the two subdomains independently perform local training and cross-domain fusion. When two adjacent autonomous systems meet the merging conditions, the master node of the merged new domain uses an attention mechanism to fuse the model parameters of the two original domains to generate a new initial domain model; the fusion weight is determined based on the task similarity and the number of nodes in the two domains. When a new node joins an autonomous region, it directly downloads the latest domain model parameters from the domain master node and immediately participates in training and task execution. When a node leaves normally or abnormally, the domain master node automatically excludes the gradient of that node during the next gradient aggregation, without affecting the overall training process.

[0009] Preferably, the sensor incorporates at least a miniature inertial measurement unit (IMU) and a temperature sensor. The Transformer model adopts a temporal Transformer encoder architecture. The input layer consists of a 32×7-dimensional tensor, representing 32 frames of 3-axis acceleration, 3-axis angular velocity, and 1-axis temperature. The embedding layer maps the 7-dimensional original features to 64-dimensional temporal features. The encoder layer consists of a 2-layer multi-head self-attention network (2 heads) and a feedforward network (128 dimensions). The output layer contains a 6-dimensional error correction, representing 3-axis acceleration deviation and 3-axis angular velocity deviation. Model pre-training: Construction of pre-training dataset: In a laboratory environment, the sensor was placed in a high and low temperature chamber to collect static data over the entire temperature range of -40℃ to +85℃; dynamic calibration data under different angular velocities and accelerations was collected using a high-precision turntable; at least 100 hours of multi-condition data were collected for each sensor model to build a general pre-training model; Transfer learning fine-tuning: Before each drone leaves the factory, it undergoes one hour of static and simple dynamic calibration at room temperature; the data from this one hour is used to fine-tune the general pre-trained model to adapt to the individual differences of individual sensors; the fine-tuned model parameters are stored locally on the drone as the initial calibration model. Real-time inference process: The sensor collects raw IMU and temperature data, which are then preprocessed to generate a 32-frame sliding window. The window data is input into a pre-trained Transformer model, which outputs a 6-dimensional error correction for the current frame. The model's prediction error is subtracted from the original measurement value to obtain the corrected IMU data. The corrected data is then output to the navigation and control system and saved for subsequent cross-validation. The cross-validation process for the same type of sensors is as follows: Trigger condition settings: Cross-validation of the same type of sensor is initiated when all of the following conditions are met: At least three drones of the same model were in the same autonomous region; The similarity of the motion states of all drones is >90%, which is determined by the cosine similarity of acceleration and angular velocity. The ambient temperature difference for all drones is less than 5°C. Distributed consensus computation: Data synchronization: Each drone broadcasts its latest corrected IMU data (up to 1 second) to all other drones within the autonomous system; Baseline value calculation: Each drone calculates the median of all data from the same type of sensors, which is used as the cluster baseline value at the current moment; Deviation calculation: Each drone calculates the deviation between its corrected data and the cluster baseline value; Anomaly detection: If the deviation of a certain drone exceeds the 3σ threshold continuously, where σ is the standard deviation of the cluster data, it is determined that the sensor has abnormal drift. Online model correction: When slow drift is detected in the sensor, with a deviation of <0.5m / s² or <5° / s, online model fine-tuning is initiated. The local Transformer model is incrementally learned using the cluster baseline value from the most recent 10 minutes as the ground truth. The learning rate is set to 1 / 100 of that used during pre-training, and only the parameters of the last two layers of the model are updated to avoid catastrophic forgetting. After fine-tuning, the old local model is automatically replaced, and real-time calibration continues.

[0010] Preferably, the sensor data includes different domains such as air, ground, sea, and underwater. A three-stage pipeline of "spatiotemporal alignment - cross-modal attention fusion - redundancy error elimination" is constructed to losslessly transform all heterogeneous data to the cluster's endogenous spatiotemporal reference. The specific method is as follows: 1) Spatiotemporal alignment of heterogeneous data a) Unified time standard All sensor data are globally synchronized using an endogenous spatiotemporal reference within the cluster, with a time synchronization accuracy of <1μs. For sensors with different sampling rates (e.g., camera 30Hz, LiDAR 10Hz, IMU 200Hz), an asynchronous time registration algorithm based on cubic spline interpolation is employed: the highest sampling rate, IMU 200Hz, is used as the reference time axis; interpolation is performed on the data from lower sampling rate sensors to generate an equally spaced data sequence aligned with the reference time axis; the interpolation weights are dynamically adjusted based on the sensor's timestamp error and historical synchronization accuracy. b. Spatial benchmark unification Each sensor has an independent local coordinate system. The transformation matrix from the local coordinate system to the cluster's endogenous reference coordinate system is calculated in real time using a dynamic extrinsic parameter calibration algorithm. The rotation matrix R and translation vector T are: Initial extrinsic parameter acquisition: A rough calibration is performed when the UAV leaves the factory to obtain the initial transformation matrix; Dynamic extrinsic parameter update: Real-time correction of extrinsic parameters using relative observations between clusters (LiDAR point cloud matching, visual feature matching, UWB ranging); External parameter reliability assessment: The reliability of the external parameter matrix is ​​calculated based on the observation quality (number of feature points, matching accuracy, ranging error); Coordinate transformation: All sensor data are transformed to the intrinsic reference coordinate system using the formula P_global = R × P_local + T; c. Single-modal data preprocessing Standardization preprocessing is performed on data from different modalities to eliminate dimensional differences between modalities: Image data: Normalized to [0,1], then processed for distortion correction and white balance; Point cloud data: downsampled to a uniform density (e.g., 10,000 points / frame), outliers removed; IMU data: Output after real-time correction by the Transformer as described above; Radar / sonar data: converted to polar coordinate grid plot, intensity values ​​normalized; 2) Multimodal feature fusion based on cross-modal attention a. Design of a lightweight cross-modal attention fusion model It adopts a dual-stream cross-attention Transformer architecture, with the total number of parameters controlled within 1 million, and runs in real time on the drone edge processor; Input layer: Inputs preprocessed multimodal data and generates respective modal feature sequences through respective modal feature extractors; Cross-attention layer: Each modality feature performs cross-attention calculations with all other modality features; Modal adaptive weighted layer: Automatically calculates the credibility weight of each modality based on the scene; Feature fusion layer: Weighted fusion of all modal features to generate a global feature representation under a unified benchmark; Output: fused 3D coordinates, semantic information, and confidence level; b. Implementation of Modality Adaptive Attention Mechanism Scene feature extraction: Extract current scene features from global features (such as light intensity, weather conditions, and terrain type). Modal credibility prediction: A credibility score of 0 to 1 is predicted for each modality based on scene features using a small MLP network; Attention weight adjustment: Multiply the modality confidence score by the attention score of the cross-attention output to obtain the final fusion weight; Weighted fusion: All modal features are summed according to their final weights to obtain the fused global features; c. Model Training and Update Pre-training: Collect multimodal datasets in a laboratory environment and train a general fusion model; Federated fine-tuning: Each autonomous region fine-tunes the model locally using real-world task data, uploading only the model parameters to the global collaboration layer; Incremental update: When a new sensor type is added, it can be quickly adapted through transfer learning without retraining the entire model; 3) Error elimination based on cluster observation redundancy: By utilizing repeated observations of the same target from multiple platforms in the cluster, the random error and systematic bias of a single sensor are eliminated simultaneously through a distributed Bayesian consensus fusion algorithm. a. Observation redundancy detection: Redundancy error elimination is initiated when the following conditions are met: The same target was observed by at least three different platforms; The time difference between all observations is <100ms; The spatial overlap of all observations is >60%; b. Distributed Bayesian Consistency Fusion Each node performs local computation: Each platform calculates the probability distribution, mean μ_i, and covariance Σ_i of the target state based on its own fusion results; Simultaneously, the reliability w_i of its own observation is calculated, which is jointly determined by the sensor accuracy, external parameter reliability, and observation distance; Global consistency computation: All platforms will broadcast their local μ_i, Σ_i, and w_i to other platforms within the autonomous system; Each platform receives observation data from all other platforms and calculates the global weighted mean and covariance: Systematic bias correction: Calculate the observation bias for each platform: ; If the deviation of a certain platform continuously exceeds the 3σ threshold, it is determined that there is a systematic deviation. The deviation values ​​are used to correct the sensor extrinsic parameters and fusion model of the platform online; c. Anomaly detection removal Robust statistical methods were used to remove outlier observations: Using the median absolute deviation (MAD) instead of the standard deviation reduces the impact of outliers; Observations with a deviation exceeding 5 times the MAD are directly identified as abnormal and removed. Platforms that exhibit abnormal observations will be flagged, and the credibility weight of their subsequent observations will be temporarily reduced.

[0011] Preferably, the autonomous domain is equipped with a local blockchain to prevent operation record data from being tampered with, and the specific method is as follows: 1) Transaction submission and block generation process The business module generates operation instructions and calculates the hash value of the original data; it signs the transaction using the node's private key and sends it to the transaction pool; the leader node collects transactions from the transaction pool, packages them into a new block, and broadcasts the new block to all follower nodes; the follower nodes verify the legality of the block: verify the leader's signature; verify the signatures of all transactions; verify whether the hash of the previous block is correct; and verify whether the transaction content conforms to the business rules. After successful verification, the followers append the block to the local blockchain and update the state database; when the leader receives confirmation from 2f+1 nodes, the block is finally confirmed, and the business module is notified that the operation is complete. 2) Data verification and traceability process When it is necessary to verify the legitimacy of an operation, the latest data is retrieved from the state database, the hash value of the data is calculated, and it is compared with the hash of the corresponding transaction data on the blockchain. If the hashes match, it means that the data has not been tampered with; otherwise, it means that the data has been tampered with. If it is necessary to trace the operation history, all related transactions on the blockchain can be queried through the transaction ID. All operations can be traced back to the specific sending node and timestamp. 3) Domain leader election synchronized with blockchain When a failure of the current domain leader is detected, an improved RAFT election is triggered. The candidate node with the highest reputation value becomes the new leader. The new leader broadcasts the latest block height and state hash, and all follower nodes synchronize blocks with the new leader to ensure ledger consistency. The new leader generates a "domain leader election" transaction, which is recorded on the blockchain. After the election is completed, the new leader begins to receive transactions and generate new blocks. 4) Safety and fault tolerance mechanisms a. Anti-tampering mechanism The blockchain uses the SHA-256 hash chain, and any modification to a block will invalidate the hashes of all subsequent blocks. All transactions must be digitally signed by the sender to prevent forgery. All blocks must be signed by the leader to prevent malicious nodes from generating invalid blocks; Each update to the state database must correspond to a transaction on the blockchain; b. Malicious Node Defense Reputation penalty mechanism: Malicious behavior, such as sending invalid transactions, forging blocks, or denial of service, will result in a permanent decrease in reputation score; Nodes with a reputation score below 0.2 will be automatically excluded from the consensus process; A multi-signature verification mechanism is adopted, and key operations, such as domain leader election and model update, require multiple nodes to sign together; Abnormal behavior detection: Automatically identify potential malicious nodes by analyzing node behavior patterns; c. Data integrity assurance Each node maintains a complete copy of the blockchain, so a single point of failure will not result in data loss. Regularly perform ledger consistency checks to ensure that the ledgers of all nodes are consistent; Important data can be backed up on multiple nodes, and backup records are also stored on the blockchain.

[0012] Preferably, the specific implementation method of the triple anti-interference technology combining frequency hopping, spread spectrum, and beamforming is as follows: 1) Adaptive frequency hopping Frequency hopping pattern generation: Dynamic frequency hopping seeds are generated based on the cluster's endogenous spatiotemporal reference, with each autonomous region using an independent frequency hopping seed; the frequency hopping sequence is generated using a chaotic algorithm with a period > 10^12, making it unpredictable; the frequency hopping bandwidth covers the entire available frequency band (e.g., Ku band 10.7~14.5GHz), and the frequency hopping rate is 1000 hops / second. Real-time spectrum sensing and interference avoidance: Each drone is equipped with a miniature spectrum analyzer that continuously scans the entire frequency hopping band; it updates the interference spectrum map every 10ms, marking the interfered frequencies; it dynamically adjusts the frequency hopping pattern to automatically avoid interfered frequencies, with a frequency hopping pattern update cycle of 1 second. Cluster synchronization frequency hopping: All nodes achieve microsecond-level time synchronization through an endogenous spatiotemporal reference; the domain master node broadcasts the latest frequency hopping seed and frequency hopping pattern every 1 second; all nodes switch to the same frequency at the same time to ensure uninterrupted cluster communication; 2) Direct sequence spread spectrum Spread code generation: A Gold code of length 1023 is used as the spreading code, with a code rate of 10 Mbps. Each UAV node uses a unique spreading code, and different autonomous systems use orthogonal spreading code groups. The spreading code seed is also generated based on an endogenous spatiotemporal reference and is updated periodically. Transmitter spread spectrum processing: The original data (rate = 1Mbps) is XORed with the spreading code to generate a spread spectrum signal; the bandwidth of the spread spectrum signal is expanded to 10MHz, and the processing gain is 10dB; the spread spectrum signal is then sent to a frequency hopping modulator for frequency hopping processing. Despreading at the receiver: The receiver first performs frequency hopping demodulation to obtain the spread spectrum signal; then it uses the locally synchronized spreading code for correlation despreading to recover the original data; finally, it uses a RAKE receiver to process the multipath signal, further improving the receiving sensitivity. 3) Digital Adaptive Beamforming Antenna array configuration: Each drone is equipped with a 4-element conformal phased array antenna, operating in the Ku band; the antenna array is integrated on the surface of the drone fuselage, without affecting aerodynamic performance; each antenna element is independently controllable, and can achieve a beam scanning range of ±60°; Adaptive beamforming algorithm: The Least Mean Square (LMS) adaptive algorithm is used to adjust the weighting coefficients of the antenna elements in real time; the main lobe is aligned with the satellite with the direction of arrival (DOA) of the satellite signal as a reference; at the same time, the DOA of the interference source is detected to form a null with a depth >50dB in the direction of interference. Multi-beam simultaneous communication: It supports the simultaneous formation of 4 independent beams, each aimed at 4 different satellites; enables simultaneous reception and transmission of signals from multiple constellations; and automatically switches to the beams of other satellites when a satellite is interfered with. The triple anti-interference technology adopts a layered and progressive approach to form an anti-interference system: First layer: Adaptive frequency hopping technology to avoid narrowband interference, making it impossible for the jammer to aim; The second layer: direct sequence spread spectrum technology, which resists broadband noise interference and extracts the signal from the noise; The third layer: digital adaptive beamforming technology, which suppresses directional interference while enhancing useful signals; Based on the consistency of the multi-constellation observation data, the implementation process for detecting spoofing attacks on navigation signals is as follows: 1) Multi-constellation observation data acquisition: Simultaneously receive navigation signals from at least 4 different satellite constellations, namely: BeiDou-3: MEO+GEO+IGSO hybrid constellation; GPS: MEO constellation; Galileo: MEO constellation; GLONASS: MEO constellation; receive signals from at least 4 satellites for each constellation, and obtain observation values ​​such as pseudorange, Doppler shift, ephemeris data and carrier-to-noise ratio; 2) Multi-dimensional consistency test calculation a) Pseudorange residual consistency test The receiver's position and clock error are calculated independently using observation data from each constellation; the calculated position is substituted into the pseudorange equations of other constellations to calculate the pseudorange residuals; if the average pseudorange residual of a constellation exceeds a threshold (e.g., 100 meters), the signal of that constellation is determined to be abnormal; when more than one constellation is abnormal, a spoofing attack alarm is triggered. b. Doppler frequency shift consistency test Based on the receiver's motion status (provided by the IMU), the expected Doppler shift is calculated; the difference between the actual Doppler shift of each constellation satellite and the expected value is compared; if the Doppler shift difference of multiple satellites in a constellation exceeds a threshold (e.g., 100Hz), it is determined to be a spoofing signal; c. Consistency check of ephemeris data Store historical ephemeris data for each constellation and build ephemeris models; compare the differences between the currently received ephemeris data and the historical models; if the ephemeris data shows abrupt changes or does not conform to the laws of orbital dynamics, it is determined to be a deceptive signal; d. Communication-Navigation Cross-validation Using the communication signal-derived navigation technology described above, the receiver position is calculated from satellite communication signals; the communication-derived position is compared with the position calculated from the navigation signals; if the difference exceeds a threshold, the navigation signal is determined to be deceptive. 3) Distributed collaborative deception detection By leveraging the distributed observation capabilities of the cluster, the accuracy of spoofing detection is further improved: each node broadcasts its spoofing detection results to other nodes within the autonomous system; the domain master node collects the detection results from all nodes and conducts a majority vote; when more than 2f+1 nodes detect a spoofing attack, a global spoofing attack is confirmed; the domain master node issues an alert to the entire network and activates the emergency response mechanism. 4) Emergency Response to Navigation Spoofing Attacks: When a navigation spoofing attack is detected, the system will automatically perform the following operations: Immediately cut off the signal from the deceived constellation and use the undeceived constellation for navigation; upgrade the anti-interference level: increase the frequency hopping rate, increase the spreading code length, and deepen the interference null trap; switch to cluster-intrinsic spatiotemporal reference navigation, completely independent of external satellite navigation signals; record the characteristic data of the deception attack and upload it to the global collaboration layer for analysis.

[0013] The beneficial effects of this invention are: By using dynamic aggregation and redundant transmission of multiple constellations, the continuity of communication and navigation services is ensured, eliminating the risk of single point of failure and geopolitical risks caused by dependence on a single satellite constellation; the distributed, decentralized architecture fundamentally avoids the bottleneck and single point of failure problem of centralized control, and the failure of any node or local area will not lead to the paralysis of the entire system, thus fundamentally enhancing the system's resilience and survivability. The dynamic cross-domain autonomous domain architecture enables task-driven cluster self-organization, self-partitioning, and self-management. It can automatically form optimal collaborative units based on task characteristics and platform status, and complete the decomposition and execution of complex tasks without manual intervention. The cluster's endogenous spatiotemporal reference technology completely eliminates the absolute dependence on ground reference stations and external navigation systems, enabling UAV clusters to operate autonomously in remote areas, signal-obscured areas, and even environments where navigation signals are completely unavailable. The real-time sensor self-calibration and cross-validation mechanism eliminates the need for pre-calibration and regular manual maintenance, significantly shortening system deployment time and reducing operation and maintenance costs. Through multi-technology integrated anti-interference methods and multi-constellation consistency deception detection, it can maintain normal operation in complex adversarial environments such as strong electromagnetic interference and navigation deception; it supports cross-domain collaboration of multiple types of unmanned platforms such as air, land, sea and submarine, and can adapt to diverse mission scenarios and geographical environments such as land, sea and air; the perfect dynamic node management and autonomous domain splitting and merging mechanism enable the system to flexibly adapt to dynamic changes in mission scale, environmental conditions and platform status. Attention-based multimodal fusion technology enables unified spatiotemporal benchmark conversion of cross-domain heterogeneous data, fully utilizing cluster observation redundancy to eliminate random errors and systematic biases of single sensors, significantly improving the accuracy and consistency of perceived data. An intrinsic security protection system is constructed, based on identity authentication using physically unclonable functions and lightweight blockchain evidence storage technology, preventing identity forgery at the hardware level and ensuring the immutability and full traceability of operation records at the data level. The federated learning architecture enables global collaboration without data leaving the domain, protecting the privacy of sensitive information while achieving cross-domain knowledge sharing and capability enhancement. Attached Figure Description

[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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.

[0015] Figure 1 This is a system structure block diagram of the present invention. Detailed Implementation

[0016] 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 embodiments of the present invention, and not all embodiments. 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.

[0017] Example 1 A satellite communication-based cross-domain control and data calibration system for unmanned aerial vehicle (UAV) swarms includes: The multi-constellation heterogeneous converged communication module includes: The multi-constellation dynamic aggregation and slicing unit is used to monitor the channel quality, bandwidth, latency, and anti-interference capability of available satellites worldwide in real time. According to actual mission requirements, the communication link is divided into low-latency and high-reliability control slices, high-bandwidth sensing slices, and high-precision calibration slices. Each slice uses links from 2 to 3 different constellations for aggregation and transmission at the same time. If a single link fails, it will not affect the completion of the overall communication transmission. The communication signal-derived navigation unit utilizes the time difference of arrival, angle of arrival, and Doppler frequency shift of the communication signal for navigation and positioning. It does not require a dedicated navigation signal, and can achieve meter-level positioning using only the communication signal even when the navigation signal is completely interfered with. After the communication signal is fused with the satellite navigation signal, the positioning accuracy is improved to the centimeter level. The satellite-machine distributed edge computing unit dynamically distributes computing tasks among the satellite's on-board processor, the UAV's edge processor, and the ground server; simple data calibration and decision-making tasks are completed on the UAV's local edge processor, while complex fusion tasks are completed on the satellite; it is configured to only transmit the final results and key data back to the ground, reducing the satellite communication bandwidth requirements. The dynamic cross-domain autonomous system control module uses an improved 3D DBSCAN algorithm to automatically divide platforms with task similarity > 85% and geographical distance < 100km into an autonomous system. An autonomous system has a minimum of 3 nodes and a maximum of 50 nodes; systems exceeding 50 nodes automatically split into multiple subdomains. The specific implementation method is as follows: 1) Input parameter preprocessing and feature vectorization: All input parameters are converted into computable numerical vectors. One-hot encoding is used for task type vectors, and complex tasks can be represented as multi-hot vectors. Geographic coordinates are converted to geocentric rectangular coordinates (X,Y,Z) to avoid errors in spherical distance calculation. Time windows are represented as [start time, end time], and the time overlap between two tasks is calculated. For platform type weights: the adaptability (0~1) of different platforms for performing different tasks is predefined, such as the weight of UAV performing reconnaissance tasks = 0.9, and the weight of unmanned vehicle performing reconnaissance tasks = 0.6. The remaining energy is normalized to the [0,1] space. 2) 3D task similarity calculation: Define a comprehensive distance metric function; the smaller the distance, the higher the similarity. in, For task type distance, use cosine distance to calculate the difference between two task vectors; For geographic distance, the spherical distance is calculated using the Haversine formula and normalized to [0,1]. The time distance is calculated as 1 minus the time overlap. When the times do not overlap at all, D_time = 1. , , The sum of the three is 1, and the default setting is... , , ; Similarity conversion formula: similarity S = 1 - D. When S > 0.85, the two nodes belong to the same potential autonomous region. 3) Improved 3D DBSCAN clustering algorithm: Initialize all nodes to an unvisited state, and randomly select an unvisited node P; Calculate the ε-neighborhood of P: the set of all nodes with a combined distance D < ε from P and a similarity S > 0.85; if the neighborhood size is ≥ 3, then P is a core node, and a new autonomous region C is created; add P and all nodes in its neighborhood to C and mark them as visited; recursively process each node in C: if the node is a core node, add the unvisited nodes in its neighborhood to C; repeat the above operation until all nodes have been visited. For isolated nodes, those with a similarity greater than 0.7 are merged into the nearest autonomous system, and the remaining isolated nodes are processed separately. 4) Dynamic partitioning and merging of autonomous regions: When the number of nodes in an autonomous region exceeds 50, partitioning is performed. The partitioning mechanism is as follows: Extract the task feature vectors of all nodes within the autonomous system; use the K-means algorithm to divide the nodes into K subsets, where K = ceil, calculated as the number of nodes / 40; reserve 10 nodes for expansion space; re-execute 3D DBSCAN on each subset to ensure that the number of nodes in each subdomain is between 3 and 50; elect a new domain master node for each subdomain. The merging mechanism is as follows: when the average similarity S between two adjacent autonomous regions is greater than 0.9 and the total number of nodes is less than 40, they are automatically merged into one autonomous region. 5) Incremental update mechanism in dynamic scenarios When a new node is added, only the average similarity between the node and all existing autonomous regions is calculated, and the node is added to the autonomous region with the highest similarity and less than 45 nodes. When a node leaves, check the size of its autonomous system. If it is less than 3, merge the remaining nodes into other similar autonomous systems. When task parameters change, only the similarity of the affected nodes is recalculated, and their respective autonomous regions are adjusted. Domain leader election is performed using an improved Practical Byzantine Fault Tolerance (PBFT) algorithm, as follows: 1) Initial elections immediately after the autonomous region is established 1.1) All nodes calculate their own and their neighbors' real-time reputation values ​​according to the formula: Reputation score R = 0.4 * Communication quality + 0.3 * Remaining energy + 0.2 * Computing power + 0.1 * Historical performance Each node maintains a globally sorted list of reputation values, which is updated synchronously via the Gossip protocol; the node with the highest reputation value automatically becomes the initial candidate node, without waiting for triggering conditions. 1.2) Nomination Phase: The initial candidate node broadcasts a Nominate message containing its own ID, current highest reputation value, and view number v=0; other nodes, upon receiving the nomination message, perform three verifications: a) verify the validity of the message signature; b) verify that the candidate node's reputation value is indeed the highest known at present; c) verify that the view number matches the local view number; if the verification passes, an Acknowledge message is sent to the candidate node; otherwise, the malicious behavior is ignored and recorded. 1.3) Confirmation Phase: Candidate nodes collect confirmation messages. When they receive valid confirmations from at least 2f+1 different nodes (f=⌊(n-1) / 3⌋), the candidate nodes use the BLS algorithm to aggregate all confirmation signatures, generate an Announce message, and broadcast it to the entire network. After receiving the Announce message, all nodes verify the aggregated signatures and update their local domain master information and view number. The election is complete, and the new domain master begins sending heartbeat messages. 2) Normal operation and heartbeat mechanism The domain owner sends a heartbeat message every 100ms, containing a summary of the current cluster status; normal nodes set a heartbeat timeout of 300ms, and if no heartbeat is received within the timeout period, the domain owner is considered to be faulty; the domain owner attaches the latest node list and reputation ranking to the heartbeat, and all nodes are updated synchronously; when the reputation value of a node changes by more than 10%, it actively broadcasts an update message, triggering a network-wide synchronization; 3) View Change If any node does not receive a heartbeat from the domain master within 300ms, a view change is immediately triggered. The node updates its local view number from v to v+1. The node sends a ViewChange message to the top-ranked node based on its latest reputation value. When the top-ranked node receives at least 2f+1 ViewChange messages, it becomes a new candidate node. The new candidate node initiates a new round of elections, and the process is exactly the same as the initial election. If the election is not completed within 1 second, the view number is automatically incremented by 1, and the next node with the highest reputation value initiates the election. 4) Handling the addition or departure of dynamic nodes When a new node joins, it sends a join request to the current domain owner. After verifying its identity, the domain owner adds it to the node list and broadcasts the updated node list and the fault tolerance threshold f. When a node leaves normally, it sends a Leave message, the domain owner updates the node list and the f value, and broadcasts the notification to the entire network. When a node leaves abnormally, the domain owner marks it as a faulty node through heartbeat timeout detection, updates the node list and f value, and broadcasts a notification to the entire network. Each time the node list changes, all nodes automatically recalculate f = ⌊(n-1) / 3⌋; Different autonomous regions (Autonomous Domains) exchange model parameters through the domain master node to achieve global collaboration. A multi-agent decision-making algorithm with federated reinforcement learning and attention mechanism is adopted. Each autonomous region trains the model locally and only uploads the model parameters, not the original data. The specific implementation method is as follows: 1) Preliminary Preparation and System Initialization: After each autonomous system is established and a domain master node is elected through the improved PBFT algorithm, the domain master node randomly initializes a set of global domain model parameters. The domain master distributes these parameters to all UAV nodes within its domain via satellite communication links to ensure that all nodes have completely consistent initial models. The domain master nodes of all autonomous systems establish point-to-point communication connections to form a global collaborative network. Each domain master node maintains a list of basic information for all other domains, including domain ID, task type, geographical range, and current model version number. 2) Local training loop within the autonomous domain: At the beginning of each training cycle, the domain master node sends the latest domain model parameters to all drones in the domain. After the drones load the new parameters, they replace the old local model. While performing actual tasks, all drones interact with the environment using the current model; each drone independently collects its own status, actions, rewards, and next status data to form a local experience trajectory; all raw experience data is stored only locally on the drone. Once each drone has collected enough empirical data, after a default of 100 steps, the gradient of the model parameters is calculated locally using the Proximal Policy Optimization (PPO) algorithm. The calculation process is entirely completed on the drone's edge processor, generating only gradient vectors and not retaining the original empirical data. All drones encrypt the calculated gradient vectors and upload them to the domain master node. After collecting the gradients of all drones in the domain, the domain master node uses the FedAvg algorithm to perform weighted aggregation, with the weights proportional to the number of experience samples collected by each drone. The domain master node updates the domain model parameters using the aggregated global gradient; after the update is completed, it enters the next training cycle and repeats the above steps; each autonomous region performs local training independently without interfering with each other; 3) Cross-Autonomous Domain Attention Fusion Process: After an autonomous domain completes 5 rounds of local training, its domain master node broadcasts the parameters of the current domain model to all other domain master nodes in the global collaborative network; at the same time, the domain master also receives model parameters sent by all other domains. Each domain master node calculates the cosine similarity between each other based on the task feature vectors of its own domain and other domains, including task type, geographical environment, target characteristics, time window, etc. The higher the similarity, the more relevant the tasks of the two domains are, and the more valuable the experience of the other is. The calculated inter-domain similarity values ​​are input into the softmax function to generate an attention weight vector; the larger the weight value, the higher the proportion of the model in the corresponding domain during fusion; the base weight of the model in this domain is fixed at 0.5, and all other domains share the remaining 0.5 weight; Each domain master node performs a weighted sum of the model parameters from all other domains based on the generated attention weights, and then fuses this sum with the model parameters from its own domain; the fusion formula is: New model parameters = 0.5 × parameters of this domain + Σ(attention weights i × parameters of other domains i) The domain master node distributes the merged new model parameters to all drones in the domain, replacing the original domain model; after the drones load the new model, they continue local training and task execution. 4) Dynamic scene adaptive processing: When the number of nodes in an autonomous system exceeds 50, it is automatically split into two subdomains. The domain master nodes of the two new subdomains inherit the latest model parameters of the original autonomous system as their respective initial domain models. Then, the two subdomains independently perform local training and cross-domain fusion. When two adjacent autonomous systems meet the merging conditions, the master node of the merged new domain uses an attention mechanism to fuse the model parameters of the two original domains to generate a new initial domain model; the fusion weight is determined based on the task similarity and the number of nodes in the two domains. When a new node joins an autonomous region, it directly downloads the latest domain model parameters from the domain master node and immediately participates in training and task execution. When a node leaves normally or abnormally, the domain master node automatically excludes the gradient of that node during the next gradient aggregation, without affecting the overall training process.

[0018] The cluster-endogenous benchmark dynamic data calibration module selects at least three nodes as benchmark nodes, and other nodes perform relative observations with the benchmark nodes. After fusing the observation data, a unified endogenous spatiotemporal benchmark for the entire cluster is generated. The sensors have built-in miniature inertial measurement units (IMUs) and temperature sensors with their own data acquisition components. The parameters of the sensors are corrected in real time using a pre-trained Transformer model, and cross-validation is performed between sensors of the same type to achieve sensor calibration. The specific implementation method is as follows: The Transformer model employs a temporal Transformer encoder architecture. The input layer consists of a 32×7-dimensional tensor, representing 32 frames of 3-axis acceleration, 3-axis angular velocity, and 1-axis temperature. The embedding layer maps the 7-dimensional original features to 64-dimensional temporal features. The encoder layer comprises two multi-head self-attention layers (2 heads) and a feedforward network (128 dimensions). The output layer contains a 6-dimensional error correction factor, representing 3-axis acceleration bias and 3-axis angular velocity bias. Model pre-training: Construction of pre-training dataset: In a laboratory environment, the sensor was placed in a high and low temperature chamber to collect static data over the entire temperature range of -40℃ to +85℃; dynamic calibration data under different angular velocities and accelerations was collected using a high-precision turntable; at least 100 hours of multi-condition data were collected for each sensor model to build a general pre-training model; Transfer learning fine-tuning: Before each drone leaves the factory, it undergoes one hour of static and simple dynamic calibration at room temperature; the data from this one hour is used to fine-tune the general pre-trained model to adapt to the individual differences of individual sensors; the fine-tuned model parameters are stored locally on the drone as the initial calibration model. Real-time inference process: The sensor collects raw IMU and temperature data, which are then preprocessed to generate a 32-frame sliding window. The window data is input into a pre-trained Transformer model, which outputs a 6-dimensional error correction for the current frame. The model's prediction error is subtracted from the original measurement value to obtain the corrected IMU data. The corrected data is then output to the navigation and control system and saved for subsequent cross-validation. The process of cross-validation of similar sensors is as follows: Trigger condition settings: Cross-validation of the same type of sensor is initiated when all of the following conditions are met: At least three drones of the same model were in the same autonomous region; The similarity of the motion states of all drones is >90%, and the similarity is determined by the cosine similarity of acceleration and angular velocity. The ambient temperature difference for all drones is less than 5°C. Distributed consensus computation: Data synchronization: Each drone broadcasts its latest corrected IMU data (up to 1 second) to all other drones within the autonomous system; Baseline value calculation: Each drone calculates the median of all data from the same type of sensors, which is used as the cluster baseline value at the current moment; Deviation calculation: Each drone calculates the deviation between its corrected data and the cluster baseline value; Anomaly detection: If the deviation of a certain drone exceeds the 3σ threshold continuously, where σ is the standard deviation of the cluster data, it is determined that the sensor has abnormal drift. Online model correction: When slow drift is detected in the sensor, with a deviation of <0.5m / s² or <5° / s, online model fine-tuning is initiated. The local Transformer model is incrementally learned using the cluster baseline value from the most recent 10 minutes as the ground truth. The learning rate is set to 1 / 100 of that used during pre-training, and only the parameters of the last two layers of the model are updated to avoid catastrophic forgetting. After fine-tuning, the old local model is automatically replaced, and real-time calibration continues. For sensor data with different domain characteristics, a multimodal fusion model based on an attention mechanism is used to transform the data to a unified endogenous spatiotemporal reference. The overall observation redundancy of the cluster is utilized to eliminate random and systematic errors from individual sensors. The implementation process is as follows: A three-stage pipeline of "spatiotemporal alignment - cross-modal attention fusion - redundancy error elimination" is constructed to losslessly transform all heterogeneous data to the cluster's endogenous spatiotemporal reference. The specific method is as follows: 1) Spatiotemporal alignment of heterogeneous data a) Unified time standard All sensor data are globally synchronized using an endogenous spatiotemporal reference within the cluster, with a time synchronization accuracy of <1μs. For sensors with different sampling rates (e.g., camera 30Hz, LiDAR 10Hz, IMU 200Hz), an asynchronous time registration algorithm based on cubic spline interpolation is employed: the highest sampling rate, IMU 200Hz, is used as the reference time axis; interpolation is performed on the data from lower sampling rate sensors to generate an equally spaced data sequence aligned with the reference time axis; the interpolation weights are dynamically adjusted based on the sensor's timestamp error and historical synchronization accuracy. b. Spatial benchmark unification Each sensor has an independent local coordinate system. The transformation matrix from the local coordinate system to the cluster's endogenous reference coordinate system is calculated in real time using a dynamic extrinsic parameter calibration algorithm. The rotation matrix R and translation vector T are: Initial extrinsic parameter acquisition: A rough calibration is performed when the UAV leaves the factory to obtain the initial transformation matrix; Dynamic extrinsic parameter update: Real-time correction of extrinsic parameters using relative observations between clusters (LiDAR point cloud matching, visual feature matching, UWB ranging); External parameter reliability assessment: The reliability of the external parameter matrix is ​​calculated based on the observation quality (number of feature points, matching accuracy, ranging error); Coordinate transformation: All sensor data are transformed to the intrinsic reference coordinate system using the formula P_global = R × P_local + T; c. Single-modal data preprocessing Standardization preprocessing is performed on data from different modalities to eliminate dimensional differences between modalities: Image data: Normalized to [0,1], then processed for distortion correction and white balance; Point cloud data: downsampled to a uniform density (e.g., 10,000 points / frame), outliers removed; IMU data: Output after real-time correction by the Transformer as described above; Radar / sonar data: converted to polar coordinate grid plot, intensity values ​​normalized; 2) Multimodal feature fusion based on cross-modal attention a. Design of a lightweight cross-modal attention fusion model It adopts a dual-stream cross-attention Transformer architecture, with the total number of parameters controlled within 1 million, and runs in real time on the drone edge processor; Input layer: Inputs preprocessed multimodal data and generates respective modal feature sequences through respective modal feature extractors; Cross-attention layer: Each modality feature performs cross-attention calculations with all other modality features; Modal adaptive weighted layer: Automatically calculates the credibility weight of each modality based on the scene; Feature fusion layer: Weighted fusion of all modal features to generate a global feature representation under a unified benchmark; Output: fused 3D coordinates, semantic information, and confidence level; b. Implementation of Modality Adaptive Attention Mechanism Scene feature extraction: Extract current scene features from global features (such as light intensity, weather conditions, and terrain type). Modal credibility prediction: A credibility score of 0 to 1 is predicted for each modality based on scene features using a small MLP network; Attention weight adjustment: Multiply the modality confidence score by the attention score of the cross-attention output to obtain the final fusion weight; Weighted fusion: All modal features are summed according to their final weights to obtain the fused global features; c. Model Training and Update Pre-training: Collect multimodal datasets in a laboratory environment and train a general fusion model; Federated fine-tuning: Each autonomous region fine-tunes the model locally using real-world task data, uploading only the model parameters to the global collaboration layer; Incremental update: When a new sensor type is added, it can be quickly adapted through transfer learning without retraining the entire model; 3) Error elimination based on cluster observation redundancy: By utilizing repeated observations of the same target from multiple platforms in the cluster, the random error and systematic bias of a single sensor are eliminated simultaneously through a distributed Bayesian consensus fusion algorithm. a. Observation redundancy detection: Redundancy error elimination is initiated when the following conditions are met: The same target was observed by at least three different platforms; The time difference between all observations is <100ms; The spatial overlap of all observations is >60%; b. Distributed Bayesian Consistency Fusion Each node performs local computation: Each platform calculates the probability distribution, mean μ_i, and covariance Σ_i of the target state based on its own fusion results; Simultaneously, the reliability w_i of its own observation is calculated, which is jointly determined by the sensor accuracy, external parameter reliability, and observation distance; Global consistency computation: All platforms will broadcast their local μ_i, Σ_i, and w_i to other platforms within the autonomous system; Each platform receives observation data from all other platforms and calculates the global weighted mean and covariance: Systematic bias correction: Calculate the observation bias for each platform: ; If the deviation of a certain platform continuously exceeds the 3σ threshold, it is determined that there is a systematic deviation. The deviation values ​​are used to correct the sensor extrinsic parameters and fusion model of the platform online; c. Anomaly detection removal Robust statistical methods were used to remove outlier observations: Using the median absolute deviation (MAD) instead of the standard deviation reduces the impact of outliers; Observations with a deviation exceeding 5 times the MAD are directly identified as abnormal and removed. Platforms that are identified as having anomalies will be flagged, and the credibility weight of their subsequent observations will be temporarily reduced. The built-in security protection module uses a Physically Unclonable Function (PUF) to give the drone a unique PUF identification fingerprint; A local blockchain is maintained for each autonomous system, and the blockchain uses an improved RAFT consensus algorithm to prevent operation record data from being tampered with; the implementation process is as follows: 1) Transaction submission and block generation process The business module generates operation instructions and calculates the hash value of the original data; it signs the transaction using the node's private key and sends it to the transaction pool; the leader node collects transactions from the transaction pool, packages them into a new block, and broadcasts the new block to all follower nodes; the follower nodes verify the legality of the block: verify the leader's signature; verify the signatures of all transactions; verify whether the hash of the previous block is correct; and verify whether the transaction content conforms to the business rules. After successful verification, the followers append the block to the local blockchain and update the state database; when the leader receives confirmation from 2f+1 nodes, the block is finally confirmed, and the business module is notified that the operation is complete. 2) Data verification and traceability process When it is necessary to verify the legitimacy of an operation, the latest data is retrieved from the state database, the hash value of the data is calculated, and it is compared with the hash of the corresponding transaction data on the blockchain. If the hashes match, it means that the data has not been tampered with; otherwise, it means that the data has been tampered with. If it is necessary to trace the operation history, all related transactions on the blockchain can be queried through the transaction ID. All operations can be traced back to the specific sending node and timestamp. 3) Domain leader election synchronized with blockchain When a failure of the current domain leader is detected, an improved RAFT election is triggered. The candidate node with the highest reputation value becomes the new leader. The new leader broadcasts the latest block height and state hash, and all follower nodes synchronize blocks with the new leader to ensure ledger consistency. The new leader generates a "domain leader election" transaction, which is recorded on the blockchain. After the election is completed, the new leader begins to receive transactions and generate new blocks. 4) Safety and fault tolerance mechanisms a. Anti-tampering mechanism The blockchain uses the SHA-256 hash chain, and any modification to a block will invalidate the hashes of all subsequent blocks. All transactions must be digitally signed by the sender to prevent forgery. All blocks must be signed by the leader to prevent malicious nodes from generating invalid blocks; Each update to the state database must correspond to a transaction on the blockchain; b. Malicious Node Defense Reputation penalty mechanism: Malicious behavior, such as sending invalid transactions, forging blocks, or denial of service, will result in a permanent decrease in reputation score; Nodes with a reputation score below 0.2 will be automatically excluded from the consensus process; A multi-signature verification mechanism is adopted, and key operations, such as domain leader election and model update, require multiple nodes to sign together; Abnormal behavior detection: Automatically identify potential malicious nodes by analyzing node behavior patterns; c. Data integrity assurance Each node maintains a complete copy of the blockchain, so a single point of failure will not result in data loss. Regularly perform ledger consistency checks to ensure that the ledgers of all nodes are consistent; Important data can be backed up on multiple nodes, and backup records are also stored on the blockchain; A triple anti-jamming technique combining frequency hopping, spread spectrum, and beamforming is employed, utilizing consistency checks of multi-constellation observation data to detect spoofing attacks on navigation signals; the implementation process is as follows: 1) Adaptive frequency hopping Frequency hopping pattern generation: Dynamic frequency hopping seeds are generated based on the cluster's endogenous spatiotemporal reference, with each autonomous region using an independent frequency hopping seed; the frequency hopping sequence is generated using a chaotic algorithm with a period > 10^12, making it unpredictable; the frequency hopping bandwidth covers the entire available frequency band (e.g., Ku band 10.7~14.5GHz), and the frequency hopping rate is 1000 hops / second. Real-time spectrum sensing and interference avoidance: Each drone is equipped with a miniature spectrum analyzer that continuously scans the entire frequency hopping band; it updates the interference spectrum map every 10ms, marking the interfered frequencies; it dynamically adjusts the frequency hopping pattern to automatically avoid interfered frequencies, with a frequency hopping pattern update cycle of 1 second. Cluster synchronization frequency hopping: All nodes achieve microsecond-level time synchronization through an endogenous spatiotemporal reference; the domain master node broadcasts the latest frequency hopping seed and frequency hopping pattern every 1 second; all nodes switch to the same frequency at the same time to ensure uninterrupted cluster communication; 2) Direct sequence spread spectrum Spread code generation: A Gold code of length 1023 is used as the spreading code, with a code rate of 10 Mbps. Each UAV node uses a unique spreading code, and different autonomous systems use orthogonal spreading code groups. The spreading code seed is also generated based on an endogenous spatiotemporal reference and is updated periodically. Transmitter spread spectrum processing: The original data (rate = 1Mbps) is XORed with the spreading code to generate a spread spectrum signal; the bandwidth of the spread spectrum signal is expanded to 10MHz, and the processing gain is 10dB; the spread spectrum signal is then sent to a frequency hopping modulator for frequency hopping processing. Despreading at the receiver: The receiver first performs frequency hopping demodulation to obtain the spread spectrum signal; then it uses the locally synchronized spreading code for correlation despreading to recover the original data; finally, it uses a RAKE receiver to process the multipath signal, further improving the receiving sensitivity. 3) Digital Adaptive Beamforming Antenna array configuration: Each drone is equipped with a 4-element conformal phased array antenna, operating in the Ku band; the antenna array is integrated on the surface of the drone fuselage, without affecting aerodynamic performance; each antenna element is independently controllable, and can achieve a beam scanning range of ±60°; Adaptive beamforming algorithm: The Least Mean Square (LMS) adaptive algorithm is used to adjust the weighting coefficients of the antenna elements in real time; the main lobe is aligned with the satellite with the direction of arrival (DOA) of the satellite signal as a reference; at the same time, the DOA of the interference source is detected to form a null with a depth >50dB in the direction of interference. Multi-beam simultaneous communication: It supports the simultaneous formation of 4 independent beams, each aimed at 4 different satellites; enables simultaneous reception and transmission of signals from multiple constellations; and automatically switches to the beams of other satellites when a satellite is interfered with. The triple anti-interference technology adopts a layered and progressive approach to form an anti-interference system: First layer: Adaptive frequency hopping technology to avoid narrowband interference, making it impossible for the jammer to aim; The second layer: direct sequence spread spectrum technology, which resists broadband noise interference and extracts the signal from the noise; The third layer: digital adaptive beamforming technology, which suppresses directional interference while enhancing useful signals; Based on the consistency of multi-constellation observation data, the implementation process of detecting navigation signal spoofing attacks is as follows: 1) Multi-constellation observation data acquisition: Simultaneously receive navigation signals from at least 4 different satellite constellations, namely: BeiDou-3: MEO+GEO+IGSO hybrid constellation; GPS: MEO constellation; Galileo: MEO constellation; GLONASS: MEO constellation; receive signals from at least 4 satellites for each constellation, and obtain observation values ​​such as pseudorange, Doppler shift, ephemeris data and carrier-to-noise ratio; 2) Multi-dimensional consistency test calculation a) Pseudorange residual consistency test The receiver's position and clock error are calculated independently using observation data from each constellation; the calculated position is substituted into the pseudorange equations of other constellations to calculate the pseudorange residuals; if the average pseudorange residual of a constellation exceeds a threshold (e.g., 100 meters), the signal of that constellation is determined to be abnormal; when more than one constellation is abnormal, a spoofing attack alarm is triggered. b. Doppler frequency shift consistency test Based on the receiver's motion status (provided by the IMU), the expected Doppler shift is calculated; the difference between the actual Doppler shift of each constellation satellite and the expected value is compared; if the Doppler shift difference of multiple satellites in a constellation exceeds a threshold (e.g., 100Hz), it is determined to be a spoofing signal; c. Consistency check of ephemeris data Store historical ephemeris data for each constellation and build ephemeris models; compare the differences between the currently received ephemeris data and the historical models; if the ephemeris data shows abrupt changes or does not conform to the laws of orbital dynamics, it is determined to be a deceptive signal; d. Communication-Navigation Cross-validation Using the communication signal-derived navigation technology described above, the receiver position is calculated from satellite communication signals; the communication-derived position is compared with the position calculated from the navigation signals; if the difference exceeds a threshold, the navigation signal is determined to be deceptive. 3) Distributed collaborative deception detection By leveraging the distributed observation capabilities of the cluster, the accuracy of spoofing detection is further improved: each node broadcasts its spoofing detection results to other nodes within the autonomous system; the domain master node collects the detection results from all nodes and conducts a majority vote; when more than 2f+1 nodes detect a spoofing attack, a global spoofing attack is confirmed; the domain master node issues an alert to the entire network and activates the emergency response mechanism. 4) Emergency Response to Navigation Spoofing Attacks: When a navigation spoofing attack is detected, the system will automatically perform the following operations: Immediately cut off the signal from the deceived constellation and use the undeceived constellation for navigation; upgrade the anti-interference level: increase the frequency hopping rate, increase the spreading code length, and deepen the interference null trap; switch to cluster-intrinsic spatiotemporal reference navigation, completely independent of external satellite navigation signals; record the characteristic data of the deception attack and upload it to the global collaboration layer for analysis.

Claims

1. A cross-domain control and data calibration system for unmanned aerial vehicle (UAV) swarms based on satellite communication collaboration, characterized in that: include: The multi-constellation heterogeneous converged communication module includes: The multi-constellation dynamic aggregation and slicing unit is used to monitor the channel quality, bandwidth, latency and anti-interference capability of available satellites around the world in real time; according to the actual mission requirements, the communication link is divided into control slice, sensing slice and calibration slice; each slice uses links from 2 to 3 different constellations for aggregation and transmission at the same time; The communication signal-derived navigation unit utilizes the time difference of arrival, angle of arrival, and Doppler frequency shift of the communication signal for navigation and positioning. The satellite-machine distributed edge computing unit dynamically distributes computing tasks among the satellite's on-board processor, the UAV's edge processor, and the ground server; simple data calibration and decision-making tasks are completed on the UAV's local edge processor, while complex fusion tasks are completed on the satellite; it is configured to only transmit the final results and key data back to the ground. The dynamic cross-domain autonomous system control module uses an improved 3D DBSCAN algorithm to automatically divide platforms with task similarity > 85% and geographical distance < 100km into an autonomous system. An autonomous system has a minimum of 3 nodes and a maximum of 50 nodes. If there are more than 50 nodes, it will automatically split into multiple subdomains. An improved practical Byzantine fault-tolerant algorithm is used for domain leader election. Different autonomous systems exchange model parameters through domain leader nodes to achieve global collaboration. The cluster-endogenous benchmark dynamic data calibration module selects at least three nodes as benchmark nodes, and other nodes perform relative observations with the benchmark nodes. After fusing the observation data, a unified endogenous spatiotemporal benchmark for the entire cluster is generated. The sensors have their own built-in data acquisition components, and the parameters of the sensors are corrected in real time using a pre-trained Transformer model. Cross-validation is performed between sensors of the same type to achieve sensor calibration. For sensor data with different domain differences, a multimodal fusion model based on an attention mechanism is used to transform the data to the unified endogenous spatiotemporal benchmark. The overall observation redundancy of the cluster is used to eliminate random and systematic errors of individual sensors. The intrinsic security protection module uses a physically unclonable function to assign a unique PUF identification fingerprint to the UAV; it maintains a local blockchain for each autonomous region, which uses an improved RAFT consensus algorithm to prevent operation record data from being tampered with; it adopts a triple anti-interference technology combining frequency hopping, spread spectrum, and beamforming, and uses consistency verification of multi-constellation observation data to detect navigation signal spoofing attacks.

2. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, The specific method for dividing the autonomous regions is as follows: 1) Input parameter preprocessing and feature vectorization: Convert all input parameters into computable numerical vectors; 2) 3D task similarity calculation: Define a comprehensive distance metric function; the smaller the distance, the higher the similarity. in, For task type distance, use cosine distance to calculate the difference between two task vectors; For geographic distance, the spherical distance is calculated using the Haversine formula and normalized to [0,1]. The time distance is calculated as 1 minus the time overlap. When the times do not overlap at all, D_time = 1. , , The sum of the three is 1, and the default setting is... , , ; Similarity conversion formula: similarity S = 1 - D. When S > 0.85, the two nodes belong to the same potential autonomous region. 3) Improved 3D DBSCAN clustering algorithm: Initialize all nodes to an unvisited state, and randomly select an unvisited node P; Calculate the ε-neighborhood of P: the set of all nodes with a combined distance D < ε from P and a similarity S > 0.85; if the neighborhood size is ≥ 3, then P is a core node, and a new autonomous region C is created; add P and all nodes in its neighborhood to C and mark them as visited; recursively process each node in C: if the node is a core node, add the unvisited nodes in its neighborhood to C; repeat the above operation until all nodes have been visited; For isolated nodes, those with a similarity greater than 0.7 are merged into the nearest autonomous system, and the remaining isolated nodes are processed separately. 4) Dynamic partitioning and merging of autonomous regions: When the number of nodes in an autonomous region exceeds 50, partitioning is performed. The partitioning mechanism is as follows: Extract the task feature vectors of all nodes within the autonomous system; use the K-means algorithm to divide the nodes into K subsets, where K = ceil, calculated as the number of nodes / 40; reserve 10 nodes for expansion space; re-execute 3D DBSCAN on each subset to ensure that the number of nodes in each subdomain is between 3 and 50; elect a new domain master node for each subdomain. The merging mechanism is as follows: when the average similarity S between two adjacent autonomous regions is greater than 0.9 and the total number of nodes is less than 40, they are automatically merged into one autonomous region. 5) Incremental update mechanism in dynamic scenarios When a new node is added, only the average similarity between the node and all existing autonomous regions is calculated, and the node is added to the autonomous region with the highest similarity and less than 45 nodes. When a node leaves, check the size of its autonomous system. If it is less than 3, merge the remaining nodes into other similar autonomous systems. When task parameters change, only the similarity of the affected nodes is recalculated, and their respective autonomous regions are adjusted.

3. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, The domain leader election method is as follows: 1) Initial elections immediately after the autonomous region is established 1.1) All nodes calculate their own and their neighbors' real-time reputation values ​​according to the formula: Reputation score R = 0.4 * Communication quality + 0.3 * Remaining energy + 0.2 * Computing power + 0.1 * Historical performance Each node maintains a globally sorted list of reputation values, which is updated synchronously via the Gossip protocol; the node with the highest reputation value automatically becomes the initial candidate node, without waiting for triggering conditions. 1.2) Nomination Phase: The initial candidate node broadcasts a Nominate message containing its own ID, current highest reputation value, and view number v=0; other nodes, upon receiving the nomination message, perform three verifications: a) verify the validity of the message signature; b) verify that the candidate node's reputation value is indeed the highest known at present; c) verify that the view number matches the local view number; if the verification passes, an Acknowledge message is sent to the candidate node; otherwise, the malicious behavior is ignored and recorded. 1.3) Confirmation Phase: Candidate nodes collect confirmation messages. When they receive valid confirmations from at least 2f+1 different nodes (f=⌊(n-1) / 3⌋), the candidate nodes use the BLS algorithm to aggregate all confirmation signatures, generate an Announce message, and broadcast it to the entire network. After receiving the Announce message, all nodes verify the aggregated signatures and update their local domain master information and view number. The election is complete, and the new domain master begins sending heartbeat messages. 2) Normal operation and heartbeat mechanism The domain owner sends a heartbeat message every 100ms, containing a summary of the current cluster status; normal nodes set a heartbeat timeout of 300ms, and if no heartbeat is received within the timeout period, the domain owner is considered to be faulty; the domain owner attaches the latest node list and reputation ranking to the heartbeat, and all nodes are updated synchronously; when the reputation value of a node changes by more than 10%, it actively broadcasts an update message, triggering a network-wide synchronization; 3) View Change If any node does not receive a heartbeat from the domain master within 300ms, a view change is immediately triggered. The node updates its local view number from v to v+1. The node sends a ViewChange message to the top-ranked node based on its latest reputation value. When the top-ranked node receives at least 2f+1 ViewChange messages, it becomes a new candidate node. The new candidate node initiates a new round of elections, and the process is exactly the same as the initial election. If the election is not completed within 1 second, the view number is automatically incremented by 1, and the next node with the highest reputation value initiates the election. 4) Handling the addition or departure of dynamic nodes When a new node joins, it sends a join request to the current domain owner. After verifying its identity, the domain owner adds it to the node list and broadcasts the updated node list and the fault tolerance threshold f. When a node leaves normally, it sends a Leave message, the domain owner updates the node list and the f value, and broadcasts the notification to the entire network. When a node leaves abnormally, the domain owner marks it as a faulty node through heartbeat timeout detection, updates the node list and f value, and broadcasts a notification to the entire network. Each time the node list changes, all nodes automatically recalculate f = ⌊(n-1) / 3⌋.

4. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, Different autonomous regions (Autonomous Domains) exchange model parameters to achieve global collaboration. A multi-agent decision-making algorithm combining federated reinforcement learning and attention mechanisms is adopted. Each Autonomous Domain trains its model locally and only uploads the model parameters, not the original data. The specific implementation method is as follows: 1) Preliminary Preparation and System Initialization: After each autonomous system is established and a domain master node is elected through the improved PBFT algorithm, the domain master node randomly initializes a set of global domain model parameters. The domain master distributes these parameters to all UAV nodes within its domain via satellite communication links to ensure that all nodes have completely consistent initial models. The domain master nodes of all autonomous systems establish point-to-point communication connections to form a global collaborative network. Each domain master node maintains a list of basic information for all other domains, including domain ID, task type, geographical range, and current model version number. 2) Local training loop within the autonomous domain: At the beginning of each training cycle, the domain master node sends the latest domain model parameters to all drones in the domain. After the drones load the new parameters, they replace the old local model. While performing actual tasks, all drones interact with the environment using the current model; each drone independently collects its own status, actions, rewards, and next status data to form a local experience trajectory; all raw experience data is stored only locally on the drone. Once each drone has collected enough empirical data, the gradient of the model parameters is calculated locally using a near-end policy optimization algorithm. The calculation process is completed entirely on the drone's edge processor, generating only gradient vectors and not retaining the original empirical data; All drones encrypt the calculated gradient vectors and upload them to the domain master node. After collecting the gradients of all drones in the domain, the domain master node uses a federated averaging algorithm to perform weighted aggregation, with the weights proportional to the number of experience samples collected by each drone. The domain master node updates the domain model parameters using the aggregated global gradient. After the update is completed, it enters the next training cycle and repeats the above steps. Each autonomous region performs local training independently without interfering with each other. 3) Cross-Autonomous Domain Attention Fusion Process: After an autonomous domain completes 5 rounds of local training, its domain master node broadcasts the parameters of the current domain model to all other domain master nodes in the global collaborative network; at the same time, the domain master also receives model parameters sent by all other domains. Each domain master node calculates the cosine similarity between each pair of tasks in its own domain and those in other domains based on the task feature vectors of other domains. The higher the similarity, the more relevant the tasks of the two domains are, and the more valuable the experience of the other is. The calculated inter-domain similarity values ​​are input into the softmax function to generate an attention weight vector; The larger the weight value, the higher the proportion of the model in the corresponding domain during fusion; Each domain master node performs a weighted sum of the model parameters from all other domains based on the generated attention weights, and then fuses this sum with the model parameters from its own domain; the fusion formula is: New model parameters = 0.5 × parameters of this domain + Σ(attention weights i × parameters of other domains i) The domain master node distributes the merged new model parameters to all drones in the domain, replacing the original domain model; after the drones load the new model, they continue local training and task execution. 4) Dynamic scene adaptive processing: When the number of nodes in an autonomous system exceeds 50, it is automatically split into two subdomains. The domain master nodes of the two new subdomains inherit the latest model parameters of the original autonomous system as their respective initial domain models. Then, the two subdomains independently perform local training and cross-domain fusion. When two adjacent autonomous systems meet the merging conditions, the master node of the merged new domain uses an attention mechanism to fuse the model parameters of the two original domains to generate a new initial domain model; the fusion weight is determined based on the task similarity and the number of nodes in the two domains. When a new node joins an autonomous domain, it directly downloads the latest domain model parameters from the domain master node and immediately participates in training and task execution. When a node leaves normally or abnormally, the domain master node will automatically exclude the gradient of that node during the next gradient aggregation, without affecting the overall training process.

5. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, The sensor incorporates at least a miniature inertial measurement unit and a temperature sensor, and the Transformer model adopts a temporal Transformer encoder architecture. Model pre-training: Construction of pre-training dataset: collect static data across the entire temperature range of -40℃ to +85℃; collect dynamic calibration data under different angular velocities and accelerations; construct a general pre-training model; Transfer learning fine-tuning: Before each drone leaves the factory, it undergoes one hour of static and simple dynamic calibration at room temperature; The data from this one hour is used to fine-tune the general pre-trained model to adapt to the individual differences of a single sensor; the fine-tuned model parameters are stored locally on the drone as the initial calibration model. Real-time inference process: The sensor collects raw IMU and temperature data, which are then preprocessed to generate a 32-frame sliding window. The window data is input into a pre-trained Transformer model, which outputs a 6-dimensional error correction for the current frame. The model's prediction error is subtracted from the original measurement value to obtain the corrected IMU data. The corrected data is then output to the navigation and control system and saved for subsequent cross-validation. The cross-validation process for the same type of sensors is as follows: Trigger condition settings: Cross-validation of the same type of sensor is initiated when all of the following conditions are met: At least three drones of the same model are located within the same autonomous system; the similarity of the motion states of all drones is > 90%; the difference in ambient temperature among all drones is < 5℃; Distributed consensus computation: Data synchronization: Each drone broadcasts its latest corrected IMU data (up to 1 second) to all other drones within the autonomous system; Baseline value calculation: Each drone calculates the median of all data from the same type of sensors, which is used as the cluster baseline value at the current moment; Deviation calculation: Each drone calculates the deviation between its corrected data and the cluster baseline value; Anomaly detection: If the deviation of a certain drone exceeds the 3σ threshold continuously, where σ is the standard deviation of the cluster data, it is determined that the sensor has abnormal drift. Online model correction: When slow drift is detected in the sensor, with a deviation of < 0.5 m / s² or < 5° / s, online model fine-tuning is initiated. The local Transformer model is incrementally learned using the cluster baseline value from the most recent 10 minutes as the ground truth. The learning rate is set to 1 / 100 of that used during pre-training, and only the parameters of the last two layers of the model are updated to avoid catastrophic forgetting. After fine-tuning, the old local model is automatically replaced, and real-time calibration continues.

6. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, The sensor data includes different domains such as air, land, sea, and underwater. A three-stage pipeline of "spatiotemporal alignment - cross-modal attention fusion - redundancy error elimination" is constructed to losslessly transform all heterogeneous data to the cluster's endogenous spatiotemporal reference. The specific method is as follows: 1) Spatiotemporal alignment of heterogeneous data a) Unified time standard All sensor data are globally synchronized using an endogenous spatiotemporal reference within the cluster, with a time synchronization accuracy of < 1μs. For sensors with different sampling rates, an asynchronous time registration algorithm based on cubic spline interpolation is used: the highest sampling rate, IMU 200Hz, is used as the reference time axis; interpolation is performed on the data from low sampling rate sensors to generate an equally spaced data sequence aligned with the reference time axis; the interpolation weights are dynamically adjusted by the sensor's timestamp error and historical synchronization accuracy. b. Spatial benchmark unification Each sensor has an independent local coordinate system. The transformation matrix from the local coordinate system to the cluster's endogenous reference coordinate system is calculated in real time using a dynamic extrinsic parameter calibration algorithm. Initial extrinsic parameter acquisition: A rough calibration is performed when the UAV leaves the factory to obtain the initial transformation matrix; Dynamic extrinsic parameter update: Real-time correction of extrinsic parameters using relative observations between clusters; External parameter reliability assessment: Calculate the reliability of the external parameter matrix based on the observation quality; Coordinate transformation: All sensor data are transformed to the intrinsic reference coordinate system using the formula P_global = R × P_local + T; c. Single-modal data preprocessing Standardization preprocessing is performed on data from different modalities to eliminate dimensional differences between modalities: Image data: normalized to [0,1]; point cloud data: downsampled to uniform density; IMU data: output after real-time correction by the Transformer; radar / sonar data: converted to polar coordinate grid map and normalized intensity values; 2) Multimodal feature fusion based on cross-modal attention a. Design of a lightweight cross-modal attention fusion model It adopts a dual-stream cross-attention Transformer architecture, with the total number of parameters controlled within 1 million, and runs in real time on the drone edge processor; Input layer: Inputs preprocessed multimodal data and generates respective modal feature sequences through respective modal feature extractors; Cross-attention layer: Each modality feature performs cross-attention calculations with all other modality features; Modal adaptive weighted layer: Automatically calculates the credibility weight of each modality based on the scene; Feature fusion layer: Weighted fusion of all modal features to generate a global feature representation under a unified benchmark; Output: fused 3D coordinates, semantic information, and confidence level; b. Implementation of Modality Adaptive Attention Mechanism Scene feature extraction: Extracting current scene features from global features; Modal credibility prediction: A credibility score of 0 to 1 is predicted for each modality based on scene features using a small MLP network; Attention weight adjustment: Multiply the modality confidence score by the attention score of the cross-attention output to obtain the final fusion weight; Weighted fusion: All modal features are summed according to their final weights to obtain the fused global features; c. Model Training and Update Pre-training: Collect multimodal datasets in a laboratory environment and train a general fusion model; Federated fine-tuning: Each autonomous region fine-tunes the model locally using real-world task data, uploading only the model parameters to the global collaboration layer; Incremental update: When a new sensor type is added, it can be quickly adapted through transfer learning without retraining the entire model; 3) Error elimination based on cluster observation redundancy: By utilizing repeated observations of the same target from multiple platforms in the cluster, the random error and systematic bias of a single sensor are eliminated simultaneously through a distributed Bayesian consensus fusion algorithm. a. Observation redundancy detection: Redundancy error elimination is initiated when the following conditions are met: The same target was observed by at least three different platforms; The time difference between all observations is < 100ms; The spatial overlap of all observations is >60%; b. Distributed Bayesian Consistency Fusion Each node performs local computation: Each platform calculates the probability distribution, mean μ_i, and covariance Σ_i of the target state based on its own fusion results; Simultaneously calculate the confidence level w_i of its own observations; Global consistency computation: All platforms will broadcast their local μ_i, Σ_i, and w_i to other platforms within the autonomous system; Each platform receives observation data from all other platforms and calculates the global weighted mean and covariance: Systematic bias correction: Calculate the observation bias for each platform: ; If the deviation of a certain platform continuously exceeds the 3σ threshold, it is determined that there is a systematic deviation. The deviation values ​​are used to correct the sensor extrinsic parameters and fusion model of the platform online; c. Anomaly detection removal Robust statistical methods were used to remove outlier observations: Using the absolute deviation of the median instead of the standard deviation reduces the impact of outliers; Observations with a deviation exceeding 5 times the MAD are directly identified as abnormal and removed. Platforms that exhibit abnormal observations will be flagged, and the credibility weight of their subsequent observations will be temporarily reduced.

7. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, The autonomous domain is equipped with a local blockchain to prevent operation record data from being tampered with. The specific method is as follows: 1) Transaction submission and block generation process The business module generates operation instructions and calculates the hash value of the original data; it signs the transaction using the node's private key and sends it to the transaction pool; the leader node collects transactions from the transaction pool, packages them into a new block, and broadcasts the new block to all follower nodes; the follower nodes verify the legality of the block: verify the leader's signature; verify the signatures of all transactions; verify whether the hash of the previous block is correct; and verify whether the transaction content conforms to the business rules. After successful verification, the followers append the block to the local blockchain and update the state database; when the leader receives confirmation from 2f+1 nodes, the block is finally confirmed, and the business module is notified that the operation is complete. 2) Data verification and traceability process When it is necessary to verify the legitimacy of an operation, the latest data is retrieved from the state database, the hash value of the data is calculated, and it is compared with the hash of the corresponding transaction data on the blockchain. If the hashes match, it means that the data has not been tampered with; otherwise, it means that the data has been tampered with. If it is necessary to trace the operation history, all related transactions on the blockchain can be queried through the transaction ID. All operations can be traced back to the specific sending node and timestamp. 3) Domain leader election synchronized with blockchain When a failure of the current domain leader is detected, an improved RAFT election is triggered. The candidate node with the highest reputation value becomes the new leader. The new leader broadcasts the latest block height and state hash, and all follower nodes synchronize blocks with the new leader to ensure ledger consistency. The new leader generates a "domain leader election" transaction and records it on the blockchain. After the election is completed, the new leader begins to receive transactions and generate new blocks. 4) Safety and fault tolerance mechanisms a. Anti-tampering mechanism The blockchain uses the SHA-256 hash chain, and any modification to a block will invalidate the hashes of all subsequent blocks. All transactions must be digitally signed by the sender to prevent forgery. All blocks must be signed by the leader to prevent malicious nodes from generating invalid blocks; Each update to the state database must correspond to a transaction on the blockchain; b. Malicious Node Defense Reputation penalty mechanism: Malicious behavior will result in a permanent decrease in reputation. Nodes with a reputation score below 0.2 will be automatically excluded from the consensus process; A multi-signature verification mechanism is adopted, and key operations require multiple nodes to sign together; Abnormal behavior detection: Automatically identify potential malicious nodes by analyzing node behavior patterns; c. Data integrity assurance Each node maintains a complete copy of the blockchain, so a single point of failure will not result in data loss. Regularly perform ledger consistency checks to ensure that the ledgers of all nodes are consistent; Important data can be backed up on multiple nodes, and backup records are also stored on the blockchain.

8. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, The specific implementation method of the triple anti-interference technology combining frequency hopping, spread spectrum, and beamforming is as follows: 1) Adaptive frequency hopping Frequency hopping pattern generation: Dynamic frequency hopping seeds are generated based on the cluster's endogenous spatiotemporal reference, with each autonomous region using an independent frequency hopping seed; the frequency hopping sequence is generated using a chaotic algorithm with a period > 10^12, making it unpredictable; the frequency hopping bandwidth covers the entire available frequency band, and the frequency hopping rate is 1000 hops / second; Real-time spectrum sensing and interference avoidance: Each drone is equipped with a miniature spectrum analyzer that continuously scans the entire frequency hopping band; it updates the interference spectrum map every 10ms, marking the interfered frequencies; it dynamically adjusts the frequency hopping pattern to automatically avoid interfered frequencies, with a frequency hopping pattern update cycle of 1 second. Cluster synchronization frequency hopping: All nodes achieve microsecond-level time synchronization through an endogenous spatiotemporal reference; the domain master node broadcasts the latest frequency hopping seed and frequency hopping pattern every 1 second; all nodes switch to the same frequency at the same time to ensure uninterrupted cluster communication; 2) Direct sequence spread spectrum Spread code generation: A Gold code of length 1023 is used as the spreading code, with a code rate of 10 Mbps. Each UAV node uses a unique spreading code, and different autonomous systems use orthogonal spreading code groups. The spreading code seed is also generated based on an endogenous spatiotemporal reference and is updated periodically. Transmitter spread spectrum processing: The original data (rate = 1Mbps) is XORed with the spreading code to generate a spread spectrum signal; the bandwidth of the spread spectrum signal is expanded to 10MHz, and the processing gain is 10dB; the spread spectrum signal is then sent to a frequency hopping modulator for frequency hopping processing. Despreading at the receiver: The receiver first performs frequency hopping demodulation to obtain the spread spectrum signal; then it uses the locally synchronized spreading code for correlation despreading to recover the original data; finally, it uses a RAKE receiver to process the multipath signal, further improving the receiving sensitivity. 3) Digital Adaptive Beamforming Antenna array configuration: Each drone is equipped with a 4-element conformal phased array antenna, operating in the Ku band; the antenna array is integrated on the surface of the drone fuselage, without affecting aerodynamic performance; each antenna element is independently controllable, and can achieve a beam scanning range of ±60°; Adaptive beamforming algorithm: The least mean square adaptive algorithm is used to adjust the weighting coefficients of the antenna elements in real time; the main lobe is aligned with the satellite with the direction of arrival of the satellite signal as a reference; at the same time, the DOA of the interference source is detected, and a null with a depth > 50dB is formed in the direction of interference. Multi-beam simultaneous communication: It supports the simultaneous formation of 4 independent beams, each aimed at 4 different satellites; enables simultaneous reception and transmission of signals from multiple constellations; and automatically switches to the beams of other satellites when a satellite is interfered with. The triple anti-interference technology adopts a layered and progressive approach to form an anti-interference system: First layer: Adaptive frequency hopping technology to avoid narrowband interference, making it impossible for the jammer to aim; The second layer: direct sequence spread spectrum technology, which resists broadband noise interference and extracts the signal from the noise; The third layer: Digital adaptive beamforming technology, which suppresses directional interference while enhancing useful signals.

9. The UAV swarm cross-domain control and data calibration system based on satellite communication collaboration according to claim 1, characterized in that, Based on the consistency of the multi-constellation observation data, the implementation process for detecting spoofing attacks on navigation signals is as follows: 1) Multi-constellation observation data acquisition: Simultaneously receive navigation signals from at least 4 different satellite constellations; receive signals from at least 4 satellites for each constellation, and acquire observation values ​​such as pseudorange, Doppler shift, ephemeris data, and carrier-to-noise ratio; 2) Multi-dimensional consistency test calculation a) Pseudorange residual consistency test The receiver's position and clock error are calculated independently using observation data for each constellation; the calculated position is substituted into the pseudorange equations of other constellations to calculate the pseudorange residuals; if the average pseudorange residual of a constellation exceeds a threshold, the signal of that constellation is determined to be abnormal. When more than one constellation shows abnormalities, a deception attack alert is triggered; b. Doppler frequency shift consistency test Calculate the expected Doppler frequency shift based on the receiver's motion state; compare the actual Doppler frequency shift of each constellation satellite with the expected value. If the Doppler frequency shift difference between multiple satellites in a constellation exceeds a threshold, it is determined to be a deception signal; c. Consistency check of ephemeris data Store historical ephemeris data for each constellation and build ephemeris models; compare the differences between the currently received ephemeris data and the historical models; if the ephemeris data shows abrupt changes or does not conform to the laws of orbital dynamics, it is determined to be a deceptive signal; d. Communication-Navigation Cross-validation Navigation technology derived from communication signals is used to calculate the receiver's position using satellite communication signals; the position derived from the communication signals is then compared with the position calculated from the navigation signals. If the difference exceeds the threshold, the navigation signal is determined to be deceptive; 3) Distributed collaborative deception detection By leveraging the distributed observation capabilities of the cluster, the accuracy of spoofing detection can be further improved: each node broadcasts its own spoofing detection results to other nodes within the autonomous system; The domain master node collects the detection results from all nodes and conducts a majority vote; when more than 2f+1 nodes detect a spoofing attack, a global spoofing attack is confirmed. The domain master node sends an alert to the entire network and activates the emergency response mechanism; 4) Emergency Response to Navigation Spoofing Attacks: When a navigation spoofing attack is detected, the system will automatically perform the following operations: Immediately cut off the signal from the deceived constellation and use the undeceived constellation for navigation; upgrade the anti-interference level: increase the frequency hopping rate, increase the spreading code length, and deepen the interference null trap; switch to cluster-intrinsic spatiotemporal reference navigation, completely independent of external satellite navigation signals; record the characteristic data of the deception attack and upload it to the global collaboration layer for analysis.

Images