System supporting automatic charging of cross-industry unmanned aerial vehicle

Abstract

The invention discloses a system supporting automatic charging of a cross-industry unmanned aerial vehicle. The system comprises an automatic charging platform deployed on a power grid tower and a background intelligent collaborative decision and operation and maintenance cloud platform. The intelligent collaborative decision and operation and maintenance cloud platform comprises a central intelligent scheduling module, a reinforcement learning module, a block chain charging auditing module and a credit hierarchical control module, and intelligent scheduling, privacy protection, credible charging and fair resource allocation of the unmanned aerial vehicle are realized through technologies such as federated learning, multi-agent reinforcement learning and a block chain. The method systematically solves the cruising ability problem which is one of core bottlenecks restricting large-scale application of cross-industry unmanned aerial vehicles under low-altitude economy, and provides key technical support for constructing open and shared low-altitude economy digital infrastructures.

Description

Technical Field

[0001] This invention belongs to the cross-application field of drone technology, power grid infrastructure, blockchain and privacy protection, and specifically relates to a system that supports automatic charging of drones across industries. Background Technology

[0002] Multi-rotor drones are currently used in many scenarios, characterized by low speed and high controllability. However, their endurance and automated operation efficiency are key factors restricting their widespread application, with charging difficulties and high costs being common problems. Traditional charging methods are no longer sufficient to meet the high-intensity operational demands of intelligent drones in complex tasks. Therefore, introducing suitable automatic charging technology can significantly improve the efficiency of drone use.

[0003] Current research on automatic charging for drones generally suffers from the following problems:

[0004] High risk of privacy breaches: In charging services involving multiple parties, centralized data training requires sharing raw data, and user behavior data is easily collected and analyzed, posing a risk of privacy breaches; Limited scheduling system: Existing automatic drone charging often only involves the management of a single system, lacking the support of a unified scheduling system across regions, and cannot adapt to the collaborative needs of multiple industries; Lack of transparency in billing: Existing platforms struggle to achieve reliable billing and auditing across organizations and industries, and credit evaluation is rigid; Insufficient compatibility: Most solutions are only compatible with drones of specific brands or uses, lacking a unified interface and standard protocol.

[0005] There is an urgent need for an automatic drone charging system that offers cross-industry versatility and strong privacy protection, while also providing convenient local charging for drones. Considering that power grid companies are key participants in the low-altitude economy (power line inspection), leveraging the widely distributed power grid tower assets, deploying an automatic drone charging platform on suitable towers would not only support power grid applications such as power line inspection but would also be particularly suitable for cross-industry drone applications in logistics, emergency rescue, and agricultural plant protection. This approach offers cost advantages and significantly expands the operational radius of drones. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a system that supports automatic charging of drones across industries, which can solve the technical problems existing in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following means:

[0008] A system supporting automatic charging of drones across industries includes:

[0009] The intelligent collaborative decision-making and operation and maintenance cloud platform, in collaboration with industry participants, is used to allocate corresponding automatic charging platforms to drones awaiting charging. Industry participants refer to drones owned by companies belonging to a specific industry that can directly access the backend cloud platform to use the automatic charging system, or indirectly, through a combination of drones using the automatic charging system and corresponding intelligent agents on the enterprise's intelligent agent, using the enterprise's deployed intelligent agent as an intermediary.

[0010] An automatic charging platform is deployed on selected power grid towers and works in conjunction with an intelligent collaborative decision-making and operation and maintenance cloud platform to receive drone charging scheduling instructions from the cloud platform and support drones to fly to the platform for automatic charging.

[0011] The intelligent collaborative decision-making and operation and maintenance cloud platform includes the following modules:

[0012] The central intelligent scheduling module adopts a hierarchical federated learning architecture for data fusion based on dynamic differential privacy, retaining more information for non-sensitive data, improving the accuracy of the global model, and is coupled with the reinforcement learning module to achieve model parameter sharing through a data interaction interface;

[0013] The reinforcement learning module employs a training framework integrated with federated learning for hierarchical decision-making systems oriented towards multi-agent collaborative decision-making, wherein federated learning is a distributed machine learning technique.

[0014] The blockchain billing audit module combines high-performance consortium blockchain, verifiable privacy computing, and tamper-resistant distributed storage to support blockchain-based drone charging billing auditing.

[0015] The credit rating control module adopts an industry-differentiated, dynamically adjusted and system-integrated control mechanism to achieve fair and efficient resource allocation and risk management.

[0016] The automatic charging platform includes the following modules;

[0017] The energy management module is used to integrate wind and solar hybrid energy storage, electromagnetic induction energy storage and fuel cell energy storage for unified scheduling, so as to achieve intelligent power supply and stable system output;

[0018] A standardized interface module is used to unify communication specifications and extend fast charging protocols, supporting adaptive charging for drones across industries. It adopts a standardized interface protocol certified by IEEE P1937.18, unifies communication specifications (such as MQTT+JSON) and extends fast charging protocols to 2kW, and is compatible with WPCQi standards (v2.0 and above) to support adaptive charging for drones across brands.

[0019] The IoT authentication module is used for automatic real-time two-way identity authentication and authorization between the drone and the platform to ensure charging safety;

[0020] The edge computing module is responsible for executing real-time control commands and performing lightweight data preprocessing and caching.

[0021] Furthermore, the data flow and calculation process of the central intelligent scheduling module are as follows:

[0022] Local training: Participants from various industries on the local dataset D k Train a local charging demand prediction model and minimize the local loss function. , where y pred This represents the predicted value y based on the model parameters w. true Representing the true value, the model parameters are optimized by iteratively minimizing the loss function using the gradient descent algorithm; after training, each participant obtains the original gradient for the t-th aggregation cycle. ;

[0023] Privacy handling and uploading: Each participant's original gradient Noise is added for perturbation, and the gradient expression after perturbation is as follows:

[0024]

[0025] Where 𝒩 represents a Gaussian distribution, Represents the identity matrix. The dynamically adjusted standard deviation is expressed as follows:

[0026]

[0027] in, Let τ represent the gradient variance calculated by participant k in round t, and τ = 0.005, which represents an empirical threshold to balance privacy and model accuracy. When the gradient variance is large (data is complex and highly sensitive), more noise is injected, and vice versa.

[0028] Cloud aggregation: Perturbation-induced gradient The data is uploaded to the cloud aggregation server via a secure channel. The cloud executes the aggregation algorithm to generate the updated global model expression as follows:

[0029]

[0030] Where η represents the learning rate, and D represents the union of the datasets of all participants. k This represents the local dataset of participant k. This represents the gradient after aggregation. This represents the global model parameters in round t+1;

[0031] Model Distribution and Application: Updated Global Model Parameters The data is distributed back to industry participants to predict the next round of local training and charging needs. The cloud uses an aggregation algorithm to generate a global model for the next round of predictions, providing the system with a heatmap and prediction of drone charging needs across the entire domain, thus enabling optimized scheduling.

[0032] Furthermore, the hierarchical decision system in the reinforcement learning module is a hierarchical reinforcement learning framework, including a high-level controller and low-level executors. The high-level controller, acting as a meta-controller, selects the optimal objective function from a dynamic multi-objective optimization library based on the system state of the global prediction model generated by federated learning. Based on the local decision suggestions uploaded by each low-level executor, it uses a constrained optimization algorithm for centralized arbitration to generate a final allocation instruction, which is then sent to the low-level executors for execution. The low-level executors are responsible for local state awareness and action suggestion generation, and receive the final action instruction from the high-level controller for execution.

[0033] Furthermore, the specific methods of the reinforcement learning module are as follows:

[0034] The system state is a hybrid state for a single drone charging request decision instance. This design aligns with a multi-agent reinforcement learning architecture of centralized training and distributed execution (CTDE). It captures all the key information needed for decision-making when processing a drone's charging request at time t, including both global and local information, as specifically expressed below:

[0035]

[0036] in, Indicates the remaining battery power of the drone. This indicates the real-time load rate of each charging platform. Indicates the urgency level of the task. Indicates the meteorological risk value. Represents the resource conflict matrix, C e (t) represents the real-time credit score;

[0037] The MADDPG (Multi-Agent Deep Deterministic Policy Gradient) algorithm framework is adopted as the multi-agent reinforcement learning algorithm used in this step to train industry participants to select charging platforms and power. Each industry participant learns a policy through an Actor-Critic network to handle collaborative decision-making among multiple agents (industry participants). During the training phase, each agent's Critic network can obtain the global system state and the action information of other agents. During the execution phase, each agent makes decision suggestions based on its own local observations and uploads them to the cloud. The cloud then performs centralized arbitration to avoid conflicts and support global optimization.

[0038] Action Space A t The specific expression is as follows:

[0039]

[0040] in, This indicates selecting a target platform from the set of available charging platforms. This represents the charging power allocated to the drone. The action generation process involves industry participants generating action suggestions based on their local strategies and uploading them to the cloud. The cloud collects all the local action suggestions uploaded by all participants and integrates the global state (such as the load rate of all charging platforms). chgProbe Resource Conflict Matrix matrix Weather data rvalue The central arbitrator (such as a high-level controller) uses a constraint optimization algorithm to fuse and weigh the various local suggestions, resolve potential target conflicts (e.g., multiple drones simultaneously suggesting the selection of the same idle charging platform), and finally generate a globally optimal final action command.

[0041] The reward function expression for the dynamic multi-objective optimization is as follows:

[0042]

[0043] in, This represents the reward for successful charging. A successful charge earns a positive reward of +1, while a failed charge earns a negative reward of -1. This indicates a time-sensitive reward, which is linked to the efficiency of task completion and encourages quick completion. The resource conflict penalty is represented by α, β, and γ, which represent the adaptive coefficients for task urgency, load balancing, and conflict penalty, respectively.

[0044] resource conflict penalty The expression is as follows:

[0045] R penalty =0.1×(1 / C e (t))×[RecentConflictFreq+HistoricalConflictRate]×K

[0046] The base penalty coefficient is 0.1, and the recent conflict frequency RecentConflictFreq = log(1 + N_recent) reflects the recent conflict frequency (logarithmically smoothed). N_recent represents the number of recent conflicts within the most recent consecutive hour sliding window. matrix The cumulative number of conflict events identified in the data measures the frequency of recent conflicts initiated by participants; the Historical Conflict Rate (Total Historical Conflicts / Total Historical Requests) reflects the frequency of historical conflicts and indicates the long-term conflict behavior patterns of participants; K represents the normalization coefficient, which is determined by setting a reasonable lower limit for the credit score C. e min And the normalization coefficient K, can make the effective R penalty The value range should be controlled within a reasonable range.

[0047] The task urgency adaptive coefficient α is used to prioritize drone missions that are high in urgency, have low battery levels, and are close in distance. The expression is as follows:

[0048]

[0049] Where α0 = 0.1 represents the scaling factor. The normalized value is set to the range [0.1, 1.0], and a charging request is forcibly triggered when the value is below 0.2. Take discrete values ​​{1, 2, 3, 4, 5}. The higher the level, the smaller the corresponding value. That is, 1 is the highest level of urgency (such as emergency rescue) and 5 is the lowest level of urgency (such as routine inspection). ˡ represents the flight distance from the drone to the charging platform, with an effective range of [0.5, 10] km. When the distance is less than 0.5 km, it is calculated as 0.5, and when it is more than 10 km, it is calculated as 10. The upper limit is that when the distance exceeds 10 km, the system needs to prioritize evaluation and decision-making.

[0050] The effective range of α is [0.02, 2.0]. When the maximum value scenario is: low battery (SOC=0.1), high urgency (Priority=1), and short distance (Distance=0.5km), α≈2.0, which corresponds to the drone receiving a higher weight for successful charging reward, and the system will try its best to ensure its charging needs.

[0051] The load balancing adaptive coefficient β is used to encourage the system to make balanced use of charging platform resources and avoid local overload. The expression is as follows:

[0052]

[0053] in, The load balancing factor, σ², represents the load balancing degree of the charging platform, and σ² represents the load rate of all charging platforms. Station The variance of the load factor of the charging platform is normalized to a range of [0,1]. A smaller variance of the load factor of all charging platforms indicates balance. When the overall load balancing factor of the system is low, the system prioritizes tasks with high timeliness. When the load balancing factor is high, the timeliness requirements of low-priority tasks are automatically reduced to prevent system overload.

[0054] The conflict penalty adaptive coefficient γ is used to penalize resource conflict behavior, especially to strengthen constraints when the system is under high load, and its expression is as follows:

[0055]

[0056] Where γ0=0.5 represents the basic penalty coefficient, used to balance the magnitude of the penalty term; This indicates a high system load status identifier; it is 1 when the system is under high load and 0 otherwise.

[0057] System high load status identifier The expression is as follows:

[0058]

[0059] Wherein, IF(•) represents an indicator function, which is 1 if the condition is met, and 0 otherwise; θ represents the average load rate of all charging platforms, used to measure the overall system load level; high =0.7, representing the overall high load threshold of the system. When the average load rate exceeds this value, the system is considered to be in a high load state; σ² represents the variance of the load rate of all charging platforms, used to measure the load balance; θ balance =0.01 represents the load balancing threshold. When the load variance is below this value, the system load is considered highly balanced. In the case of global saturation—that is, "overall high load" and "load balanced"— The value is 1, and γ takes the value of 0.65, thus increasing the severity of the punishment.

[0060] Cloud-based arbitration is a core component ensuring centralized management of charging platform allocation by the power grid company. Its constrained optimization algorithm employs integer linear programming, modeling the allocation problem as a constrained optimization problem. The objective function maximizes the global reward (such as weighted charging success rate and load balancing), while constraints include each charging platform serving only one drone and the load rate not exceeding a limit. The expression is as follows:

[0061]

[0062] ∀i (Each drone is assigned a charging platform)

[0063] ∀j (Each charging platform can serve at most one drone)

[0064] x ij x ∈{0,1} ij Indicates whether drone i is assigned to charging platform j, U ij =f(S) t， A t Θ) is the utility function, which is the expected utility value calculated by the system for each "drone-charging platform" pair (i,j), reflecting the expected benefit of this allocation decision. Where S t Let A represent the state space. t Θ represents the action space, and Θ represents the system parameter set, which includes parameters that are pre-set, relatively fixed, or adjusted over a longer time scale in the reinforcement learning algorithm and related modules.

[0065] Furthermore, the specific methods of the blockchain billing audit module are as follows:

[0066] 1) A three-layer hybrid architecture and data flow of main chain-side chain-IPFS

[0067] Based on the Hyperledger Fabric main chain-side chain design, the main chain is the core chain in the consortium blockchain, responsible for final notarization and auditing, and storing data summaries (such as transaction Merkle roots), but with a low processing frequency to ensure security; the side chain is an auxiliary chain connected to the main chain, based on the Lightning Network protocol, specifically for handling high-frequency, small-amount charging transactions, improving performance to support processing ≥1000 transactions per second, and periodically anchoring data summaries to the main chain.

[0068] Transaction Generation: The edge computing module performs calculations on charging event data, which includes power E, time T, power P, and user identifier UID. The expression for generating the transaction billing voucher is as follows:

[0069]

[0070] Where H represents the hash function, Indicates a digital signature;

[0071] Privacy-Preserving & Off-Chain Storage: The raw data uses homomorphic encryption to implement verifiable privacy computations, supporting billing calculations on the ciphertext, as shown in the following expression:

[0072]

[0073] The encrypted data C is uploaded to IPFS, and a content identifier is returned. The Content Identifier (CID) is used as an index and stored on the blockchain through subsequent processes;

[0074] Sidechain Processing & High-Frequency Settlement: Billing transactions are routed to the sidechain, which uses the Hash Time Lock Contract (HTLC) protocol for processing. Data is stored on the sidechain to achieve high-performance settlement, as shown in the following expression:

[0075]

[0076] Where H(R) represents the hash of the payment preimage R generated by the payer. This indicates the timeout period, and Amount indicates the payment amount.

[0077] Mainchain Notarization & Auditing: The sidechain will record a batch of transactions {Tx1, Tx2, ..., Tx} every 10 minutes within a period. n The Merkle root of this batch of transactions is generated by constructing a Merkle tree. m The leaf nodes are the hash values ​​of each transaction; Root storage The root hash of the Merkle tree constructed for all content identifiers (CIDs) corresponding to this batch of transactions, whose generation algorithm is the same as that of the Root. m Similarly, this ensures that the integrity of off-chain stored data can be verified. The sidechain periodically stores a batch of transactions in the Merkle Root. m Merkle root of the corresponding content identifier (CID) set Anchoring to the main chain, the expression is as follows:

[0078]

[0079] in, This represents the main chain transaction ID, Timestamp represents the timestamp, and Block_Hash represents the block hash value, used for recording and auditing.

[0080] 2) Secure storage and audit traceability mechanism

[0081] Data sharding and threshold secret sharing:

[0082] To securely store the decryption key, the encrypted data C is fragmented using Shamir's Secret Sharing scheme before storage, generating n fragments (l, f(l)), where l = 1, 2, ..., n; given the secret SK, a polynomial is constructed as follows:

[0083]

[0084] Among them, a i represents the polynomial coefficients, x represents the variable, and mod p represents the modular operation;

[0085] The formula (k,n)=(3,4) is used, where k is the threshold value, representing the minimum number of fragments required to recover the original key, and n is the total number of fragments generated after splitting the secret key. The data fragments are stored in four different edge computing modules. During auditing and tracing, only any three fragments need to be collected to recover the original key SK using the Lagrange interpolation formula, as shown below:

[0086]

[0087] After reading the encrypted data from IPFS and recovering the original data using SK via homomorphic decryption, the audit node needs to perform data integrity verification. This process involves calculating the hash value of the recovered transaction data and comparing it with the Merkle root of the corresponding transaction stored on the main chain. m The hash value recorded in the database is compared with the data. If they match, it proves that the encrypted data recovered from IPFS has not been tampered with during storage and transmission, and the data integrity has been verified. After verification, a final audit report is generated.

[0088] Quantum-resistant cryptography and encryption: The fragments are encrypted using a post-quantum cryptography (PQC) algorithm, as shown in the following expression:

[0089]

[0090] It adopts CRYSTALS-Dilithium-3 (signature security level 5) or Kyber-1024 (key encapsulation security level 5), with a key length of ≥2048 bits, and an expected security period of ≥20 years against quantum computing attacks.

[0091] Billing and Smart Contracts: Billing is performed through smart contracts deployed on the main chain, as shown in the following expression:

[0092]

[0093] Where E represents electricity (kWh), This indicates the real-time electricity price (yuan / kWh). This represents the industry discount factor. This represents the peak-valley coefficient (e.g., 1.2 for peak periods and 0.8 for valley periods).

[0094] Furthermore, the specific process of the credit rating control module is as follows:

[0095] Industry-specific credit score quantification model: Real-time credit score of an enterprise is denoted as C. e (t). Here, the subscript e indicates that the subject of the credit score is an enterprise, the smallest unit for precise resource allocation, and t indicates that the credit score is a real-time dynamic value, updated in real time according to the enterprise's behavior. The basic credit score is denoted as C. e Used to calculate a company's real-time credit score, each company e calculates its basic credit score C. e The expression is as follows:

[0096]

[0097] in, This represents the weight of the j-th indicator specific to industry i, belonging to company e, reflecting the differences between industries, and

[0098] , This represents the normalized score of company e on indicator j (usually mapped to the interval [0, 100]), reflecting the different performances of different companies within the same industry; the following companies are examples:

[0099] Power grid company: C1 = 0.6•X1 (equipment reliability) + 0.3•X2 (task completion rate) + 0.1•X3 (timeliness of payment);

[0100] Logistics company: C2 = 0.5 x 1 (on-time delivery rate) + 0.4 x 2 (package integrity rate) + 0.1 x 3 (complaint rate)

[0101] This design allows different companies, such as power grid companies and logistics companies, to receive fair evaluations under their respective standards, avoiding fairness issues caused by mismatched metrics.

[0102] Dynamically adjust charging quota allocation: Calculate the charging quota that can be obtained in time period t based on the enterprise's credit score. The specific expression is as follows:

[0103]

[0104] in, This represents the quota allocated to company e during time period t. C represents the total available system quota for time period t. e (t) represents the real-time credit score of company e in time period t, and its specific expression is as follows:

[0105]

[0106] Where, λ e (t) represents a dynamic adjustment factor used to reward good behavior or punish bad behavior in the company recently. It is dynamically calculated based on real-time events recorded by the system (such as resource conflicts, task completion status, etc.) and is defined as follows:

[0107] λ e (t)>1: Reward good behavior; if there are no task defaults for 30 consecutive days, resource conflict rate <5%, and payment delay rate <3%, then λ e (t) is set to 1.05 (i.e., credit score increases by 5%).

[0108] λ e (t)=1: Normal behavior, no adjustment;

[0109] λ e (t)<1: Punish misbehavior; such as recent resource conflicts or task timeouts, then λ e (t) can be set to 0.9 (i.e., a 10% reduction in credit score); in the event of serious breach of trust, λ e (t) is immediately set to a very small value, triggering a credit rating downgrade.

[0110] Corporate Credit Score C e The quantitative standard for (t) is enterprise-differentiated credit evaluation, which divides user credit rating into four levels: A / B / C / D. The dynamic credit assessment model is as follows:

[0111]

[0112] Dual access control: Attribute-based access control and real-time system status-based access control are combined to form the final permission decision; access requests are set as a tuple, the expression of which is as follows:

[0113]

[0114] Subject represents user ID, industry type, and credit score C. e (t), where Action represents requesting charging or querying status, Resource represents the charging platform ID and available power, and Environment represents the current time t and platform load rate L;

[0115] The output of the access control policy is {Permit, Deny}, and its expression is as follows:

[0116]

[0117] in, E represents the minimum credit score threshold. est Indicates the drone's battery capacity B based on the requested charging. cap and current remaining power SOC uv The estimated amount of electricity required for this charge is expressed as follows: , Indicates the rated capacity of the drone battery. Indicates the remaining battery power of the drone. Indicates the amount of charge the target has.

[0118] An access request will only be approved if all of the following conditions are met:

[0119] Condition 1: The requester's credit score (C) e (t) must meet or exceed the minimum threshold ;

[0120] Condition 2: The remaining charging quota of the requester during this time period t. ≥E est This condition ensures that the allocated charging quota is sufficient to complete the charging task, thus avoiding the fragmentation of resource allocation.

[0121] Condition 3: The load rate L of the target charging platform does not exceed its maximum safety limit L. max ;

[0122] Credit rating and real-time power binding: When the system load rate L≥80%, a dynamic strategy is automatically triggered, and the charging power of C-level users is limited to 0.5C; Priority downgrade: Under the same high load conditions, the priority score of C-level users' tasks in the scheduling queue is multiplied by a decay factor δ (e.g., δ=0.5).

[0123] Collaborative innovation with reinforcement learning:

[0124] Credit Score Ce (t) is used as part of the reinforcement learning reward function R to calculate the resource conflict penalty. The expression is as follows:

[0125]

[0126] The credit mechanism guides the entire multi-agent system toward a fairer and more efficient direction by changing the optimization environment of the reinforcement learning agent.

[0127] Compared with the prior art, the present invention has the following beneficial effects:

[0128] This invention employs a hierarchical federated architecture, allowing participants from various industries to train models locally. It innovatively makes noise amplitude a variable dynamically determined by the local gradient variance, uploading only gradient noise processed with dynamic differential privacy to strictly ensure the original data is not leaked. Simultaneously, a dynamic noise adjustment mechanism maximizes the global model utility. The global prediction model generated by federated learning is embedded into a multi-agent deep deterministic policy gradient framework as the initial state space for reinforcement learning. Resource allocation records generated during reinforcement learning decision-making feed back into the training data updates for federated learning, achieving cross-layer coupling between federated learning and reinforcement learning. This enables each UAV's local scheduling decisions to be based on global optimization information, achieving system-level dynamic multi-objective optimal scheduling. The combination of MADDPG and constraint optimization algorithms in reinforcement learning constitutes an enhanced CTDE architecture, which not only retains all the advantages of CTDE (such as privacy protection and scalability) but also solves the most challenging conflict and hard constraint satisfaction problems in multi-agent systems by introducing a lightweight centralized arbitration layer.

[0129] Simultaneously, a three-layer hybrid architecture of "main chain-side chain-IPFS" is designed. The side chain efficiently processes high-frequency, small-value transactions, while the main chain periodically stores transaction summaries, achieving a perfect balance between performance, security, and cost. It innovatively integrates homomorphic encryption (HE) and threshold secret sharing (TSS). Raw billing data is homomorphically encrypted and stored in IPFS, with its access keys distributed and sharded. Quantum-resistant cryptographic algorithms are used to encrypt these shards, providing a forward-looking guarantee for the system's long-term data security.

[0130] Unique credit score quantification formulas and indicator weights are designed for different industries such as power, logistics, and agriculture, enabling fair and impartial evaluation across industries. Real-time credit scores are dynamically linked to the total system resources through a weighted allocation model, supporting dynamic credit quota allocation and automatically calculating and allocating charging quotas to achieve precise resource allocation that "rewards the good and punishes the bad." At the same time, credit scores are directly embedded as factors into the reward function of reinforcement learning, enabling the scheduling system to proactively avoid allocating scarce resources to low-credit users and intelligently guide the system to evolve in a fairer and more efficient direction.

[0131] By organically integrating the above-mentioned innovations, several key challenges that have constrained the large-scale cross-industry application of low-altitude economic drones in terms of improving endurance have been systematically addressed. These challenges include collaborative scheduling under privacy and security, reliable and efficient value settlement, fair and accurate resource allocation, and standardized and seamless facility access. This provides crucial technical support for building an open and shared digital infrastructure for the low-altitude economy. Attached Figure Description

[0132] Figure 1 This is a system architecture diagram of a specific embodiment of the present invention;

[0133] Figure 2 A schematic diagram of the drone charging process of this invention:

[0134] Figure 3 This is a schematic diagram of the entire process of the billing audit module of the present invention, from the occurrence of charging to the completion of billing audit. Detailed Implementation

[0135] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and specific examples.

[0136] System Architecture

[0137] like Figure 1 As shown, a system supporting automatic charging of drones across industries demonstrates the interaction between an intelligent collaborative decision-making and operation and maintenance cloud platform, an automatic charging platform, drones, and enterprise intelligent agents:

[0138] The intelligent collaborative decision-making and operation and maintenance cloud platform, in collaboration with industry participants, is used to allocate corresponding automatic charging platforms to drones waiting to be charged, and is deployed and managed by the power grid company.

[0139] The automatic charging platform is deployed by the power grid company on selected power grid towers to work in conjunction with the intelligent collaborative decision-making and operation and maintenance cloud platform. It receives drone charging scheduling instructions from the latter and supports drones to fly to the platform for automatic charging.

[0140] Enterprise intelligent agents, deployed by the corresponding enterprise, are computing devices deployed on the enterprise side of the drone's owner. They serve as a unified agent node for multiple drones within the enterprise, running a corresponding intelligent agent for each drone lacking computing power, and interacting with the intelligent collaborative decision-making and operation and maintenance cloud platform in the background.

[0141] The intelligent collaborative decision-making and operation and maintenance cloud platform includes the following modules:

[0142] The central intelligent scheduling module adopts a hierarchical federated learning architecture for data fusion based on dynamic differential privacy, retaining more information for non-sensitive data, improving the accuracy of the global model, and is coupled with the reinforcement learning module to achieve global prediction and intelligent scheduling based on federated learning.

[0143] The reinforcement learning module adopts a training framework integrated with federated learning for hierarchical decision-making systems for multi-agent collaborative decision-making, solving dynamic multi-objective optimization of cross-industry drone charging scheduling;

[0144] The blockchain billing audit module combines consortium blockchains, verifiable privacy computing, and tamper-resistant distributed storage to support blockchain-based drone charging billing auditing.

[0145] The credit rating control module adopts an industry-differentiated, dynamically adjusted and system-integrated control mechanism to achieve fair and efficient resource allocation and risk management.

[0146] The automatic charging platform includes the following modules;

[0147] The energy management module is used to integrate wind and solar hybrid energy storage, electromagnetic induction energy storage and fuel cell energy storage for unified scheduling, so as to achieve intelligent power supply and stable system output;

[0148] A standardized interface module is used to unify communication specifications and extend fast charging protocols, supporting adaptive charging for drones across industries. It adopts a standardized interface protocol certified by IEEE P 1937.18, unifies communication specifications (such as MQTT+JSON) and extends fast charging protocols to 2kW, and is compatible with WPCQi standards (v2.0 and above) to support adaptive charging for drones across brands.

[0149] The IoT authentication module is used for automatic real-time two-way identity authentication and authorization between the drone and the platform to ensure charging safety;

[0150] The edge computing module is responsible for executing real-time control commands and performing lightweight data preprocessing and caching.

[0151] This application introduces an enterprise intelligent agent, enabling the system architecture to support a hybrid direct / indirect connection mode to adapt to drone devices with varying computing capabilities. For drones with powerful onboard computing capabilities, they can directly access the backend intelligent collaborative decision-making and operation and maintenance cloud platform as industry participants, as previously described. However, for drones with limited computing power, it is recommended to deploy a corresponding enterprise intelligent agent within the enterprise to access the backend intelligent collaborative decision-making and operation and maintenance cloud platform. In this mode, the drone is only responsible for local state perception and command execution, while its corresponding intelligent agent runs within the enterprise intelligent agent. This agent represents the drone as an industry participant, responsible for gradient exchange in federated learning and collaborative decision-making in reinforcement learning with the cloud platform. This design significantly reduces the computing power requirements on the drone while ensuring the privacy of enterprise data and the scalability of the system.

[0152] System Deployment

[0153] A smart collaborative decision-making and operation and maintenance cloud platform is deployed in the cloud. Automatic charging platforms are then installed on selected power grid towers, connected to wind-solar hybrid power sources and / or electromagnetic induction power extraction devices, and equipped with energy storage devices, embedded control systems, communication modules, and digital identity recognition modules (such as DID). After installation, each charging platform registers its availability and resource status with the cloud platform. Users register accounts through the blockchain platform, uploading information such as drone model and battery capacity to obtain a unique identifier for subsequent billing and access control. The system is deployed in a layered manner: Cloud: global optimization and model training; Edge: automatic charging platforms perform local energy management, charging control execution, and edge computing; End-side: local perception and decision-making by industry participants (drones or enterprise intelligent agents). After system deployment, each automatic charging platform monitors its own operating status in real time through its edge computing module and continuously reports the data to the backend smart collaborative decision-making and operation and maintenance cloud platform.

[0154] Charging task scheduling process

[0155] like Figure 2 As shown, in this invention, the modules are connected through data flow and control flow to form a collaborative closed-loop system. The specific process is as follows:

[0156] Initiation point (charging request): When the drone detects that its battery level is below a preset threshold, it will send a charging request to the intelligent collaborative decision-making and operation and maintenance cloud platform through itself or its corresponding enterprise intelligent agent. The request includes the drone's real-time status data, such as remaining battery power (SOC) and mission urgency, but does not include private data.

[0157] Cloud platform intelligent scheduling decision-making: After receiving a request, the cloud platform makes collaborative decisions with industry participants through the central intelligent scheduling module and reinforcement learning module.

[0158] 1) Federated learning and privacy protection: Based on a hierarchical federated learning architecture, participants from various industries add dynamic differential privacy noise to the gradient data locally before uploading it. The cloud platform only integrates this protected gradient information to ensure data privacy and security.

[0159] 2) Reinforcement learning optimization: Combining the MADDPG algorithm, participants from various industries generate local charging platform selection suggestions based on their local conditions and upload them to the cloud platform. The cloud platform uses a constrained optimization algorithm to integrate the global conditions (including charging platform load rate, meteorological risk value, resource conflict matrix, etc.) with local suggestions and centrally output the optimal allocation scheme.

[0160] 3) Output action: After the cloud platform selects the optimal automatic charging platform, it generates a scheduling instruction (including target platform ID and authorization information).

[0161] Command issuance: The cloud platform sends dispatch instructions (including target charging platform ID, authorization information, etc.) to the drone and the corresponding automatic charging platform through a secure channel; the drone flies to the target charging platform according to the instructions, and uses RTK (real-time dynamic positioning) technology to achieve centimeter-level precise landing during the process.

[0162] Authentication and Charging: After the drone lands, the automatic charging platform initiates the following process:

[0163] 1) Two-way authentication: Through the IoT authentication module, the drone and the platform conduct real-time two-way authentication to ensure charging safety;

[0164] 2) Physical Fixation and Connection: The platform's physical locking device activates wind-resistant fixation and establishes a charging connection with the drone through a standardized interface module;

[0165] 3) Energy Management: The energy management module provides a stable power output during the charging process. The charging power is dynamically adjusted according to the user's credit level (e.g., Level A users are allowed 2C fast charging, while Level C users are limited to 0.5C).

[0166] Recording and settlement: such as Figure 3 The diagram illustrates the entire process of the billing audit module from the start of charging to the completion of the billing audit. Data generated during charging (such as electricity level E and time T) is initially processed by the edge computing module and then uploaded to the blockchain billing audit module for reliable recording and billing. After charging is completed, the smart contract triggers a payment instruction, deducting the fee and updating the user's balance. The credit rating control module dynamically updates the user's credit score based on the charging behavior (this affects the fairness of subsequent scheduling; for example, users with low credit scores will face higher penalties). The drone starts, flies away from the automatic charging platform, and the task ends.

[0167] The above description is merely an embodiment of the present invention. It should be reiterated that those skilled in the art can make several improvements to the present invention without departing from the principle of the present invention, and these improvements are also included within the scope of protection of the claims of the present invention.

Claims

1. A system supporting automatic charging of drones across industries, characterized in that: include: The intelligent collaborative decision-making and operation and maintenance cloud platform, in collaboration with industry participants, is used to allocate corresponding automatic charging platforms to drones waiting to be charged. Industry participants can access the backend cloud platform directly to use the automatic charging system, or indirectly access the backend cloud platform through an enterprise-deployed intelligent agent as an intermediary. An automatic charging platform is deployed on selected power grid towers and works in conjunction with an intelligent collaborative decision-making and operation and maintenance cloud platform to receive drone charging scheduling instructions from the cloud platform and support drones to fly to the platform for automatic charging. The intelligent collaborative decision-making and operation and maintenance cloud platform includes the following modules: The central intelligent scheduling module adopts a hierarchical federated learning architecture for data fusion based on dynamic differential privacy, retains more information for non-sensitive data, and is coupled with the reinforcement learning module to achieve model parameter sharing through a data interaction interface. The reinforcement learning module employs a training framework integrated with federated learning for hierarchical decision-making systems oriented towards multi-agent collaborative decision-making, wherein federated learning is a distributed machine learning technique. The blockchain billing audit module combines consortium blockchains, verifiable privacy computing, and tamper-resistant distributed storage to support blockchain-based drone charging billing auditing. The credit rating control module adopts an industry-differentiated, dynamically adjusted, and system-integrated control mechanism. The automatic charging platform includes the following modules; The energy management module is used to integrate wind and solar hybrid energy storage, electromagnetic induction energy storage and fuel cell energy storage for unified scheduling, so as to achieve intelligent power supply and stable system output; A standardized interface module is used to unify communication standards and extend fast charging protocols, supporting adaptive charging for drones across industries; The IoT authentication module is used for automatic real-time two-way identity authentication and authorization between the drone and the platform to ensure charging safety; The edge computing module is responsible for executing real-time control commands and performing lightweight data preprocessing and caching.

2. The system for supporting automatic charging of drones across industries according to claim 1, characterized in that: The data flow and calculation process of the central intelligent scheduling module are as follows: Local training: Participants from various industries on the local dataset D k Train a local charging demand prediction model and minimize the local loss function. , where y pred This represents the predicted value y based on the model parameters w. true The model parameters are optimized by iteratively minimizing the loss function to represent the true value using the gradient descent algorithm; After training, each participant obtains the original gradient for the t-th aggregation cycle. ; Privacy handling and uploading: Each participant's original gradient Noise is added for perturbation, and the gradient expression after perturbation is as follows: Where 𝒩 represents a Gaussian distribution, Represents the identity matrix. The dynamically adjusted standard deviation is expressed as follows: in, Let τ represent the gradient variance calculated for participant k in round t, and τ represent the empirical threshold. Cloud aggregation: Perturbation-induced gradient The data is uploaded to the cloud aggregation server via a secure channel. The cloud executes the aggregation algorithm to generate the updated global model expression as follows: Where η represents the learning rate, and D represents the union of the datasets of all participants. k This represents the local dataset of participant k. This represents the gradient after aggregation. This represents the global model parameters in round t+1; Model Distribution and Application: Updated Global Model Parameters The model is distributed back to industry participants to predict the next round of local training and charging needs, and the cloud generates a global model for the next round of predictions through an aggregation algorithm.

3. The system for supporting automatic charging of drones across industries according to claim 1, characterized in that: The hierarchical decision system in the reinforcement learning module is a hierarchical reinforcement learning framework, including a high-level controller and low-level executors. The high-level controller, as a meta-controller, selects the optimal objective function from a dynamic multi-objective optimization library based on the system state of the global prediction model generated by federated learning. Based on the local decision suggestions uploaded by each low-level executor, it uses a constrained optimization algorithm for centralized arbitration to generate a final allocation instruction, which is then sent to the low-level executors for execution. The low-level executors are responsible for local state perception and action suggestion generation, and receive the final action instruction from the high-level controller for execution.

4. A system for supporting automatic charging of drones across industries according to claim 3, characterized in that: The specific methods of the reinforcement learning module are as follows: The system state is a hybrid state for a single drone charging request decision instance, capturing all key information at time point t, including global and local information, as shown in the following expression: ； in, Indicates the remaining battery power of the drone. This indicates the real-time load rate of each charging platform. Indicates the urgency level of the task. Indicates the meteorological risk value. Represents the resource conflict matrix, C e (t) represents the real-time credit score; Action Space A t The specific expression is as follows: ； in, This indicates selecting a target platform from the set of available charging platforms. This refers to the charging power allocated to the drone. The action generation process involves participants from various industries generating action suggestions based on local strategies and uploading them to the cloud. The cloud then performs centralized optimization before issuing the final action command for execution. The reward function expression for the dynamic multi-objective optimization is as follows: ； in, This represents the reward for successful charging. A successful charge earns a positive reward of +1, while a failed charge earns a negative reward of -1. This indicates a reward for timely completion of the task; The resource conflict penalty is represented by α, β, and γ, which represent the adaptive coefficients for task urgency, load balancing, and conflict penalty, respectively. The expression for the task urgency adaptive coefficient α is as follows: ； Where α0 = 0.1 represents the scaling factor. The normalized value is taken to be in the range [0.1, 1.0]. Take discrete values ​​{1, 2, 3, 4, 5}, with higher levels corresponding to smaller values. This indicates the flight distance from the drone to the charging platform, with an effective range of [0.5, 10] km; The expression for the load balancing adaptive coefficient β is as follows: ； in, The load balancing factor represents the degree of load balancing of the charging platform, σ² represents the variance of the load rate of all charging platforms, and the normalized value range of the charging platform load rate is [0,1]. The expression for the conflict penalty adaptive coefficient γ is as follows: ； Where γ0 = 0.5 represents the base penalty coefficient. Indicates a system high load status identifier; System high load status identifier The expression is as follows: ； Where IF(•) represents an indicator function, θ represents the average load rate of all charging platforms. high =0.7 represents the overall high load threshold of the system. When the average load rate exceeds this threshold, the system is considered to have entered a high load state. σ² represents the variance of the load rate of all charging platforms, and θ balance =0.01 indicates the load balancing threshold.

5. A system for supporting automatic charging of drones across industries according to claim 1, characterized in that: The specific methods of the blockchain billing and auditing module are as follows: 1) A three-layer hybrid architecture and data flow of main chain-side chain-IPFS Based on the Hyperledger Fabric main chain-side chain design, the main chain is the core chain in the consortium blockchain, responsible for final notarization and auditing, and storing data summaries; the side chain is an auxiliary chain connected to the main chain, which is dedicated to processing more than 1,000 charging transactions per second with a single transaction amount of less than 20 yuan, and periodically anchors the data summaries to the main chain. Transaction Generation: The edge computing module calculates charging event data, which includes power E, time T, power P, and user identifier UID; the expression for generating a transaction billing voucher is as follows: ； Where H represents the hash function, Indicates a digital signature; Privacy processing and off-chain storage: The raw data (Data) uses homomorphic encryption to achieve verifiable privacy computation, supporting billing calculations on the ciphertext, as shown in the following expression: ； The encrypted data C is uploaded to IPFS, and a content identifier is returned. The Content Identifier (CID) is used as an index and stored on the blockchain through subsequent processes; Sidechain processing and high-frequency settlement: Billing transactions are routed to the sidechain, which uses a hash time-locked contract protocol for processing. Data is stored on the sidechain for settlement, as shown in the following expression: ； Where H(R) represents the hash of the payment preimage R generated by the payer. This indicates the timeout period, and Amount indicates the payment amount. Main chain notarization and auditing: The side chain periodically stores a batch of Merkle root data for each transaction. m Merkle root of the corresponding content identifier (CID) set storage Anchoring to the main chain, the expression is as follows: ； in, This represents the main chain transaction ID, Timestamp represents the timestamp, and Block_Hash represents the block hash value, used for recording and auditing. 2) Secure storage and audit traceability mechanism Data sharding and threshold secret sharing: Before storage, the encrypted data C is fragmented using Shamir's Secret Sharing scheme to generate n fragments (l, f(l)), where l = 1, 2, ..., n; given the secret SK, a polynomial is constructed as follows: ； Among them, a i represents the polynomial coefficients, x represents the variable, and mod p represents the modular operation; Data is fragmented and stored in four different edge computing modules using (k,n)=(3,4). During auditing and tracing, only any three fragments need to be collected to recover the original key SK using the Lagrange interpolation formula, as shown below: Quantum-resistant cryptography and encryption: Quantum-resistant cryptographic algorithms are used to encrypt the fragments, as shown in the following expression: ； Billing and Smart Contracts: Billing is performed through smart contracts deployed on the main chain, as shown in the following expression: ； Where E represents electricity (kWh), This indicates the real-time electricity price (yuan / kWh). This represents the industry discount factor. This represents the peak-valley coefficient.

6. A system for supporting automatic charging of drones across industries according to claim 1, characterized in that: The specific process of the credit rating control module is as follows: Industry-specific credit score quantification model: Each enterprise e calculates its basic credit score C. e The expression is as follows: ;in, This represents the weight of the j-th indicator specific to industry i, to which company e belongs, and , This represents the normalized score of firm e on indicator j: Dynamically adjust charging quota allocation: Calculate the charging quota that can be obtained in time period t based on the enterprise's credit score. The specific expression is as follows: ;in, This represents the quota allocated to company e during time period t. C represents the total available system quota for time period t. e (t) represents the real-time credit score of company e in time period t, and its specific expression is as follows: ; where λ e (t) represents a dynamic adjustment factor used to reward good behavior or punish bad behavior of the company in the near future; Dual access control: Attribute-based access control and real-time system status-based access control are combined to form the final permission decision; access requests are set as a tuple, the expression of which is as follows: ； Subject represents user ID, industry type, and credit score C. e (t), where Action represents requesting charging or querying status, Resource represents the charging platform ID and available power, and Environment represents the current time t and platform load rate L; The output of the access control policy is {Permit, Deny}, and its expression is as follows: ； in, This indicates the minimum threshold for credit score. This represents the quota allocated to company e during time period t. , Indicates the rated capacity of the drone battery. Indicates the remaining battery power of the drone. Indicates the amount of charge the target has received; Collaborative innovation with reinforcement learning: Credit Score C e (t) is used as part of the reinforcement learning reward function R to calculate the resource conflict penalty R. penalty The expression is as follows: 。

Images