Distributed task allocation and path coordination optimization algorithm for heterogeneous unmanned aerial vehicles

Abstract

This invention relates to the field of UAV swarm collaborative control technology, and discloses a distributed task allocation and path collaborative optimization algorithm for heterogeneous UAVs. The key technical points are: adopting a full-process collaborative optimization architecture with five modules connected in sequence and feedback in a closed loop: "generative neural symbol digital twin modeling, quantum-enhanced distributed causal matching, reinforcement learning-enhanced dual-system hierarchical game optimization, multimodal fusion heterogeneous adaptation obstacle avoidance, and time-series-aware active alliance evolution". This achieves an overall performance improvement of heterogeneous UAV swarms in terms of task allocation, path planning, dynamic obstacle avoidance, fault tolerance, and collaborative evolution, and ultimately achieves the goal of efficient, reliable, green, and adaptive swarm collaborative operation.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for unmanned aerial vehicle (UAV) swarms, and more specifically, to a distributed task allocation and path collaborative optimization algorithm for heterogeneous UAVs. Background Technology

[0002] With the widespread application of heterogeneous UAVs in complex scenarios such as inspection, plant protection, and emergency response, traditional centralized task allocation and path planning methods face significant challenges: centralized architectures suffer from high communication overhead and poor real-time performance, making them difficult to adapt to dynamic environments; task allocation is often based on greedy or heuristic rules, ignoring the deep causal constraints between UAV capabilities and task requirements, making it prone to getting trapped in local optima and slow convergence; path planning often relies on single-modal perception, is sensitive to environmental pseudo-associations, and does not fully consider the dynamic differences and real-time energy consumption coupling of heterogeneous UAVs such as rotorcraft and fixed-wing aircraft, resulting in low obstacle avoidance efficiency and short endurance; fault handling is mostly a passive response, lacking proactive fault tolerance and coalition reorganization mechanisms based on time-series prediction, which seriously affects operational continuity.

[0003] For example, see Chinese invention application CN120428742A, "An Efficient Cooperative Control Method for Multi-UAV Formation," which discloses that "the method uses digital twin dynamics simulation to predict UAV status and correct errors, and a real-time closed-loop optimization mechanism can quickly replan the path when flight is abnormal. It also reduces the difference between virtual and real data through a domain adaptive module. It integrates optimized dynamic window method and deep reinforcement learning for local obstacle avoidance, and uses distributed auction mechanism, spatiotemporal corridor technology and leader-follower model to achieve multi-UAV cooperative planning and formation stability." However, it still has shortcomings such as "ignoring the deep causal constraints of UAV capabilities and mission requirements, easily getting trapped in local optima and having slow convergence; path planning often relies on single-modal perception, is sensitive to environmental pseudo-correlation, and does not fully consider the coupling of dynamic differences and real-time energy consumption."

[0004] In summary, existing technologies lack a distributed collaborative closed loop that can integrate high-precision dynamic modeling, intelligent matching, hierarchical game optimization, heterogeneous adaptive obstacle avoidance, and active fault tolerance. This results in significant deficiencies in the system's dynamism, energy efficiency, reliability, and overall optimization performance. Existing heterogeneous UAV scheduling technologies are unable to meet the demands of large-scale, dynamic, and highly reliable operations in complex scenarios.

[0005] Therefore, this invention provides a distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs, which improves the above-mentioned technical problems. Summary of the Invention

[0006] This disclosure aims to address the shortcomings of existing technologies by providing a distributed task allocation and path collaborative optimization algorithm for heterogeneous UAVs. The invention employs a closed-loop collaborative architecture consisting of five modules: generative neural symbolic digital twin modeling, quantum-enhanced distributed causal matching, reinforcement learning-enhanced dual-system hierarchical game optimization, multimodal fusion heterogeneous adaptive obstacle avoidance, and time-series perception active alliance evolution. This architecture achieves fully coupled dynamic optimization of task-path-environment-energy consumption-carbon footprint, solving the stability, real-time performance, reliability, and green energy-saving issues in large-scale heterogeneous UAV swarm scheduling.

[0007] The above-mentioned technical objective of this invention is achieved through the following technical solution: a distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs, comprising the following steps:

[0008] S1. Generative Neural Symbol Digital Twin Modeling: Construct and dynamically update a neural symbol map representing the full-dimensional state and multi-dimensional mission requirements of the UAV. Through an adaptive causal incremental synchronization mechanism and edge computing hierarchical preprocessing, provide highly reliable and low-redundancy digital mapping inputs for subsequent modules.

[0009] S2, Quantum-enhanced distributed causal matching: Based on the graph output in step S1, through causal logic pruning, quantum feature extraction and encoding, distributed singularity conflict resolution and lightweight quantum approximation optimization, the UAV and the mission are accurately matched, and the optimal matching matrix is ​​output.

[0010] S3. Reinforcement learning enhances dual-system hierarchical game optimization: Based on the output of step S2, a dual-system hierarchical game framework of "fast response-global optimization" is constructed. By integrating reinforcement learning and game theory, the five-dimensional fully coupled optimization of task-path-environment-energy consumption-carbon footprint is achieved, and a structured decision rule chain and optimal action instructions are output.

[0011] S4. Multimodal Fusion Heterogeneous Adaptive Obstacle Avoidance: Based on the output of step S3 and environmental perception data, through multimodal topology modeling, federated learning-specific path generation, real-time energy consumption model coupling, and causal-driven dynamic obstacle avoidance, a safe and efficient flight path is generated and optimized in real time for the UAV.

[0012] S5. Time-series-aware proactive alliance evolution: Fault warning is achieved based on dynamic causal fault tree, supplementary drones are selected through multi-dimensional trust model, and seamless supplementation and dynamic updating of alliance structure are achieved by combining digital twin and path generation.

[0013] Steps S1 to S5 are executed sequentially, and a closed-loop optimization system is formed through multi-layer feedback calibration.

[0014] As a preferred embodiment of the present invention, S1 includes:

[0015] S1.1 Dynamic temporal prediction: A lightweight generative diffusion model is adopted to predict the future state and mission requirements of UAVs based on historical time series data;

[0016] S1.2 Data-driven feature extraction: Lightweight neural networks are used to perform nonlinear transformations on real-time state features, predicted state features, and task requirement features to generate high-dimensional discriminative features;

[0017] S1.3 First-order predicate logic reasoning: Establish logical association rules between UAV status and mission requirements through atomic predicates and compound predicates;

[0018] S1.4, Neural Symbol Atlas Construction: Integrate prediction, feature and rule information to construct a heterogeneous UAV neural symbol capability atlas and a task neural symbol requirement atlas;

[0019] S1.5 Adaptive Causal Incremental Synchronization and Edge Preprocessing: Based on communication load and energy consumption status, the synchronization threshold is dynamically adjusted to trigger key data synchronization, and outlier detection, key feature selection and dimensionality reduction preprocessing are performed at edge nodes.

[0020] S1.6 Closed-loop feedback calibration: Receive feedback results from subsequent modules and update model parameters and map association weights.

[0021] As a preferred embodiment of the present invention, S2 includes:

[0022] S2.1 Causal logic pruning: Based on first-order predicate logic rules and deep causal relationship analysis, drone-task pairs with mismatched capabilities or deep conflicts are eliminated to form a set of compatible matching pairs;

[0023] S2.2 Quantum Feature Extraction and Encoding: For compatible matching pairs, high-dimensional quantum features are extracted using a lightweight quantum random forest, and non-Hermitian encoding matrices are used to encode the features into quantum states or classical equivalent vectors.

[0024] S2.3 Distributed Singularity Conflict Resolution: Adjacent UAVs exchange coding features and utilize the singularity evolution characteristics of non-Hermitian systems to achieve self-driven conflict resolution;

[0025] S2.4 Lightweight Quantum Approximation Optimization: Construct an objective function that integrates matching fitness, resource utilization, and energy consumption balance, and solve for the optimal matching matrix using a lightweight quantum approximation optimization algorithm;

[0026] S2.5 Closed-loop feedback calibration: The matching results are fed back to the digital twin model to dynamically adjust the algorithm parameters.

[0027] As a preferred embodiment of the present invention, S3 includes:

[0028] S3.1 Construct a dual-system decision architecture of "rapid response - global optimization", in which the rapid response layer is based on a lightweight deep Q network to cope with millisecond-level environmental changes, and the global optimization layer is based on trust region policy optimization to achieve second- to minute-level global target balance.

[0029] S3.2 Training a fast response layer whose state vector contains key parameters of environmental change, outputting instantaneous actions through a deep separable convolutional network, and the reward function integrates obstacle avoidance success, energy consumption control and task delay;

[0030] S3.3 Training the global optimization layer, whose state vector contains all-dimensional system parameters, outputs global optimization actions through the policy function, and the update process is constrained by KL divergence to ensure stability;

[0031] S3.4 Construct a multi-dimensional hierarchical payment function that integrates local immediate benefits, global long-term benefits, energy consumption costs, and carbon footprint;

[0032] S3.5. Using drone swarms as game participants, solve the Nash equilibrium iteratively through strategy evaluation and strategy improvement;

[0033] S3.6. Feed the optimized strategy and rule chain back to the digital twin model and update the model parameters.

[0034] As a preferred embodiment of the present invention, S4 includes:

[0035] S4.1 Multimodal Topology Graph Construction: Integrating visual, lidar, meteorological, and location features, a high-precision environmental topology graph is constructed through Granger causality tests and attention mechanisms, filtering out pseudo-associative edges;

[0036] S4.2, Federated Learning Dedicated Path Generation: Using the federated learning framework, a meta-path generator adapted to the flight characteristics of different types of UAVs such as rotorcraft and fixed-wing aircraft is trained and path clusters are generated online in real time;

[0037] S4.3 Real-time energy consumption model coupling and path optimization: Couple path length, number of turns, flight speed and UAV real-time energy consumption status to optimize path cost while meeting endurance constraints;

[0038] S4.4 Causal-driven dynamic obstacle avoidance adjustment: Based on the causal relationship of the topology graph, only the future path of the associated UAVs is predicted, and the optimal obstacle avoidance adjustment amount is solved by model predictive control.

[0039] S4.5 Dynamic Iterative Closed Loop: When the environmental state changes drastically and exceeds the threshold, the entire process of updating the topology and adjusting obstacle avoidance is re-triggered.

[0040] As a preferred embodiment of the present invention, S5 includes:

[0041] S5.1 Fault Timing Prediction and Early Warning: The temporal deep features of fault precursors are extracted by LSTM, and the fault probability is calculated by combining dynamic causal fault tree and logistic regression model. When the threshold is exceeded, an early warning is triggered.

[0042] S5.2 Multidimensional Trust Model Construction: Integrating four dimensions—cooperative causal compatibility, failure probability of replacement drones, historical mission completion quality, and energy consumption stability—to quantify the trustworthiness of drone replacement.

[0043] S5.3, Rapid Selection of Candidate UAVs for Replacement: Based on trust threshold, failure probability upper limit and mission capability predicate, the candidate set is quickly selected at the edge node, and an expansion mechanism is provided;

[0044] S5.4, Replacement Path and Task Handover Optimization: Simulate the handover scenario in the digital twin model, generate an optimized path for the replacement UAV through the meta-path generator, and use model predictive control for dynamic correction;

[0045] S5.5 Dynamic Evolution of Alliance Structure: Update the alliance adjacency matrix and trust model parameters to achieve adaptive adjustment of the alliance;

[0046] S5.6 Closed-loop feedback calibration: Feedback the alliance evolution results to the digital twin model to update the fault prediction parameters and associated weights.

[0047] As a preferred embodiment of the present invention, in S1.3, the defined compound predicate The logical expression is: in, For drones Execute the task Sufficient energy consumption; For drones Execute the task The load capacity meets the requirements; For drones The accuracy of the operation meets the task requirements. Require; For drones Adaptation task Environmental requirements; For drones The equipment is in good health condition.

[0048] As a preferred technical solution of the present invention, in S2.2, quantum feature extraction and encoding are adapted to three engineering operation modes: classical platform quantum circuit simulation, small-scale commercial quantum hardware deployment, and quantum-inspired classical algorithm implementation; according to different modes, the corresponding implementation method is selected to generate high-dimensional quantum features or quantum-inspired classical features, and they are encoded through a non-Hermitian encoding matrix.

[0049] As a preferred embodiment of the present invention, in step S3.4, the constructed payment function is as follows: in, The time decay coefficient increases with time. For scene adaptation coefficients, and These are the penalty coefficients for energy consumption and carbon footprint, respectively. For the local immediate benefit of drone i performing action a, For drone i in the time window The overall long-term benefits within, The energy cost of drone i performing action a. The carbon footprint of drone i performing action a.

[0050] As a preferred embodiment of the present invention, in step S5.2, the comprehensive calculation formula for multi-dimensional trust level is as follows:

[0051] in, to These are the weight coefficients for each dimension, and their sum is 1; To coordinate the degree of causal compatibility, To supplement the predicted failure probability of drone y, To compensate for the historical mission completion quality of drone Y, To compensate for the energy consumption stability of drone y.

[0052] In summary, the present invention has the following beneficial effects:

[0053] Firstly, through the generative neural symbol digital twin modeling module, a high-precision, low-redundancy digital mapping of the UAV's full-dimensional state and multi-dimensional mission requirements was achieved. Combined with adaptive causal incremental synchronization and edge preprocessing, the communication load and core node pressure were effectively reduced, providing a highly reliable real-time data foundation for subsequent optimization. This solved the problems of insufficient accuracy, high data redundancy, and poor real-time performance of traditional modeling methods.

[0054] Secondly, by using a quantum-enhanced distributed causal matching module, and based on causal logic pruning, lightweight quantum computing or heuristic classical optimization is introduced to achieve efficient and accurate matching between UAVs and missions. This balances the global optimality of the matching with the algorithm's convergence speed, and solves the technical defects of traditional allocation methods, such as being prone to getting trapped in local optima and having slow convergence.

[0055] Third, by enhancing the dual-system hierarchical game optimization module through reinforcement learning, a collaborative decision-making mechanism of "rapid response-global optimization" was constructed, which integrates the ability to cope with immediate environmental changes and the ability to balance long-term benefits. This achieves multi-dimensional coupled optimization of tasks, paths, energy consumption, and carbon footprint, thereby improving the system's dynamic adaptability, interpretability, and global optimization effect.

[0056] Fourth, by integrating multimodal perception data to construct a high-precision topology environment model through a multimodal fusion heterogeneous adaptation obstacle avoidance module, and generating exclusive paths adapted to the characteristics of different aircraft types such as rotorcraft and fixed-wing aircraft based on federated learning, and combining real-time energy consumption models for path optimization and causal-driven obstacle avoidance, the flight energy consumption is significantly reduced while ensuring obstacle avoidance safety, solving the pain points of traditional methods such as false association misjudgment, poor heterogeneous adaptability, and high energy consumption.

[0057] Fifth, through the time-series-aware proactive alliance evolution module, early fault warning based on dynamic causal fault trees and rapid replacement screening based on multi-dimensional trust models are realized. Combined with digital twin simulation and path optimization, fault handling is upgraded from passive response to proactive seamless connection. Furthermore, through the dynamic evolution of the alliance structure and closed-loop feedback, the continuity and overall reliability of cluster operations are greatly improved. Attached Figure Description

[0058] Figure 1 A flowchart illustrating a distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs provided in an embodiment of the present invention. Detailed Implementation

[0059] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application. These all fall within the protection scope of the present application.

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0061] Unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0062] Furthermore, the technical features involved in the various embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0063] This disclosure aims to address the technical problems of existing heterogeneous UAV swarms in task allocation and path coordination under dynamic and complex environments, such as insufficient global optimization, poor real-time performance, high energy consumption, weak fault tolerance, and poor heterogeneous adaptability. Therefore, this disclosure proposes a distributed task allocation and path coordination optimization algorithm for heterogeneous UAVs, applicable to various types of UAV swarms including rotorcraft, fixed-wing aircraft, compound-wing aircraft, and vertical takeoff and landing fixed-wing aircraft. It can be applied to complex operational scenarios such as low-altitude logistics, emergency rescue, urban inspection, agricultural plant protection, maritime search and rescue, and power line inspection, and possesses cross-scenario migration capabilities. It employs a full-process collaborative optimization architecture with five sequentially connected and closed-loop feedback modules: "generative neural symbolic digital twin modeling—quantum-enhanced distributed causal matching—reinforcement learning-enhanced dual-system hierarchical game optimization—multimodal fusion heterogeneous adaptation obstacle avoidance—time-sequence-aware active alliance evolution." This architecture achieves overall performance improvement in heterogeneous UAV swarms in task allocation, path planning, dynamic obstacle avoidance, fault tolerance, and collaborative evolution, ultimately achieving efficient, reliable, green, and adaptive swarm collaborative operations.

[0064] Please refer to Figure 1 , Figure 1 The flowchart of the distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs described in this disclosure embodiment is shown. Based on a three-layer closed-loop architecture of "hardware layer-algorithm layer-application layer", and using a distributed self-organizing hypernetwork as the core scheduling carrier, the entire process optimization is completed collaboratively through the following five steps:

[0065] S1. Generative Neural Symbol Digital Twin Modeling Module: This module integrates the dynamic temporal prediction capabilities of a lightweight generative diffusion model, the data-driven feature extraction capabilities of a lightweight neural network, and the rule-based reasoning capabilities of first-order predicate logic to construct a neural symbolic atlas representing the full-dimensional state and multi-dimensional mission requirements of the UAV. Through an adaptive causal incremental synchronization mechanism and edge computing hierarchical preprocessing, a closed loop of "data acquisition - dynamic prediction - feature extraction - logical association - atlas construction - synchronous preprocessing - feedback calibration" is achieved, providing highly reliable and low-redundancy digital mapping inputs for subsequent modules.

[0066] Variable definition: Let the set of heterogeneous drone swarms be . ( (Total number of drones in the swarm); Individual drone ( The real-time state feature set of ) is ( (Total number of UAV state characteristics), including real-time energy consumption. (Characterizing the current remaining usable energy of the UAV), payload status (Characterizing the current payload mass and remaining payload capacity of the UAV), flight performance degradation trend (Value range [0,1], the closer the value is to 1, the smaller the performance degradation), real-time flight speed Attitude angle (respectively, roll angle, pitch angle, and yaw angle), equipment health status (Value range [0,1], the closer the value is to 1, the better the device status) and other key characteristics; the set of tasks to be executed is ( (Total number of tasks); Single task ( The demand feature set is ( (Total number of task requirement features), including dynamic priority. (Value range [1, 10], the larger the value, the higher the urgency of the task), time window constraint (Earliest start time and latest end time of the task, respectively), and accuracy requirements for the task. (Spatial positioning accuracy required for task execution), operating range (Size of the spatial region covered by the task), environmental adaptation requirements (Values ​​range [0,1], representing the task's tolerance to complex environments) and other key features; the parameter set of the lightweight generative diffusion model is... (Drone status) (Task Requirements); The network parameter set for the data-driven module is as follows: (Drone) (Task); The adjacency matrix of the neural symbolic map is (Drone capabilities) (Task requirements); Synchronization threshold related parameters are: (Dynamic synchronization threshold) (Communication load associated threshold) (Energy consumption status associated threshold); Features after edge preprocessing are (Dimensional reduction characteristics of drones).

[0067] S1.1 Dynamic Temporal Prediction: The dynamic prediction module of the lightweight generative diffusion model is trained based on the historical state temporal dataset of UAVs and the dynamic change dataset of mission requirements. It outputs the predicted values ​​of future state and requirements, providing temporal enhancement input for the data-driven module.

[0068] Specifically, the time-series dataset of drone historical states is defined as follows: ( (total length of time series data sampling), where For drones In the The state feature vector at each sampling time; the dataset shows dynamic changes in task requirements. ,in For the task In the The module requires a feature vector at each sampling time. It introduces Gaussian noise to perturb the original data through forward diffusion, and then learns the data distribution and reconstructs the predicted value through a reverse generation process. The specific process satisfies:

[0069]

[0070]

[0071]

[0072]

[0073] in, This represents the forward diffusion conditional probability distribution; The mean ,variance Gaussian distribution; The noise coefficient sequence for the UAV status data (value range [0.0001, 0.02]). The sequence of noise coefficients for the task requirement data (value range [0.0001, 0.02]). It is the identity matrix; , These are lightweight generative diffusion prediction functions for UAV status and mission requirements, respectively. , Noise prediction networks for drone status and mission requirements (lightweight U-Net architecture, parameter set) , ); To predict the time step.

[0074] The model parameters are iteratively optimized using the AdamW optimizer, with the optimization objective being to minimize the prediction error loss function.

[0075]

[0076] in, For drones No. The actual noise vector at the sampling time; For the task No. The actual noise vector at the sampling time; It is the square of the L2 norm.

[0077] S1.2 Data-driven feature extraction: The data-driven module of the lightweight neural network adopts a six-level architecture of "standardization-convolutional feature extraction-batch normalization-activation-random deactivation-fully connected mapping" to perform nonlinear transformations on real-time state features, predicted state features and task requirement features to generate high-dimensional discriminative features.

[0078] Specifically, the UAV state feature mapping network is defined as follows: (parameter set) , For the first Layer weight matrix, For the first Layer bias vector); task requirement feature mapping network is (parameter set) , For the first Layer weight matrix, For the first Layer bias vector).

[0079] First, the input features are standardized to eliminate dimensional differences:

[0080]

[0081]

[0082] in, For drones No. Standardization results of class state features; This is the global mean of the feature; This is the global standard deviation of the feature; To prevent extremely small values ​​where the denominator is zero (range [1e-6, 1e-4]); For the task No. Standardization results of class-specific requirements characteristics; This is the global mean of the feature; This is the global standard deviation of this feature.

[0083] The standardized features are concatenated with the predicted features output by the dynamic prediction module and then input into the network for feature extraction and mapping.

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] in, For drones The standardized state feature vector; For the task Standardized demand feature vector; For feature splicing operations; For one-dimensional convolution operations ( , The number of convolution kernels, , (where kernel size is the number of convolutional kernels). For batch normalization operation; For activation functions; For random deactivation operation ( The inactivation probability ranges from [0.1, 0.3]. For fully connected layer operations ( , (The output dimension is set according to the task complexity, representing the high-dimensional feature output dimension). For drones High-dimensional feature representation; For the task High-dimensional feature representation.

[0097] S1.3 First-order predicate logic reasoning: Establish logical association rules between UAV status and mission requirements through atomic predicates and compound predicates, providing rule constraints for neural symbolic graphs.

[0098] Specifically, define the set of atomic predicates:

[0099]

[0100] ,

[0101] The logical definitions of each atomic predicate are as follows:

[0102] Drones Execute the task Energy consumption is sufficient, and the truth value determination rule is: ,in For the task Estimated energy demand This is the energy redundancy coefficient (value range [1.05, 1.2]).

[0103] Drones Execute the task The load capacity meets the requirements, and the truth value judgment rule is as follows: ,in For drones Maximum load capacity, For the task Load requirements, This is the load redundancy factor (value range [1.05, 1.2]).

[0104] Drones The accuracy of the operation meets the task requirements. The truth-testing rule is as follows: ,in For drones Maximum operating accuracy.

[0105] Drones Adaptation task Environmental requirements, truth value determination rules are as follows ,in For drones Environmental adaptability (value range [0,1]).

[0106] Drones The equipment is in good health condition, and the truth value determination rule is as follows: ,in This is the threshold for the health status of the equipment (the value ranges from [0.6, 0.8], and anything below this threshold is considered an abnormality).

[0107] Constructing compound predicates based on atomic predicates Characterizing drones Capable of performing tasks The comprehensive ability, logically expressed as:

[0108]

[0109] in, For logical AND operation, a compound predicate is true only if all atomic predicates are true.

[0110] S1.4, Neural Symbolic Atlas Construction: Based on the output of the dynamic prediction module, the high-dimensional features of the data-driven module, and the predicate rules of the rule reasoning module, a heterogeneous UAV neural symbolic capability atlas is constructed. Neural Symbol Demand Mapping for Tasks :

[0111] The set of high-dimensional feature representations for all UAVs;

[0112] A set of high-dimensional feature representations for all tasks;

[0113] A set of predicates specific to drones (including device status, performance level, etc.);

[0114] A set of predicates specific to the task (including time urgency, accuracy criticality, etc.);

[0115] For drone capability-related adjacency matrix, elements ,in For feature similarity, This is the predicate for inter-drone collaborative capabilities (truth value 1 or 0).

[0116] To meet the task requirements, an adjacency matrix is ​​associated, with elements... ,in For feature similarity, This is the predicate for inter-task conflict (a truth value of 1 indicates no conflict, and 0 indicates a conflict).

[0117] S1.5 Adaptive Causal Incremental Synchronization and Edge Preprocessing:

[0118] An adaptive causal incremental synchronization mechanism dynamically adjusts the synchronization threshold based on the real-time communication load and energy consumption status of the UAV, synchronizing key data only when triggering conditions are met. This, combined with edge computing preprocessing, reduces the pressure on core scheduling nodes. The synchronization triggering conditions are: ① The change in UAV state characteristics reaches the dynamic synchronization threshold; ② First-order predicate logic rules trigger (e.g., ...). The truth value changes from true to false.

[0119] Dynamic synchronization threshold For drones At any moment The synchronous triggering threshold is calculated using the following model:

[0120]

[0121] in, It is a dynamic weighting coefficient that can be adaptively adjusted according to the ratio of real-time communication load to remaining energy consumption (value range [0,1]). It increases when the communication load is high and decreases when the energy consumption is low. The threshold is associated with the communication load; The threshold is associated with the energy consumption status.

[0122] Communication load association threshold calculation:

[0123]

[0124] in, This is the communication load adjustment factor (value range [0.1, 0.5]). For a moment Used communication bandwidth; Maximum communication bandwidth; For a moment The amount of data to be transmitted; To estimate transmission time; This represents the maximum amount of data that can be transmitted in a single transaction. This is the minimum transmission time threshold.

[0125] Energy consumption status associated threshold calculation:

[0126]

[0127] in, Energy consumption adjustment factor (value range [0.2, 0.6]); For a moment Real-time energy consumption; Maximum battery life energy consumption; This represents the real-time energy consumption change rate. The time constant is the weight of the rate of change in energy consumption.

[0128] UAV state characteristic changes The Euclidean distance between the current state feature vector and the previous sampling time:

[0129]

[0130] in, The data sampling interval; , The first Class features at time The value compared to the previous sampling time.

[0131] when When a predicate rule is triggered, the drone synchronizes key data via the LoRaWAN / NB-IoT protocol. ,in The set of predicates that trigger synchronization. To synchronize timestamps.

[0132] After the edge computing layered preprocessing and synchronous data transmission to the edge computing nodes, the "outlier detection and correction - key feature selection - feature dimensionality reduction" process is executed:

[0133] Outlier Detection and Correction: The Isolation Forest algorithm is used to identify outliers. The outlier detection function is as follows:

[0134]

[0135] in, The feature value to be detected; For isolated forest anomaly scores; This is the threshold for abnormal scores (range [0.3, 0.5], scores higher than this threshold are considered abnormal); when When necessary, interpolation is used for correction: ( , (These are the normal characteristic values ​​before and after the outlier).

[0136] Key feature selection: Features strongly relevant to the task are selected based on the mutual information criterion. The mutual information between the feature and the task is as follows:

[0137]

[0138] in, Features The marginal probability; For task type The marginal probability; Features With task type Joint probability; before selecting mutual information value ( The key feature set consists of the features of the key feature set. .

[0139] Feature Dimensionality Reduction: PCA is used to reduce the dimensionality of the key feature set.

[0140]

[0141]

[0142]

[0143] in, The covariance matrix of the key feature set; The number of feature samples; The mean vector of the key feature set; For eigenvalue decomposition operations ( For the eigenvalue matrix, (e.g., eigenvector matrix) For the front The eigenvector matrix corresponding to the largest eigenvalues ​​( ); This represents the feature representation after dimensionality reduction.

[0144] S1.6 Closed-loop feedback calibration: Edge computing nodes will perform preprocessed feature calibration. The system transmits key data synchronously to the core scheduling node, and simultaneously receives feedback results from subsequent task matching and path optimization stages, updating model parameters and neural symbol map association weights to form a closed loop.

[0145]

[0146]

[0147]

[0148] in, , These are the parameters of the UAV state noise prediction network before and after the update; , These are the network parameters for predicting task requirements and noise, before and after the update. The learning rate (range [0.001, 0.01]). For feedback loss function; , These are reference values ​​for the characteristics and needs reported in subsequent stages; The step size for updating the map weights (value range [0.01, 0.1]). The score is given for the collaborative effect between drones (the value range is [0,1]).

[0149] The preprocessed features and neural symbol maps output by edge computing nodes provide highly reliable and low-redundancy input data for subsequent quantum-enhanced distributed causal matching, achieving synergistic optimization of modeling accuracy, communication efficiency, and real-time performance.

[0150] S2, Quantum-Enhanced Distributed Causal Matching Module: Based on the UAV Neural Symbolic Ability Map Output of S1 Neural Symbol Demand Mapping for Tasks Through a collaborative process encompassing "causal logic pruning, quantum feature enhancement, non-Hermitian quantum encoding, distributed singularity conflict resolution, lightweight quantum secondary optimization, and closed-loop feedback calibration," this system achieves efficient and accurate matching between heterogeneous UAVs and missions. It balances the flexibility of distributed architecture, the engineering feasibility of quantum computing, and the reliability of causal inference, addressing the slow convergence and insufficient global optimality issues of traditional matching algorithms. The quantum enhancement algorithms are compatible with three engineering operation modes: "classical platform quantum circuit simulation," "small-scale commercial quantum hardware deployment," and "quantum-inspired classical algorithm implementation." They are also compatible with existing commercial equipment and open-source toolchains, enabling deployment and execution under current engineering conditions.

[0151] Let the set of UAV-task pairs initially screened by the digital twin model be . ,in:

[0152] For heterogeneous drone swarms ( (Number of drones) For the set of tasks to be executed ( (Number of tasks) For the first A drone, For the first One task;

[0153] Causal conflict determination function Used to characterize drones With the task Deep-seated causal conflicts (such as the irreversible conflict between performance degradation trends and task duration, and the inherent incompatibility between environmental adaptability and task environment).

[0154] Set of compatible matching pairs This is a set of conflict-free drone-mission pairs after causal pruning.

[0155] drones The high-dimensional classical eigenvectors are ( (The classic feature dimension is taken from the output of the data-driven module of the digital twin model). (Task) The high-dimensional classical eigenvectors are ( (The classic feature dimensions of the task are taken from the output of the data-driven module of the digital twin model).

[0156] Quantum feature space dimension (Adapting to the computational constraints of existing classical simulations and small-scale quantum hardware), the non-Hermitian quantum state space is (dimensions and) Consistent);

[0157] Adjacent UAV sets within a distributed communication domain ,in For drones and Real-time physical distance, Effective communication distance;

[0158] Pruning Calculator This is used to remove causally incompatible matching pairs.

[0159] Lightweight Quantum Random Forest (LQRF) includes A quantum decision tree (controlling computational complexity), its parameter set is: ( );

[0160] Quantum operating mode identifier (0 represents classical platform quantum circuit simulation, 1 represents small-scale commercial quantum hardware deployment, and 2 represents quantum-inspired classical algorithm implementation).

[0161] Classical simulation accuracy threshold (Ensure that the simulation results and the operational errors of the quantum hardware are within an acceptable range for engineering purposes).

[0162] Non-Hermitian coding matrix (The diagonal elements are real numbers, and the off-diagonal elements are complex numbers, satisfying...) , for Hermetic conjugation);

[0163] Conflict Residual Function This is used to determine whether the conflict has been resolved.

[0164] Number of quantum circuit iterations in the Lightweight Quantum Approximation Optimization Algorithm (LQAOA) (Reduce hardware and simulation overhead), parameter set is , ;

[0165] Matching matrix ,in Indicates drone Assigned to task , Indicates unallocated;

[0166] Feedback calibration related parameters: weight update step size Learning rate Cost weighting coefficient .

[0167] S2.1 Causal Logic Pruning: Using first-order predicate logic rules (compound predicates) of digital twin models Based on deep causal relationship analysis, and through pruning operators... For the initial set of drone-mission pairs Conflict pruning significantly reduces the number of invalid matching pairs, thereby lowering the computational overhead of subsequent quantum computing. The pruning process satisfies the following:

[0168]

[0169] in: For logical NOT operation, The compound predicate defined in Chapter (I) is used to characterize unmanned aerial vehicles (UAVs). Capable of performing tasks Comprehensive abilities; This is a causal conflict determination function. A value of 1 indicates the existence of a deep causal conflict, and a value of 0 indicates no conflict. Its specific calculation formula is as follows:

[0170]

[0171] in, The number of causal conflict types (such as energy consumption-duration conflict, load-weight conflict, performance degradation-operational accuracy conflict, etc.). For the first The weight coefficient for class conflict (values ​​range [0,1], and satisfy) ); For the first The degree of class conflict quantification (value range [0,1], calculated by the causal association edge strength of the neural symbolic graph); The conflict determination threshold (value range [0.3, 0.6]); This is a symbolic function; it outputs 1 if the input is greater than 0, and 0 otherwise.

[0172] S2.2 Quantum Feature Extraction and Encoding: For compatible matching pairs, high-dimensional quantum features are extracted using a lightweight quantum random forest, and non-Hermitian encoding matrices are used to encode the features into quantum states or classical equivalent vectors.

[0173] Lightweight Quantum Random Forest Feature Extraction: Targeting Compatible Matching Pairs First, a joint classical feature vector is obtained through feature concatenation:

[0174]

[0175] in Used to extract high-dimensional features of drones High-dimensional features of the task Structurally concatenated Dimensional vector; then according to the quantum operating mode By selecting the appropriate implementation method, deep extraction is performed on the joint classical features to generate high-dimensional quantum features or quantum-inspired classical features. The specific implementation is as follows:

[0176] when (Classical Platform Quantum Circuit Simulation): A lightweight quantum random forest is implemented using open-source quantum simulation libraries such as Qiskit and Cirq. Each quantum decision tree encodes features using 1-2 qubits. Single / double quantum gates such as H-gates and CNOT gates are used to implement branch decisions. Redundant quantum gate circuits are removed, and the quantum circuit depth of a single decision tree is controlled to not exceed 8 layers. The feature extraction output is as follows:

[0177]

[0178] when (Small-scale commercial quantum hardware deployment): Adapted to commercial 5-20 qubit processors such as IBM Quantum Experience and Alibaba Cloud quantum computing platform, the quantum circuits are compiled into an adapted instruction set through a hardware interface, enabling real-time reading of quantum measurement results and reconstruction of quantum characteristics. .

[0179] when (Quantum-inspired classical algorithm implementation): Employing a classical approximation model of quantum decision trees, the system simulates quantum superposition properties through stochastic Fourier transform to generate high-dimensional classical features equivalent to quantum features. To ensure engineering availability in the absence of quantum hardware.

[0180] Parameter set of lightweight quantum random forest (Parameters such as quantum gate rotation angle and measurement basis selection) are obtained through optimization using the quantum gradient descent algorithm (quantum operation mode 0, 1) or the classical approximate optimization algorithm (quantum operation mode 2), with the unified optimization objective being to minimize the feature discrimination loss.

[0181]

[0182] in: For quantum features or quantum-inspired classical features, the projection function onto the classical feature space; It is the square of the L2 norm.

[0183] The classic simulation mode needs to meet the following requirements. Quantum hardware deployment and quantum heuristics must meet the following requirements .

[0184] Non-Hermitian quantum state encoding: based on quantum operating modes Adapting the encoding logic, the extracted high-dimensional features are encoded into non-Hermitian quantum states or classical equivalent encoding vectors, improving encoding efficiency and feature distinguishability while reducing resource consumption.

[0185] when or At that time, the encoding process satisfies:

[0186]

[0187] in, For drones - mission pairs The corresponding non-Hermitian quantum state vector; non-Hermitian encoding matrix The optimal solution is obtained by maximizing the distinguishability between quantum states:

[0188]

[0189] in, Quantum state vector The corresponding right arrow, For the left arrow, For trace operations.

[0190] In classical simulations, non-Hermitian quantum states are simulated through complex matrix operations, while quantum hardware deployments store quantum state information through quantum registers.

[0191] when At that time, classical complex matrix encoding is used to simulate the characteristics of non-Hermitian quantum states, generating classical equivalent encoding vectors. Its calculation process is mathematically equivalent to quantum coding, and it is implemented only through classical computing power, ensuring the consistency of matching logic under different operating modes.

[0192] S2.3 Distributed Singularity Conflict Resolution: Neighboring UAVs exchange encoded feature information (quantum state measurement results or classical equivalent vectors) via commercial low-power communication protocols such as LoRaWAN and NB-IoT. The conflict resolution logic is adapted according to the quantum operating mode, utilizing the singularity evolution characteristics of non-Hermitian systems (or classical equivalent simulation) to achieve self-driven conflict resolution without the need for manually preset adjustment coefficients.

[0193] Define the Hamiltonian (or classical equivalent Hamiltonian) of a non-Hermitian system:

[0194]

[0195] in, For drones and The communication weight coefficient (with values ​​ranging from [0,1], and positively correlated with the communication signal strength) is used. The classical equivalent Hamiltonian approximates the non-Hermitian properties through a real matrix, ensuring consistent evolutionary behavior.

[0196] The evolution of non-Hermitian quantum states (or classical equivalent vectors) follows the non-Hermitian extension of the Schrödinger equation (or the classical dynamical equation):

[0197]

[0198] in, To reduce Planck's constant, For evolution time, For a moment The non-Hermitian quantum state vector (or classical equivalent vector); in classical simulation, the equation is solved by numerical integration; in quantum hardware deployment, it is realized through quantum circuit evolution; and in quantum heuristic mode, it is simulated by a classical differential equation solver.

[0199] When the system evolves to a singularity When a quantum state (or classical equivalent vector) undergoes a phase transition, the conflict-related component is suppressed, while the conflict-free component is enhanced. The conflict resolution criterion is:

[0200]

[0201] in: For conflict residue; To perform the real part operation; It is the inner product of two quantum states (or classical equivalent vectors) at the singular point time; Let be the conflict residual threshold (range [0, 0.2]). When the conflict residual is less than this threshold, the conflict is considered resolved. The set of compatible matching pairs after conflict resolution is denoted as . .

[0202] S2.4 Lightweight Quantum Approximation Optimization: For the set of compatible matching pairs after conflict resolution First, construct an optimization objective function that integrates matching adaptability, resource utilization, and energy consumption balance:

[0203]

[0204] in: The weight coefficients of the objective function (within the range [0,1], and satisfying...) ); For encoded features and optimal matching features The first term is the similarity (value range [0,1]); the second term is the UAV resource utilization constraint; and the third term is the task allocation uniqueness constraint.

[0205] Based on the quantum operating mode Choose the appropriate optimization implementation method:

[0206] when or At that time, a lightweight quantum circuit structure with "hybrid layer + problem layer" (the number of quantum gates does not exceed 50) is adopted, and the quantum circuit expression is as follows:

[0207]

[0208] in: For the unitary operator at the problem level, the corresponding energy term of the optimization objective function is represented. For mixed-layer unitary operators, It is a revolving door around the Y-axis.

[0209] Optimize quantum circuit parameters using the Nelder-Mead classical optimization algorithm:

[0210]

[0211] in, It is the initial quantum state (uniform superposition state). Its corresponding left arrow; for Hermitian conjugate; The objective function constructed above is used for optimization. In classical simulation, the parameters are solved directly through optimization algorithms, while in quantum hardware deployment, the parameters are iteratively updated through a classical-quantum hybrid optimization framework.

[0212] when At that time, the classical equivalent implementation of the quantum approximation optimization algorithm is adopted. By combining the greedy algorithm with the classical simulation of the quantum superposition principle, the matching matrix is ​​obtained by directly solving the optimization objective function.

[0213] Finally, the optimal matching matrix is ​​obtained through quantum measurement (mode 0, 1) or classical solution (mode 2). Its expression is:

[0214]

[0215] in For quantum measurement operations (quantum operation mode), in classical / heuristic modes, the allocation scheme with the highest probability among the optimization results is directly selected. .

[0216] S2.5, Closed-loop feedback calibration: The optimal matching matrix... Feedback is fed into the generative neural symbol digital twin model. Simultaneously, based on the actual operating effect of the quantum operating mode (including simulation accuracy, hardware measurement error, and heuristic algorithm efficiency), the parameters of the quantum algorithm and the digital twin model are dynamically adjusted, forming an engineering closed loop of "digital twin modeling - matching optimization - feedback calibration - parameter adjustment".

[0217]

[0218]

[0219]

[0220] in: , The drones before and after the update With the task Graph association weights; This is the parameter set for the digital twin model (including parameters for the generative diffusion model and data-driven modules). , These are the parameters before and after the update, respectively. This is the gradient operator with respect to the model parameters; To calibrate the loss function, The matching matrix predicted by the digital twin model; The engineering costs (including hardware usage costs, simulation time costs, and algorithm deployment complexity) for different operating modes. To ensure the selection of the engineering implementation path with the optimal "accuracy-cost" ratio for the dynamically optimized quantum operation mode.

[0221] The optimal matching matrix output in this step To enhance learning and provide accurate task allocation results for hierarchical game optimization in dual-system learning, all quantum-related modules have clear engineering implementation paths, are compatible with existing commercial quantum hardware, classical simulation tools and open-source algorithm libraries, ensuring the feasibility and practicality of the technology.

[0222] S3. Reinforcement Learning-Enhanced Dual-System Hierarchical Game Optimization Module: Constructs a "rapid response-global optimization" dual-system hierarchical game framework, integrating the dynamic decision-making capabilities of reinforcement learning with the global equilibrium characteristics of game theory to achieve five-dimensional fully coupled optimization of task, path, environment, energy consumption, and carbon footprint. Through a dual-layer decision-making mechanism inspired by the basal ganglia and prefrontal cortex, it balances rapid response to environmental changes with long-term balance of global benefits, outputting a structured decision rule chain. Furthermore, through a feedback-calibrated back-connected generative neural symbolic digital twin model, it forms a closed loop of "matching result input - dual-system optimization - decision output - feedback calibration," addressing the technical shortcomings of traditional optimization algorithms, such as poor globality, insufficient interpretability, and weak dynamic adaptability.

[0223] Let: System state space ;

[0224] in: Environmental status space (including wind speed, wind direction, visibility, location of temporary no-fly zones, etc.) Class parameters); State space for drone swarms (including real-time energy consumption, location, payload, performance degradation, etc.) Class parameters); For the task state space (including completion progress, remaining time window, dynamic priority, etc.) (Class parameters).

[0225] Let: Action space ;

[0226] in: This is a set of rapid-layer actions (including real-time response actions such as path fine-tuning, speed adjustment, and temporary obstacle avoidance, totaling...) (one action); For the slow layer action set (including global optimization actions such as task reassignment, global path planning, and energy-balanced scheduling, a total of...) (One action).

[0227] Payoff function: Local instantaneous payoff function (Representing the immediate benefit of a single action); Global long-term benefit function (characterizing time window) (cumulative global benefit within).

[0228] Cost function: Real-time energy consumption cost function (Characterizing the energy consumption of the i-th UAV in performing its actions); Carbon footprint function (Characterizing the carbon emission intensity of the i-th drone's actions).

[0229] Dynamic coefficient: Time decay coefficient (The value range is [0,1], increasing as the task progresses, satisfying...) , The initial attenuation coefficient, (Total task duration); Scene adaptation coefficient (Value range [0,1], [0.8,1.0] for emergency rescue scenarios, and [0.2,0.5] for agricultural plant protection scenarios).

[0230] Game theory related: Set of game participants ( (Number of drones, consistent with the drone swarm set mentioned earlier); participant strategy set. ( For drones (Decision-making strategy); payment function (Characterizing drones) The game payout).

[0231] Algorithm parameters:

[0232] Rapid Response Layer: Decision Cycle Discount factor Network parameters Experience replay pool ;

[0233] Global optimization layer: decision cycle ( ), discount factor Strategy parameters Value network parameters KL divergence threshold ;

[0234] Feedback calibration: Weight update step size Learning rate Feedback loss function .

[0235] S3.1 Dual-system decision-making architecture: Based on the functional division of the basal ganglia and prefrontal cortex, a "rapid response-global optimization" collaborative decision-making mechanism is constructed to adapt to optimization needs at different time scales, and the two-layer decision-making achieves collaboration through state synchronization.

[0236] Rapid Response Layer (Basal Node Inspiration): Targeting millisecond-level environmental changes (sudden strong winds, temporary no-fly zones), with the goal of "rapid adjustment and immediate loss mitigation," the decision-making cycle... Output real-time actions And synchronize the state after the action is executed to the global optimization layer.

[0237] Global Optimization Layer (Prefrontal Heuristic): Targeting global objectives (task progress, energy consumption balance) at the second-to-minute level, with the goal of "long-term balance and global optimum," the decision cycle... Based on the state of the rapidly updated layer and its own optimization goals, a global action is output. It also provides motion constraints (such as maximum speed adjustment range) for the fast layer.

[0238] S3.2 Fast Response Layer Training: A lightweight deep Q-network (LDQN) is used to achieve a fast response to environmental changes, requiring only the input of key states related to the environmental change, thus reducing computational complexity.

[0239] Fast layer state vector This includes the location of sudden obstacles, the magnitude of sudden changes in wind speed, and the extent of temporary no-fly zones. Class key parameters;

[0240] LDQN architecture: It adopts the "convolutional feature extraction - bottleneck layer compression - fully connected output" approach. It replaces standard convolutions with depthwise separable convolutions and removes redundant fully connected layers. The Q-value output is:

[0241]

[0242] in, For state standardization operations, For depthwise separable convolution, (Same as the number of actions in the fast layer).

[0243] Training mechanism: Target network and experience replay, experience replay pool Stored samples The training loss function is:

[0244]

[0245] in, For the target network parameters (from every 100 iterations) synchronous).

[0246] Instant reward calculation:

[0247]

[0248] in, ( (for obstacle avoidance weights) A reward is given for successful obstacle avoidance (1 point for success, 0 points for failure). This represents the maximum possible energy consumption of the action. The time delay caused by the action. This represents the remaining time window for the task.

[0249] S3.3 Global Optimization Layer Training: Achieving long-term balance of global benefits based on Trust Region Policy Optimization (TRPO), with the complete system state as input:

[0250] Slow layer state vector It includes full-dimensional state parameters of the environment, drones, and missions.

[0251] Policy function: Output the probability distribution of actions and update the stability using the KL divergence constraint strategy.

[0252] Surrogate objective function (measures the magnitude of policy improvement):

[0253]

[0254] in, The strategy parameters before the update. This is the dominance function for the slow layer.

[0255] Dominance function calculation:

[0256]

[0257]

[0258]

[0259] in, State-value function (via value network) Fitting).

[0260] Strategy update constraints (to avoid profit fluctuations):

[0261]

[0262] in, Update the KL divergence constraint threshold for the strategy;

[0263] The optimal policy parameters are obtained by solving the constrained optimization problem using the conjugate gradient method. .

[0264] S3.4 Construction of a Multi-Dimensional Layered Payment Function: Integrating two layers of revenue, energy consumption cost, and carbon footprint as the core principle of game theory optimization, adapting to different operational scenarios:

[0265]

[0266] in: , Penalty coefficients for energy consumption and carbon footprint (increased for green work scenarios) ).

[0267] Localized immediate benefits:

[0268]

[0269] in, For drones Task progress increment, For drones Dynamic priority allocation for tasks.

[0270] Global long-term benefits:

[0271]

[0272] in, To predict the time window, For a specific moment within the time window, For the kth drone in task progress increment at any time

[0273] The second item is the energy consumption balance reward, which aims to suppress excessive differences in drone energy consumption.

[0274] S3.5, Hierarchical Game Optimization: Using drone swarms as game participants, Nash equilibrium is achieved through iterative "strategy evaluation - strategy improvement" to ensure global optimality.

[0275] Nash equilibrium solution definition: strategy combination Satisfying any participant With strategy :

[0276]

[0277] in, The game termination time distribution (task completion or iteration count reached) ).

[0278] Strategy iteration process:

[0279] a. Strategy Evaluation: Based on the current strategy portfolio The expected payoffs of each participant were estimated using Monte Carlo Tree Search (MCTS). Update value network parameters .

[0280] b. Strategy Improvement:

[0281] Fast layer: adopts - Greedy strategy ( (Updated linearly from 0.9 to 0.1) ;

[0282] Slow layer: Updated via TRPO's trusted region constraints .

[0283] c. Termination condition: When the expected iterative fluctuation range of the payment value... When the Nash equilibrium is reached, the optimal strategy combination is output. .

[0284] S3.6, Structured Rule Chain Extraction and Feedback: Extracting and feeding back the optimization results of the dual system (optimal strategy). Structured rule chain Action commands Feedback is sent to the generative neural symbol digital twin model to update the model parameters and the association weights between the neural symbol atlas, forming a closed-loop technology:

[0285]

[0286]

[0287] in: , These are the dynamic correlation weight matrices before and after the update, respectively; , These are the parameter sets for the digital twin model (including parameters for the generative diffusion model and the data-driven module). This is the gradient operator for the dynamic correlation weight matrix; Gradient operator for parameter set of digital twin model The optimal strategy combination predicted by the digital twin model; For feedback loss function; To provide feedback for calibrating the learning rate; Update the step size for dynamically associated weights.

[0288] Synchronous updates: Structured rule chain Injecting a rule-based reasoning module into the digital twin model enhances the dynamic adaptability of subsequent modeling.

[0289] The optimal action command output in this stage Structured rule chain It provides a decision-making basis for multimodal fusion, heterogeneous adaptation, and obstacle avoidance. The feedback calibration mechanism ensures real-time collaboration between dual-system optimization and digital twin modeling, thereby improving the dynamic adaptability and global optimization effect of the entire technical solution.

[0290] S4. Multimodal Fusion Heterogeneous Adaptation Obstacle Avoidance Module: Constructs a full-process obstacle avoidance system encompassing "multimodal topology modeling - federated dedicated path generation - dynamic energy consumption coupling - causal-driven dynamic obstacle avoidance". It enhances environmental perception accuracy through multimodal feature fusion, adapts to the flight characteristics of heterogeneous UAVs through federated learning, optimizes path efficiency by combining real-time energy consumption constraints, achieves precise obstacle avoidance based on causal relationships, and triggers dynamic iterative updates when the environment changes abruptly, forming a closed loop of "perception-modeling-generation-optimization-obstacle avoidance-iteration". This addresses the technical shortcomings of traditional path planning, such as false correlation misjudgment, poor heterogeneous adaptability, high obstacle avoidance energy consumption, and insufficient safety.

[0291] Let: a heterogeneous drone swarm set ;

[0292] in: For rotorcraft assembly, For fixed-wing aircraft assembly, For compound wing aircraft assembly, A collection of vertical takeoff and landing fixed-wing aircraft; a single unmanned aerial vehicle (UAV) ( , (Total number of drones).

[0293] Let: Multimodal feature set ;

[0294] in: Visual features (pixel-level environmental semantic information) Features of LiDAR (3D point cloud distance information); Meteorological characteristics (wind speed, wind direction, visibility) Location features (latitude, longitude, altitude).

[0295] Let: Topological graph ;

[0296] in: It is a set of nodes (including UAV location nodes, obstacle nodes, and task nodes). This represents the total number of nodes. It is a set of edges (representing the relationships between nodes). For a set of tags (including time slice tags) Causal label ).

[0297] Federated learning related: Client collection Edge aggregation node Local model parameters ( ), global model parameters Single UAV personalized model parameters .

[0298] UAV status and path parameters: Real-time energy consumption status Path parameter set ;in, For path length, For the number of turns, This refers to the real-time flight speed.

[0299] Obstacle avoidance related parameters: causal correlation prediction model LSTM model parameters Model Predictive Control (MPC): Prediction Time Domain Control time domain ; Safety distance threshold Fill in the safety distance Causal correlation of drone ensemble .

[0300] S4.1 Multimodal Topology Graph Construction: Integrating multimodal environmental perception data, causal verification, and attention mechanisms, a high-precision topology graph is constructed, filtering out false related edges, and providing a reliable environmental model for path planning and obstacle avoidance.

[0301] Node feature vector definition: Topology graph nodes ( The feature vector is obtained by concatenating and mapping multimodal features:

[0302]

[0303] in, , nodes Corresponding features, For feature dimension, It is a 3-layer fully connected lightweight chemical neural network.

[0304] Granger causality test: To determine the effective causal relationship between nodes, the test statistic is:

[0305]

[0306] in, The sum of squared residuals under the assumption of no causality. This is the sum of squared residuals under the assumption of causality. The length of the time series data. The lag order; when ( When the threshold for causality testing is reached, a valid causal relationship is determined to exist.

[0307] Edge weight calculation: Combining attention mechanisms to strengthen effective associations and suppress spurious associations.

[0308]

[0309] in, This is the attention weight matrix. When the edge is a pseudo-related edge, it should be filtered out. The threshold for Granger causality testing.

[0310] Topological graph adjacency matrix: defined as ,satisfy:

[0311] .

[0312] S4.2 Federated Learning Dedicated Path Generation: While protecting local data privacy, train meta-path generators adapted to the flight characteristics of different types of heterogeneous UAVs, and generate path clusters online in real time.

[0313] Federated Training Mode: Employs a three-tiered model of "local training - global aggregation - personalized fine-tuning," with the local training objective being to minimize path adaptation loss.

[0314]

[0315] in, For the client, To generate the path, To adapt to the ideal path of this type of drone (rotorcraft focus on vertical obstacle avoidance, while fixed-wing aircraft focus on smooth flight paths). , The loss weight coefficient is used to generate the estimated energy consumption of the path.

[0316] Global model aggregation: After the local model training is completed, the parameters are uploaded to the edge aggregation node, and the global model is updated using a weighted federated average.

[0317]

[0318] in, For the client The number of drones in China The total number of drones, These are the local model parameters for client c.

[0319] Personalized fine-tuning: After the global model is deployed, it is adapted to the characteristics of a single drone.

[0320]

[0321] in, To fine-tune the learning rate, For individual drone-specific losses.

[0322] Online path generation: The meta-path generator is based on the current topology graph. Output path cluster ( (where is the number of path candidates), the path generation probability distribution is:

[0323]

[0324] in, This is a comprehensive evaluation function for the path (integrating adaptability, energy consumption, and safety).

[0325] S4.3 Real-time Energy Consumption Model Coupling and Path Optimization: Correlate path parameters with the real-time energy consumption status of the UAV to optimize path energy efficiency and ensure that endurance constraints are met.

[0326] Real-time energy consumption model: Integrating path length, number of turns, flight speed, and real-time energy consumption status to construct an energy consumption model:

[0327]

[0328] in, For execution path The remaining energy consumption after that, Energy consumption coefficient per unit length The energy consumption coefficient for a single turn. The speed-related energy consumption coefficient, For path Flight duration.

[0329] Path optimization objective: Minimize path cost while satisfying energy consumption constraints.

[0330]

[0331] In the formula,

[0332] in, , , To optimize the weighting coefficients, To reserve energy for emergency use.

[0333] S4.4 Causal-Driven Dynamic Obstacle Avoidance Adjustment: Based on the causal relationship edges of the topology graph, the path requirements of adjacent UAVs are predicted, and obstacle avoidance adjustment is performed only for causally related UAVs. Dynamic obstacle avoidance is achieved in conjunction with MPC.

[0334] Neighbor Path Prediction: Predicting the future path trajectory of causally related drones using an LSTM model.

[0335]

[0336] in, For drones future The predicted path of the step. For the length of the historical trajectory, These are the parameters for the LSTM model.

[0337] MPC optimization objective: Minimize obstacle avoidance adjustment costs and trajectory deviation.

[0338]

[0339] In the formula,

[0340]

[0341] in, For path adjustment amount, The minimum distance between paths. , To optimize weights, For speed adjustment amount, This is the maximum flight speed of the drone.

[0342] Actual path correction: After solving the MPC optimization problem, the obstacle avoidance adjustment amount is obtained. The corrected actual flight path is as follows:

[0343] .

[0344] S4.5 Dynamic Iterative Closed Loop: When the environmental state changes abruptly (such as sudden obstacles or sudden changes in wind speed), the entire process of "multimodal topology update → path cluster generation → energy consumption coupling → obstacle avoidance adjustment" is re-triggered to ensure the dynamic adaptation of the obstacle avoidance strategy. The iteration trigger condition is: ① The change in multimodal features exceeds a threshold. ② The change in the weight of the causal relationship edge exceeds the threshold. ③ The obstacle avoidance adjustment exceeds the maximum threshold. .

[0345] The optimal obstacle avoidance path output in this step It provides real-time flight commands for drones, and combines heterogeneous adaptation characteristics with energy consumption constraints to reduce energy consumption while ensuring obstacle avoidance safety, and provides real-time operational status data support for the evolution of time-series perception active alliances.

[0346] S5, Time-Sequence Aware Active Alliance Evolution Module: Constructs a full-process active alliance mechanism encompassing "fault timing prediction - multi-dimensional trust modeling - proactive replacement scheduling - alliance dynamic evolution - feedback calibration". It achieves early fault warning through a dynamic causal fault tree, quantifies UAV collaborative reliability based on a multi-dimensional causal trust model, and ensures seamless replacement by combining digital twins and a meta-path generator. Ultimately, it achieves adaptive updates to the alliance structure and feeds back to the digital twin model to form a closed loop, addressing the technical shortcomings of traditional fault handling such as passive lag, poor replacement adaptability, and insufficient alliance stability.

[0347] Let: a heterogeneous drone swarm set ( (This refers to the total number of drones, consistent with the cluster size mentioned earlier).

[0348] single drone ( Fault precursor feature set ( (The number of pre-fault warning features), specifically including:

[0349] Motor vibration amplitude Battery voltage fluctuations Motor temperature Flight attitude fluctuation .

[0350] Dynamic Causal Fault Tree (DCFT): Node Set Where Top represents the top event "drone malfunction", Inter represents intermediate events, and Basic represents basic events. Minimum number of cut sets. , No. Minimum cut set Cut set weight coefficient (satisfy ).

[0351] Fault-related parameters: Fault prediction time window Failure probability (Value range [0,1]); Fault warning threshold Basic event trigger probability .

[0352] Multidimensional Causal Trust Model: Dimension Set , respectively, correspond to the degree of cocausal compatibility Fault probability correction Quality of completion of historical tasks Energy consumption stability Trust score Dimension weight coefficient (satisfy ).

[0353] Replacement-related parameters: Set of candidate drones for replacement Optimal replacement drone Trust threshold Upper limit of candidate drone failure probability ; Optimization objective function for fill-in path Fill in the safety distance .

[0354] Alliances and Update Parameters: Alliance Structure Adjacency Matrix ( (Indicates alliance); Trust update step size Weight update step size Trust prediction loss function Correction amount for fill effect .

[0355] Model and Feedback Parameters: LSTM Model Parameters (Specifically for extracting early warning features of faults), Logistic regression model parameters Feedback calibration learning rate Graph weight update step size .

[0356] S5.1 Fault Timing Prediction and Early Warning: Integrating fault precursor features from neural symbolic ability maps, timing data processing, and symbolic rules, to achieve accurate early fault prediction:

[0357] Deep temporal feature extraction: Feature mining of fault precursor time series data using LSTM:

[0358]

[0359] in, For a moment The feature vector of the precursor to the fault, The sampling length of the time series data. The sampling interval is denoted as .

[0360] Basic event triggering rules: Triggering conditions are defined based on symbol rules, such as basic events. The rules for (abnormal motor vibration) are as follows: ( (Vibration threshold).

[0361] Failure probability calculation: The probability of the top event occurring is obtained by combining the minimum cut set and the trigger probability of the basic event with weighting coefficients.

[0362]

[0363] in, From temporal deep features Output from the logistic regression model ( ).

[0364] Fault warning triggered: When At that time, the drone was determined Entering fault warning state triggers the active compensation process.

[0365] S5.2, Construction of a Multidimensional Trust Model: Integrating collaborative characteristics, fault states, historical performance, and energy consumption stability to quantify the trustworthiness of inter-UAV replacement:

[0366] Co-causal compatibility: Based on the strength of association edges in neural symbolic graphs and causal compatibility.

[0367]

[0368] in, For drones capability feature vector For cosine similarity, It is a causal compatible predicate (truth value 0 or 1).

[0369] Historical task completion quality: weighted calculation based on historical task data.

[0370]

[0371] in, For the number of historical missions, , , The first Normalized values ​​for the accuracy, timeliness, and energy consumption control of each task completion. .

[0372] Energy consumption stability: Calculated based on historical energy consumption fluctuation variance.

[0373]

[0374] in, The variance of historical energy consumption fluctuations. (To avoid the denominator being zero).

[0375] Multi-dimensional trust level comprehensive calculation:

[0376] .

[0377] S5.3 Rapid Selection of Candidate UAVs for Replacement: Edge computing nodes efficiently select suitable candidate replacement sets based on fault warning and trust models.

[0378]

[0379] in, For drones The task in progress Compound predicates (representations) defined for the chapter on generative neural symbolic digital twin models Capable of execution (ability).

[0380] Candidate set expansion mechanism: If Lower the trust threshold And introduce redundant drone ensembles :

[0381] .

[0382] S5.4, Optimization of Completion Path and Task Handover: Based on the generative neural symbolic digital twin model and meta-path generator, seamless completion is achieved:

[0383] Complementary Scene Mapping: Building Maps in Digital Twin Models ,in , , These are digital mappings of drones, missions, and environments, respectively.

[0384] Task handover parameter output: Simulates the handover process and outputs the optimal handover time. Data to be transferred Task progress status .

[0385] Path completion optimization: Metapath generator is Generate path clusters The optimization objective is:

[0386]

[0387] In the formula,

[0388] in, To fill in the path length, For path energy consumption, For flight duration, .

[0389] Dynamic path correction: Environmental disturbances and schedule deviations are compensated for through MPC, resulting in the optimal replacement path after correction.

[0390]

[0391] in, Real-time path adjustment (based on MPC prediction time domain) With control time domain (Solve)

[0392] S5.5 Dynamic Evolution of Alliance Structure: After the replacement is completed, the alliance structure and trust model are updated to achieve adaptive adjustment.

[0393] Alliance adjacency matrix update:

[0394]

[0395] in, To update the alliance adjacency matrix, add alliance associations between the replacement drone and the original drone.

[0396] Trust model update:

[0397]

[0398]

[0399] in, The correction amount for the fill-in effect (calculated based on the smoothness of task connection and energy consumption control level). This represents the gradient of the loss function with respect to the dimension weights.

[0400] Drone status update: If Once the fault is fixed, the member will be re-added to the available set and its initial trust level will be reset. If the fault cannot be fixed immediately, the member will be removed from the alliance and marked as "pending maintenance".

[0401] S5.6 Closed-loop feedback calibration: This involves adjusting the coalition evolution results (updated coalition structure) The trust model parameters and supplementary execution data are fed back to the generative neural symbol digital twin model to update the model parameters and the association weights between the neural symbol atlas.

[0402]

[0403]

[0404] in, Associating drone capabilities with weights For fault prediction parameters in digital twin models, ( (To predict the failure probability of the model), forming a closed loop of "failure early warning - supplementary optimization - alliance update - model calibration".

[0405] This step upgrades fault handling from passive backup to proactive intervention through proactive early warning and dynamic alliance evolution, ensuring the continuity of cluster operations; the closed-loop feedback mechanism improves the fault prediction accuracy and trust model adaptability of the digital twin model.

[0406] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs, characterized in that, Includes the following steps: S1. Generative Neural Symbol Digital Twin Modeling: Construct and dynamically update a neural symbol map representing the full-dimensional state and multi-dimensional mission requirements of the UAV. Through an adaptive causal incremental synchronization mechanism and edge computing hierarchical preprocessing, provide highly reliable and low-redundancy digital mapping inputs for subsequent modules. S2, Quantum-enhanced distributed causal matching: Based on the graph output in step S1, through causal logic pruning, quantum feature extraction and encoding, distributed singularity conflict resolution and lightweight quantum approximation optimization, the UAV and the mission are accurately matched, and the optimal matching matrix is ​​output. S3. Reinforcement learning enhances dual-system hierarchical game optimization: Based on the output of step S2, a dual-system hierarchical game framework of "fast response-global optimization" is constructed. By integrating reinforcement learning and game theory, the five-dimensional fully coupled optimization of task-path-environment-energy consumption-carbon footprint is achieved, and the structured decision rule chain and optimal action instructions are output. S4. Multimodal Fusion Heterogeneous Adaptive Obstacle Avoidance: Based on the output of step S3 and environmental perception data, through multimodal topology modeling, federated learning-specific path generation, real-time energy consumption model coupling, and causal-driven dynamic obstacle avoidance, a safe and efficient flight path is generated and optimized in real time for the UAV. S5. Time-series-aware proactive alliance evolution: Fault warning is achieved based on dynamic causal fault tree, supplementary drones are selected through multi-dimensional trust model, and seamless supplementation and dynamic updating of alliance structure are achieved by combining digital twin and path generation. Steps S1 to S5 are executed sequentially, and a closed-loop optimization system is formed through multi-layer feedback calibration.

2. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 1, characterized in that, S1 includes: S1.1 Dynamic temporal prediction: A lightweight generative diffusion model is adopted to predict the future state and mission requirements of UAVs based on historical time series data; S1.2 Data-driven feature extraction: Lightweight neural networks are used to perform nonlinear transformations on real-time state features, predicted state features, and task requirement features to generate high-dimensional discriminative features; S1.3 First-order predicate logic reasoning: Establish logical association rules between UAV status and mission requirements through atomic predicates and compound predicates; S1.4, Neural Symbol Atlas Construction: Integrate prediction, feature and rule information to construct a heterogeneous UAV neural symbol capability atlas and a task neural symbol requirement atlas; S1.5 Adaptive Causal Incremental Synchronization and Edge Preprocessing: Based on communication load and energy consumption status, the synchronization threshold is dynamically adjusted to trigger key data synchronization, and outlier detection, key feature selection and dimensionality reduction preprocessing are performed at edge nodes. S1.6 Closed-loop feedback calibration: Receive feedback results from subsequent modules and update model parameters and map association weights.

3. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 1, characterized in that, S2 includes: S2.1 Causal logic pruning: Based on first-order predicate logic rules and deep causal relationship analysis, drone-task pairs with mismatched capabilities or deep conflicts are eliminated to form a set of compatible matching pairs; S2.2 Quantum Feature Extraction and Encoding: For compatible matching pairs, high-dimensional quantum features are extracted using a lightweight quantum random forest, and non-Hermitian encoding matrices are used to encode the features into quantum states or classical equivalent vectors. S2.3 Distributed Singularity Conflict Resolution: Adjacent UAVs exchange coding features and utilize the singularity evolution characteristics of non-Hermitian systems to achieve self-driven conflict resolution; S2.4 Lightweight Quantum Approximation Optimization: Construct an objective function that integrates matching fitness, resource utilization, and energy consumption balance, and solve for the optimal matching matrix using a lightweight quantum approximation optimization algorithm; S2.5 Closed-loop feedback calibration: The matching results are fed back to the digital twin model to dynamically adjust the algorithm parameters.

4. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 1, characterized in that, S3 includes: S3.1 Construct a dual-system decision architecture of "rapid response - global optimization", in which the rapid response layer is based on a lightweight deep Q network to cope with millisecond-level environmental changes, and the global optimization layer is based on trust region policy optimization to achieve second- to minute-level global goal balance; S3.2 Training a fast response layer whose state vector contains key parameters of environmental changes, outputting instantaneous actions through a deep separable convolutional network, and the reward function integrates obstacle avoidance success, energy consumption control and task delay; S3.3 Training the global optimization layer, whose state vector contains all-dimensional system parameters, outputs global optimization actions through the policy function, and the update process is constrained by KL divergence to ensure stability; S3.4 Construct a multi-dimensional hierarchical payment function that integrates local immediate benefits, global long-term benefits, energy consumption costs, and carbon footprint; S3.

5. Using drone swarms as game participants, solve the Nash equilibrium iteratively through strategy evaluation and strategy improvement; S3.

6. Feed the optimized strategy and rule chain back to the digital twin model and update the model parameters.

5. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 1, characterized in that, S4 includes: S4.1 Multimodal Topology Graph Construction: Integrating visual, lidar, meteorological, and location features, a high-precision environmental topology graph is constructed through Granger causality tests and attention mechanisms, filtering out pseudo-associative edges; S4.2, Federated Learning Dedicated Path Generation: Using the federated learning framework, a meta-path generator adapted to the flight characteristics of different types of UAVs such as rotorcraft and fixed-wing aircraft is trained and path clusters are generated online in real time; S4.3 Real-time energy consumption model coupling and path optimization: Couple path length, number of turns, flight speed and UAV real-time energy consumption status to optimize path cost while meeting endurance constraints; S4.4 Causal-driven dynamic obstacle avoidance adjustment: Based on the causal relationship of the topology graph, only the future path of the associated UAVs is predicted, and the optimal obstacle avoidance adjustment amount is solved by model predictive control. S4.5 Dynamic Iterative Closed Loop: When the environmental state changes drastically and exceeds the threshold, the entire process of updating the topology and adjusting obstacle avoidance is re-triggered.

6. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 1, characterized in that, S5 includes: S5.1 Fault Timing Prediction and Early Warning: The temporal deep features of fault precursors are extracted by LSTM, and the fault probability is calculated by combining dynamic causal fault tree and logistic regression model. When the threshold is exceeded, an early warning is triggered. S5.2 Multidimensional Trust Model Construction: Integrating four dimensions—cooperative causal compatibility, failure probability of replacement drones, historical mission completion quality, and energy consumption stability—to quantify the trustworthiness of drone replacement. S5.3, Rapid Selection of Candidate UAVs for Replacement: Based on trust threshold, failure probability upper limit and mission capability predicate, the candidate set is quickly selected at the edge node, and an expansion mechanism is provided; S5.4, Replacement Path and Task Handover Optimization: Simulate the handover scenario in the digital twin model, generate an optimized path for the replacement UAV through the meta-path generator, and use model predictive control for dynamic correction; S5.5 Dynamic Evolution of Alliance Structure: Update the alliance adjacency matrix and trust model parameters to achieve adaptive adjustment of the alliance; S5.6 Closed-loop feedback calibration: Feedback the alliance evolution results to the digital twin model to update the fault prediction parameters and associated weights.

7. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 2, characterized in that, In S1.3, the compound predicate is defined. The logical expression is: in, For drones Execute the task Sufficient energy consumption; For drones Execute the task The load capacity meets the requirements; For drones The accuracy of the operation meets the task requirements. Require; For drones Adaptation task Environmental requirements; For drones The equipment is in good health condition.

8. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 3, characterized in that, In S2.2, quantum feature extraction and encoding are adapted to three engineering operation modes: classical platform quantum circuit simulation, small-scale commercial quantum hardware deployment, and quantum-inspired classical algorithm implementation. According to different modes, the corresponding implementation method is selected to generate high-dimensional quantum features or quantum-inspired classical features, and they are encoded through a non-Hermitian encoding matrix.

9. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 4, characterized in that, In S3.4, the payment function is constructed as follows: ;in, The time decay coefficient increases with time. For scene adaptation coefficients, and These are the penalty coefficients for energy consumption and carbon footprint, respectively. For the local immediate benefit of drone i performing action a, For drone i in the time window The overall long-term benefits within, The energy cost of drone i performing action a. The carbon footprint of drone i performing action a.

10. The distributed task allocation and path cooperative optimization algorithm for heterogeneous UAVs according to claim 6, characterized in that, In S5.2, the comprehensive calculation formula for multidimensional trust level is as follows: ; in, to These are the weight coefficients for each dimension, and their sum is 1; To coordinate the degree of causal compatibility, To supplement the predicted failure probability of drone y, To compensate for the historical mission completion quality of drone Y, To compensate for the energy consumption stability of drone y.

Images