Unmanned aerial vehicle flight authority management method and platform based on block chain technology

Abstract

The invention provides an unmanned aerial vehicle flight authority management method and platform based on a block chain technology, and relates to the field of unmanned aerial vehicle flight management, and the method comprises the steps: constructing a block chain network, and encrypting and chaining flight authority; deploying the student model after knowledge distillation to an edge computing device for risk prediction before flight and abnormal flight identification in flight; before the unmanned aerial vehicle takes off, the flight legality is verified, pre-flight risk prediction is executed through a student model, and then a smart contract is called to perform matching verification of on-chain permission records; during execution of the flight task, an abnormal score multi-model combination mechanism is introduced to assist the student model to dynamically identify abnormal flight, and an intelligent contract execution response mechanism is triggered when an abnormal behavior is detected; after the unmanned aerial vehicle lands, the flight record is encrypted and stored in the off-chain storage system, and the flight record hash value, the off-chain storage index and the related access credential are linked. The whole-process authority management of the flight task of the unmanned aerial vehicle is realized, and the supervision efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of drone flight management, and in particular to a drone flight rights management method and platform based on blockchain technology. Background Art

[0002] With the widespread application of drones in logistics and distribution, inspection and monitoring, agricultural plant protection, emergency rescue, and other fields, the number of drones in low-altitude airspace is increasing, placing higher demands on flight permission management, airspace scheduling, and flight safety. Flight permission management is a core component of drone operational safety. Its goal is to ensure that drones fly in legal airspace and during legal time periods according to their permitted plans, preventing safety incidents caused by permission conflicts, overflights, and illegal flights. Existing drone flight permission management systems are primarily based on centralized management platforms, using manual or semi-automatic methods for permission application, approval, and supervision. This traditional model has the following obvious shortcomings in actual application: most of the permission application, approval and execution data are stored in a centralized database, which lacks an anti-tampering mechanism and makes it difficult to ensure the credibility of the data throughout its life cycle; the permission approval response is slow and there is a lack of an automated review mechanism; the in-flight supervision mechanism is generally based on post-audit, the supervision mechanism lags behind, and abnormal behavior is difficult to detect in a timely manner; the current permission approval system has weak data linkage capabilities with the meteorological system, airspace situational awareness system, and historical flight data analysis system, making it difficult to achieve risk prediction and dynamic adjustment of permissions based on multi-dimensional data; the existing management platform still relies on static rules or manual experience for flight plan approval and abnormal behavior identification, and it is difficult to adapt to the dynamically changing airspace environment and complex mission requirements.

[0003] To sum up, there is an urgent need for a new method for drone flight permission management that can prevent tampering of permission data, intelligently and automatically execute approval processes, and detect real-time anomalies in flight behavior. It can also integrate multi-source data such as meteorological, airspace, and historical behavior for intelligent risk prediction and intelligently support permission decision-making. Summary of the Invention

[0004] This application provides a drone flight permission management method and platform based on blockchain technology, which solves the technical problems of existing technologies such as easy tampering of permission data, slow response to permission approval, lagging supervision mechanism, weak linkage between permission and risk assessment system, and lack of intelligent model to support permission decision-making, achieving the technical effect of improving supervision efficiency and flight safety.

[0005] In view of the above problems, this application provides a method for managing drone flight rights based on blockchain technology, which includes: Build a blockchain network and pre-deploy multiple smart contracts corresponding to drone flight rights management services. After the received flight rights are verified for compliance by the smart contracts, they are encrypted and stored on the blockchain. The student model, which has undergone knowledge distillation, is deployed to edge computing devices for pre-flight risk prediction and in-flight abnormal flight identification. The teacher model is a multimodal dynamic network structure model that integrates multi-source heterogeneous historical data such as drone flight records, meteorological factor data, and airspace congestion to learn flight risk assessment knowledge. This knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework. Before the drone takes off, it receives a flight application and calls a smart contract for compliance verification. If the verification passes, it obtains real-time weather data and permitted airspace congestion information and combines it with the flight application. It calls the student model deployed on the edge computing device to perform pre-flight risk prediction, obtain the flight risk level, and meet the flight requirements. The smart contract is called to verify the matching between the flight application and the on-chain permission record. After verification, an encrypted flight permit is generated, and the permit information and its hash value are stored in the blockchain. During the drone's flight mission, it continuously collects flight data, meteorological data, and airspace congestion information. An anomaly scoring multi-model combination mechanism is introduced to assist the student model in dynamically identifying abnormal flight behavior. Once abnormal behavior is detected, the smart contract is triggered to execute the early warning response mechanism. After the drone lands, the collected flight data, path trajectory, execution time and flight application content are compared and verified to generate a complete flight record. After encryption, it is stored in the off-chain storage system. The flight record hash value is calculated, and the smart contract is called to perform format verification on the on-chain data. After verification, the flight record hash value, off-chain storage index and related access credentials are written to the blockchain to realize full-process permission management of drone flight missions.

[0006] This application also provides a drone flight rights management platform based on blockchain technology, which includes: The blockchain network construction module is used to build a blockchain network, pre-deploy multiple smart contracts corresponding to the drone flight permission management business, and verify the compliance of the received flight permissions through the smart contracts before encrypting and storing them on the blockchain; The flight risk prediction module deploys the knowledge-distilled student model to edge computing devices for pre-flight risk prediction and in-flight abnormal flight identification. The teacher model is a multimodal dynamic network structure model that integrates multi-source heterogeneous historical data such as drone flight records, meteorological factors, and airspace congestion to learn flight risk assessment knowledge. This knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework. The pre-flight permission matching module is used to receive flight applications before the drone takes off, call smart contracts for compliance verification, obtain real-time meteorological data and permitted airspace congestion information in combination with the flight application, call the student model deployed on the edge computing device, perform pre-flight risk prediction, obtain the flight risk level, meet the flight requirements, call smart contracts to verify the matching of the flight application with the on-chain permission record, generate an encrypted flight permit after verification, and store the permit information and its hash value on the blockchain; The smart contract response mechanism module is used to continuously collect flight data, meteorological data, and airspace congestion during the UAV's flight mission. The anomaly scoring multi-model combination mechanism is introduced to assist the student model in dynamically identifying abnormal flight behavior. Once abnormal behavior is detected, the smart contract execution warning response mechanism is triggered. The post-landing data blockchain storage module is used to compare and verify the collected flight data, path trajectory, execution time and flight application content after the drone lands, generate a complete flight record, and store it in the off-chain storage system after encryption. It calculates the flight record hash value, calls the smart contract to perform format verification on the on-chain data, and after verification, writes the flight record hash value, off-chain storage index and related access credentials to the blockchain to realize full-process permission management of drone flight missions.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: Build a blockchain network, encrypt and store flight permission data on the chain, ensure the integrity, security and non-tamperability of permission information during transmission and storage, and provide a trusted permission infrastructure for flight management; by training a multimodal dynamic network structure model as a teacher model, integrating multi-source heterogeneous historical data such as flight records, meteorological factor data, and airspace congestion, achieve deep modeling and knowledge extraction of flight risks, improve the risk perception and generalization capabilities of the model, build a teacher-student heterogeneous knowledge distillation framework, transfer flight risk assessment knowledge to a lightweight student model, and deploy it to edge computing devices, achieve efficient risk prediction and anomaly detection in resource-constrained environments, improve model reasoning efficiency, and be suitable for the actual use of drones. Timely needs; in the pre-flight stage, the flight application is risk assessed through the edge student model to achieve intelligent risk control decisions before takeoff; the application is automatically compared and reviewed with the on-chain permissions in combination with the smart contract mechanism, which improves approval efficiency, reduces manual intervention, and ensures compliance of flight behavior; during the flight, the edge student model is used to identify abnormal flight behavior. Once an abnormality is identified, the smart contract automatically triggers an early warning and records the violation to the blockchain, enhancing the initiative and traceability of in-process supervision; after the flight mission is completed, the actual flight data is compared with the original flight application, and a flight record is automatically generated and encrypted and stored on the chain to form a complete, reliable, and tamper-proof flight behavior file, which is convenient for flight compliance assessment and subsequent responsibility tracing. The present invention realizes the intelligence, credibility and automation of drone authority management, and improves supervision efficiency and flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 A flowchart of a method for managing drone flight rights based on blockchain technology provided in an embodiment of the present application is provided. Figure 2 A schematic diagram of the structure of a drone flight rights management platform based on blockchain technology provided in an embodiment of the present application.

[0009] Explanation of the accompanying symbols: blockchain network construction module 11, flight risk prediction module 12, pre-flight authority matching module 13, execution of smart contract response mechanism module 14, post-landing data blockchain storage module 15. DETAILED DESCRIPTION

[0010] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0011] In order to make the purpose, technical solutions and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limiting this application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0012] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict, and the terms “first\second” involved are merely used to distinguish similar objects and do not represent a specific ordering of the objects. The terms “including” and “having” and any variations are intended to cover non-exclusive inclusions. For example, a process, method, platform, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application belongs. The terms used herein are for the purpose of describing the embodiments of this application only.

[0013] The present application embodiment provides a method for managing drone flight rights based on blockchain technology, such as Figure 1 As shown, the method includes: Build a blockchain network and pre-deploy multiple smart contracts corresponding to drone flight rights management services. After the received flight rights are verified for compliance by the smart contracts, they are encrypted and stored on the blockchain.

[0014] Specifically, blockchain technology is used to manage drone flight permissions. Its decentralized, tamper-proof, traceable, and smart contract-based nature addresses challenges inherent in traditional centralized management methods, such as trust, tampering, collaboration, and auditing. It also addresses the need for a trusted, secure, transparent, and collaborative data infrastructure within the drone operating environment. Before flight, permissions are logged on-chain. During flight, deviations from on-chain permissions trigger an on-chain violation record. After the flight, actual flight records are compared with permissions to generate a flight log, which is encrypted and uploaded to the chain for automatic traceability.

[0015] Furthermore, the construction of blockchain network includes: Based on the alliance chain architecture, adopt appropriate consensus algorithms to build the blockchain network and define node roles and their permissions; Pre-deploy multiple smart contracts corresponding to drone flight permission management services, and call smart contracts to verify the compliance of flight permissions; The flight permission information is compiled into a labeled data packet. The hash value of the standard data packet is calculated using a hash algorithm. The standard data packet is encrypted and stored in an independent off-chain storage system. The data packet hash value, off-chain storage location, access credentials, and transaction hash are written into the block record of the blockchain. A three-level composite key index table is established in an independent off-chain storage system. The first-level key is the hash prefix of the drone's identity information, the second-level key is the flight record, and the third-level key is the time range. The transaction hash on the chain is located through the third-level composite key, and the off-chain data access information is obtained from the on-chain record.

[0016] Specifically, when building a blockchain network, an appropriate consortium chain architecture, such as Fabric or FISCOBCO, should be selected. Consensus algorithms, such as PBFT, Raft, or HotStuff, can be chosen, with high efficiency and low latency mechanisms suitable for consortium chain scenarios. This ensures the real-time requirements of pre-flight drone reviews and in-flight records. Node roles include at least supervisory nodes, data provider nodes, and verification nodes. These nodes submit relevant data to the blockchain network in compliance with the Blockchain Data Format Specification. Pre-deployed smart contracts are used to verify the compliance of flight permissions uploaded to the blockchain, verify the compliance of flight applications before takeoff, verify the match between flight applications and on-chain permission records, and verify the format of flight records uploaded to the blockchain. These contracts are core tools for achieving automation, security, and compliance in drone flight management. By encoding rules into the blockchain and combining them with off-chain storage and indexing technologies, data privacy and performance can be guaranteed while ensuring regulatory compliance, providing reliable technical support for the large-scale deployment of drones. Verification nodes, represented by relevant industry organizations or trusted third-party organizations, are responsible for verifying the authenticity and compliance of data. A standardized drone data packet is constructed. This data packet serves as the fundamental information carrier for a series of operations, including flight applications, permission authorization, and behavior auditing. Its standardized structure facilitates on-chain smart contract processing. The data packet includes drone identity information, including a unique identifier and public key certificate; permitted flight areas, which use spatial grid coding to represent airspace boundaries; takeoff point coordinates, represented by latitude and longitude or grid numbers; mission types, such as inspection and surveying, using predefined classification codes; and mission flight time, which includes takeoff time and planned flight duration or end time for time conflict detection. The data in the standard data packet is hashed using SHA-256 to standardize the data. The unique identifier of the drone identity is used; permitted flight areas are encoded into fixed-format strings, such as GeoHash; takeoff point coordinates are standardized to uniform decimal points for longitude and latitude; mission types are represented using standard classification codes; and time fields are formatted as a unified UTC timestamp or ISO8601 string. The standardized fields are concatenated into a string in a fixed order, with a unified delimiter between fields, to form a concatenated hash string. The concatenated hash string is used as input and encrypted using the SHA-256 hash algorithm to generate a fixed-length (256-bit / 32-byte) hash value. This hash value serves as the unique identifier for the standard data packet and can be written along with off-chain storage location, access credentials, transaction hashes, and other information. On-chain, this ensures that on-chain records are strongly bound to off-chain data and are verifiable and traceable. Blockchain networks primarily record lightweight information such as hash values, data pointers, and status summaries. They are not suitable for directly storing large or sensitive data. Real data is typically stored off-chain, and the constructed three-level composite key index table serves as a bridge to efficiently and securely locate off-chain data and associate it with on-chain records.The three-level composite key index table has a first-level key, an identity information hash prefix, a second-level key, and a flight airspace code. The third-level key is the time range. Through the three-layer nested structure, the search space is gradually narrowed, and data is searched in the form of drone-airspace-time. The index ultimately points to the location of the off-chain data and its on-chain anchor point.

[0017] The student model that has undergone knowledge distillation is deployed to edge computing devices for pre-flight risk prediction and in-flight abnormal flight identification. The teacher model is a multimodal dynamic network structure model, which is used to integrate multi-source heterogeneous historical data such as drone flight records, meteorological factor data, and airspace congestion to learn flight risk assessment knowledge. The learned flight risk assessment knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework.

[0018] Specifically, a lightweight student model is deployed on edge computing devices to perform pre-flight risk prediction and in-flight anomaly detection. The risk prediction results serve as smart contract invocation conditions to assist in the intelligent review of flight permissions. During anomaly detection, flight anomalies are identified, triggering smart contract alerts and recording the anomaly data to the blockchain. Before takeoff, the model integrates flight applications with real-time environmental data to predict the mission's risk level. The risk prediction results directly serve as one of the criteria for the subsequent smart contract authorization review to determine whether to grant an encrypted flight permit. The model continuously analyzes the drone's flight status data and real-time weather / airspace change data to identify abnormal behavior or potential flight violations. Upon detecting an anomaly, the edge device can directly invoke a smart contract to automatically record the anomaly on-chain, trigger an alert mechanism, or engage relevant control systems. A multimodal dynamic network structure model is constructed, integrating structured information such as the drone's flight trajectory, weather trends and anomaly indicators, airspace congestion, mission type, and execution time into multimodal input data. This integrated input data is organized into a dynamic graph to simulate the dynamic evolution of the drone's status and environmental factors, learning the drone's risk evolution path and enabling modeling and dynamic prediction of flight risks. The network structure is complex and the inference cost is high, making it unsuitable for direct deployment on edge devices. The multimodal dynamic network structure model is the teacher model. The teacher-student heterogeneous knowledge distillation framework is used to transfer the knowledge of the teacher model to the student model deployed on the edge device. The multimodal dynamic network structure model is the knowledge source, the teacher-student heterogeneous knowledge distillation framework is the "transmission channel", and the edge student model is a lightweight executor, achieving high-performance and efficient flight risk monitoring.

[0019] Furthermore, the teacher model is a multimodal dynamic network structure model that is used to integrate multi-source heterogeneous historical data such as drone flight records, meteorological factor data, and airspace congestion to learn flight risk assessment knowledge. This knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework, including: Define drone nodes in a multimodal dynamic network structure model, obtain drone flight records based on historical data, and extract drone node features based on the flight records. Static features include take-off and landing points, historical mission types, historical mission flight times, and historical abnormal events. Dynamic features are generated by sampling historical flight trajectories at preset time intervals to generate a historical flight trajectory time series, which contains the spatial position of each time point. The spatial position includes the drone's geographic coordinates and flight altitude at that time point. The meteorological nodes in the multimodal dynamic network structure model are defined. The static characteristics of the meteorological nodes are the geographical locations of the meteorological monitoring points within the permitted flight airspace. Based on the 12-hour window before the start time of the historical flight mission, the historical meteorological trend feature sequences of the meteorological monitoring points associated with the historical flight trajectory of the UAV are extracted. At the same time, the type, occurrence time, and duration of abnormal weather are identified and recorded to form an abnormal weather time series feature sequence. The above two feature sequences are aligned with the time series of the UAV's historical flight trajectory to generate a historical meteorological trend feature sequence and a historical abnormal weather sequence synchronized with the time series of the UAV's historical flight trajectory. These are the dynamic characteristics of the meteorological nodes. Airspace nodes are defined in a multimodal dynamic network structure model. The static characteristics of airspace nodes are grid cells divided by a combination of space and altitude in the permitted flight airspace. Based on the historical flight trajectories of drones, the congestion degree of each grid cell in each time period is counted within a 24-hour rolling sliding window. The airspace congestion trend characteristics are extracted using a time series modeling method to form a time series feature sequence of airspace congestion trends covering multiple time points. This sequence is aligned with the time series of drone flight trajectories to generate a historical airspace congestion trend feature sequence synchronized with the drone's historical flight trajectories, which serves as the dynamic characteristic of the airspace node. A multimodal dynamic network structure model consisting of drone nodes, meteorological nodes, and airspace nodes was constructed. Spatial edges, meteorological influence edges, and task-related edges were set between nodes. The network model was trained using reinforcement learning to establish spatiotemporal dependencies between various nodes and obtain the spatiotemporal risk characteristics of drone nodes. The trained multimodal dynamic network structure model is used as a teacher model to generate pseudo labels and risk score vectors for flight risk identification. The IsolationForest model is used as the student model. Through the heterogeneous knowledge distillation mechanism, the pseudo labels and risk score vectors in the teacher model are transferred to the student model to complete the construction of the student model.

[0020] Specifically, a teacher model trained on drone flight history and multi-source environmental (weather, airspace) data transfers flight risk knowledge to a lightweight student model deployed at the edge, enabling the migration and execution of edge-side risk assessment capabilities. The static features of drone nodes include take-off and landing points, historical mission types, historical mission flight times, and abnormal event records. These features are stable and unchanging, making them suitable as static node features and the basic background for risk assessment. The take-off and landing points are used to determine the type of airspace in which historical missions were located. Different historical mission types have different requirements for routes, flight altitudes, and speeds, and thus different risk patterns. These can be used to model the relationship between missions and behavioral and environmental risks. Historical mission flight times determine whether a flight is at peak times and can assist in determining whether it is susceptible to airspace congestion or weather interference. Abnormal event records reflect whether the drone has been prone to problems in history and have an "implicit risk" label, which helps provide prior information for flight risk prediction. The dynamic features of a drone node are historical flight trajectory time series, obtained by sampling historical flight trajectories at set intervals. A drone flight trajectory is a continuous record of the drone's actual movement path in time and space from takeoff to landing. It consists of a series of time-stamped spatial location information, describing the drone's flight process and behavior. The generated historical flight trajectory time series can reflect flight speed, directional changes, stability, whether it is executing according to the planned trajectory, whether it is flying within the permitted airspace, and whether there are any altitude anomalies or drift. It is crucial for in-flight anomaly detection and deviation prediction. Static features provide "risk context" before flight, while dynamic features support "behavior monitoring" during flight. At meteorological nodes, historical meteorological trend features reflect the speed and direction of weather changes within a specific historical time period (e.g., wind speed gradients, rapid humidity increases, etc.), helping to identify potential flight risks in advance. Abnormal weather conditions (e.g., thunderstorms, low visibility, and severe convection) directly constitute potential or actual flight risk sources and can be used to train models to identify "high-risk flight windows." By aligning the historical dynamic features of meteorological nodes with the time points of historical UAV flight trajectories, the model can identify moments when a flight trajectory passes through hazardous weather segments, enabling "joint risk reasoning under spatiotemporal conditions." The introduction of airspace nodes addresses the dimension of spatial resource competition and traffic conflict risk that impacts flight safety. Flight density influences flight risk. Airspace congestion is a significant source of flight conflict, delays, or mission interruptions for UAVs. Static features are spatial-altitude grid cells derived from the division of permitted flight airspace. This enables flight-to-space mapping, uniformly mapping the trajectories of different historical flight missions to the same grid, facilitating the accumulation and comparison of statistical features. Dynamic features are time-series features of historical airspace congestion trends extracted based on statistical changes in flight density over a rolling time window. Congestion status varies over time, reflecting current or impending airspace load conditions.

[0021] Furthermore, a historical meteorological trend feature sequence and a historical abnormal meteorological sequence synchronized with the UAV's historical flight trajectory time series are generated, including: Based on historical data, the historical flight missions of each UAV are obtained. A time window of 12 hours before the start time of the mission is set. The time series data of meteorological factors of each meteorological monitoring point within the window are extracted and standardized. The meteorological factors include wind speed, wind direction, temperature, precipitation and visibility. Based on the spatial position corresponding to each time point in the historical flight trajectory time series of each UAV, combined with the monitoring range of each meteorological monitoring point, the meteorological monitoring points covering the spatial position are selected. If multiple meteorological monitoring points cover the same position, the meteorological monitoring point closest to the spatial position is selected to establish the corresponding association relationship between the meteorological monitoring point and each time point in the historical flight trajectory time series of the UAV; The meteorological factor time series data of the meteorological monitoring points associated with each time point in the historical flight trajectory time series of the UAV are input into the LSTM time series model, the meteorological factor change trend embedding vector of each time point is extracted, and the historical meteorological trend feature sequence of each associated meteorological monitoring point is generated; Set an abnormal judgment threshold for each type of meteorological factor, traverse the meteorological factor time series data of the associated meteorological monitoring points, mark abnormal meteorological conditions, record the abnormal meteorological type, occurrence time and duration, and construct a historical abnormal meteorological sequence for the associated meteorological monitoring points; The historical meteorological time series trend characteristics and historical abnormal meteorological time series characteristics of each associated meteorological monitoring point are aligned with the UAV historical flight trajectory time series in time, forming a historical meteorological trend feature sequence and historical abnormal meteorological sequence synchronized with the UAV historical flight trajectory time series.

[0022] Specifically, according to documents from the Civil Aviation Administration of China, such as the "Guidance on Airworthiness Certification of UAVs Based on Operational Risks" and the "Guidelines for Airworthiness Safety Assessment of Civil Unmanned Aircraft", the Civil Aviation Administration of China's meteorological management of UAV operations mainly focuses on requiring operators to evaluate the real-time meteorological conditions of the flight area, such as wind speed, precipitation, visibility, etc., based on mission requirements and UAV performance, to avoid flying in extreme weather such as strong winds, rain, snow, and heavy fog. Therefore, the wind speed, wind direction, temperature, precipitation and visibility time series data within the time window of 12 hours before the start time of each UAV's historical flight mission are extracted. Time series data refers to a data set recorded in chronological order that reflects the changes in meteorological factors over time. For example, a monitoring point records meteorological factors once an hour within 12 hours to form a time series data set containing 12 data points. To build and train an LSTM time series model, the meteorological factor time series data from meteorological monitoring points associated with each time point in the drone's historical flight trajectory time series is fed into the LSTM. The continuous time series data is segmented into fixed-length subsequences. For example, if the time series is 12 hours, the data can be divided into multiple sliding windows (with a step size of 1 hour) to capture local trends. For each subsequence, the LSTM unit considers the input data at the current time step and the hidden state at the previous time step. For example, the meteorological factor time series data is X = {x1, x2, ⋯, xt}, where xt represents the meteorological factor data vector at time t (including standardized values of wind speed, wind direction, temperature, precipitation, visibility, etc.). The LSTM unit controls the flow of information through its internal gate structure (forget gate, input gate, and output gate). The forget gate determines which information is forgotten from the previous hidden state, the input gate determines which new information is added to the current cell state, and the output gate determines which information from the current cell state is output as the current hidden state. As time steps progress, the LSTM continuously updates its hidden state, capturing long-term dependencies in the time series data of meteorological factors. The output is an embedding vector representing the changing trend of the meteorological factors at each time point. These embedding vectors form the historical meteorological trend feature sequence for each associated meteorological monitoring point. For example, if the embedding vector yt is obtained for each time point t after LSTM processing for a certain meteorological monitoring point, then the embedding vectors {y1, y2, ⋯, yT} for all time points constitute the historical meteorological trend feature sequence for that meteorological monitoring point, where T is the total duration of the time series. Given that the sampling frequency of drone trajectories is 10-20 Hz, the sampling frequency of meteorological monitoring is 1-5 Hz, and the sampling frequency of airspace congestion is 1-2 Hz, to achieve unified alignment and synchronous modeling of multimodal features, a unified sequence based on the time axis of drone trajectories is constructed for model processing. With the time series of the drone's historical flight trajectories as the main axis, the unified temporal resolution is set to the lowest common multiple frequency (e.g., 20 Hz).On this basis, linear interpolation is performed on the historical meteorological trend feature sequences and the historical airspace congestion trend feature sequences to complete them to high-frequency time steps consistent with the trajectory time points. The interpolated feature sequences are synchronized with the trajectory time series, forming a unified high-frequency multimodal input sequence for model processing, effectively preserving the dynamic trends of various temporal features. For the historical abnormal meteorological sequence, this sparse discrete event sequence uses the nearest value method to match each flight trajectory time point ti with the nearest meteorological observation time point tj. If the time window belonging to tj contains an abnormal event, it is marked as an anomaly. The historical abnormal meteorological sequence and the historical UAV flight trajectory are synchronized on the same time axis. After alignment, data verification and optimization are performed. The changing trends of the aligned trajectories, meteorological factors, and airspace congestion on the same time axis are plotted. First, the temporal synchronization and logical consistency of the three types of data are compared visually (for example, whether sudden changes in wind speed affect flight trajectories or congestion fluctuations). The alignment error is quantified using Euclidean distance to identify outliers or areas with large deviations. Then, based on the accuracy of the anomaly detection results, the data alignment method is optimized, such as upgrading the interpolation strategy or introducing dynamic time warping.

[0023] Furthermore, a historical airspace congestion trend feature sequence synchronized with the historical flight trajectory of the UAV is generated, including: Divide the permitted flight airspace into spatial and altitude dimensions, and construct a spatial-altitude grid structure, where each grid cell corresponds to a latitude, longitude, and altitude layer; Based on the set 24-hour rolling sliding window, the number of drones appearing in each grid cell within each window is counted, and the average frequency of drone appearance within the window step is calculated to obtain the congestion degree of each grid cell in each time period, forming a historical congestion time series for each grid cell in different windows; The historical congestion time series corresponding to each grid unit is input into the LSTM network to extract the trend of congestion change over time and obtain the historical congestion trend feature sequence of each grid unit; According to the spatial position of each time point in the drone's historical flight trajectory time series, the grid unit to which it belongs is located, and the historical congestion trend feature sequence of the grid unit is aligned with the drone's historical flight trajectory time series in time to generate a historical airspace congestion trend feature sequence synchronized with the drone's historical flight trajectory time.

[0024] Specifically, a geographic coordinate grid (e.g., longitude and latitude grid) is used, with the grid size (e.g., 1 km × 1 km) set according to airspace management requirements. The altitude dimension is divided by flight level (e.g., every 300 meters as a layer), forming a three-dimensional grid cell (e.g., grid ID = (longitude grid, latitude grid, altitude level). Each grid cell contains a unique identifier for its spatial location, altitude range, and level, which is used for subsequent statistics and mapping. A 24-hour rolling sliding window is used for time windowing, with a step size of 1 hour (i.e., statistics are updated every hour). Within each time window, the number of drone occurrences in all grid cells is counted. The average drone occurrence frequency within the window step size (e.g., hourly mean) is calculated as the congestion indicator for that grid cell. The historical congestion time series for each grid cell is obtained, e.g., [congestion_t1, congestion_t2, ..., congestion_t24], representing the historical airspace congestion. The historical congestion time series for each grid cell is normalized to ensure that data from different grid cells are in the same dimension. The historical spatial congestion time series is reshaped into the format of (time step length × feature dimension). For example, if the time series length is 24 hours and the congestion dimension is 1, the input dimension of the LSTM network is 24 hours × 1. This input dimension setting means that each time point corresponds to a congestion value. The LSTM controls the flow of information through forget gates, input gates, and output gates. For example, the congestion time series for a grid cell is [0.5, 0.6, 0.8, 1.0, 0.9, 0.7, ...] (24 hours). After normalization, the LSTM input is fed into the model. At the third hour (congestion level 0.8), the input gate amplifies the "increasing congestion" signal. The output hidden state vector may be a high-dimensional feature (e.g., [0.32, -0.15, 0.87, ...]), encoding the trend of "rapidly increasing congestion." Ultimately, the collection of vectors from all 24 time steps constitutes the historical congestion trend feature sequence for that grid cell, which can be used for subsequent spatiotemporal analysis or trajectory alignment tasks.

[0025] Furthermore, the multimodal dynamic network structure model includes: Construct a multimodal dynamic network structure model as follows: The historical flight trajectory time series is divided into multiple continuous time slices according to the time dimension. A corresponding static network graph is constructed in each time slice. Each static network graph consists of nodes and edges. The nodes include drone nodes, weather nodes and airspace nodes, and the edges include space edges, weather impact edges and mission-related edges. Spatial edges are used to connect spatially adjacent nodes and obtain the airspace congestion trend of adjacent node areas. The edge weight is obtained by jointly calculating the inverse of the geographical distance between nodes and the historical airspace congestion trend feature sequence of the airspace unit in which they are located. Among them, adjacent nodes include drone nodes and drone nodes, drone nodes and airspace nodes, and airspace nodes and airspace nodes. The meteorological influence edge is used to connect the meteorological node with the UAV node within its monitoring range. The edge weight is determined by calculating the matching degree between the meteorological time series trend characteristics of the meteorological node and the meteorological factors at the corresponding UAV location time point. Task association edges are used to connect UAV nodes of the same or similar task types. The edge weight is calculated by combining the semantic similarity between task types and the proximity of task execution time. Connect the same nodes in adjacent time slices to form time edges, which are used to model time evolution relationships; In each time slice, the graph attention network performs type-aware attention aggregation on various nodes of each static network graph, learns the representation of each type of node, and realizes the propagation of various node representations in the time dimension through time edges. At the same time, the node state evolution mechanism is introduced to obtain the spatiotemporal feature representation of the drone node in each historical time slice, completing the construction of the entire multimodal dynamic network structure model.

[0026] Specifically, the flight risk prediction model built on a multimodal dynamic network integrates multi-source heterogeneous data such as historical flight trajectories, weather trends, and airspace congestion to construct a spatiotemporal correlation network graph structure, which can capture the complex interactive relationship between drones and the environment. For example, the meteorological impact edge between meteorological nodes and drone nodes (edge weight is based on the degree of meteorological matching) can identify the impact of adverse weather on flight paths in advance (such as the risk of deviation caused by strong winds), and the spatial edge of airspace nodes (edge weight combines geographical distance and congestion trend) can predict the collision risk in high-density airspace, and has the ability to predict risks before flight; through the graph attention network (GAT), type-aware attention aggregation of nodes (drones, weather, airspace) can dynamically update node representations (such as when the drone trajectory deviates, the weight of the meteorological impact edge changes). Combined with the temporal evolution mechanism of time edge propagation, abnormal behaviors (such as sudden changes in trajectories and sudden increases in airspace congestion) can be captured in real time, and have real-time anomaly detection capabilities. The spatial edge weight is jointly calculated by the inverse of the geographical distance between nodes and the historical airspace congestion trend feature sequence. The formula is: , Among them, d ij For node u i with u j Geographical distance, Congestion(u i ,u j ) is the node u in the spatial unit i with u jThe time series of airspace congestion trends in the region, where Norm(∙) is the normalization function and α∈[0,1] is the adjustment parameter used to balance the weights of the inverse distance and congestion trend (for example, α=0.6 means that the distance influence accounts for 60% and the congestion trend accounts for 40%). The inverse of geographic distance reflects spatial proximity (the closer the distance, the greater the influence), and the airspace congestion trend quantifies the regional flight density (the higher the congestion, the stronger the correlation). The two are calculated jointly to take into account spatial layout and dynamic environment. The meteorological influence edge weight is based on the degree of matching of meteorological factors between meteorological nodes and drone nodes, and the formula is: , Here, Simmeteo(mk, dL) is the similarity (cosine similarity is optional) between the historical weather trend feature sequence of meteorological node mk and the flight trajectory time series of drone node dL. AnomalyScore(mk, dL) is the score of the impact of abnormal weather (such as strong winds and heavy rain) on the flight of meteorological node mk during the flight time of drone node dL (for example, if strong winds are present and last for >2 hours, the score is set to 1.5; if there are no abnormalities, the score is set to 1). β∈[0,1] is a tuning parameter used to balance the weight of normal weather matching and abnormal weather impact (for example, β=0.7 indicates a 70% normal match and a 30% abnormal impact). Weather impact edges match normal weather trends (such as wind speed and temperature) through similarity, while introducing abnormal weather scores (such as strong winds and low visibility) as a penalty term to enhance sensitivity to extreme weather. The task association edge weight combines the semantic similarity of task types and the proximity of task execution times. The formula is: , Here, SemSim(tp, tq) is the semantic similarity between task types tp and tq of drone nodes dp and dq (cosine similarity can be calculated using pre-trained task embedding vectors), Tp and Tq are the execution times of tasks tp and tq, respectively, |Tp-Tq| represents the time difference, γ is the normalization coefficient for semantic similarity, and λ is the time decay coefficient (for example, λ=0.1 means that the larger the time difference, the faster the weight decays). For task-related edges, semantic similarity captures the logical associations between task types (e.g., "inspection" and "surveying and mapping" are similar), while the time decay function reinforces the relevance of recent tasks (e.g., the smaller the time difference, the higher the weight). A multimodal dynamic network architecture model combines a graph attention network (GAT) with a temporal evolution mechanism to learn and dynamically model the spatiotemporal features of drone, weather, and airspace nodes. In each time slice, the GAT performs a self-attention mechanism based on node type (drone, weather, airspace), dynamically aggregates information from neighboring nodes, and generates a node representation that integrates type features with local interactions (for example, drone nodes focus on mission type and airspace congestion, while weather nodes focus on wind speed and abnormal weather). By connecting identical nodes in adjacent time slices (temporal edges), the node representation from the previous moment is transferred to the current slice, forming a temporal evolution path. A state evolution mechanism (e.g., a time series model predicts node state changes) is also introduced to update node features and feed them back to the graph attention network for continuous optimization of dynamic features. Ultimately, the model outputs a spatiotemporal feature representation of the drone node within each time slice. This representation captures both static attributes (e.g., mission type and historical trajectory) and local environmental interactions (e.g., airspace congestion and weather impacts), while also modeling dynamic evolution processes (e.g., the real-time impact of sudden weather events on trajectories) through the temporal dimension.

[0027] Furthermore, the construction of the student model is completed, including: The trained multimodal dynamic network structure model is used as the teacher model. Historical multimodal data is input to extract the spatiotemporal feature representation of drone nodes. The classification head outputs the probability distribution of normal and abnormal flight. High-confidence pseudo-labels are filtered through a threshold as the initial abnormal samples. At the same time, the regression head outputs a risk score vector to quantify the contribution of each risk dimension. High-confidence pseudo-labels are used as strong supervision samples. The label propagation algorithm is used to propagate feature similarity on low-confidence samples to generate extended abnormal samples. The original features and risk score vectors are combined to construct a complete semi-supervised training set. The original features come from the feature sequences extracted for each node by constructing a multimodal dynamic network structure model. A feature adapter is introduced to map the risk score vector to the feature space of the student model. Soft labels are generated based on the probability distribution of normal and abnormal flights. KL divergence loss is used to preserve the uncertainty of the soft labels, guiding the student model to learn richer decision boundaries. Furthermore, the risk score is used to adjust the weights of IsolationForest training samples, so that the model prioritizes high-risk samples during training. The student model parameters are optimized using a joint loss function to complete knowledge distillation and obtain the distilled parameters. In the process of training the IsolationForest model based on the semi-supervised training set and distilled parameters, a three-level progressive sample selection mechanism is introduced to expand the training sample set. This mechanism consists of three stages: first, the initial abnormal samples are used as core abnormal samples; second, the expanded abnormal samples are used as supplementary abnormal samples; finally, during the model training process, samples with short path lengths and high risk scores are dynamically identified and included in the training as dynamic abnormal samples; On the expanded training sample set, the IsolationForest model prioritizes high-risk feature dimensions for node splitting. It also introduces a path length penalty mechanism based on risk scores, imposing penalty weights proportional to the risk scores on abnormal samples to improve the anomaly detection performance of the training model. The trained student model is quantized and compressed and then deployed to the edge device to complete the construction of the student model.

[0028] Specifically, the multimodal dynamic network architecture model extracts spatiotemporal features of drone nodes, such as location, velocity, mission status, and weather interaction, from historical time slices using a graph attention network. These features form the model's underlying representation of drone behavior and are used to capture spatiotemporal dynamic patterns. The classification head's pseudo-label generation directly relies on the spatiotemporal feature representations. These features provide the classification head with a discriminative basis, while pseudo-labels are the result of the classification head's interpretation of these features. Through a fully connected layer or attention mechanism, the decision boundary for distinguishing normal from abnormal flight is learned. For example, the spatiotemporal feature vector hi (including location, velocity, and spatial density) is input and the probability distribution P(yi∈{normal, abnormal}) is output. High-confidence samples (e.g., confidence > 0.8) are filtered using a threshold to generate pseudo-labels, which serve as initial abnormal samples. The regression head, similarly based on the same spatiotemporal feature representations, such as the drone node's location, velocity, and weather interaction, aims to quantify the contribution of risk dimensions, such as collision probability and weather impact. The risk dimension model learns the contribution of each risk dimension through the regression head and normalizes the contribution of each risk dimension to generate a standardized risk score vector. High-confidence pseudo-labels (discrete) and standardized risk score vectors (continuous) together constitute the supervisory signal, guiding the student model to learn the discriminative and risk-aware capabilities of the teacher model. High-confidence pseudo-labels (e.g., confidence > 0.8) generated by the teacher model are used as strong supervisory examples to provide a clear discrimination boundary for the model. A label propagation algorithm is used to infer labels for low-confidence examples based on feature similarity, generating extended anomaly samples to supplement the training set. Raw features (e.g., spatiotemporal characteristics of drone nodes) are further concatenated with risk score vectors (e.g., standardized contributions such as collision probability and meteorological impact) to form an enhanced feature space, thereby improving the model's sensitivity to key risks. Ultimately, by jointly optimizing the supervisory loss (cross-entropy loss for strong supervision) and the weak supervision loss (KL divergence loss for label propagation), the student model (e.g., Isolation Forest) is guided to learn the discriminative and risk-aware capabilities of the teacher model. Dynamically adjusting weight coefficients during training balances the influence of strong and weak supervisory signals. Feature adapters are introduced to mitigate distribution differences and optimize pseudo-label reliability to enhance model robustness. The core goal of feature adapters is to align the feature spaces of different models or modalities and solve compatibility issues caused by differences in feature distribution.The core logic of building a feature adapter is to convert the original feature space (such as a high-dimensional risk score vector) into a feature space compatible with the student model by designing a mapping relationship between the target interface (Target) and the source interface (Adapter). The specific steps include: defining the source interface (original feature dimension) and the target interface (student model input dimension), designing the adapter class (inheritance or combination) to implement the feature conversion logic (such as full-connection layer compression / expansion, nonlinear activation, normalization and residual connection), and jointly optimizing the adapter parameters through supervision loss (such as categorical cross entropy) and alignment loss (such as KL divergence or MSE). At the same time, dynamic adjustment strategies (such as hidden layer size adaptation) and lightweight design (such as sparse connection) are introduced to improve compatibility and efficiency. Finally, by freezing the source model parameters and only training the adapter, seamless alignment of the feature space and enhanced model performance are achieved. The training set is expanded through a three-level progressive abnormal sample screening mechanism. First, the initial abnormal samples with high-confidence pseudo-labels are used as the core benchmark. Second, the expanded abnormal samples generated by label propagation supplement data diversity. Finally, during the iterative process of IsolationForest model training, the path length of each sample in the forest structure is dynamically monitored. Because IsolationForest's judgment of the degree of abnormality depends on the path length of the sample (abnormal samples are more easily isolated and have shorter paths), the system will identify in real time samples with abnormally short path lengths and corresponding higher risk scores. As dynamic abnormal samples, such samples may be emerging potential risk points, helping to complete the boundary conditions missed in the training set and form a multi-stage coverage abnormal sample system. After the expansion of the complete training sample set is completed, the IsolationForest model enters the training phase. Two key mechanisms are introduced in this process: one is the feature dimension priority splitting strategy, where the model prioritizes high-risk feature dimensions that have a significant impact on flight safety, such as collision probability and wind speed mutation, when constructing the decision tree, in order to enhance the model's sensitivity to key risk factors; the other is the path length penalty mechanism, which assigns a penalty weight proportional to its risk score to each abnormal sample. That is, high-risk samples are given greater importance in training, and their recognition ability is enhanced through the path structure.

[0029] Before the drone takes off, it receives a flight application and calls a smart contract for compliance verification. If the verification is passed, it obtains real-time meteorological data and congestion information of the permitted airspace and combines it with the flight application. It calls the student model deployed on the edge computing device to perform pre-flight risk prediction, obtain the flight risk level, meet the flight requirements, and call the smart contract to verify the matching of the flight application and the on-chain permission record. After verification, an encrypted flight permit is generated, and the permit information and its hash value are stored in the blockchain.

[0030] Specifically, the flight application must be actively submitted by the applicant. The main body of the flight application includes: the drone's unique ID (such as registration number, public key) representing the drone's identity; the operator's identity (such as operator ID, certification certificate signature) representing the legitimacy of the operator; and the flight mission description, including takeoff and landing points (GPS coordinates or geographic grid), flight path (path coordinate sequence or reference route number), mission type (such as logistics, inspection, emergency response, etc.), payload information (such as cameras, cargo, sensors, etc.), and flight time (start and end times). Before the drone takes off, after receiving the flight application, the system calls the first smart contract to automatically verify the legitimacy of the drone's identity, the legitimacy of the operator, the integrity and semantic legitimacy of the flight mission content, and airspace and time compliance. This step is a precursor to flight risk assessment and permit generation, ensuring that the submitted application is compliant at the logical and policy levels. Once verified, the system will proceed to the next stage of risk prediction. After the pre-flight risk prediction is passed, the second smart contract is called to perform on-chain permission matching verification on the flight application. This process relies on the registration permissions and dynamic announcements on the blockchain to realize automatic review of the compliance of flight permits, ensuring that all flight behaviors are carried out within the established policies and authorization scope. Main verification content: Flight airspace authority verification, field source: flight path in the application content, verification method: call the airspace permission range stored on the chain, check whether the flight path is completely within the authorized airspace range, check whether the path crosses or approaches sensitive areas, check whether there are time-limited temporary prohibited areas, output: path compliance or path out of bounds; flight time authority verification, field source: take-off time and duration in the application content, verification method: compare with the available time window of the drone registered on the chain, check whether the selected time period conflicts with the temporary no-fly time, and can be sharded according to the flight time span to verify whether each time period is within the permitted range, output: time compliance or time violation; mission type and payload authority verification, field source: in the application The task type (such as logistics, inspection, emergency) and payload information (camera equipment, sensor type, etc.) in the content, verification method: query the on-chain registration information to see whether the current drone / operator has been authorized to perform this type of task, check whether it has permission to carry specific payloads (such as whether it is allowed to carry thermal imaging equipment, broadcasting equipment), and for specific sensitive tasks (such as emergency rescue), determine whether it has special flight qualifications, output: task authority match or mismatch; operator and drone binding authority verification, field source: operator ID (or public key) carried in the application content, verification method: verify whether the operator is a drone-bound user registered on the chain, check whether the user has the authority to control the drone of this model or mission type, output: consistent permissions or unauthorized operation.Verification passed: Generate an encrypted flight permit (including flight ID, permitted time period, mission summary, risk score, path hash, etc.) and write it into the blockchain; Verification failed: Return the reason for failure (path out of bounds, time violation, illegal payload, etc.), refuse to generate the permit, and trigger notification or audit record.

[0031] Furthermore, a pre-flight risk prediction is performed to obtain the flight risk level, including: Extract the pre-flight mission type, flight path, flight time, and take-off and landing points from the flight application to form static mission parameters; Based on the flight path and take-off and landing points, meteorological factor data of the associated meteorological monitoring points are collected in real time, and the congestion of the permitted airspace network units passed by the flight path is counted in real time. The meteorological factor data and congestion degree form dynamic mission parameters; The static and dynamic task parameter features are fused into a unified structured feature vector and input into the student model deployed on the edge computing device. Based on the student model's inference calculation of the fused features, a continuous potential risk score is output. Preset risk level threshold and compare it with the output potential risk score, When the potential risk score is greater than the first threshold, it is judged as high risk. The system automatically rejects the flight application and calls the smart contract to record the high-risk attempt in the blockchain. When the potential risk score is between the second threshold and the first threshold, it is determined to be medium risk, and the operator is prompted to supplement risk mitigation measures and resubmit the application; When the potential risk score is lower than the second threshold, it is judged as low risk and enters the subsequent process.

[0032] Specifically, during the pre-takeoff risk prediction phase for drones, the system collects real-time meteorological data (such as wind speed, wind direction, visibility, temperature, and precipitation probability) and congestion within the permitted airspace (such as the location, number, and altitude of other drones in the airspace). This data is then quantified as the congestion level for each grid cell in the airspace. Subsequently, the flight mission type, flight path, flight time, takeoff and landing points, meteorological data, and congestion level are fused into a unified structured feature vector. The fusion process is as follows: 1) Continuous numerical features, such as wind speed, visibility, and congestion level of airspace grid cells, are processed using Z-score normalization or Min-Max normalization to eliminate dimensionality differences. 2) Categorical features, such as logistics / inspection / emergency, are converted into computable form using one-hot encoding or embedding vectors, e.g., mission type "logistics" → [1, 0, 0]. 3) Path complexity extraction: Flight path data (e.g., longitude and latitude sequences) is algorithmically converted into quantifiable complexity metrics, such as total path length, number of turns, and turn frequency, which are then normalized and incorporated into a feature vector. 4) Fixed-length vector fusion: All normalized values and encoded categorical information are concatenated into a one-dimensional vector, a unified structured feature vector, such as x = [wind speed, wind direction, visibility, mission type code, congestion, path complexity, etc.]. This unified structured feature vector is then fed as input to a student model optimized by knowledge distillation, deployed on an edge computing device, for low-latency flight risk prediction. During the student model's training phase, it receives supervisory signals from a teacher model, including high-confidence pseudo-labels, a standardized risk score vector, and a soft label probability distribution for normal and abnormal flight states. By introducing a KL divergence loss function to preserve the uncertainty in the soft labels and combining it with a sample weighting strategy based on the risk score, a joint loss objective is constructed to guide the student model in learning the relative contribution of various risk dimensions (e.g., collision probability, wind speed mutation, airspace congestion, etc.) to anomaly detection during training. Ultimately, during the distillation optimization process, the student model established a nonlinear mapping capability from the input unified structured feature vector to a continuous potential risk score, enabling abnormal identification of flight status and risk quantification. In the pre-flight prediction stage, the model outputs a potential risk score (e.g., 0.0–1.0) and divides it into first, second, and third risk levels based on the set threshold. If the risk score exceeds the first risk threshold (e.g., >0.7), the system automatically rejects the flight application, and the smart contract records the violation attempt on the chain to ensure traceability; if it is the second risk (e.g., 0.4–0.7), the system will prompt the operator to supplement the risk mitigation measures (e.g., optimize the path, adjust the time), and resubmit the application; if it is the third risk (<0.4), it will automatically pass the review.The risk score output by the constructed student model is a comprehensive continuous risk score value, which is used to represent the potential abnormality of the current input feature under the combined effect of multidimensional risk factors. It integrates the multidimensional risk score vector information provided by the teacher model during the training phase. The final output is a single-dimensional risk score, which is used for risk judgment and flight mission management decisions.

[0033] During the flight mission, the drone continuously collects flight data, meteorological data and airspace congestion conditions. The anomaly scoring multi-model combination mechanism is introduced to assist the student model in dynamically identifying abnormal flight behavior. Once abnormal behavior is detected, the smart contract is triggered to execute the early warning response mechanism.

[0034] Specifically, pre-flight risk prediction only calls upon the deployed student model. High complexity is not required before flight. The student model already integrates risk perception and knowledge distillation, enabling on-device prediction capabilities. It inputs structured features and outputs a potential risk score, enabling a single decision. In-flight scenarios require real-time monitoring and rapid response to various complex behaviors, such as yaw, derailment, and hovering. A single model cannot capture all these patterns. Therefore, a multi-model combination mechanism for anomaly scoring is introduced. This multi-scoring mechanism assesses risk from different perspectives, improving recognition robustness. The student model provides the main risk trends, while the multi-model mechanism provides detailed anomaly analysis and error correction. This multi-model combination mechanism for anomaly scoring is deployed as an auxiliary module in parallel with the student model.

[0035] Furthermore, an abnormal scoring multi-model combination mechanism is introduced to assist the student model in dynamically identifying abnormal flight behaviors, including: During the UAV's flight mission, it continuously collects its flight trajectory, meteorological factor data from associated meteorological monitoring points, and the congestion level of the permitted airspace network units that the flight trajectory passes through. The collected multi-source information is fused into a unified structured feature vector and input into the student model to obtain the potential risk score of the current flight mission. The potential risk score of the current flight mission falls within the preset confidence region, triggering the anomaly scoring multi-model combination mechanism. The confidence region is the risk score interval, which lies between the high risk threshold and the low risk threshold, and is used to indicate that the student model has a high degree of uncertainty in its judgment. The anomaly scoring multi-model combination mechanism includes multiple sub-models for independent risk factor assessment, which score the current flight mission on different risk dimensions. Among them, the risk factors include at least the degree of trajectory deviation and mission behavior consistency. The weighting coefficient of each risk factor is set based on its historical importance, and the scoring results of each risk factor are weighted and summed to generate the fused risk score of the current task; When the fusion risk score exceeds the preset abnormal flight judgment threshold, an abnormal flight judgment result is generated, triggering the smart contract execution response mechanism.

[0036] Specifically, the anomaly scoring multi-model combination mechanism is a risk scoring integration method based on multiple sub-models. It provides additional decision-making support when student models face uncertainty judgment during drone flight missions. By independently evaluating risk indicators such as trajectory deviation, behavioral consistency, and isolation in different dimensions, and constructing a more stable and interpretable final judgment result through weighted fusion, it improves the intelligence and credibility of the overall system. The selection of multiple sub-models in the anomaly scoring multi-model combination mechanism follows the following principles: select features in historical data that have a significant impact on flight safety, such as trajectory deviation, mission deviation, and group behavioral isolation; each model must focus on different risk dimensions to avoid high redundancy between dimensions and improve overall judgment diversity; the score output by each model is traceable, facilitating the formation of readable and verifiable multi-dimensional anomaly explanations. In summary, the trajectory deviation degree scoring model and the flight behavior consistency scoring model were selected. The trajectory deviation degree scoring model is used to measure the degree of deviation between the current flight trajectory and the flight path in the flight application, and calculate the spatial distance between the real-time flight trajectory sequence and the flight path sequence in the flight application. The model uses the dynamic time warping (DTW) method to calculate the Euclidean distance between the two sequences, normalizes the distance, and presets the maximum allowable deviation threshold D max , ; Among them, S 偏离 is the normalized trajectory deviation score, D DTW is the calculated Euclidean distance, and Dmax is the preset maximum allowable deviation threshold. 偏离∈[0, 1], the larger the value, the more serious the deviation. The mission behavior consistency scoring model determines whether the current state of the drone is consistent with the mission plan, such as whether it suddenly accelerates and rises, changes the flight altitude, returns early, etc. when performing a cruise mission. Input the current flight state vector (heading angle, flight speed, flight altitude) and the mission stage to which the current time point belongs (such as "route segment 1", "circling segment", etc.), score the current input, and obtain the anomaly score. The model construction method is as follows: obtain the type and planning process of the flight mission from the flight application, and pre-build a mission state transition model to characterize the state change relationship of the drone during normal execution. The state transition model includes multiple mission execution state nodes and their corresponding legal transfer paths. Each state node corresponds to a predefined flight parameter tolerance. The range includes heading angle range, flight altitude range, and flight speed range. During flight, the UAV's current flight state information, including current heading, flight altitude, speed, and mission phase identifier, is collected in real time. The current flight state is matched with the expected state in the mission state transition model. If the current state does not belong to any legal state node or an illegal state transition path exists, it is marked as inconsistent behavior and an inconsistent event is recorded. All state detection results during the flight mission are statistically analyzed to calculate a mission behavior consistency score. This score is a function of the ratio of the number of inconsistent events to the total number of detections. After normalization, it generates a consistency score in the interval [0, 1], where higher scores indicate greater mission behavior deviation. In scenarios such as single-aircraft flight, routine route inspections, and regional mission flights, the trajectory deviation degree scoring model and the flight behavior consistency scoring model are the most cost-effective identification methods. The trajectory deviation degree scoring model can quickly identify spatial violations such as deviation from the route, exceeding the inspection area, and illegal detours. The mission behavior consistency scoring model identifies behavioral anomalies such as abnormal landings, premature terminations, and out-of-chronological execution. Abnormal flight determination is divided into two parts. One is when the potential risk score of the current flight mission falls within a preset confidence region, triggering the abnormal scoring multi-model combination mechanism. If the risk score exceeds the preset abnormal flight determination threshold, an abnormal flight determination result is generated. The other is when the potential risk score of the current flight mission falls within the preset high-risk threshold. During the drone's flight mission, the model outputs a potential risk score (e.g., 0.0–1.0), which is divided into high-risk threshold and low-risk threshold according to the pre-set threshold. The high-risk threshold directly triggers the third smart contract, while the low-risk threshold indicates that the drone can continue to perform the mission. The confidence region between the high-risk threshold and the low-risk threshold is the risk score interval, triggering the abnormal scoring multi-model combination mechanism.Triggering the smart contract and executing the response mechanism is to automatically execute on-chain operations and system feedback once an abnormal flight is identified, including: recording the relevant parameters of the abnormal behavior on the chain, including the score value, degree of deviation, inconsistent behavior information, trigger time, etc., which are packaged and uploaded to the chain; calling the permission management contract to freeze or restrict the subsequent flight permissions of the drone, and issuing warnings to the operator or unit account based on the blockchain identity mechanism; using the off-chain interface to notify the regulator in real time; and through asynchronous communication from on-chain to off-chain, the flight behavior label is sent back to the model training system as a feedback sample.

[0037] After the drone lands, the collected flight data, path trajectory, execution time and flight application content are compared and verified to generate a complete flight record. After encryption, it is stored in the off-chain storage system. The flight record hash value is calculated, and the smart contract is called to perform format verification on the on-chain data. After verification, the flight record hash value, off-chain storage index and related access credentials are written to the blockchain to realize full-process permission management of drone flight missions.

[0038] Specifically, after the drone lands, the blockchain + off-chain collaborative evidence storage model is executed to achieve closed-loop management of the drone mission. After the drone completes the flight mission and lands safely, the system automatically retrieves the actual flight data of the mission, such as the flight trajectory, take-off and landing time, flight duration, and mission tags, and compares and verifies it with the previous flight application content to determine whether there are abnormal conditions such as path deviation, time window exceeded, or mission incomplete. After the verification is completed, a standardized structured flight record document is generated, which contains the entire process data of the mission execution and the comparison results, and the record file is encrypted and protected and stored in the off-chain storage system. At the same time, after calculating the hash value of the encrypted flight record, the smart contract pre-deployed in the blockchain network is called to verify the format of the task identifier and the data upload subject; after the verification is passed, the smart contract automatically completes the write operation of the hash value of the encrypted flight record, the index address of the off-chain storage, and the access control credentials, and stores them in the blockchain.

[0039] This process ensures that the flight mission execution process is tamper-proof and traceable, enabling regulators or relevant authorized entities to verify the authenticity and integrity of flight data based on blockchain, thereby achieving closed-loop trusted management of flight missions and meeting the needs of audit accountability and behavioral compliance verification.

[0040] The present application embodiment provides a UAV flight rights management platform based on blockchain technology, such as Figure 2 As shown, the platform includes: The blockchain network construction module 11 is used to build a blockchain network, pre-deploy multiple smart contracts corresponding to the drone flight permission management business, and perform compliance verification on the received flight permissions through the smart contracts before encrypting and storing them in the blockchain; The flight risk prediction module 12 is used to deploy the student model after knowledge distillation to the edge computing device for pre-flight risk prediction and in-flight abnormal flight identification. The teacher model is a multimodal dynamic network structure model used to integrate multi-source heterogeneous historical data such as drone flight records, meteorological factor data, and airspace congestion to learn flight risk assessment knowledge. The learned flight risk assessment knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework. The pre-flight permission matching module 13 is used to receive the flight application before the UAV takes off, call the smart contract for compliance verification, obtain real-time meteorological data and permitted airspace congestion information in combination with the flight application, call the student model deployed on the edge computing device, perform pre-flight risk prediction, obtain the flight risk level, meet the flight requirements, call the smart contract to verify the matching between the flight application and the on-chain permission record, generate an encrypted flight permit after verification, and store the permit information and its hash value in the blockchain; The smart contract response mechanism module 14 is used to continuously collect flight data, meteorological data, and airspace congestion information during the UAV's flight mission. The anomaly scoring multi-model combination mechanism is introduced to assist the student model in dynamically identifying abnormal flight behavior. When abnormal behavior is detected, the smart contract execution early warning response mechanism is triggered. The post-landing data blockchain storage module 15 is used to compare and verify the collected flight data, path trajectory, execution time and flight application content after the drone lands, generate a complete flight record, and store it in the off-chain storage system after encryption. It calculates the flight record hash value, calls the smart contract in the blockchain, performs format verification on the on-chain data, and writes the flight record hash value, off-chain storage index and related access credentials to the blockchain after verification to realize the full-process permission management of the drone flight mission.

[0041] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present application. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. The UAV flight rights management method based on blockchain technology is characterized by: The method includes: Build a blockchain network and pre-deploy multiple smart contracts corresponding to drone flight rights management services. After the received flight rights are verified for compliance by the smart contracts, they are encrypted and stored on the blockchain. The student model, which has undergone knowledge distillation, is deployed to edge computing devices for pre-flight risk prediction and in-flight abnormal flight identification. The teacher model is a multimodal dynamic network structure model that integrates multi-source heterogeneous historical data such as drone flight records, meteorological factor data, and airspace congestion to learn flight risk assessment knowledge. This knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework. Before the drone takes off, it receives a flight application and calls a smart contract for compliance verification. If the verification passes, it obtains real-time weather data and permitted airspace congestion information and combines it with the flight application. It calls the student model deployed on the edge computing device to perform pre-flight risk prediction, obtain the flight risk level, and meet the flight requirements. The smart contract is called to verify the matching between the flight application and the on-chain permission record. After verification, an encrypted flight permit is generated, and the permit information and its hash value are stored in the blockchain. During the UAV's flight mission, it continuously collects flight data, meteorological data, and airspace congestion information. It introduces an anomaly scoring multi-model combination mechanism to assist the student model in dynamically identifying abnormal flight behavior. Once abnormal behavior is detected, it triggers the call of the smart contract to execute the early warning response mechanism. After the drone lands, the collected flight data, path trajectory, execution time and flight application content are compared and verified to generate a complete flight record. After encryption, it is stored in the off-chain storage system. The flight record hash value is calculated, and the smart contract is called to perform format verification on the on-chain data. After verification, the flight record hash value, off-chain storage index and related access credentials are written to the blockchain to realize full-process permission management of drone flight missions.

2. The method for managing drone flight rights based on blockchain technology according to claim 1, characterized in that: Build a blockchain network, including: Based on the alliance chain architecture, adopt appropriate consensus algorithms to build the blockchain network and define node roles and their permissions; Pre-deploy multiple smart contracts corresponding to drone flight permission management services, and call smart contracts to verify the compliance of flight permissions; The flight permission information is compiled into a labeled data packet. The hash value of the standard data packet is calculated using a hash algorithm. The standard data packet is encrypted and stored in an independent off-chain storage system. The data packet hash value, off-chain storage location, access credentials, and transaction hash are written into the block record of the blockchain. A three-level composite key index table is established in an independent off-chain storage system. The first-level key is the hash prefix of the drone's identity information, the second-level key is the flight record, and the third-level key is the time range. The transaction hash on the chain is located through the third-level composite key, and the off-chain data access information is obtained from the on-chain record.

3. The method for managing drone flight rights based on blockchain technology according to claim 1, wherein: teacher The model is a multimodal dynamic network structure model that is used to integrate multi-source heterogeneous historical data such as drone flight records, meteorological factor data, and airspace congestion to learn flight risk assessment knowledge. Through the constructed teacher-student heterogeneous knowledge distillation framework, the learned flight risk assessment knowledge is transferred to the student model, including: Define drone nodes in a multimodal dynamic network structure model, obtain drone flight records based on historical data, and extract drone node features based on the flight records. Static features include take-off and landing points, historical mission types, historical mission flight times, and historical abnormal events. Dynamic features are generated by sampling historical flight trajectories at preset time intervals to generate a historical flight trajectory time series, which contains the spatial position of each time point. The spatial position includes the drone's geographic coordinates and flight altitude at that time point. The meteorological nodes in the multimodal dynamic network structure model are defined. The static characteristics of the meteorological nodes are the geographical locations of the meteorological monitoring points within the permitted flight airspace. Based on the 12-hour window before the start time of the historical flight mission, the historical meteorological trend feature sequences of the meteorological monitoring points associated with the historical flight trajectory of the UAV are extracted. At the same time, the type, occurrence time, and duration of abnormal weather are identified and recorded to form an abnormal weather time series feature sequence. The above two feature sequences are aligned with the time series of the UAV's historical flight trajectory to generate a historical meteorological trend feature sequence and a historical abnormal weather sequence synchronized with the time series of the UAV's historical flight trajectory. These are the dynamic characteristics of the meteorological nodes. Airspace nodes are defined in a multimodal dynamic network structure model. The static characteristics of airspace nodes are grid cells divided by a combination of space and altitude in the permitted flight airspace. Based on the historical flight trajectories of drones, the congestion degree of each grid cell in each time period is counted within a 24-hour rolling sliding window. The airspace congestion trend characteristics are extracted using a time series modeling method to form a time series feature sequence of airspace congestion trends covering multiple time points. This sequence is aligned with the time series of drone flight trajectories to generate a historical airspace congestion trend feature sequence synchronized with the drone's historical flight trajectories, which serves as the dynamic characteristic of the airspace node. A multimodal dynamic network structure model consisting of drone nodes, meteorological nodes, and airspace nodes was constructed. Spatial edges, meteorological influence edges, and task-related edges were set between nodes. The network model was trained using reinforcement learning to establish spatiotemporal dependencies between various nodes and obtain the spatiotemporal risk characteristics of drone nodes. The trained multimodal dynamic network structure model is used as a teacher model to generate pseudo labels and risk score vectors for flight risk identification. The IsolationForest model is used as the student model. Through the heterogeneous knowledge distillation mechanism, the pseudo labels and risk score vectors in the teacher model are transferred to the student model to complete the construction of the student model.

4. The method for managing drone flight rights based on blockchain technology as claimed in claim 3, characterized in that: generate The historical meteorological trend feature sequence and historical abnormal meteorological sequence synchronized with the UAV historical flight trajectory time series include: Based on historical data, the historical flight missions of each UAV are obtained. A time window of 12 hours before the start time of the mission is set. The time series data of meteorological factors of each meteorological monitoring point within the window are extracted and standardized. The meteorological factors include wind speed, wind direction, temperature, precipitation and visibility. Based on the spatial position corresponding to each time point in the historical flight trajectory time series of each UAV, combined with the monitoring range of each meteorological monitoring point, the meteorological monitoring points covering the spatial position are selected. If multiple meteorological monitoring points cover the same position, the meteorological monitoring point closest to the spatial position is selected to establish the corresponding association relationship between the meteorological monitoring point and each time point in the historical flight trajectory time series of the UAV; The meteorological factor time series data of the meteorological monitoring points associated with each time point in the historical flight trajectory time series of the UAV are input into the LSTM time series model, the meteorological factor change trend embedding vector of each time point is extracted, and the historical meteorological trend feature sequence of each associated meteorological monitoring point is generated; Set an abnormal judgment threshold for each type of meteorological factor, traverse the meteorological factor time series data of the associated meteorological monitoring points, mark abnormal meteorological conditions, record the abnormal meteorological type, occurrence time and duration, and construct a historical abnormal meteorological sequence for the associated meteorological monitoring points; The historical meteorological time series trend characteristics and historical abnormal meteorological time series characteristics of each associated meteorological monitoring point are aligned with the UAV historical flight trajectory time series in time, forming a historical meteorological trend feature sequence and historical abnormal meteorological sequence synchronized with the UAV historical flight trajectory time series.

5. The method for managing drone flight rights based on blockchain technology as claimed in claim 3, characterized in that: Generate a historical airspace congestion trend feature sequence synchronized with the historical flight trajectory of the drone, including: Divide the permitted flight airspace into spatial and altitude dimensions, and construct a spatial-altitude grid structure, where each grid cell corresponds to a latitude, longitude, and altitude layer; Based on the set 24-hour rolling sliding window, the number of drones appearing in each grid cell within each window is counted, and the average frequency of drone appearance within the window step is calculated to obtain the congestion degree of each grid cell in each time period, forming a historical congestion time series for each grid cell in different windows; The historical congestion time series corresponding to each grid unit is input into the LSTM network to extract the trend of congestion change over time and obtain the historical congestion trend feature sequence of each grid unit; According to the spatial position of each time point in the drone's historical flight trajectory time series, the grid unit to which it belongs is located, and the historical congestion trend feature sequence of the grid unit is aligned with the drone's historical flight trajectory time series in time to generate a historical airspace congestion trend feature sequence synchronized with the drone's historical flight trajectory time.

6. The method for managing drone flight rights based on blockchain technology as claimed in claim 3, characterized in that: Multimodal dynamic network structure model, including: Construct a multimodal dynamic network structure model as follows: The historical flight trajectory time series is divided into multiple continuous time slices according to the time dimension. A corresponding static network graph is constructed in each time slice. Each static network graph consists of nodes and edges. The nodes include drone nodes, weather nodes and airspace nodes, and the edges include space edges, weather impact edges and mission-related edges. Spatial edges are used to connect spatially adjacent nodes and obtain the airspace congestion trend of adjacent node areas. The edge weight is obtained by jointly calculating the inverse of the geographical distance between nodes and the historical airspace congestion trend feature sequence of the airspace unit in which they are located. Among them, adjacent nodes include drone nodes and drone nodes, drone nodes and airspace nodes, and airspace nodes and airspace nodes. The meteorological influence edge is used to connect the meteorological node with the UAV node within its monitoring range. The edge weight is determined by calculating the matching degree between the meteorological time series trend characteristics of the meteorological node and the meteorological factors at the corresponding UAV location time point. Task association edges are used to connect UAV nodes of the same or similar task types. The edge weight is calculated by combining the semantic similarity between task types and the proximity of task execution time. Connect the same nodes in adjacent time slices to form time edges, which are used to model time evolution relationships; In each time slice, the graph attention network performs type-aware attention aggregation on various nodes of each static network graph, learns the representation of each type of node, and realizes the propagation of various node representations in the time dimension through time edges. At the same time, the node state evolution mechanism is introduced to obtain the spatiotemporal feature representation of the drone node in each historical time slice, completing the construction of the entire multimodal dynamic network structure model.

7. The method for managing drone flight rights based on blockchain technology as claimed in claim 3, characterized in that: Complete the construction of the student model, including: The trained multimodal dynamic network structure model is used as the teacher model. Historical multimodal data is input to extract the spatiotemporal feature representation of drone nodes. The classification head outputs the probability distribution of normal and abnormal flight. High-confidence pseudo-labels are filtered through a threshold as the initial abnormal samples. At the same time, the regression head outputs a risk score vector to quantify the contribution of each risk dimension. High-confidence pseudo-labels are used as strong supervision samples. The label propagation algorithm is used to propagate feature similarity on low-confidence samples to generate extended abnormal samples. The original features and risk score vectors are combined to construct a complete semi-supervised training set. The original features come from the feature sequences extracted for each node by constructing a multimodal dynamic network structure model. A feature adapter is introduced to map the risk score vector to the feature space of the student model. Soft labels are generated based on the probability distribution of normal and abnormal flights. KL divergence loss is used to preserve the uncertainty of the soft labels, guiding the student model to learn richer decision boundaries. Furthermore, the risk score is used to adjust the weights of IsolationForest training samples, so that the model prioritizes high-risk samples during training. The student model parameters are optimized using a joint loss function to complete knowledge distillation and obtain the distilled parameters. In the process of training the IsolationForest model based on the semi-supervised training set and distilled parameters, a three-level progressive sample selection mechanism is introduced to expand the training sample set. This mechanism consists of three stages: first, the initial abnormal samples are used as core abnormal samples; second, the expanded abnormal samples are used as supplementary abnormal samples; finally, during the model training process, samples with short path lengths and high risk scores are dynamically identified and included in the training as dynamic abnormal samples; On the expanded training sample set, the IsolationForest model prioritizes high-risk feature dimensions for node splitting. It also introduces a path length penalty mechanism based on risk scores, imposing penalty weights proportional to the risk scores on abnormal samples to improve the anomaly detection performance of the training model. The trained student model is quantized and compressed and then deployed to the edge device to complete the construction of the student model.

8. The method for managing drone flight rights based on blockchain technology according to claim 1, wherein: Perform pre-flight risk prediction and obtain the flight risk level, including: Extract the pre-flight mission type, flight path, flight time, and take-off and landing points from the flight application to form static mission parameters; Based on the flight path and take-off and landing points, meteorological factor data of the associated meteorological monitoring points are collected in real time, and the congestion of the permitted airspace network units passed by the flight path is counted in real time. The meteorological factor data and congestion degree form dynamic mission parameters; The static and dynamic task parameter features are fused into a unified structured feature vector and input into the student model deployed on the edge computing device. Based on the student model's inference calculation of the fused features, a continuous potential risk score is output. Preset risk level threshold and compare it with the output potential risk score, When the potential risk score is greater than the first threshold, it is judged as high risk. The system automatically rejects the flight application and calls the smart contract to record the high-risk attempt in the blockchain. When the potential risk score is between the second threshold and the first threshold, it is determined to be medium risk, and the operator is prompted to supplement risk mitigation measures and resubmit the application; When the potential risk score is lower than the second threshold, it is judged as low risk and enters the subsequent process.

9. The method for managing drone flight rights based on blockchain technology according to claim 1, wherein: The introduction of anomaly scoring multi-model combination mechanism assists the student model in dynamically identifying abnormal flight behaviors, including: During the UAV's flight mission, it continuously collects its flight trajectory, meteorological factor data from associated meteorological monitoring points, and the congestion level of the permitted airspace network units that the flight trajectory passes through. The collected multi-source information is fused into a unified structured feature vector and input into the student model to obtain the potential risk score of the current flight mission. The potential risk score of the current flight mission falls within the preset confidence region, triggering the anomaly scoring multi-model combination mechanism. The confidence region is the risk score interval, which lies between the high risk threshold and the low risk threshold, and is used to indicate that the student model has a high degree of uncertainty in its judgment. The anomaly scoring multi-model combination mechanism includes multiple sub-models for independent risk factor assessment, which score the current flight mission on different risk dimensions. Among them, the risk factors include at least the degree of trajectory deviation and mission behavior consistency. The weighting coefficient of each risk factor is set based on its historical importance, and the scoring results of each risk factor are weighted and summed to generate the fused risk score of the current task; When the fusion risk score exceeds the preset abnormal flight judgment threshold, an abnormal flight judgment result is generated, triggering the smart contract execution response mechanism.

10. The drone flight rights management platform based on blockchain technology is characterized by: The platform is used to execute the blockchain-based drone flight permission method according to any one of claims 1 to 9, comprising: The blockchain network construction module is used to build a blockchain network, pre-deploy multiple smart contracts corresponding to the drone flight permission management business, and verify the compliance of the received flight permissions through the smart contracts before encrypting and storing them on the blockchain; The flight risk prediction module deploys the knowledge-distilled student model to edge computing devices for pre-flight risk prediction and in-flight abnormal flight identification. The teacher model is a multimodal dynamic network structure model that integrates multi-source heterogeneous historical data such as drone flight records, meteorological factors, and airspace congestion to learn flight risk assessment knowledge. This knowledge is then transferred to the student model through the constructed teacher-student heterogeneous knowledge distillation framework. The pre-flight permission matching module is used to receive flight applications before the drone takes off, call smart contracts for compliance verification, obtain real-time meteorological data and permitted airspace congestion information in combination with the flight application, call the student model deployed on the edge computing device, perform pre-flight risk prediction, obtain the flight risk level, meet the flight requirements, call smart contracts to verify the matching of the flight application with the on-chain permission record, generate an encrypted flight permit after verification, and store the permit information and its hash value on the blockchain; The smart contract response mechanism module is used to continuously collect flight data, meteorological data, and airspace congestion during the UAV's flight mission. The anomaly scoring multi-model combination mechanism is introduced to assist the student model in dynamically identifying abnormal flight behavior. Once abnormal behavior is detected, the smart contract execution warning response mechanism is triggered. The post-landing data blockchain storage module is used to compare and verify the collected flight data, path trajectory, execution time and flight application content after the drone lands, generate a complete flight record, and store it in the off-chain storage system after encryption. It calculates the flight record hash value, calls the smart contract to perform format verification on the on-chain data, and after verification, writes the flight record hash value, off-chain storage index and related access credentials to the blockchain to realize full-process permission management of drone flight missions.

Images