Unmanned aerial vehicle adaptive task planning method and system fusing real-time environment perception

Abstract

The application relates to the technical field of unmanned planes, and discloses an unmanned plane adaptive task planning method and system fusing real-time environment perception, which has the technical scheme as follows: inputting unmanned plane cluster physical parameters and task target parameters, completing system hardware calibration, software environment and key module parameter initialization setting; based on a full-closed loop technical architecture of "perception-cognition-modeling-decision-cooperation-feedback", the following sub-steps are executed until the task termination condition is met; through the full-closed loop technical architecture, deep understanding of environment essential laws, dynamic self-evolution of planning schemes and autonomous emergence of cluster global intelligence are realized, and core technical support is provided for efficient, safe and intelligent task execution of unmanned planes in a complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically, to a method and system for adaptive mission planning of UAVs that integrates real-time environmental perception. Background Technology

[0002] Existing UAV mission planning algorithms mostly rely on static constraints, single-step optimization logic, or preset rules, making it difficult to effectively cope with environmental uncertainties in extreme scenarios such as dynamic changes, complex terrain, and strong interference (e.g., terrain undulations, electromagnetic interference fluctuations, and dynamic movement of obstacles). They lack in-depth exploration of the inherent causal relationships between environmental elements and forward-looking prediction of potential risks. The dynamic adaptive adjustment capability of the planning scheme is weak, and it is difficult to achieve efficient global information sharing and cluster global intelligent autonomous emergence in multi-UAV collaborative scenarios. This results in low UAV mission execution efficiency, high safety risks, and insufficient robustness, failing to meet the core requirements of efficient, safe, and intelligent UAV mission execution in fields such as civilian surveying and mapping, emergency rescue, public safety, and military collaboration.

[0003] For example, see Chinese Patent CN120276468A, "A Collaborative Control System and Method for Unmanned Aerial Vehicle (UAV) Clusters," which discloses "constructing a safety cross-section and generating an optimized path to allow the UAV cluster to avoid obstacles as a whole; achieving real-time reconstruction of the three-dimensional trajectory and millimeter-wave-level motion state update of dynamic obstacles in complex environments through multi-modal sensor fusion of multi-source sensing units; improving the accuracy of collision time interval prediction through the anthropomorphic trajectory prediction algorithm of the collaborative decision-making unit; and enabling large-scale clusters to synchronously complete obstacle avoidance path updates and morphological reshaping in a short time through an adaptive switching architecture of 5G and LoRa dual-mode communication." However, it still has shortcomings such as "a lack of in-depth exploration of the inherent causal relationships between environmental elements and forward-looking prediction of potential risks, weak dynamic adaptive adjustment capability of the planning scheme, and difficulty in achieving efficient global information sharing and global intelligent autonomous emergence of the cluster in multi-UAV collaborative scenarios."

[0004] Therefore, the present invention provides an adaptive mission planning method and system for unmanned aerial vehicles that integrates real-time environmental perception, thereby improving the above-mentioned technical problems. Summary of the Invention

[0005] This disclosure aims to address the shortcomings of existing technologies by providing an adaptive mission planning method and system for unmanned aerial vehicles (UAVs) that integrates real-time environmental perception. Through a fully closed-loop technical architecture of "perception-cognition-modeling-decision-coordination-feedback", this invention achieves a deep understanding of the essential laws of the environment, dynamic self-evolution of planning schemes, and autonomous emergence of global intelligence in the cluster, providing core technical support for the efficient, safe, and intelligent execution of UAV missions in complex and dynamic environments.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: an adaptive mission planning method for unmanned aerial vehicles (UAVs) that integrates real-time environmental perception, comprising the following steps:

[0007] S1. Initialization Phase: Input the physical parameters of the UAV cluster and the mission target parameters, and complete the initialization settings of system hardware calibration, software environment and key module parameters;

[0008] S2, Closed-Loop Execution Phase: Based on the "Perception-Cognition-Modeling-Decision-Collaboration-Feedback" closed-loop technical architecture, the following sub-steps are executed until the task termination conditions are met:

[0009] S2.1 Multi-source sensing and data preprocessing: The intelligent metasurface-enhanced multi-source sensing module collects multi-dimensional environmental, spatial, and task data in real time, and uses the nucleated causal mutual information maximization algorithm to perform redundancy removal and false association filtering, outputting a standardized environmental sensing dataset.

[0010] S2.2 Environmental Causal Cognition and Risk Prediction: Based on the environmental perception dataset, a dynamic causal cognition map is constructed using an improved PC algorithm, potential risks are predicted through a counterfactual reasoning model, and a high-risk event set is generated;

[0011] S2.3 Uncertainty Evolution Modeling: Based on the dynamic causal cognitive map and high-risk event set, the uncertain cellular automata are used to simulate the future multi-path evolution trend of the environment, and a "task-resource-environment" co-evolution constraint equation is constructed.

[0012] S2.4 Generative Evolutionary Game Decision-Making: Taking the future multi-path evolution trend of the environment and the set of high-risk events as input, multiple candidate planning schemes are generated through generative adversarial networks, and the globally optimal planning scheme is output based on the multi-participant evolutionary game mechanism combined with the quantum phase evolution algorithm.

[0013] S2.5, Cluster Awareness Emergence, Coordination and Execution: For multi-drone scenarios, local cognitive data is shared through a secure communication link to build a collective unconscious pool. When the conditions of cognitive consistency and policy consensus are met, cluster awareness emerges to achieve dynamic division of labor and collaborative task execution. For single-drone scenarios, the optimal planning scheme is executed directly.

[0014] S2.6 Feedback Iteration: Based on task execution feedback data, the dynamic causal cognitive graph, evolution rules, and game weight parameters are optimized in reverse, and the collective unconscious pool and solution evolution knowledge base are updated.

[0015] S3. Task Termination Phase: When one of the following conditions is met—subtask completion rate reaches the target, cluster survival probability is too low, or task timeout—the task is terminated, the task execution report and algorithm optimization parameter set are output, and the system is shut down in an orderly manner.

[0016] As a preferred technical solution of the present invention, in step S2.1, the intelligent metasurface enhanced multi-source sensing module performs adaptive sensing optimization of the environment by dynamically adjusting the beam direction, sensing frequency band and signal gain. Specifically, it includes: calculating the scene adaptation coefficient based on the environmental scene feature vector, dynamically adjusting the electromagnetic detector signal gain in a strong interference environment, and dynamically adjusting the lidar emission power in a foggy environment.

[0017] As a preferred technical solution of the present invention, in step S2.2, the dynamic causal cognitive graph adopts a triplet structure, including: a set of causal nodes, a set of directed causal edges, and causal strength weights that are dynamically updated based on real-time trigger strength, historical confidence, mechanism adaptability, and complexity.

[0018] As a preferred technical solution of the present invention, in step S2.3, the uncertain cellular automaton divides the task area into a dynamic cellular grid. Each cell is defined including a basic state, an evolution rule generated based on a causal chain, and a cellular uncertainty intensity that integrates perceived noise, causal uncertainty, and environmental disturbance. The credibility of the environmental evolution path is calculated based on the conformity between the evolution scenario and the causal chain and the average uncertainty intensity of the path.

[0019] As a preferred technical solution of the present invention, in step S2.4, the generator of the generative adversarial network is trained using a multi-objective loss function, the loss term including scene adaptability loss, constraint satisfaction loss, task gain loss and regularization loss; the quantum phase evolution algorithm maps the game payoff of the candidate scheme to the quantum phase, and solves the Nash equilibrium through the phase evolution equation.

[0020] As a preferred technical solution of the present invention, in step S2.5, the triggering condition for the emergence of cluster consciousness is that the cluster cognitive consistency and the global optimal strategy consensus both exceed a preset threshold; after emergence, the cluster automatically differentiates into a task execution group, a risk avoidance group, and a collaborative support group to achieve dynamic functional division of labor.

[0021] As a preferred technical solution of the present invention, step S2.5 also includes a robustness guarantee mechanism: by monitoring the status of the UAV in real time, when an individual fault is detected, the task migration and cluster reconstruction are quickly carried out based on the replacement adaptability; when encountering extreme environments, the "survival priority" mode is switched to ensure the survival of the cluster through the strategy of distributed detection and concentrated breakthrough.

[0022] An adaptive mission planning system for unmanned aerial vehicles (UAVs) that integrates real-time environmental perception, the system comprising:

[0023] A smart metasurface-enhanced multi-source sensing module is used to collect and preprocess multi-dimensional environmental data in real time.

[0024] The Environmental Causal Cognition and Counterfactual Reasoning module is used to construct a dynamic causal cognition map and make risk predictions;

[0025] The uncertainty evolution modeling module is used to simulate the multi-path evolution trend of the environment and construct cooperative evolution constraints;

[0026] The generative evolutionary game decision-making module is used to generate candidate solutions and output the optimal plan through game optimization.

[0027] The cluster consciousness emergence collaboration module is used to realize cognitive sharing, consciousness emergence, and collaborative execution among multiple drones;

[0028] The feedback iterative optimization module is used to optimize system parameters based on execution feedback.

[0029] The initialization and task management module is used for system initialization, task flow control, and termination processing.

[0030] As a preferred technical solution of the present invention, the intelligent metasurface enhanced multi-source sensing module integrates an intelligent metasurface unit, an RGB-D vision sensor, a lidar, a meteorological sensor, an electromagnetic interference detector, and a combined navigation module, forming a multi-source data input system of "local fine-grained sensing + global coarse-grained assistance".

[0031] As a preferred technical solution of the present invention, the system further includes: a scheme evolution knowledge base, used to store historical optimal planning schemes and their corresponding environmental characteristics, evolution scenarios and game payoffs, and to mine the mapping relationship between environmental characteristics and optimal schemes through a self-supervised learning model in order to optimize the parameters of the generative evolutionary game decision module.

[0032] In summary, the present invention has the following beneficial effects:

[0033] Firstly, through the "Environmental Causal Cognition and Counterfactual Reasoning Module," this invention not only collects environmental data but also uncovers the inherent causal relationships between environmental elements, resource status, and task objectives, constructing a dynamic causal cognitive map. This enables the system to understand "why" environmental changes affect tasks, rather than merely perceiving "what" is changing, thereby achieving forward-looking predictions of the essential laws of the environment and potential risks, providing a solid logical foundation for planning and decision-making.

[0034] Secondly, through the synergy of the "uncertainty evolution modeling module" and the "generative evolutionary game decision-making module," this invention can simulate the multi-path evolution trend of the future environment and generate and optimize planning schemes in real time based on evolutionary game theory. This method eliminates the dependence on static rules and single-step optimization, enabling UAVs to quickly and automatically adjust their action strategies when the external environment (such as obstacle movement and interference fluctuations) and their own state (such as endurance) change dynamically, ensuring the timeliness and optimality of the planning.

[0035] Third, through the "cluster consciousness emergence collaboration module," this invention enables multiple drone swarms to share cognition through secure communication and spontaneously "emerge" with global collaborative intelligence when conditions are met, automatically differentiating into functionally complementary groups (such as execution, obstacle avoidance, and backup). This collaboration mechanism based on collective subconsciousness allows the swarm to dynamically divide labor and cooperate efficiently like an organic whole, greatly improving the execution efficiency and overall robustness of complex tasks.

[0036] Fourth, a complete closed-loop optimization system of "perception-decision-execution-feedback" has been constructed. This invention is not a simple accumulation of isolated modules, but rather a complete closed-loop technology chain that includes feedback and iteration. The system can optimize perception parameters, causal models, evolution rules, and decision weights in reverse based on the actual effect of task execution, enabling the entire system to have the ability to learn online and continuously improve, becoming more intelligent with use and maintaining a high level of task adaptability and execution efficiency in the long term.

[0037] Fifth, this invention designs a multi-layered protection mechanism from individual to cluster level, and from normal to extreme conditions. At the individual level, rapid recovery is achieved through fault monitoring and adaptive backup; at the cluster level, a "survival-first" mode and a collaborative breakthrough strategy are used to cope with extreme environments. These mechanisms ensure that the system can still maximize mission completion or safe evacuation under abnormal conditions such as equipment failure and strong interference, significantly improving the system's reliability and survivability.

[0038] Sixth, the architecture proposed in this invention, centered on "intelligent metasurface-enhanced perception, causal cognition, evolutionary modeling, game theory decision-making, and cluster emergence," features a clear modular design. This architecture is not only applicable to the specific tasks and sensor combinations described above, but its concepts and methods can also be adapted to other unmanned systems or automated task scenarios requiring real-time perception, intelligent decision-making, and multi-agent collaboration, demonstrating good versatility and technological extensibility. Attached Figure Description

[0039] Figure 1 This is a framework diagram of an UAV adaptive mission planning system that integrates real-time environmental perception, provided in an embodiment of the present invention.

[0040] Figure 2A flowchart of an adaptive mission planning method for unmanned aerial vehicles (UAVs) that integrates real-time environmental perception, provided in an embodiment of the present invention. Detailed Implementation

[0041] 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.

[0042] 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.

[0043] 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.

[0044] 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.

[0045] This disclosure aims to address the problems of low mission execution efficiency, high safety risks, and insufficient robustness in existing UAV mission planning algorithms under extreme scenarios such as dynamic changes, complex terrain, and strong interference. These problems include insufficient understanding of the essential laws of the environment, weak dynamic adaptive capability of planning schemes, and lack of global intelligent collaboration among the swarm. Therefore, this disclosure proposes a UAV adaptive mission planning method and system that integrates real-time environmental perception to support efficient, safe, and intelligent mission execution by UAVs. This method employs a fully closed-loop technical architecture consisting of "intelligent metasurface-enhanced multi-source perception, environmental causal cognition and counterfactual reasoning, uncertainty evolution modeling, generative evolutionary game decision-making, and swarm awareness emergence collaboration." It deeply integrates cognitive science and next-generation AI technology, achieving deep environmental understanding, self-evolution of planning schemes, and swarm awareness emergence through dynamic adaptation mechanisms and uniquely identified variables. This reduces the system's dependence on static constraints and rule presets, and improves adaptability to extreme scenarios.

[0046] Please refer to Figure 1 , Figure 1The diagram illustrates the framework of the UAV adaptive mission planning system integrating real-time environmental perception as described in this embodiment. It adopts a closed-loop technical architecture of "perception-cognition-modeling-decision-coordination-feedback," and achieves adaptive mission planning through five core modules, as detailed below:

[0047] Intelligent metasurface-enhanced multi-source sensing module: As the foundation for end-to-end data acquisition and preprocessing, it builds a multi-source data input system of "local fine-grained sensing + global coarse-grained assistance" through multi-dimensional hardware configuration. Combined with the metasurface adaptive sensing enhancement mechanism, it optimizes sensing parameters, simultaneously collects physical environment, spatial rules and task-related data, and then uses the causal mutual information maximization algorithm to remove redundancy and false correlations, outputting a standardized, high-precision environmental sensing dataset, providing low-latency data support for subsequent modules.

[0048] The Environmental Causal Cognition and Counterfactual Reasoning module: Receives perceived data and uses an improved PC algorithm to mine directed causal relationships between environmental elements, constructing a dynamic causal cognitive graph and updating causal strength weights based on real-time data. It predicts potential risks using a counterfactual reasoning model, generates a set of high-risk events, and synchronizes it to subsequent modules. Coupled with an environmental metacognition mechanism, it quickly identifies environmental metatypes, invokes corresponding cognitive paradigms, and ensures cognitive accuracy through cognitive confidence calibration, thus realizing the transformation from data perception to logical cognition.

[0049] Uncertainty Evolution Modeling Module: Based on the causal relationships and risk information output by the cognitive layer, the task region is divided into dynamic cellular grids through an uncertain cellular automaton. The module defines the basic state of the cells, evolution rules, and uncertainty intensity, simulates the future multi-path evolution trend of the environment, and calculates the credibility of each path. It constructs a "task-resource-environment" co-evolution constraint equation, distinguishes between hard and soft constraints, and dynamically adjusts the constraint threshold based on the credibility of the evolution scenario, transforming environmental uncertainty and dynamic evolution characteristics into a computable model representation.

[0050] Generative evolutionary game decision-making module: Taking evolutionary scenarios and risk events as input, it generates multiple sets of candidate planning schemes through generative adversarial networks, introduces a multi-participant evolutionary game mechanism, and combines quantum phase to solve Nash equilibrium to output the globally optimal planning scheme; it constructs a scheme evolution knowledge base to reuse historical experience, mines the mapping relationship between the environment and the optimal scheme through self-supervised learning, and updates the generator and discriminator parameters in real time to achieve scheme self-evolution and improve decision-making efficiency.

[0051] The swarm consciousness emergence collaboration module: For multi-drone scenarios, it shares local cognitive data through a secure communication link of "millimeter wave + quantum key distribution" to build a collective unconscious pool containing global characteristics; based on collaborative game theory and cognitive consistency, global strategy consensus, it triggers swarm consciousness emergence, automatically differentiates functional groups to achieve dynamic division of labor and global collaboration; through individual fault adaptive replacement and extreme environment "survival first" collaborative breakthrough mechanism, it ensures swarm robustness, and each module iterates and optimizes parameters through real-time feedback to maintain the efficient and stable operation of the entire closed-loop technology chain.

[0052] Specifically, the intelligent metasurface-enhanced multi-source sensing module.

[0053] The intelligent metasurface-enhanced multi-source sensing module is used to achieve high-precision, real-time acquisition and preprocessing of multi-dimensional environmental data, providing data support for subsequent cognition and modeling. In terms of sensing hardware configuration, the UAV is equipped with an intelligent metasurface unit, an RGB-D visual sensor, a lidar, a meteorological sensor (wind speed / air pressure / precipitation detection), an electromagnetic interference detector, and a GPS / BeiDou + IMU integrated navigation module. It also accesses global airspace control data from ground stations / satellites, forming a multi-source data input system of "local fine-grained perception + global coarse-grained assistance." Local perception data focuses on high-precision environmental detail capture within a 1-100 meter range within the mission area, while global assistance data covers macroscopic information such as airspace rules and large-scale meteorological trends within a 1-10 kilometer range. The two are synchronized temporally and spatially through a data fusion gateway, ensuring comprehensive and consistent data coverage.

[0054] The metasurface sensing enhancement mechanism dynamically adjusts the beam direction, sensing frequency band, and signal gain through intelligent metasurface units, adaptively optimizing sensing parameters for different environmental scenarios. It defines environmental scene feature vectors. ,in, This refers to the actual wind speed. This represents the actual electromagnetic interference intensity. For actual visibility, The probability of precipitation (range 0-1). Standard feature vectors are obtained through feature normalization. ;

[0055] The normalization formula is:

[0056]

[0057]

[0058]

[0059] ;

[0060] in, , For extreme values ​​of wind speed statistics, ,

[0061] These are the statistical extreme values ​​of electromagnetic interference intensity. , These are the extreme values ​​for visibility statistics.

[0062] Constructing environmental scene adaptation coefficients based on standard feature vectors The calculation formula is:

[0063]

[0064] in, , , , Weighting coefficients (satisfying) The influence weights of different environmental factors are adaptively determined using the entropy weight method to ensure that the influence weights match the scene complexity. The weighting of wind speed influence; The weighting is determined by the intensity of electromagnetic interference. The visibility factor influences the weighting; The weighting is determined by the probability of precipitation.

[0065] Under strong interference environment ( ), electromagnetic detector signal gain The dynamic adjustment formula is:

[0066]

[0067] in, The reference signal gain of the electromagnetic detector (determined by the sensor hardware specifications and initial calibration, with a value range of 10-20dB). It is a natural exponential function, and the gain is adjusted exponentially to achieve a nonlinear fit with the interference intensity. The impact of enhancing the overall adaptability of the scene on the gain.

[0068] In foggy weather ( ), lidar transmission power The dynamic adjustment formula is:

[0069]

[0070] in, The reference transmit power of the lidar (determined by sensor hardware specifications and initial calibration, ranging from 5-15W) is used to ensure the detection distance and accuracy in foggy and precipitation scenarios through dual-factor adjustment. Simultaneously, "physical environment data + spatial rule data + task-related data" are collected, and the collection frequency is matched with the sensor type (30fps for visual sensors, 10Hz for lidar, and 5Hz for meteorological sensors).

[0071] The data preprocessing adopts the "kernelized causal mutual information maximization" algorithm to remove redundancy and filter false associations in multi-source data. This algorithm transforms linearly inseparable data into linearly separable data in a high-dimensional feature space through kernel function mapping, thereby improving the accuracy of causal relationship identification.

[0072] First, calculate any two types of sensing data. , nucleated causal mutual information value The calculation formula is:

[0073]

[0074] in, For the data sample size, For matrix trace operations, , They are respectively , The kernel matrix (constructed using the Gaussian kernel function, Gaussian kernel parameters) K-fold cross-validation was used to optimize the initialization phase from a preset range [0.1, 10] to maximize the accuracy of causal discovery. , These are the diagonal elements of the kernel matrix. , They are respectively , The kernel entropy, the formula for calculating kernel entropy is:

[0075]

[0076]

[0077] in, This involves performing the natural logarithm operation. A data dependency model is constructed based on a Bayesian network, and a redundancy threshold is set. (Value range 0.05-0.2, adaptively adjusted by data dimension) When When data is deemed redundant, it is removed. This method effectively preserves the nonlinear causal relationships between data while eliminating invalid and redundant information.

[0078] Subsequently, spurious associations were identified by improving the Pearson correlation coefficient and using causal orientation tests. Introducing the time lag factor, the calculation formula is as follows:

[0079]

[0080] in, This is the time lag (range 0-50ms, determined based on the data sampling frequency). , They are respectively At any moment , At any moment By introducing a time lag factor, the sampled values ​​can be effectively used to identify spurious associations that are time-synchronized but have no causal relationship.

[0081] Simultaneously, a causal direction test algorithm based on conditional mutual information is employed to calculate the causal direction index. :

[0082]

[0083] in, For the set of condition variables (including except) , Other sensory data), For conditional causal mutual information, the calculation formula is:

[0084]

[0085] when and If the data is found to be spurious (time synchronization but no causal relationship), only the core causal features are retained to ensure that the preprocessed data can accurately reflect the intrinsic relationships between environmental elements.

[0086] The final output is a standardized environmental perception dataset. The standardization process uses a modified Z-score method, and the calculation formula is as follows:

[0087]

[0088] in, For a moment The original data, For a moment The moving average of the data within the previous second. For a moment The sliding standard deviation of the data in the first second is used to achieve real-time standardization of the data through sliding statistics, avoiding the lag in adapting to dynamic scenarios caused by fixed statistics.

[0089] Dataset Includes: terrain elevation Obstacle status (existence) Movement speed ,

[0090] size Meteorological parameters (wind speed) ,wind direction Precipitation intensity Electromagnetic interference intensity airspace permissions (0 for no restrictions, 1 for low restrictions, 2 for medium restrictions, 3 for no-fly zones), Target dynamics (Location ),state Core data such as {stationary, moving, target locked, out of contact} are updated at a frequency of 50ms / time and transmitted to the subsequent cognition and modeling modules via a high-speed data bus.

[0091] Specifically, the module on environmental causal cognition and counterfactual reasoning.

[0092] The Environmental Causal Cognition and Counterfactual Reasoning module is used to deeply explore the intrinsic causal relationships among environmental, resource, and task elements, achieving a precise understanding of the essential laws of the environment and a forward-looking prediction of potential risks, providing core logical support for subsequent modeling and decision-making. The Dynamic Causal Cognition Graph (D-CCG) constructs a standardized environmental perception dataset output by the intelligent metasurface-enhanced multi-source sensing module. Using Gaussian kernels as the core input, an improved PC causal discovery algorithm is employed to mine directed causal relationships. This algorithm introduces kernel function mapping and adaptive threshold adjustment mechanisms on the basis of traditional PC algorithms. It maps nonlinear sensing data to a high-dimensional feature space through Gaussian kernel functions, and then eliminates false associations based on conditional independence tests (CI-test). It can accurately identify multi-order and weakly correlated causal chains, such as complex causal relationships such as "activation of electromagnetic interference source → decrease in signal-to-noise ratio of communication link → increase in data transmission error rate → delay in task instruction execution → risk of task execution timeout" and "increased terrain undulation → increased flight drag → faster energy consumption of endurance → insufficient remaining endurance → reduced task coverage".

[0093] Define the triplet structure of D-CCG ,in, This is a set of causal nodes, encompassing three core types of nodes: environment nodes (such as...) Represents sudden changes in wind speed, Represents enhanced electromagnetic interference, Represents reduced visibility, Representing increased terrain undulation), resource nodes (such as...) Indicating insufficient battery life This indicates a decline in communication quality. Indicates insufficient load power. Represents navigation accuracy offset), task-type nodes (such as...) This indicates that the task execution timed out. This indicates that the target detection failed. Represents data transmission interruption, (This indicates incomplete task coverage). All nodes use a unique encoding rule of "type identifier-serial number" to ensure the uniqueness of node differentiation. This is a set of directed causal edges, representing the logical relationship between cause and effect. The edge identifier corresponds to the node code (e.g., ...). correspond causal relationship correspond The causal relationship), each causal edge is associated with a corresponding causal mechanism description (such as... The causal mechanism is that "when the electromagnetic interference intensity exceeds -60dBm, for every 1dB decrease in the signal-to-noise ratio of the communication link, the data transmission error rate increases by 5%".

[0094] The causal strength weight is used to quantify the real-time reliability and impact of causal relationships. It is dynamically updated based on real-time sensing data, historical statistical information, and the adaptability of causal mechanisms. The calculation formula is as follows:

[0095]

[0096] in, , , , Weighting coefficients (satisfying) The initial values ​​were set to 0.3, 0.25, 0.25, and 0.2 respectively, and can be dynamically optimized through online learning. Real-time trigger strength weight for causal chains; Weights for historical confidence in causal relationships; Weights for the fit of the causal mechanism; Weights are used to represent the complexity coefficients of the causal chain. The real-time trigger strength of the causal chain (range 0-1) is calculated based on the activation level of causal nodes in the current sensed data. For example, for a causal chain... , ,in, It is the Sigmoid activation function. For real-time electromagnetic interference intensity, and These are the electromagnetic interference activation threshold and the communication quality activation threshold, which are preset according to the sensor range and mission safety requirements. For real-time communication quality; The historical confidence level of the causal relationship (range 0-1) is based on the statistical analysis of the triggering frequency and accuracy of the causal chain in historical data. ,in, This represents the number of times this causal chain has been successfully triggered throughout history. This represents the total number of times this causal chain has been triggered throughout history. To ensure smoothness and avoid a denominator of 0; The causal mechanism fit (range 0-1) is used to quantify the degree of matching between the current environmental state and the description of the causal mechanism. For example, for the causal mechanism of "electromagnetic interference → communication quality degradation", The attenuation coefficient κ controls the sensitivity of the fit to the deviation, with a reference value set to 0.1. This serves as the reference interference strength in the causal mechanism, set based on typical historical interference scenarios. The complexity coefficient of the causal chain (ranging from 0 to 1) is determined by the length of the causal chain and the node association density. ,in, The length of the causal chain is (number of nodes - 1). Node association density (the causal chain involves the number of associations between a node and other nodes / the total number of nodes). Based on the latest dataset every 30ms. Update the node activation status (activated / inactive) and causal strength weight of D-CCG. For inactive nodes, the strength weight of their associated causal edges is temporarily set to 0 to ensure that the graph can adapt to the dynamic changes of the environment and tasks in real time.

[0097] Counterfactual reasoning risk prediction focuses on key causal chains in D-CCG. The selection criterion for key causal chains is "an impact weight ≥ 0.6 on the core mission objective." The impact weight is calculated using the Analytic Hierarchy Process (AHP), i.e., by constructing an impact matrix of mission objective-causal chains and solving for the eigenvectors to obtain the impact weight of each causal chain. First, a counterfactual reasoning model based on a structural causal model (SCM) is constructed. The structural causal model is defined as... ,in, These are exogenous variables (such as random environmental disturbances and sensor noise). These are endogenous variables (i.e., causal nodes in D-CCG). This is a set of causal functions (corresponding to the causal mechanism of causal edges). A counterfactual query is defined as "if the policy is executed..." (Such as path adjustment, payload power adjustment, flight speed adjustment, etc.), intervening in endogenous variables The value is Will it trigger a causal chain? Risk events ?

[0098] Strength weights based on causal chains Combined strategy Intervention effects on causal nodes, calculating risk events probability of occurrence The calculation formula is:

[0099]

[0100] in, For the first in the causal chain The strength weight of each causal edge; For strategy For the first causal nodes The intervention intensity (range -1 to 1) is calculated as follows: positive intervention (increasing the node activation probability) takes a positive value, negative intervention (decreasing the node activation probability) takes a negative value, and no intervention takes a 0 value. For example, the strategy "increase the communication antenna gain" affects the node. The intervention intensity for (degraded communication quality) is ; End node of the causal chain Triggering risk events The inherent probability (determined based on historical risk event statistics, such as...) trigger The inherent probability of (communication interruption risk) is 0.7). This is a correction factor for environmental uncertainty (range 0.8~1.2), determined by the environmental dynamics. calculate, It is used to quantify the impact of dynamic environmental changes on risk probability.

[0101] Generate a set of potential risk events ,in, The codes are unique and clearly categorized, including collision risks. Communication interruption risk Risk of insufficient power Navigation failure risk Task timeout risk Risk of data loss Each risk event is associated with a corresponding risk level (low / medium / high) and impact scope (local / global). Risk thresholds are set. (Values ​​range from 0.6 to 0.8, adaptively adjusted based on the task's safety level; for example, 0.75 for emergency rescue tasks and 0.65 for routine surveying tasks.) High-risk events are marked as "key avoidance events" and synchronized to the uncertainty evolution modeling module and the generative evolutionary game decision-making module. At the same time, the causal chain, triggering conditions and intervention suggestions corresponding to the risk event are recorded to provide a basis for subsequent planning scheme optimization.

[0102] The environmental meta-cognition and adaptation mechanism achieves rapid adaptation to different scenarios through a meta-cognitive learner model. This model adopts the MAML (Model-Agnostic Meta-Learning) framework in meta-learning, quickly learning the cognitive patterns of different environmental scenarios with a small number of samples, achieving scenario adaptation without retraining. First, the global feature vector of D-CCG is extracted. The feature vectors encompass 16 core dimensions, comprehensively characterizing the structure and state features of the graph. The calculation formula is as follows:

[0103]

[0104]

[0105] in, The causal node activation density is the ratio of the number of activated nodes to the total number of nodes. The average strength of the causal chain is the average of the strength weights of all causal edges. The proportion of high-risk causal chains is the ratio of the number of high-risk causal chains (intensity weight ≥ 0.7) to the total number of causal chains. The node association complexity is calculated based on the average path length and clustering coefficient of the graph. ; This represents the average length of the causal chain, i.e., the average length of all causal chains. Node association density is the ratio of the average number of associated edges of all nodes to the total number of nodes. , , These represent the average activation intensity of environment, resource, and task nodes, respectively. , These are the correlation coefficients for environment-resource nodes and resource-task nodes, respectively. The graph stability coefficient is the average rate of change of node activation state and causal strength weight between two consecutive updates. This refers to the dynamic coefficients of the spectrum, i.e., the proportion of highly dynamic causal chains (intensity weight change rate ≥ 0.1 / 30ms); , These represent the average risk probability and the maximum risk probability of the potential risk event set, respectively. The scene adaptation coefficient is the average similarity between the current map features and the historical scene features.

[0106] The feature vector is processed through a meta-learning model. For classification, the model employs a two-layer optimization structure: the inner layer optimizes for a single environment scenario, updating model parameters through gradient descent to minimize inference error; the outer layer optimizes across multiple environments scenario, updating meta-parameters to maximize the generalization ability of the inner layer. This model allows for rapid identification of the "meta-type" of the current environment. Including urban dynamic environment Complex mountainous terrain Strong interference countermeasure environment Open sea environment Indoor confined environment The metatype uses a unique identifier encoding (e.g., correspond , correspond ).

[0107] Based on the identified environmental metatypes Invoke pre-stored "cognitive paradigms". A cognitive paradigm is a set of cognitive strategies optimized for a specific environment or scenario, including causal chain monitoring priority, feature extraction weights, and inference threshold adjustments. For example... Cognitive Paradigm Corresponding to (Strong Interference and Counter-Environment) This includes: prioritizing and enhancing the monitoring frequency of the causal chain "electromagnetic interference → communication quality → data transmission → task execution" (from 30ms / time to 15ms / time), increasing the extraction weight of electromagnetic interference-related features (from 0.2 to 0.4), and reducing the threshold for determining the risk of communication interruption. (From 0.7 to 0.6), increasing the weight of causal strength The coefficient (increased from 0.25 to 0.35); Cognitive Paradigm for (Complex Mountainous Topography) This includes: prioritizing the monitoring of the causal chain of "terrain undulation → flight drag → endurance consumption → mission coverage"; increasing the weighting of terrain elevation and wind speed features in the extraction process; raising awareness of insufficient power risks; and optimizing the adaptability of causal mechanisms. The calculation function is adapted to the terrain variation pattern.

[0108] To ensure cognitive accuracy, cognitive confidence is calculated. This is used to quantify the degree of matching between the current cognitive outcome and the actual environmental state. The calculation formula is:

[0109]

[0110] in, The number of currently active causal chains; For the first The strength weight of each activated causal chain; For the current sensing data and the first The result of the causal chain reasoning The matching degree (ranging from 0 to 1) is obtained by calculating the cosine similarity of the data distribution. ; For the first The correlation between the risk events corresponding to each causal chain and the current sensed data (value range 0-1). For example, if the intensity of electromagnetic interference in the sensed data increases significantly, the correlation coefficient of the communication interruption risk increases. A confidence threshold is set. (Value range 0.7~0.85, dynamically adjusted based on environment complexity; set to 0.75 for high-dynamic environments and 0.8 for low-dynamic environments). When this happens, the "metasurface sensing parameter reconfiguration + causal chain reconstruction" process is automatically initiated:

[0111] (1) Metasurface Sensing Parameter Reconfiguration: Based on the current high-risk causal chain and low-matching inference results, determine the environmental elements that need enhanced sensing (e.g., if the risk of communication interruption is high, enhance electromagnetic interference and communication quality sensing), and adjust the beam direction of the smart metasurface (aligning it with the direction of the interference source or communication target), sensing frequency band (switching to the anti-interference frequency band), and signal gain (increasing the gain parameters of the corresponding sensors). For example, for electromagnetic interference sensing, adjust the metasurface beam direction to... Gain increased to ;

[0112] (2) Causal chain reconstruction: Causal chains with a matching degree lower than 0.5 are re-mined. An improved PC algorithm is used in conjunction with the reconfigured perception data to re-identify causal relationships, remove false causal edges, supplement missing causal edges, and update the calculation parameters of causal strength weight (such as adjusting...). , (Equal coefficients) ensure that the cognitive system can quickly correct deviations and adapt to the actual environmental conditions.

[0113] Specifically, the uncertainty evolution modeling module.

[0114] The uncertainty evolution modeling module is used to accurately simulate the multi-path evolution trend of the environment in the future, construct a dynamically adaptable environmental model, and transform environmental uncertainty and dynamic evolution characteristics into a computable and optimizable quantitative representation, providing comprehensive and accurate environmental prediction support for the subsequent decision-making module. This module utilizes the Dynamic Causal Cognition Map (D-CCG) and the set of potential risk events output by the Environmental Causal Cognition and Counterfactual Reasoning module. Using standardized environmental perception datasets as the core input. By using Uncertainty Cellular Automata (U-CA) and Co-evolutionary Constraint Modeling, a multi-dimensional and forward-looking characterization of environmental evolution can be achieved.

[0115] Uncertainty cellular automata (U-CA) modeling is the core tool for environmental evolution simulation. First, the task region is divided into dynamic cellular grids, with the grid size... Based on environmental dynamics Adaptive adjustments ensure a dynamic balance between modeling accuracy and computational efficiency. Environmental dynamism. Taking into account multiple factors such as the dynamic changes of obstacles, fluctuations in electromagnetic interference intensity, and variations in meteorological parameters, the calculation formula is as follows:

[0116]

[0117] in, , , Weighting coefficients (satisfying) The initial values ​​are set to 0.4, 0.3, and 0.3 respectively, and can be dynamically optimized through online learning. , , The contribution ratios of obstacle dynamics, electromagnetic interference fluctuations, and meteorological variability to the overall environmental dynamics are respectively characterized, and the initial values ​​are set according to the degree of influence of each factor on the flight safety of the UAV. For a moment The average moving speed of obstacles within the mission area; The maximum moving speed threshold for obstacles (set to 10m / s); unit of time (Take 30ms, consistent with the D-CCG update cycle) the change in electromagnetic interference intensity; This is the maximum electromagnetic interference intensity (set to -40dBm). unit of time Changes in internal wind speed; This represents the maximum wind speed (set to 20 m / s). Based on The grid size adjustment rule is: when (In a highly dynamic environment) ;when (In a dynamic environment) ;when In (low dynamic environment), .

[0118] Define each cell The three-dimensional state comprehensively characterizes the environmental attributes, evolutionary laws, and uncertainty features of the cell:

[0119] Basic state : Characterizing the real-time environmental attributes of cells, it is the fundamental data for evolutionary simulation, specifically including terrain elevation. Electromagnetic interference intensity The existence of obstacles (0 indicates no obstacles, 1 indicates obstacles) Obstacle movement speed (0 when there are no obstacles), airspace permission level (0 for no restrictions, 1 for low restrictions, 2 for medium restrictions, 3 for no-fly zones), all attribute data are from... Extract and assign values ​​after spatial registration;

[0120] Evolutionary rules Based on D-CCG, dynamic generation of causal chains ensures that evolutionary laws are consistent with the essential causal relationship of the environment. The rule expression is a condition-action pair structure, that is, "if a cell satisfies a certain causal triggering condition, then the corresponding evolutionary action is executed." For example, for the causal chain "increased electromagnetic interference → decreased communication quality," the evolutionary rule is "if a cell satisfies a certain causal triggering condition, then the corresponding evolutionary action is executed." electromagnetic interference intensity ( The threshold for triggering the interference evolution is set to a value ranging from -80 to -60 dBm (dynamically adjusted based on the causal chain strength weight). Then, at the next moment, the interference diffuses to the three adjacent cells, with a diffusion intensity attenuation coefficient. ",in, The strength weight of this causal chain is determined by the interference strength; the higher the interference strength, the more reliable the causal chain, and the better the diffusion range and attenuation coefficient match the actual evolution law. For the causal chain of "obstacle movement → risk area expansion", the evolution rule is "if the cell..." The existence of obstacles And movement speed ( If the threshold for triggering obstacle movement evolution is set (ranging from 0.5 to 1 m / s), then the neighboring cells along the movement direction at the next moment... The obstacle presence is set to 1, and the movement speed is inherited from 80% of the original cell.

[0121] Uncertainty intensity The formula for calculating the impact of multiple uncertainties, including comprehensive quantitative sensing noise, causal chain uncertainty, and random environmental disturbances, on the evolution results is as follows:

[0122]

[0123] in, , , Weighting coefficients (satisfying) The initial values ​​are set to 0.3, 0.4, and 0.3 respectively. For sensor-sensed noise weights; Weights representing the uncertainty of the causal chain; Weights for global environmental random perturbation coefficients; For cells The sensor sensing noise (value range 0-1) is calculated from the noise model of each sensor, such as the sensing noise of lidar. ( (This refers to the inherent noise variance of the lidar). For cells The uncertainty of the causal chain (ranging from 0 to 1) is calculated from the variance of the causal strength weights. ( The number of causal chains associated with a cell. For the first The strength weight of each causal chain, (As average intensity weight). The global environmental random disturbance coefficient (range 0-1) is calculated based on random changing factors such as meteorological parameters and airspace control instructions.

[0124] Based on the parallel computing characteristics of U-CA, simulate the environment 5-15 seconds ahead. Possible evolutionary paths and their durations Adaptive adjustment based on the intensity of environmental uncertainty: hour, (Shorten the prediction time to ensure accuracy in environments with high uncertainty); hour, ;

[0125] hour, (Extending the forecast period in low uncertainty environments to improve forward-looking capabilities). The number of evolutionary paths is determined by the global average uncertainty intensity:

[0126] ( (Total number of cells / grids).

[0127] Adaptive adjustment: hour ; hour ;

[0128] hour Each path corresponds to an evolutionary environment scenario:

[0129] ( , (Consistent with the cell state update cycle), the scene identifier adopts a unique encoding rule of "timestamp-path sequence number" (e.g., ... (Represents the evolution scenario of the first timestamp and the third path).

[0130] To quantify the reliability of each evolutionary path, the credibility of each path was calculated based on the causal inference results of D-CCG. The calculation formula is:

[0131]

[0132] in, The number of core causal chains in D-CCG (core causal chains are defined as strength weights) (causal chain). For the first The strength weight of the core causal chain; The degree of conformity between the evolutionary scenario and the causal chain (value range 0-1) is obtained by calculating the matching degree between the activation sequence of causal nodes in the scenario and the logic of the causal chain. ( (where the sequence length is...) For the first Average uncertainty intensity of each evolution path:

[0133] ;

[0134] For the first Cells in the path (the intensity of uncertainty); This is an uncertainty attenuation coefficient (range 0.8-1.2), used to weaken the credibility weight of highly uncertain paths. Credibility The value ranges from 0 to 1. A higher value indicates that the evolutionary path fits the causal law better, the uncertainty is lower, and the priority is higher in subsequent decisions.

[0135] Co-evolutionary constraint modeling is used to construct a dynamic constraint system integrating "task-resource-environment" to ensure the feasibility and optimizability of planning schemes under various evolutionary scenarios. The core idea of ​​the constraint equations is to tightly couple the UAV state, task requirements, and environmental evolution to achieve real-time adaptation of constraints to environmental changes. First, the general form of the co-evolutionary constraint equations is constructed:

[0136]

[0137] in, , The weighting coefficients for hard and soft constraints (satisfying) , , (to ensure the core guarantee status of hard constraints). For the first Hard constraint aggregate values ​​under each evolutionary scenario; For the first Soft constraint aggregation values ​​under an evolutionary scenario.

[0138] Hard constraints These are key constraints ensuring the safety and basic feasibility of UAV missions, and constitute a set of inequality constraints, specifically including:

[0139] Collision avoidance constraints: ,in, Location of the drone With evolutionary scenarios Middle obstacle area The shortest distance; The dynamic safety threshold is adjusted based on the evolving scenario's credibility and the presence of obstacles.

[0140] ( The baseline safety threshold is set at 2-5m. For the first (Average moving speed of obstacles in each scenario)

[0141] Battery life constraints: ,in, This refers to the remaining energy consumption of the drone; For the first Minimum energy consumption required to complete the remaining tasks in each evolutionary scenario:

[0142] ( (Calculated based on path length and load power to represent the minimum energy consumption under ideal conditions).

[0143] Airspace permission constraints: ,in, For the first The position of drones in an evolutionary scenario The corresponding airspace permission level ensures that drones do not enter the no-fly zone. ).

[0144] Hard constraint aggregation value The calculation method is as follows:

[0145]

[0146] This means taking the minimum value of all hard constraints to ensure that all hard constraints are satisfied.

[0147] soft constraints These are constraints for achieving task optimization goals, transforming optimization objectives such as "optimal energy consumption, optimal task efficiency, and minimum risk" into constraints, specifically including:

[0148] Energy consumption optimality constraint: ,in, The optimal energy consumption under the current state (based on the scheme evolution knowledge base) (Found by query) For the first Predicted energy consumption in various scenarios;

[0149] Task efficiency constraints: ,in, For the first Task efficiency (task completion progress / time) in a specific scenario; The minimum efficiency required for the task;

[0150] Minimum risk constraint: ,in, This represents the theoretical minimum risk value. For the first Predicted risk value in each scenario (based on) (Calculation of the probability of occurrence of risk events in the evolutionary scenario).

[0151] Transforming soft constraints into a "constrained game" The weights of each player in the game:

[0152] (satisfy Based on the evolutionary scenario credibility and task status, the adjustment is dynamically adjusted using the following formula:

[0153]

[0154]

[0155]

[0156] in, This represents the maximum energy consumption of the drone. This represents the current progress of the task. For the first The average risk probability in each scenario; This represents the degree of matching between the set of potential risk events and the evolutionary scenario. Soft constraint aggregation value. The calculation method is as follows:

[0157]

[0158] Based on the credibility of the evolutionary scenario Dynamically adjust constraint thresholds to adapt constraints to scenario reliability: when In high-confidence scenarios, hard constraint thresholds are strictly enforced. Taking 1.1 times the baseline value, the soft constraint weights are tilted towards task efficiency; when In medium-confidence scenarios, the thresholds for non-core hard constraints should be appropriately relaxed. Reduce by 10%, and the soft constraint weights are evenly distributed; when In low-confidence scenarios, only the core hard constraints of collision avoidance and endurance are retained, while the soft constraint weights are tilted towards the lowest risk, ensuring that drone safety is prioritized when environmental prediction uncertainty is high.

[0159] Specifically, the generative evolutionary game decision-making module.

[0160] This module uses the multi-path evolution scenario output by the uncertainty evolution modeling module. Potential risk event set output by the environmental causal cognition and counterfactual reasoning module Using the core input, multiple highly adaptable candidate planning schemes are generated through a generative adversarial network (GAN). A multi-participant evolutionary game mechanism is introduced, combined with the quantum phase evolution algorithm to find the global optimal solution. At the same time, the scheme self-evolution is achieved by relying on the scheme evolution knowledge base and self-supervised learning, ensuring the real-time performance, optimality and efficiency of the planning scheme in dynamic and complex environments.

[0161] The self-evolution mechanism of this scheme is as follows:

[0162] Step 1: Real-time iteration triggering. Every 30ms, the latest standardized environmental sensing dataset is generated based on the output of the intelligent metasurface-enhanced multi-source sensing module. and the uncertainty evolution modeling module generates Individual evolutionary environment scenarios:

[0163] ( Trigger generator Initiate candidate solution generation. The generator employs a Deep Convolutional Generative Adversarial Network (DCGAN) architecture. The input layer fuses environmental feature vectors, risk event vectors, and task state vectors, generating the solution through feature encoding, multi-scale feature fusion, and decoding. Group candidate planning scheme Each set of solutions includes path planning: Execution strategy:

[0164] (These are flight speed, payload power, and communication gain, respectively).

[0165] The generator loss function is designed as a multi-objective optimization, taking into account scenario adaptability, constraint satisfaction, and task reward. The formula is as follows:

[0166]

[0167] In the formula, , , , Weighting coefficients (satisfying) The initial values ​​were set to 0.3, 0.3, 0.25, and 0.15 respectively. Assign weights to scene adaptability loss; Weights are used to constrain satisfaction loss. Weighting of task reward and loss; The weights are used for regularization loss. The scene adaptability loss is calculated by using the feature similarity between candidate solutions and the evolving scene:

[0168]

[0169] To determine the constraint satisfaction loss, the degree of violation of the co-evolution constraint equations is quantified:

[0170]

[0171] The task reward loss is negatively correlated with the game payoff function:

[0172]

[0173] in, For the first The game payoffs of the candidate solutions. To achieve the theoretical maximum benefit; To regularize the loss, L2 regularization is used to avoid model overfitting:

[0174]

[0175] in, This represents the total number of parameters in the generator network. For the first One parameter.

[0176] After candidate solutions are generated, a multi-participant evolutionary game is initiated. The game participants include individual drones, environmental evolution agents, and mission objective agents. The game payoff function is defined. Taking into account the degree of task completion, resource consumption, risk avoidance, and constraint satisfaction:

[0177]

[0178] In the formula, , , , Weighted by returns (satisfying) ), , , , These represent the importance of task completion, resource conservation, risk avoidance, and constraint compliance in decision-making, respectively, with initial values ​​set to [0.4, 0.2, 0.3, 0.1], and are optimized based on feedback. Rewards for task completion (calculated based on task progress and the quality of goal achievement); Resource consumption costs (including energy consumption and range loss):

[0179]

[0180] in, To predict energy consumption, For maximum energy consumption, For path length, Average flight speed, The maximum driving range is calculated based on its energy capacity and average energy consumption per unit distance. Risk cost (calculated based on the probability of potential risk events occurring):

[0181]

[0182] in, The total number of risk events. For the first The probability of such risk events occurring The severity level of a risk event (value 1-10) is predefined based on the potential damage level of the risk event to task success and platform security. The constraint satisfaction degree (obtained by normalizing the calculation results based on the constraint equation, with a value of 0-1).

[0183] The Nash equilibrium is solved using a quantum phase evolutionary game algorithm, mapping the payoffs of each candidate solution to a quantum phase. The formula is:

[0184]

[0185] in, To achieve the theoretical minimum gain, the quantum phase evolution equation is:

[0186]

[0187] In the formula, The phase evolution rate (values ​​range from 0.8 to 1.2). Noise intensity (values ​​range from 0.01 to 0.05). Gaussian white noise (following the rules) The optimal phase is extracted every 100ms via quantum measurement. The corresponding optimized planning scheme is output. With execution strategy This ensures that the plan is synchronized with the real-time evolution of the environment.

[0188] Step 2: Reusing historical experience. Constructing a "Solution Evolution Knowledge Base". It adopts a distributed storage architecture to store the optimal solution for each game. , ), environmental evolution laws ( , ), constraint satisfaction status ( ) and game payoff The solution evolution knowledge base adopts a key-value pair storage structure, where the key is the environment metadata type. With global average uncertainty intensity Combined logo The value is the set of corresponding optimal solutions. ( (This represents the number of optimal solutions in similar scenarios).

[0189] To improve the retrieval efficiency of the solution evolution knowledge base, a hierarchical index structure is constructed: the first-level index is based on the environment meta-type. Classification, secondary index based on uncertainty intensity range During partitioning and retrieval, binary search is used to quickly locate the set of optimal solutions for similar scenarios, keeping the retrieval time complexity under time complexity. Simultaneously, a knowledge base update mechanism for scheme evolution is established. When the similarity between a newly added optimal scheme and an existing scheme exceeds a threshold... When the number of solutions in the same scenario exceeds a threshold, replace the original solution with the one that has the lowest similarity. At that time, retain the highest-yielding top performer. Each approach ensures the effectiveness and conciseness of the solution evolution knowledge base.

[0190] Step 3: Self-supervised learning optimization. This involves using a self-supervised learning model. right The data in the model is used for feature extraction and pattern discovery. The model adopts the Transformer architecture, and the input is the environmental feature vector.

[0191] The output is the feature vector of the optimal solution:

[0192] .

[0193] The self-supervised learning pretext task is designed as a "environmental feature - optimal solution" matching prediction, and the model is trained through contrastive learning to capture the mapping relationship between the two. The contrastive loss function is defined as follows:

[0194]

[0195] In the formula, This refers to the batch sample size. For cosine similarity, The temperature parameter (values ​​range from 0.05 to 0.1) is used. The model parameters are iteratively optimized using gradient descent. After training, the model can directly output the optimal solution prediction based on real-time environmental characteristics. .

[0196] Based on the difference between the predicted results and the actual optimal solution, a parameter optimization vector is generated. ,in, Adjustments to generator network parameters:

[0197]

[0198] In the formula, The learning rate (ranging from 0.001 to 0.01). These are the generator network parameters; For discriminator Adjustment amount of scoring criteria The weights are adjusted for the game payoff function. The generator is updated in real-time. Initial weights and discriminator The scoring criteria ensure that subsequent candidate solutions are generated more closely to the optimal solution distribution, reducing the number of game iterations by more than 60% and significantly improving decision-making efficiency. Simultaneously, the self-supervised learning model is incrementally trained every 500ms, incorporating the latest solution evolution knowledge base data to ensure continuous optimization of the model's generalization ability.

[0199] Specifically, a cluster-based awareness emerges in collaborative modules.

[0200] This module enables global intelligent collaboration among multiple drones, enhancing cluster robustness and task execution capabilities. It is suitable for multi-drone scenarios, while automatically skipping collaboration-related steps in single-drone scenarios. It builds global information support through distributed cognitive sharing, achieves dynamic division of labor by triggering consciousness emergence through cluster game theory, and combines robustness protection mechanisms to cope with faults and extreme environments, ensuring efficient cluster collaboration and task achievement in complex and dynamic scenarios.

[0201] (1) Distributed cognitive sharing and construction of a collective unconscious pool:

[0202] Step 1: Employing a secure communication link based on "millimeter wave + quantum key distribution" Develop a distributed cognitive sharing protocol This ensures the security and real-time nature of data transmission. The drone acts as a distributed node, sharing four core types of local data in real time: dynamic causal cognitive graph fragments. (No. Local dynamic causal cognition map of drones), environmental evolution scenario (Locally predicted environmental evolution path), candidate planning schemes (Locally generated candidate paths and execution strategies), game payoffs (Game payoff value of the local solution).

[0203] The quantum key update frequency of the communication link is synchronized with the sensing data update frequency (50ms / time). Key generation adopts a quantum random number generation mechanism based on the BB84 protocol to ensure the unbreakability of the key. Communication delay control is achieved through dynamic allocation of link bandwidth and data compression optimization, defining communication delay constraints:

[0204]

[0205] In the formula, The amount of data transmitted in a single transmission is reduced by more than 30% after compression using adaptive Huffman coding; To be assigned to the The communication bandwidth of the drone is dynamically adjusted based on data priority.

[0206]

[0207] in, This represents the total bandwidth of the link. For the first The data priority of the drone (calculated based on mission importance and data timeliness, with a value of 0-1); For signal propagation delay, calculations are based on the relative distance between UAVs:

[0208]

[0209] in, For the first The relative distance between the drone and the communication center. For the speed of light ( ).

[0210] Step 2: Aggregate shared data from all drones and extract global common features through a cross-attention mechanism. This feature vector encompasses 12 core dimensions, comprehensively depicting the global state of the cluster:

[0211]

[0212]

[0213] The calculation methods for each dimension are as follows:

[0214] Global environmental evolution trend : A weighted fusion based on all local evolution scenarios, with the weight being the scenario credibility. :

[0215] ;

[0216] Overall cluster resource status Including total range ,

[0217] Load availability (Values ​​range from 0 to 1)

[0218] Communication coverage ;

[0219] Global progress of the task ,in, For the first Priority weights for each subtask For the first The completion progress of each subtask (value 0-1);

[0220] Overall risk level This represents the highest probability of risk faced by the cluster.

[0221] Cluster size It reflects the number of drones currently operating normally in real time;

[0222] Overall task efficiency ,in, For the first The mission time for deploying the drone;

[0223] Total energy cost of the cluster After normalization, the value ranges from 0 to ;

[0224] Cluster survival probability The value ranges from 0 to 1.

[0225] Constructing a "collective unconscious pool" based on global features Among them, the global causal cognitive graph A graph fusion algorithm is used to merge all local fragments, combining node matching and edge strength weighting.

[0226]

[0227] In the formula, For nodes in the global graph arrive Causal strength weights For the first The corresponding weights of the local map of the drone. This serves as a local data validity indicator (1 for valid, 0 for invalid). The number of drones with valid data;

[0228] Global Environment Evolution Scenario Collection By deduplication and

[0229] The top-K most reliable evolutionary paths were selected based on their credibility ranking. This ensures the representativeness and reliability of the overall scenario.

[0230] (2) Cluster game and consciousness emergence mechanism:

[0231] Step 1: Multi-agent cooperative game. Each drone acts as an independent player:

[0232] ( , Given the number of drones in the cluster, and with the constraint of "maximizing the overall benefit of the cluster," the system calls [a specific function] when generating its own candidate solutions. Based on global data, it dynamically adjusts its own reward function to achieve synergistic optimization of individual and cluster interests:

[0233]

[0234] In the formula, Weighted by returns (satisfying) The initial values ​​are set to 0.3, 0.2, 0.2, 0.1, and 0.2 respectively, and can be dynamically optimized through online learning. Weighting of task completion rewards; Weighted by resource consumption cost; Risk cost weighting; Weights for the deviation between individual and global solutions; The contribution weights to the cluster are defined as follows:

[0235] For the first The reward for completing a drone mission (value 0-1) is calculated based on mission completion progress and the quality of goal achievement:

[0236] in, Values ​​range from 0 to 1, representing the quality of mission achievement (such as detection accuracy, data integrity, etc.).

[0237] Resource consumption cost (value 0-1), combining energy consumption and range loss:

[0238]

[0239] in, To predict energy consumption, For maximum energy consumption, For path length, For maximum driving range;

[0240] Risk cost (value 0-1), calculated based on the probability of potential risk events occurring:

[0241]

[0242] in, For the first The probability of such risk events occurring The severity of the risk (value 1-10);

[0243] The deviation between the individual solution and the globally optimal solution (values ​​0-1) is calculated using path similarity:

[0244]

[0245] in, To determine the optimal path for the cluster at the current moment, the path that maximizes the total cluster benefit is selected by comparing the contributions of each candidate drone solution to the global benefit.

[0246] The contribution of an individual to the cluster (value 0-100) is positively correlated with willingness to fill gaps and degree of resource sharing:

[0247]

[0248] in, The willingness to fill in the gap (calculated based on remaining resources and task compatibility, with a value of 0-1). This represents the degree of resource sharing (such as communication bandwidth, the proportion of shared sensing data, with a value of 0-1).

[0249] Step 2: Triggering Conditions for Emergent Consciousness. Two core indicators are calculated to determine whether clustered emergent consciousness has been triggered: clustered cognitive consistency. : Quantify the degree of matching between the local dynamic causal cognition graph of all drones and the global causal cognition graph, with a value range of 0-1:

[0250]

[0251] in, The graph matching degree is calculated by weighting node overlap rate and edge strength similarity:

[0252]

[0253] For the map The set of nodes, Let be the set of edges. The strength weight of the corresponding edge;

[0254] Global optimal strategy consensus : Quantify the degree of acceptance of the globally optimal strategy by all drones, with a value range of 0-100:

[0255]

[0256] in, For the first The drone scores the global optimal path (calculated based on scene adaptability and constraint satisfaction, with a value of 0-100).

[0257] when and When the cluster consciousness emerges, it enters a global collaborative decision-making mode; otherwise, it maintains a distributed decision-making mode, continues to optimize the local solution, and shares updates.

[0258] Step 3: Post-emergence Collaborative Behavior. After the emergence of cluster consciousness, it automatically differentiates into three functional groups, achieving dynamic division of labor and global collaboration:

[0259] Task execution group : Responsible for core task execution (such as target detection, data acquisition), and the selection criteria for members are as follows: and The number of drones in the group accounts for 60%-70% of the total number of drones in the cluster; the dynamic division of labor within the group is achieved through local game theory, and the sub-task allocation for each drone is based on capability adaptability.

[0260]

[0261] in, For the first The drone carried out the first The capability values ​​of each sub-task (calculated based on payload type and flight performance) are used to prioritize assigning the corresponding sub-tasks to the UAV with the highest capability adaptability.

[0262] Risk avoidance group : Detect and mark risk areas, plan safe routes, and select members based on the following criteria. and The obstacle avoidance success rate is ≥95%, and the number of drones in the group accounts for 20%-30% of the total number of drones in the cluster; the group adopts a "partition detection + path fusion" strategy, dividing the task area into... Each drone is responsible for risk detection in one sub-region, and the detection results are used to generate a global safe path through a path fusion algorithm.

[0263]

[0264] in, Points on the path Real-time risk value (calculated based on obstacles and interference intensity);

[0265] Collaborative replacement group Redundant resource reserves and fault coverage; member selection criteria are as follows: and The number of drones in the group accounts for 10% of the total number of drones in the cluster; real-time monitoring of drones in the group. and Based on the resource consumption rate, predict the replacement demand:

[0266]

[0267] in, As the energy consumption early warning threshold, This is an indicator function (1 if the condition is met, 0 otherwise), so prepare for filling in the gaps in advance.

[0268] Each group based on The system autonomously executes global state strategies, and groups synchronize key information through a cognitive sharing network at a synchronization frequency of 100ms per instance to ensure collaborative consistency. For example... After updating the risk area, the risk boundary data is synchronized to [the relevant data source] in real time. , Dynamically adjust flight paths to mitigate risks; according to Monitor task progress and resource consumption, plan backup paths in advance, and ensure seamless integration.

[0269] (3) Robust emergence guarantee:

[0270] Step 1: Individual Fault Adaptation. Through... Real-time monitoring of the operational status of each drone:

[0271] {Normal, Fault}, status monitoring indicators include communication link connectivity, sensor operating status, abnormal power system energy consumption, etc., with a monitoring frequency of 50ms / time. When a fault is detected in a certain drone ( (fault), cluster through The failed data of the drone is automatically removed, and the remaining drones are quickly redistributed for tasks based on cooperative game theory:

[0272] Task migration priority calculation: unfinished tasks for the faulty UAV Calculate the fit of each candidate replacement drone:

[0273]

[0274] In the formula, Weighting coefficients (satisfying) The initial values ​​are 0.4, 0.3, and 0.3 respectively. Compatibility weights between drones and faulty missions; Weight of remaining resource sufficiency; Weighting of distance fit to the faulty task area; For the first Compatibility of the UAV with the mission that failed (calculated based on payload capacity and mission type matching degree, with a value of 0-1). For the first The remaining resources of the drones ( (Values ​​range from 0 to 1). For the first Distance fit between the drone and the faulty mission area ( , This is the actual distance. The attenuation coefficient has a value of 0.01-0.05 and a value of 0-1.

[0275] Filling the gap decision: selection middle highest and The drone, acting as a backup, takes over the task from the faulty drone within ≤5ms and updates the globally optimal path. With execution strategy :

[0276]

[0277]

[0278] in, This is the original path of the malfunctioning drone. To fill in the gaps in the planned path of the drone, For strategy fusion operators (based on task priority weighted fusion execution strategy parameters);

[0279] Cluster Reconstruction: After the replacement is completed, the composition of each group is recalculated, starting from... or Select drones with sufficient resources and high contribution to supplement the existing ones. This ensures the redundancy reserve capacity of the backup group and maintains the robustness of the cluster.

[0280] Step 2: Collaborative Breakthrough in Extreme Environments. When extreme environments (such as global electromagnetic interference) are detected... A large number of obstacles emerged. The cluster's awareness automatically switches to "survival first" mode, and the benefit function for all drones is uniformly adjusted to:

[0281]

[0282] In the formula, Survival weight ( ), The individual's survival probability (value 0-1) is calculated based on the distance from the risk area and the effectiveness of avoidance strategies.

[0283]

[0284] in, The shape parameter is used to adjust the curves of the impact of distance and risk level on the probability of survival (values ​​range from 0.5 to 1.0). For the first The shortest distance between a drone and a high-risk area. Environmental risk level (value 0-1); The energy cost for breaking through (value 0-100) is calculated based on the breakout path length and energy consumption rate:

[0285]

[0286] in, The length of the breakout path. Energy consumption rate per unit distance.

[0287] A "dispersed avoidance - concentrated breakthrough" strategy is generated through multi-party game theory collaboration:

[0288] Distributed detection phase: Distributed at a uniform angle (angle between adjacent UAV detection directions) Disperse to different directions to detect the safety factor of the breakout route:

[0289]

[0290] In the formula, For the first The length of the channel, For the channel position The existence of obstacles (0 for none, 1 for present). For the channel position electromagnetic interference intensity, The maximum interference intensity threshold;

[0291] Concentrated Breakthrough Phase: and Break out along the safest route, maintaining formation spacing during the breakout. To avoid collisions; at the same time The rearguard, by adjusting flight attitude and releasing jamming signals, covers the main force's breakout until all drones leave the extreme environment area;

[0292] Post-breakthrough restructuring: After escaping the extreme environment, based on Reassess the cluster's status, repair damaged cognitive maps and evolutionary scenario data, adjust the composition and task allocation of each group, restore normal task execution mode, and ensure the overall survival probability of the cluster. .

[0293] Through the aforementioned distributed cognitive sharing, cluster game theory, and robustness guarantee mechanisms, this module enables global intelligent collaboration among multiple UAV clusters in complex and dynamic environments. This not only ensures the high efficiency of task execution but also enhances the cluster's adaptability to faults and extreme environments, providing core technical support for the large-scale application of multi-UAV systems.

[0294] Please refer to Figure 2 , Figure 2 A flowchart illustrating the UAV adaptive mission planning method integrating real-time environmental perception as described in this embodiment is shown. The overall process mainly includes the following three steps:

[0295] S1, Initialization Phase.

[0296] The initialization phase is the foundation for the closed-loop execution of the entire algorithm process. By accurately inputting core parameters, completing the calibration of the system hardware and software environment, and setting the initial parameters of the modules, it provides standardized startup conditions for subsequent real-time perception, cognitive modeling, and decision-making collaboration, ensuring that the algorithm has the ability to adapt to tasks and environments from the very beginning.

[0297] Input parameters need to cover two core types of information: the physical characteristics of the drone swarm and the mission objectives. All parameters are imported through a standardized interface, supporting both manual configuration and file import modes. The parameter format is uniformly JSON, which facilitates subsequent module calls and parsing.

[0298] Drone swarm parameters The specific definitions and quantification standards are as follows:

[0299] Cluster size : Refers to the total number of drones participating in the mission (value range 1- , (To the maximum cluster size supported by the hardware), it must meet the following requirements. Ensure that cluster communication and collaborative computing power are compatible;

[0300] Model type Based on function, they are divided into detection type, operation type, communication relay type, etc., and each type corresponds to a unique code. (For example, the detection type is coded as 01 and the operation type is coded as 02), and the mapping relationship between the code and the machine type is stored in the system configuration library;

[0301] Load capacity The maximum payload capacity of a single drone; payload capabilities of different drone models meet the requirements. ,in, The minimum load threshold is 0.1 kg. Maximum load threshold (5kg);

[0302] physical constraints It includes three types of core constraint parameters, all of which have been determined through hardware performance testing and manufacturer parameter calibration:

[0303] Battery life limit The maximum usable energy consumption of a drone when fully charged is calculated using the following formula: ,in, This is the battery's rated voltage. For battery capacity, Battery energy conversion efficiency (value ranges from 0.85 to 0.95);

[0304] Speed ​​range : The minimum stable flight speed (valued at 1-2 m / s). For the maximum flight speed (valued at 10-20 m / s), the speed constraint must satisfy aerodynamic principles, namely:

[0305] ,in, For the maximum thrust of the drone, air density, For windward area, This refers to the air drag coefficient;

[0306] Load limit The maximum payload weight that a drone can carry is positively correlated with the strength of its fuselage structure, and must meet certain requirements. ,in, For the drone's own weight, The acceleration due to gravity is 9.8 m / s². For the structural strength of the fuselage, To allow for stress.

[0307] Mission Objectives The global task needs to be finely decomposed and standardized in definition, as follows:

[0308] Subtask decomposition: The global task is decomposed into spatial regions and functional modules. Sub-tasks ( Each subtask contains the target location. Task Type (Such as exploration, surveying, and material delivery) Completion standards (such as detection accuracy, data integrity, etc.);

[0309] Time window : For subtask start time, The deadline for the subtask must be met. , Furthermore, the global time window must cover all subtasks, i.e. , (Total deadline for the task);

[0310] Priority weight Quantify the importance of subtasks to meet the following requirements. The value was determined using the Analytic Hierarchy Process (AHP), and the calculation process is as follows:

[0311] a. Construct a subtask importance judgment matrix ,in, Subtasks relative subtasks Importance (values ​​1-9 and their reciprocals);

[0312] b. Calculate the largest eigenvalue of the judgment matrix. With corresponding feature vectors ;

[0313] c. Perform a consistency check: ,like ( If the average random consistency index is used, then the normalized eigenvector becomes the priority weight. .

[0314] The system startup process requires hardware calibration and software environment initialization to ensure consistency and reliability in the collaborative operation of all modules.

[0315] The intelligent metasurface-enhanced multi-source sensing module is activated sequentially: the RGB-D visual sensor, lidar, weather sensor, electromagnetic interference detector, and GPS / BeiDou + IMU integrated navigation module are activated, and sensor self-tests are completed (such as lidar transmit power calibration and visual sensor white balance adjustment). The intelligent metasurface unit initiates beamforming initialization, sets the reference beam direction to the center of the mission area, the reference sensing frequency band to the default operating frequency band, and the reference signal gain to [value missing]. dB;

[0316] Communication Link Calibration: Activating a Secure Communication Link for Millimeter Wave + Quantum Key Distribution Link connectivity testing and delay calibration were completed, and the quantum key initialization used the BB84 protocol to generate the initial key pair. The key length is 256 bits, and the communication delay calibration formula is: ,in, To calibrate the distance between points, For the speed of light ( m / s), ensuring actual communication latency ms;

[0317] Solution evolution knowledge base initialization: Load the historical best solution template library, with templates categorized by environment meta type. The uncertainty intensity range is categorized and stored. Each template contains information such as the planning path, execution strategy, and constraint satisfaction status under typical scenarios. The solution evolution knowledge base index structure is as follows: ,in, This is the template storage address;

[0318] Initialization of the collective unconscious pool: ,in, This is the initial global feature vector (all dimensions are set to 0). Empty graph An empty set of scenes.

[0319] Parameter initialization requires setting key thresholds, weight coefficients, and conversion coefficients for each module. All parameters are stored in the system parameter configuration file, supporting online adjustment and iterative optimization.

[0320] Threshold parameter:

[0321] Redundancy Removal Threshold The value ranges from 0.05 to 0.2, depending on the data dimension. Adaptive adjustment, the formula is:

[0322] The higher the data dimension, the smaller the threshold, to ensure the accuracy of redundant recognition;

[0323] Risk assessment threshold Values ​​range from 0.6 to 0.8, based on the task security level. (Levels 1-5) confirmed. The higher the security level, the larger the threshold and the stricter the risk assessment.

[0324] Cognitive confidence threshold The value ranges from 0.7 to 0.85, depending on the environmental complexity. (0-1) Calculation, The more complex the environment, the lower the threshold, adapting to dynamic cognitive needs;

[0325] Interference evolution trigger threshold Value range: -80 to -60 dBm, based on reference interference intensity. set up, dBm ensures sensitivity to interference evolution triggering;

[0326] Benchmark security threshold Values ​​range from 2-5m, depending on the size of the drone. Positive correlation , This is the largest model size;

[0327] Extreme electromagnetic interference threshold The value is set to -40dBm, which is the maximum statistical value of electromagnetic interference intensity.

[0328] Large-scale obstacle threshold Values ​​range from 0.6 to 0.8, based on the area of ​​the task region. Sure, , This represents the total area of ​​the obstacles.

[0329] Weighting coefficients:

[0330] : The weighting coefficients for the intelligent metasurface sensing enhancement mechanism, satisfying The initial values ​​were 0.4, 0.3, and 0.3, respectively, and adaptive optimization was performed using the entropy weight method.

[0331] : These are the weight coefficients of the cluster game payoff function, with initial values ​​of 0.3, 0.2, 0.2, 0.1, and 0.2, respectively, and can be dynamically adjusted based on task type;

[0332] : This is the weighting coefficient for the survival-first mode, fixed at 0.9, to ensure the core position of the survival objective;

[0333] Phase conversion coefficient The value ranges from 0.8 to 1.2, used in the quantum phase evolution game algorithm. The initial value is 1.0, and it is dynamically optimized based on the game convergence speed. , The convergence speed of the game is (0-2).

[0334] S2, Closed-loop execution phase.

[0335] The closed-loop execution phase is the core link in the algorithm's realization of dynamic iteration across the entire chain of "perception-cognition-modeling-decision-coordination-feedback". Each step is executed in an orderly manner according to a fixed cycle or real-time response mechanism. Through data interaction and parameter optimization between modules, the planning scheme is continuously adapted to dynamic environment and task changes, ensuring the efficiency, safety and robustness of UAV mission execution.

[0336] Step 1: Multi-source sensing and preprocessing (50ms / time). The intelligent metasurface-enhanced multi-source sensing module synchronously collects multi-dimensional data at a preset frequency, covering physical environment data (terrain, obstacles, weather, electromagnetic interference), airspace rule data (airspace permissions, control instructions), and task-related data (target dynamics, task progress), forming a multi-source data input of "local fine-grained + global coarse-grained". The data acquisition frequency is precisely matched with the sensor type: visual sensor 30fps, lidar 10Hz, weather sensor 5Hz, electromagnetic interference detector 20Hz, ensuring data timeliness and integrity.

[0337] Intelligent metasurface units based on environmental scene feature vectors Dynamically optimize the sensing parameters and obtain a standard feature vector through feature normalization. :

[0338]

[0339]

[0340]

[0341]

[0342] in, This refers to the actual wind speed. This represents the actual electromagnetic interference intensity. For actual visibility, This represents the probability of precipitation (value 0-1). m / s m / s, dBm, dBm, km km is the statistical extreme value. Based on Build environment scene adaptation coefficient :

[0343]

[0344] In the formula, The weighting coefficients adaptively determined by the entropy weighting method (satisfying) Under strong interference environment ( ), electromagnetic detector signal gain Dynamic adjustment:

[0345]

[0346] In foggy weather ( ), lidar transmission power Adjustment:

[0347]

[0348] in, The reference gain is 10-20dB. The reference power is 5-15W.

[0349] Data preprocessing employs a "nuclearized causal mutual information maximization" algorithm to remove redundancy and spurious associations. First, it calculates the data for any two classes. nucleated causal mutual information :

[0350]

[0351] In the formula, For sample size, For matrix trace operations, It is a Gaussian kernel matrix. For kernel entropy:

[0352]

[0353]

[0354] Set redundancy threshold (0.05-0.2), when Redundant data was removed in a timely manner. This was achieved by improving the Pearson correlation coefficient. With causal direction index Identifying spurious associations:

[0355]

[0356]

[0357] in, This is the time lag (0-50ms). For the set of condition variables, For conditional causal mutual information. When and The association was initially identified as spurious. Ultimately, the environment-aware dataset was standardized by improving the Z-score. :

[0358]

[0359] In the formula, This is the original data. , These are the moving mean and standard deviation of the data from the previous second, respectively.

[0360] Step 2: Environmental causal cognition (30ms / time). Based on By using an improved PC algorithm to mine directed causal relationships, a dynamic causal cognitive graph is constructed. . It covers core nodes related to environment, resources, and tasks. For directed causal edges and related causal mechanisms, The causal strength weight is dynamically updated based on real-time trigger strength, historical confidence, causal mechanism fit, and complexity.

[0361]

[0362] In the formula, Weighting coefficients (satisfying) ); The real-time trigger strength of the causal chain (calculated from the node activation level). ( (smoothing coefficient) ( ), ( The length of the causal chain. (Node association density).

[0363] Counterfactual reasoning, based on structural causal models (SCM), predicts potential risks and calculates risk events. Probability of occurrence:

[0364]

[0365] In the formula, The intensity of the strategic intervention is (-1 to 1). This is the inherent trigger probability. ( (For environmental dynamics). Set risk thresholds. (0.6-0.8), marked High-risk event set .

[0366] Extracting global feature vectors from the graph using a meta-learning model (Including 16 core dimensions) Identify environmental meta-types And invoke the corresponding cognitive paradigm Calculate cognitive confidence. :

[0367]

[0368] when When the value is between 0.7 and 0.85, the reconfiguration of the metasurface sensing parameters and the reconstruction of the causal chain are initiated to correct cognitive biases.

[0369] Step 3: Uncertainty Evolution Modeling (30ms / cycle). First, calculate the environmental dynamics. Adaptive adjustment of the mesh size of the Uncertainty Cellular Automaton (U-CA) :

[0370]

[0371] In the formula, Weighting coefficients (satisfying) ), The average moving speed of the obstacle. , This represents the change in disturbance intensity and wind speed per unit time. according to (5m×5m) (10m×10m) (20m×20m) Adaptive adjustment.

[0372] Define cell Three-dimensional states: basic state (Attributes such as terrain, disturbances, and obstacles), evolution rules (Condition-action pairs generated based on causal chains), Uncertainty intensity :

[0373]

[0374] In the formula, Weighting coefficients (satisfying) ), To sense noise, Due to the uncertainty of causal chains, For global random perturbation coefficients.

[0375] Based on U-CA simulation of the next 5-15 seconds Evolutionary path ( Depend on (Adaptive adjustment), calculating the credibility of each path. :

[0376]

[0377] Construct a co-evolutionary constraint equation for "task-resource-environment":

[0378]

[0379] In the formula, , , This is the aggregate value of hard constraints (the minimum value of collision avoidance, endurance, and airspace permission constraints). The aggregated value of soft constraints (a weighted sum of constraints for optimal energy consumption, task efficiency, and minimum risk) is based on Dynamically adjust constraint thresholds.

[0380] Step 4: Generative Evolutionary Game Decision Making (output every 100ms). To the generator... (DCGAN architecture) Input and Inject Gaussian noise generate Group of candidate schemes (Including path and execution strategy). The generator loss function is:

[0381]

[0382] In the formula, Weighting coefficients (satisfying) );

[0383] , , , .

[0384] After candidate solutions are generated, a multi-participant evolutionary game is initiated, with the payoff function as follows:

[0385]

[0386] in, Weighting of task completion rewards; Weighted by resource consumption cost; Risk cost weighting; The weights are used to constrain satisfaction.

[0387] The Nash equilibrium is solved using a quantum phase evolution game algorithm, and the payoffs are mapped to quantum phases. :

[0388]

[0389] The quantum phase evolution equation is:

[0390]

[0391] Extract the optimal phase every 100ms Output and Store the relevant data in the scheme evolution knowledge base. Through self-supervised learning models Generate parameter optimization vector ,renew With discriminator parameter.

[0392] Step 5: Swarm awareness emerges and collaboration (real-time execution). In a single drone scenario, execution is performed directly. and Real-time feedback on task execution status (Completed / In Progress / Abnormal).

[0393] In multi-drone scenarios, drones pass through Share local cognitive data (causal graph fragments, evolutionary scenarios, candidate solutions, game payoffs) to update the collective unconscious pool. Calculating cognitive consistency Consensus with global strategy :

[0394]

[0395]

[0396] when and At that time, a cluster consciousness emerged, and the cluster differentiated into (Task Execution Group) (Risk Avoidance Group) (Collaborative support group), each group dynamically divides tasks and synchronizes key information.

[0397] pass Real-time monitoring of drone operating status When a fault is detected, the suitability of the candidate replacement drone is calculated. ,choose middle highest and The drone completes task takeover and updates the global plan within ≤5ms. In extreme environments, it switches to "survival-first" mode and adjusts the reward function. The "dispersed detection-concentrated breakthrough" strategy ensures a cluster survival probability of ≥75%.

[0398] Step 6: Feedback Iteration. Each time a subtask is completed ( Completed), collect task execution data. Including task completion rate Energy consumption Obstacle avoidance success rate Risk event occurrence .based on Reverse optimization of module parameters:

[0399]

[0400]

[0401]

[0402] In the formula, Update the parameter step size (0.01-0.05). For the total benefit of the cluster, For overall task efficiency. Update. and Determine the completion rate of all subtasks. Finish If yes, proceed to the task termination stage; otherwise, return to step 1 to continue the closed-loop iteration.

[0403] S3, Task Termination Phase.

[0404] The task termination phase is the final stage of the algorithm's closed-loop execution. The termination timing is determined through multi-dimensional quantitative verification. The system sorts out the data of the entire task execution process and outputs standardized results. Hardware and software resources are shut down according to the standard procedures to ensure the integrity, traceability and security of the task execution, while providing optimization reference for subsequent similar tasks.

[0405] The termination condition determination adopts a real-time parallel verification mechanism. The task will terminate when any of the following quantification conditions are met. All determination logic is calculated independently and the results are synchronized in real time:

[0406] Subtask completion rate target achieved: The overall completion rate is calculated based on the weighted average of subtask priorities, using the following formula:

[0407]

[0408] In the formula, For the first The completion rate of each subtask at the termination time (values ​​range from 0 to 1, calculated based on the task completion standard and the actual achievement). For the first The priority weights of each subtask (satisfying) );when When the subtask completion rate reaches the target, it is determined that the subtask completion rate has met the target.

[0409] The cluster survival probability is too low: The overall survival probability of the currently operating drones is calculated using the following formula:

[0410]

[0411] In the formula, The number of drones that were operating normally at the time of termination. For the first The real-time survival probability of a normal drone (valued between 0 and 1, calculated based on remaining resources, distance from the risk area, and effectiveness of avoidance strategies); when At that time, to ensure equipment safety, the termination condition was determined to be met.

[0412] Task Timeout: Extract the deadline for all subtasks. Determine the total deadline for the task. Set timeout tolerance threshold (Value range 0-30s, emergency rescue mission) s, routine tasks s), when the actual termination time If the timeout occurs, the task is deemed to have timed out.

[0413] The output stage generates two types of standardized output files, stored in JSON format and supporting local encrypted storage and cloud backup. The file structure is standardized and the field definitions are uniform.

[0414] Task execution report It comprehensively records key information throughout the entire task execution process, and its core includes four major modules:

[0415] a. Summary of planning schemes: in chronological order ( , (To record the optimal planning path at each stage, representing the number of decision-making steps) With execution strategy (Representing flight speed, payload power, and communication gain), and simultaneously labeling the evolution scenario identifiers corresponding to each scheme. Explanation of risk avoidance;

[0416] b. Quantitative data on execution performance: Statistics on core performance indicators, including average task completion quality. (Values ​​range from 0 to 1) Total energy consumption of the cluster Average obstacle avoidance success rate (Values ​​range from 0 to 1), average decision delay ( For the first (time of triggering the next decision);

[0417] c. Risk Event Handling Log: Record the time of occurrence of each risk event. Type encoding (e.g., collision risk code 01, communication interruption risk code 02), probability of occurrence Response strategies and treatment effect (Values ​​range from 0 to 1, where 1 indicates complete avoidance and 0 indicates no avoidance).

[0418] d. Description of abnormal situations: Record abnormal events such as equipment failure, communication interruption, and sudden environmental changes, including the time of occurrence, scope of impact (local / global), handling measures, and recovery results.

[0419] Algorithm optimization parameter set : Stores the final parameters of each module after feedback and iterative optimization, covering:

[0420] a. Threshold parameter: (Redundancy removal threshold) (Risk assessment threshold) (Cognitive confidence threshold) (Interference evolution trigger threshold) (Baseline safety threshold) etc.;

[0421] b. Weighting coefficients: , , (Supersurface sensing enhancement weights) (Cluster game payout weights) (Survival priority mode weights), etc.;

[0422] c. Conversion and adaptation parameters: (Phase conversion coefficient) (Self-supervised learning rate), intelligent metasurface beam direction and gain adaptation parameters, causal intensity weight calculation coefficients, etc.

[0423] The system shutdown process is executed in an orderly manner according to a fixed procedure to ensure the safe release of resources and data integrity.

[0424] Sensing system shutdown: Sequentially shut down the lidar, RGB-D vision sensor, weather sensor, electromagnetic interference detector, and GPS / BeiDou + IMU integrated navigation module. Reset the intelligent metasurface unit to its initial state (beam direction aligned with the center of the mission area, signal gain adjusted to the reference value). dB, sensing frequency band switched to standby frequency band);

[0425] Communication link shutdown: Terminate the secure communication link for "millimeter wave + quantum key distribution" The quantum key destruction process is executed, sensitive data in the link cache is cleared, and the power to the communication module is turned off.

[0426] Data verification and storage: and Data is synchronously stored in a local encrypted storage unit and a cloud backup server. Data integrity is ensured through hash verification, using the following formula:

[0427]

[0428] Compare the hash values ​​calculated locally Hash value stored in the cloud If both are consistent, the data storage is considered complete.

[0429] 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. An adaptive mission planning method for unmanned aerial vehicles (UAVs) that integrates real-time environmental perception, characterized in that, The method includes the following steps: S1. Initialization Phase: Input the physical parameters of the UAV cluster and the mission target parameters, and complete the initialization settings of system hardware calibration, software environment and key module parameters; S2. Closed-loop execution phase: Based on the "perception-cognition-modeling-decision-collaboration-feedback" closed-loop technical architecture, the following sub-steps are executed until the task termination conditions are met: S2.1 Multi-source sensing and data preprocessing: The intelligent metasurface-enhanced multi-source sensing module collects multi-dimensional environmental, spatial, and task data in real time, and uses the nucleated causal mutual information maximization algorithm to perform redundancy removal and false association filtering, outputting a standardized environmental sensing dataset. S2.2 Environmental Causal Cognition and Risk Prediction: Based on the environmental perception dataset, a dynamic causal cognition map is constructed using an improved PC algorithm, potential risks are predicted through a counterfactual reasoning model, and a high-risk event set is generated; S2.3 Uncertainty Evolution Modeling: Based on the dynamic causal cognitive map and high-risk event set, the uncertain cellular automata are used to simulate the future multi-path evolution trend of the environment, and a "task-resource-environment" co-evolution constraint equation is constructed; S2.4 Generative Evolutionary Game Decision-Making: Taking the future multi-path evolution trend of the environment and the set of high-risk events as input, multiple candidate planning schemes are generated through generative adversarial networks, and the globally optimal planning scheme is output based on the multi-participant evolutionary game mechanism combined with the quantum phase evolution algorithm. S2.5, Cluster Awareness Emergence, Coordination and Execution: For multi-drone scenarios, local cognitive data is shared through a secure communication link to build a collective unconscious pool. When the conditions of cognitive consistency and policy consensus are met, cluster awareness emerges to achieve dynamic division of labor and collaborative task execution. For single-drone scenarios, the optimal planning scheme is executed directly. S2.6 Feedback Iteration: Based on task execution feedback data, the dynamic causal cognitive graph, evolution rules, and game weight parameters are optimized in reverse, and the collective unconscious pool and solution evolution knowledge base are updated. S3. Task Termination Phase: When one of the following conditions is met—subtask completion rate reaches the target, cluster survival probability is too low, or task timeout—the task is terminated, the task execution report and algorithm optimization parameter set are output, and the system is shut down in an orderly manner.

2. The UAV adaptive mission planning method integrating real-time environmental perception as described in claim 1, characterized in that, In step S2.1, the intelligent metasurface enhanced multi-source sensing module performs adaptive sensing optimization of the environment by dynamically adjusting the beam direction, sensing frequency band and signal gain. Specifically, this includes: calculating the scene adaptation coefficient based on the environmental scene feature vector, dynamically adjusting the electromagnetic detector signal gain in a strong interference environment, and dynamically adjusting the lidar emission power in a foggy environment.

3. The UAV adaptive mission planning method integrating real-time environmental perception as described in claim 1, characterized in that, In step S2.2, the dynamic causal cognitive graph adopts a triplet structure, including: a set of causal nodes, a set of directed causal edges, and causal strength weights that are dynamically updated based on real-time trigger strength, historical confidence, mechanism adaptability, and complexity.

4. The UAV adaptive mission planning method integrating real-time environmental perception according to claim 1, characterized in that, In step S2.3, the uncertain cellular automaton divides the task region into a dynamic cellular grid. Each cell is defined with a basic state, an evolutionary rule generated based on a causal chain, and a cellular uncertainty intensity that integrates perceived noise, causal uncertainty, and environmental disturbance. The credibility of the environmental evolution path is calculated based on the conformity between the evolution scenario and the causal chain and the average uncertainty intensity of the path.

5. The UAV adaptive mission planning method integrating real-time environmental perception according to claim 1, characterized in that, In step S2.4, the generator of the generative adversarial network is trained using a multi-objective loss function, the loss term of which includes scene adaptability loss, constraint satisfaction loss, task payoff loss and regularization loss; the quantum phase evolution algorithm maps the game payoff of candidate schemes to quantum phase, and solves the Nash equilibrium through the phase evolution equation.

6. The UAV adaptive mission planning method integrating real-time environmental perception according to claim 1, characterized in that, In step S2.5, the triggering condition for the emergence of cluster consciousness is that both the cluster cognitive consistency and the global optimal strategy consensus exceed a preset threshold; after emergence, the cluster automatically differentiates into task execution group, risk avoidance group and collaborative support group to achieve dynamic functional division of labor.

7. The UAV adaptive mission planning method integrating real-time environmental perception according to claim 1, characterized in that, Step S2.5 also includes a robustness guarantee mechanism: by monitoring the status of the UAV in real time, when an individual failure is detected, the task migration and cluster reconstruction are carried out quickly based on the replacement adaptability; when encountering extreme environments, the "survival first" mode is switched to ensure the survival of the cluster through distributed detection and concentrated breakthrough strategies.

8. An adaptive mission planning system for unmanned aerial vehicles (UAVs) integrating real-time environmental perception, characterized in that: The system is used to implement the UAV adaptive mission planning method that integrates real-time environmental perception as described in any one of claims 1 to 7, and the system includes: A smart metasurface-enhanced multi-source sensing module is used to collect and preprocess multi-dimensional environmental data in real time. The Environmental Causal Cognition and Counterfactual Reasoning module is used to construct a dynamic causal cognition map and make risk predictions; The uncertainty evolution modeling module is used to simulate the multi-path evolution trend of the environment and construct cooperative evolution constraints; The generative evolutionary game decision-making module is used to generate candidate solutions and output the optimal plan through game optimization. The cluster consciousness emergence collaboration module is used to realize cognitive sharing, consciousness emergence, and collaborative execution among multiple drones; The feedback iterative optimization module is used to optimize system parameters based on execution feedback. The initialization and task management module is used for system initialization, task flow control, and termination processing.

9. The UAV adaptive mission planning system integrating real-time environmental perception according to claim 8, characterized in that, The intelligent metasurface-enhanced multi-source sensing module integrates an intelligent metasurface unit, an RGB-D vision sensor, a lidar, a meteorological sensor, an electromagnetic interference detector, and a combined navigation module, forming a multi-source data input system of "local fine-grained sensing + global coarse-grained assistance".

10. The UAV adaptive mission planning system integrating real-time environmental perception according to claim 8, characterized in that, The system also includes: a scheme evolution knowledge base, used to store historical optimal planning schemes and their corresponding environmental characteristics, evolution scenarios and game payoffs, and to mine the mapping relationship between environmental characteristics and optimal schemes through a self-supervised learning model in order to optimize the parameters of the generative evolutionary game decision module.

Images