Distributed system for dynamic perception and adaptation of uav swarm environment

Abstract

The application discloses a distributed system for dynamic perception and adaptation of a UAV cluster environment, and belongs to the technical field of computers.The distributed system comprises a perception layer, a data processing layer, a decision layer, an execution layer and a feedback layer.The perception layer is used for collecting state information of each node in the UAV cluster in real time.The data processing layer is used for preprocessing and analyzing the collected UAV node state information.The decision layer generates an adjustment decision for the UAV cluster environment based on a preset UAV cluster adaptation strategy library and the dynamic characteristic data.The execution layer performs corresponding adjustment operations on the UAV cluster nodes according to the adjustment decision.The feedback layer is used for collecting new state information after the adjustment operations on the UAV nodes.The application realizes innovative integration and layered cooperation of multiple technical modules, and maintains the order of the cluster through dynamic balance of the strategy library, so that a self-growth closed loop of 'perception-decision-execution-feedback' is formed.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a distributed system for dynamic perception and adaptation in unmanned aerial vehicle (UAV) swarm environments. Background Technology

[0002] In recent years, drone swarms have been widely used in emergency rescue, power line inspection, and other fields due to their advantages such as collaborative operation and flexible task allocation. Drone swarms, through the coordinated efforts of multiple drones, can achieve efficient operations in complex environments. For example, they can quickly build 3D maps of disaster areas in emergency rescue and automate the monitoring of long-distance power transmission lines in power line inspections.

[0003] However, existing drone swarm systems have many shortcomings in terms of dynamic environmental perception and adaptation capabilities. At the environmental perception level, traditional perception methods often focus only on single-dimensional data, such as monitoring only the drone's location information. It is difficult to comprehensively collect multi-source information on drone hardware resources (CPU, memory, disk I / O, etc.), software operating status (service processes, response time, etc.), and network communication status (latency, throughput, etc.), resulting in an incomplete understanding of the overall operating status of the swarm and an inability to promptly detect potential resource bottlenecks or fault hazards.

[0004] In terms of data processing, with the expansion of drone swarm scale and the increase in task complexity, the amount of collected data has increased dramatically and exhibits characteristics of multi-source and heterogeneity. Existing data processing technologies are struggling to efficiently clean, analyze, and integrate massive amounts of data. Especially in federated learning, it is difficult to balance data privacy protection and collaborative efficiency. At the same time, there is a lack of effective means to deeply mine data to accurately extract dynamic environmental characteristics, thus failing to provide reliable basis for decision-making.

[0005] In the decision-making and execution phases, traditional decision-making algorithms suffer from low efficiency and poor optimization results when faced with complex NP-hard problems such as resource scheduling and task allocation in drone swarms, making it difficult to adapt to dynamically changing environmental requirements. During execution, the management, service control, and network adjustments of drone nodes lack flexibility and intelligence, failing to achieve precise resource allocation and efficient task collaboration. Furthermore, most existing systems lack a robust feedback mechanism, making it impossible to adjust strategies promptly based on execution results and form a closed-loop dynamic optimization system.

[0006] In summary, the shortcomings of existing UAV swarm systems in terms of dynamic environmental perception and adaptation capabilities severely limit their application effectiveness and scope of application in complex scenarios. Therefore, we propose a distributed system for dynamic environmental perception and adaptation in UAV swarms. Summary of the Invention

[0007] The purpose of this invention is to provide a distributed system for dynamic perception and adaptation in unmanned aerial vehicle (UAV) swarm environments, in order to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A distributed system for dynamic perception and adaptation in UAV swarm environments comprises a perception layer, a data processing layer, a decision-making layer, an execution layer, and a feedback layer distributed across the nodes of the UAV swarm, forming a closed-loop dynamic adaptation architecture; wherein:

[0010] The perception layer consists of sensor modules distributed across each UAV node, used to collect node status information in real time and transmit it to the data processing layer via a distributed network. The node status information includes hardware resource information, software running status information, network communication status information, network adaptation parameters of UAV nodes participating in federated learning, and multi-dimensional basic data required by the digital twin engine, and is transmitted via a distributed network.

[0011] The data processing layer, connected to the perception layer, is used to preprocess, protect privacy, fuse and analyze raw state information, and generate dynamic feature data. It includes a data cleaning module, a federated learning privacy protection module, a digital twin fusion module, and a data analysis module. The federated learning privacy protection module employs a differential privacy unit that dynamically adjusts noise intensity based on data sensitivity, using Laplace noise for highly sensitive data and Gaussian noise for low-sensitivity data, while meeting a preset privacy budget. A homomorphic encryption unit encrypts local model parameters and transmits them to the aggregation node, supporting encrypted aggregation and encrypted distribution of global parameters. A dynamic node management unit allocates temporary keys, adjusts noise and encryption parameters when a new node joins, and cancels the key, clears temporary parameters, and triggers incremental encrypted updates to the global model when a node exits.

[0012] The decision layer, connected to the data processing layer, generates decisions on adjustment, service migration, or negative entropy injection based on the strategy library and dynamic feature data; it includes a strategy library for storage resource adjustment / service migration / negative entropy injection strategies, and a quantum-inspired optimizer for solving NP-hard problems based on the quantum annealing algorithm;

[0013] An execution layer, connected to the decision layer, is used to execute the adjustment decisions generated by the decision layer; it includes a hardware resource adjustment module, a service control module, and a network communication adjustment module.

[0014] The feedback layer, which is connected to the execution layer and the data processing layer, is used to collect new state information after the execution decision and feed it back to the data processing layer. If the deviation between the new state and the prediction exceeds the threshold, the data processing layer is triggered to perform incremental training of the model and update the policy library, forming a self-growing closed loop.

[0015] Preferably, the sensor module of the perception layer includes hardware sensors and software sensors; the hardware sensors are used to collect hardware resource information such as CPU utilization and memory usage; the software sensors are used to collect software operating status and network communication status information such as service response time and network latency.

[0016] Preferably, the data cleaning module employs differentiated processing for different data types: median filtering is used to denoise continuous time series data, and interpolation is used to complete and repair discrete abnormal data, ultimately unifying them into the Protocol Buffers standard format.

[0017] Preferably, the federated learning privacy protection module adds dynamic noise to the data through a differential privacy algorithm and combines it with homomorphic encryption to achieve secure ciphertext transmission, supporting edge nodes to dynamically join / leave the training cluster.

[0018] Preferably, the digital twin fusion module is based on the LSTM-PPO hybrid model, which integrates historical 1-hour load data and real-time status data to generate a load trend prediction for the next 15 minutes and simulate strategy risks.

[0019] Preferably, the quantum-inspired optimizer is based on the quantum parallelism of the quantum annealing algorithm to solve task allocation problems involving 100+ nodes and outputs fine-grained instructions with resource allocation precision. The quantum-inspired optimizer is used to efficiently solve NP-hard problems of cluster resource allocation, task scheduling, and service migration path planning. Specifically, it includes a problem modeling module that transforms the actual problem into an energy function containing decision variables, objective function, and constraint penalty terms.

[0020] The core module of quantum annealing achieves global optimization through quantum state initialization, tunneling process simulation, and annealing scheduling.

[0021] The classical-quantum interface module adopts a hybrid architecture of quantum inspiration and classical computing, which is adapted to the limited computing power of drone nodes;

[0022] The result verification and correction module performs feasibility verification and local correction on the output optimal solution.

[0023] It also features lightweight, distributed collaboration, and dynamic response capabilities, allowing the annealing process to be interrupted and restarted based on changes in cluster topology.

[0024] Preferably, the feedback data processing module of the feedback layer compares the new state information with the digital twin prediction results. If the load prediction error is > 15%, it triggers incremental training of the LSTM-PPO model; if the policy execution deviation exceeds the threshold, it automatically optimizes the policy library parameters.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] (1) This invention constructs a distributed system for dynamic perception and adaptation of UAV swarms. Through the collaborative collection of five categories of state information by multi-source sensors in the perception layer, combined with differential privacy protection in the data processing layer, LSTM-PPO digital twin modeling, and a decision layer driven by quantum annealing algorithm, it realizes intelligent processing of the entire process from data acquisition, privacy enhancement, trend prediction to millisecond-level quantum optimization decision-making. The core breakthrough lies in introducing a system entropy change model to maintain the swarm's orderliness, using quantum parallelism to shorten the decision time of NP-hard problems by more than 90%, and using a digital twin "trial and error sandbox" to avoid strategy risks in advance, forming a self-growing closed loop of "perception-decision-execution-feedback".

[0027] (2) The system of this invention achieves high anomaly detection accuracy through spatiotemporal cross-validation and multi-source data fusion. It automatically matches strategies for different task types (inspection / rescue), reducing the frequency of manual intervention, improving emergency response speed, and achieving a high cluster task completion rate. Furthermore, it achieves "usable but invisible" sensitive data through differential privacy and homomorphic encryption. Its innovative architecture demonstrates unique value in scenarios such as disaster relief and intelligent inspection, meeting the requirements for low latency and anti-interference, and improving the efficiency of civilian inspections through dynamic task allocation, providing an efficient and secure technical foundation for the large-scale application of distributed UAV clusters. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Example

[0030] A distributed system for dynamic perception and adaptation in UAV swarm environments comprises a perception layer, a data processing layer, a decision-making layer, an execution layer, and a feedback layer distributed across the nodes of the UAV swarm, forming a closed-loop dynamic adaptation architecture; wherein:

[0031] The perception layer, composed of sensor modules distributed across each UAV node, is used to collect node status information in real time and transmit it to the data processing layer via a distributed network. This node status information includes hardware resource information (CPU utilization, memory usage, disk I / O rate, GPU computing load), software runtime status information (service process status, service response time, request error rate, container resource quota), network communication status information (inter-node network latency, throughput, packet loss rate, TCP connection count), network adaptation parameters of the UAV nodes participating in federated learning (communication bandwidth threshold, encrypted transmission latency tolerance), and multi-dimensional basic data required by the digital twin engine (hardware configuration fingerprint, software environment mirror hash value). By collecting multi-dimensional status information in real time through sensor modules distributed across each UAV node, the system provides raw data support for decision-making and transmits the data via a distributed network.

[0032] Federated Learning Overall Architecture Design

[0033] This system adopts the FedAvg framework to build a two-level architecture of "central node - edge node", which is adapted to the dynamic and distributed characteristics of drone swarms. The specific architecture is as follows:

[0034] Central node: Deployed on the ground command center server (CPU: Intel Xeon 8375C, GPU: NVIDIA A100), responsible for global model initialization, parameter aggregation, encryption key management, and edge node admission / exit control;

[0035] Edge nodes: These are the individual drone nodes, which are divided into "training nodes" (20% computing power nodes, equipped with NVIDIA Jetson AGX Orin) and "data nodes" (80% communication / task nodes) based on their functions. Training nodes are responsible for local data training, while data nodes only provide cleaned data (and do not participate in model training).

[0036] Communication Link: The system adopts a dual-link approach of "Mesh self-organizing network + 5G". The transmission of model parameters between the training node and the central node is prioritized through 5G (bandwidth ≥ 100Mbps). When the 5G signal is weak, it automatically switches to the Mesh network (bandwidth ≥ 20Mbps) to ensure communication reliability.

[0037] Federated Learning Core Process (Complete Training Cycle)

[0038] Taking "Unmanned Aerial Vehicle (UAV) Cluster Load Prediction Model Training" as an example, the complete training cycle (1 round) consists of 5 stages, with a total time of ≤10 minutes. The specific steps are shown in Table 1 below:

[0039] Table 1 Training of the UAV swarm load prediction model

[0040] ;

[0041] The hardware resource information of the drone nodes includes CPU utilization, memory usage, disk I / O rate, and GPU computing load. The software runtime status information includes the process status of the services carried by the drone, service response time, request error rate, and container resource quota. The network communication status information includes network latency, throughput, packet loss rate, and number of TCP connections between drone nodes. Through proactive monitoring mechanisms such as GPU computing load anomaly warnings and container resource quota overrun detection, passive fault handling is transformed into proactive risk avoidance, improving cluster availability. This provides high-quality raw input to the data processing layer, directly impacting the accuracy of resource scheduling at the decision-making level.

[0042] Specifically, the sensor module of the perception layer includes hardware sensors and software sensors. The hardware sensors are used to collect hardware resource information such as CPU utilization and memory usage. The software sensors are used to collect software operation status and network communication status information such as service response time and network latency. The hardware sensors (such as GPU temperature sensors) and software sensors (computing load monitoring plugins) adopt a spatiotemporal cross-verification mechanism: when the GPU temperature exceeds 85℃ and the computing load is >90% for 5 consecutive minutes, it is judged as an overload risk; if the data from a single sensor is abnormal (such as normal temperature but sudden increase in load), it triggers a comparison of data from multiple nodes within 10 seconds to eliminate occasional noise interference.

[0043] The data acquisition process includes: software runtime status information for drone nodes (service process status, service response time, request error rate, and container resource quota); network communication status information for drone nodes (inter-node network latency, throughput, packet loss rate, and number of TCP connections); network adaptation parameters for drone nodes participating in federated learning (communication bandwidth threshold and encrypted transmission latency tolerance); and multi-dimensional basic data required by the digital twin engine (drone hardware configuration fingerprint and software environment mirror hash value). Data acquisition frequency: 100Hz for hardware status, 20Hz for network status, and 5Hz for federated learning parameters; median filter window size: 5 sampling points; linear interpolation is used for completion; 100 decision trees are constructed using the isolated forest algorithm; and the anomaly detection threshold is set to 3 times the standard deviation.

[0044] The data processing layer, connected to the perception layer, is used for preprocessing, privacy protection, data fusion, and analysis of raw state information to generate dynamic feature data. It includes a data cleaning module (denoising, error correction, and format standardization), a federated learning privacy protection module (differential privacy perturbation + homomorphic encryption transmission), a digital twin fusion module (LSTM-PPO hybrid model to construct a virtual image), and a data analysis module (extracting core parameters such as load balancing status and resource bottlenecks). It provides "clean" raw data to the federated learning privacy protection module, avoiding noise interference with the differential privacy perturbation effect; and provides uniformly formatted time-series data to the digital twin fusion module, improving LSTM model training efficiency and ensuring the consistency of the virtual image with the physical cluster's state. Among them, the federated learning privacy protection module includes: a differential privacy unit that dynamically adjusts noise intensity according to data sensitivity, using Laplace noise for highly sensitive data and Gaussian noise for low-sensitivity data, to meet the preset privacy budget; a homomorphic encryption unit that encrypts local model parameters and transmits them to the aggregation node, supporting encrypted aggregation and encrypted distribution of global parameters; and a dynamic node management unit that allocates temporary keys, adjusts noise and encryption parameters when a new node is added, and cancels keys, clears temporary parameters and triggers incremental encrypted updates of the global model when a node leaves.

[0045] Specifically, the data cleaning module employs differentiated processing for different data types: median filtering is used to denoise continuous time-series data (such as CPU utilization and network latency), while interpolation is used to repair discrete outlier data (such as request error rates and container resource quotas), ultimately unifying the data into the Protocol Buffers standard format. Median filtering and interpolation algorithms are used to denoise and correct errors in the raw data from the perception layer, unifying the data format to the Protocol Buffers standard. Through preprocessing such as cleaning and privacy enhancement, the accuracy, security, and usability of the data input to the decision layer are ensured. Sliding window filtering and the Isolation Forest algorithm are used to identify and remove outliers, ensuring data reliability.

[0046] The federated learning privacy protection module adds dynamic noise to the data using a differential privacy algorithm (the noise intensity is positively correlated with the data sensitivity), and combines it with homomorphic encryption (such as the CKKS scheme) to achieve secure ciphertext transmission. It supports the dynamic joining / leaving of edge nodes in the training cluster; allows cross-node collaborative training of the global model while protecting sensitive user data; and provides a secure communication foundation for the federated learning layer, supporting the dynamic joining / leaving of edge nodes in the training cluster and ensuring the robustness of distributed training.

[0047] 1. Differential privacy (dynamic noise injection)

[0048] (1) Algorithm principle and parameters

[0049] Differential privacy ensures that deleting or adding one data point does not significantly alter the data distribution by adding noise (Laplace noise) to the data. The core formula is: f'(x) = f(x) + Lap(Δf / ε), where:

[0050] f(x): Raw data (such as drone location coordinates, hardware utilization);

[0051] Lap(·): Laplace distribution, with a scale parameter of Δf / ε (Δf is the data sensitivity, and ε is the privacy budget);

[0052] Sensitivity calculation: dynamically set according to data type—UAV location information (Δf=0.8, significantly impacting accuracy), service response time (Δf=0.3, less impacting accuracy);

[0053] Privacy budget allocation: Total budget ε=1.0, evenly distributed according to the number of nodes (ε=0.01 per node when there are 100 nodes) to avoid excessive noise from a single node causing data invalidation.

[0054] (2) Noise dynamic adjustment logic

[0055] When the number of cluster nodes is ≥50, the noise scale is reduced by 15% (e.g., location information noise is reduced from 0.05 to 0.0425), because the participation of more nodes can distribute the pressure of privacy protection.

[0056] When the data is used to train the federated learning model, the noise scale decreases with each training epoch (by 5% every 10 epochs) to ensure higher data accuracy when the model converges in later stages.

[0057] 2. CKKS Homomorphic Encryption (Secure Ciphertext Transmission)

[0058] (1) Algorithm implementation steps

[0059] CKKS (Cheon-Kim-Kim-Song) is an algorithm that supports near-homomorphic encryption, allowing addition, subtraction, and multiplication operations to be performed in ciphertext. It is well-suited for the "model parameter aggregation" requirement in federated learning. The implementation steps are as follows:

[0060] Key generation: The Federated Learning Center generates a 2048-bit public key (for encryption) and a private key (for decryption), and the public key is distributed to each drone node;

[0061] Data encryption: The node encrypts local data (such as model gradients) with a public key to generate ciphertext (approximately 4KB in length per record), with encryption time ≤10ms;

[0062] Ciphertext transmission and aggregation: Ciphertext is transmitted to the center through a Mesh network. The center calculates the gradient average in the ciphertext state (supporting homomorphic addition) without decryption, thus avoiding data leakage.

[0063] Decryption of results: The center uses its private key to decrypt the aggregated gradients, generates global model parameters, and then encrypts and distributes them to each node to complete one round of training.

[0064] (2) Performance optimization

[0065] Batch encryption: Packing 10 data entries into one batch processing unit reduces encryption time from 10ms / entry to 15ms / 10 entries, improving efficiency by 6.7 times;

[0066] Key caching: Nodes cache public keys locally (valid for 1 hour) to avoid re-acquiring them for each transmission, reducing network interaction time (from 50ms to 5ms).

[0067] The digital twin fusion module, based on the LSTM-PPO hybrid model, integrates historical one-hour load data with real-time status data (including node location and task type) to generate a load trend prediction for the next 15 minutes (error rate ≤12%), and provides a virtual "trial and error sandbox" to simulate strategy risks. By integrating historical time-series data and real-time status data, a virtual image of a drone swarm is constructed by training the LSTM-PPO hybrid prediction model, generating dynamic feature data containing load trend predictions. This provides a "trial and error sandbox" for decision-makers: it simulates service migration and resource adjustment strategies in a virtual space (e.g., mimicking the migration of 5 microservices to a backup node), exposing risks such as network congestion and resource conflicts in advance, and avoiding real-world swarm instability. The generated dynamic feature data, including "load trends for the next 15 minutes," gives decision-makers "foresight."

[0068] LSTM-PPO hybrid model

[0069] 1. Model Structure Design

[0070] LSTM (Long Short-Term Memory) is responsible for extracting features from time-series data, while PPO (Proximal Policy Optimizer) is responsible for optimizing the prediction strategy. Together, they generate the load trend for the next 15 minutes, as shown in Table 2 below:

[0071] Table 2 LSTM-PPO Hybrid Model Structure Design

[0072] ;

[0073] 2. Training and Prediction Logic

[0074] Training data preparation: Cluster load data from the past 7 days (approximately 6 million records) were used and divided into a training set (4.2 million records) and a validation set (1.8 million records) in a 7:3 ratio. Each data record contains 5 types of features (CPU, GPU, memory, network latency, and number of tasks).

[0075] LSTM feature extraction: Redundant features are filtered out through gating units (input gate, forget gate, output gate), with a focus on retaining load peaks (such as high CPU load at 10:00 and 16:00) and trend features (such as linear growth of GPU load with increasing task).

[0076] PPO policy optimization: Using the features output by LSTM as input, the prediction parameters are adjusted through "policy gradient". The goal is to minimize the "MSE (mean squared error) between the predicted load and the actual load". Each training iteration is 100 times, and the MSE converges to ≤0.01 (corresponding to an error rate ≤10%).

[0077] Prediction execution: During real-time prediction, the model takes the load data of the most recent hour as input and outputs 90 prediction points (10 points per second) within 500ms, with an error rate of ≤12%, and marks high-risk intervals (such as the time period when the GPU load is ≥85%).

[0078] The data analysis module, based on the fused cluster environment data, extracts dynamic feature data through statistical analysis and model calculations. This includes core parameters such as cluster load balancing status, resource bottleneck information, and service quality indicators. It transforms basic indicators like CPU utilization and network latency into "decision factors" that directly guide action (e.g., a load balancing index > 0.8 triggers task rebalancing), reducing the complexity of the decision-making algorithm and providing a "lightweight feature vector" for the quantum-inspired optimizer. Focusing on key dimensions such as resource bottlenecks and service quality, it shortens the quantum annealing algorithm's solution time by 50%, making it more efficient at handling NP-hard scheduling problems.

[0079] The decision-making layer, connected to the data processing layer, generates decisions on adjustments, service migrations, or negative entropy injection based on a strategy library and dynamic feature data. This includes a strategy library for storage resource adjustment / service migration / negative entropy injection strategies, and a quantum-inspired optimizer for solving NP-hard problems based on the quantum annealing algorithm. It deeply integrates the dynamic features output from the data processing layer (including digital twin prediction results and federated learning node capability assessments) to form a "data-driven" decision-making mechanism. This provides the execution layer with high-precision, fine-grained adjustment instructions (such as CPU quotas for each container and bandwidth allocation for each network link), ensuring operational efficiency and security. It introduces thermodynamic entropy theory and quantum computing technology into cluster decision-making, solving the problems of "coarse decision-making, low efficiency, and poor scalability" in traditional distributed systems. The quantum-inspired optimizer is used to efficiently solve NP-hard problems related to cluster resource allocation, task scheduling, and service migration path planning, specifically including:

[0080] The problem modeling module transforms real-world problems into energy functions containing decision variables, objective functions, and constraint penalty terms;

[0081] The core module of quantum annealing achieves global optimization through quantum state initialization, tunneling process simulation, and annealing scheduling.

[0082] The classical-quantum interface module adopts a hybrid architecture of quantum inspiration and classical computing, which is adapted to the limited computing power of drone nodes;

[0083] The result verification and correction module performs feasibility verification and local correction on the output optimal solution.

[0084] It also features lightweight, distributed collaboration, and dynamic response capabilities, allowing the annealing process to be interrupted and restarted based on changes in cluster topology.

[0085] Algorithms for solving NP-hard problems (quantum annealing)

[0086] Algorithm Principles and Adaptation Scenarios

[0087] Quantum annealing utilizes the quantum tunneling effect to quickly traverse the solution space of NP-hard problems such as task allocation and resource scheduling. Compared with traditional genetic algorithms, it can avoid getting trapped in local optima. In this system, it is used for the scenario of "100+ node task allocation" (objective: minimize total task latency + maximize resource utilization).

[0088] 2. Algorithm Implementation Steps

[0089] Problem modeling: Transform task assignment into an "integer programming problem" and define variables:

[0090] x_ij: 1 indicates that task i is assigned to node j, 0 indicates that it is not assigned;

[0091] Objective function: min(ΣΣx_ij×t_ij) - max(ΣΣx_ij×r_ij) (t_ij is the execution delay of task i on node j, and r_ij is the resource utilization rate);

[0092] Constraints: Each task is assigned to only 1 node, and node resource usage is ≤ hardware limit (e.g., CPU ≤ 75%).

[0093] Quantum annealing process:

[0094] Initialization: Set the initial temperature T0=100 (simulating the energy of the quantum system), and randomly initialize the solution space (tasks are randomly assigned).

[0095] Tunneling search: Through the quantum tunneling effect, jump to a new solution with probability P=exp(-ΔE / T) (adjust the value of x_ij), where ΔE is the energy difference between the new solution and the current solution;

[0096] Temperature decay: Every 100 iterations, the temperature decays at a rate of T = T0 × 0.95, gradually reducing the tunneling probability and stabilizing to the optimal solution;

[0097] Termination condition: The temperature drops to T=1 or the objective function shows no improvement after 20 consecutive iterations, output the optimal solution (task allocation scheme).

[0098] 3. Performance Indicators

[0099] Solution efficiency: In a scenario with 100 nodes and 50 tasks, the solution time is reduced from 100ms in the traditional genetic algorithm to less than 40ms, and the decision latency is reduced by 60%.

[0100] Solution quality: The output task allocation scheme reduces the total latency by 25% compared to traditional algorithms, and the resource utilization variance decreases from ±15% to ±8%;

[0101] Fine-grained output: The solution includes resource allocation precision (0.1 CPU cores, 100MB memory), such as "Task 1 (image recognition) is allocated to node A: 0.8 CPU cores + 2GB memory, execution latency ≤300ms".

[0102] In the power grid inspection scenario: the strategy library prioritizes the use of the "PV panel identification task-specific strategy" and plans charging nodes in advance based on the remaining battery capacity of the drone predicted by the digital twin; the quantum-inspired optimizer optimizes the inspection path, thereby shortening the task completion time.

[0103] Extreme weather response: When the sensing layer detects wind speed > 15m / s, the decision-making layer automatically triggers "wind-resistant mode":

[0104] Hardware layer: Reduce the drone's flight altitude to 50 meters and limit the maximum roll angle to 20°;

[0105] Network layer: Switch to Mesh self-organizing network protocol, and adjust the communication frequency band to 5.8GHz (stronger anti-interference).

[0106] Mission level: Pause high-precision imaging mission and switch to radar scanning mode.

[0107] The task instructions output by the decision generation module include geofencing information (such as "Do not enter the area with a radius of 100 meters at coordinates (x,y)"), which are synchronized to the command center GIS platform in real time via a Kafka message queue.

[0108] Specifically, the strategy library stores drone node resource adjustment strategies, service migration strategies, and negative entropy injection strategies designed based on the system entropy change model. It maintains cluster order by dynamically balancing system entropy. The strategy library dynamically balances three types of strategies: pre-stored resource adjustment, service migration, and negative entropy injection. It maintains cluster order based on the system entropy change model, supports custom strategy plugins, and automatically matches task types (such as inspection and rescue), reducing manual intervention costs.

[0109] The quantum-inspired optimizer, based on the quantum annealing algorithm, solves task allocation problems and outputs fine-grained instructions with precise resource allocation. The quantum-inspired optimizer solves NP-hard problems such as multidimensional resource scheduling using the quantum annealing algorithm, reducing decision latency from 100ms to below 40ms (a 60% reduction) compared to traditional genetic algorithms; it generates refined instructions with precise resource allocation (0.1 CPU cores, 100MB memory) (decision layer - quantum-inspired optimizer).

[0110] The decision generation module combines dynamic feature data output from the data processing layer with adaptive strategies in the strategy library to generate specific instructions for adjusting UAV node resource allocation, migrating services, and rescheduling tasks, providing operational basis for the execution layer. It transforms feature parameters output from the data analysis module (such as resource bottleneck node ID and service quality degradation level) into specific instructions (such as "migrate 3 GPU tasks from node A to node B" or "expand database service replicas to 5"), with an accuracy rate of ≥95%. This supports fine-grained operations at the execution layer: generating refined instructions that include precision in resource allocation (0.1 CPU cores, 100MB memory) and service migration order (migrate stateless services first, then stateful services), ensuring the efficiency and safety of adjustment operations.

[0111] The execution layer, connected to the decision layer, is used to execute the adjustment decisions generated by the decision layer. It includes a hardware resource adjustment module (dynamically allocating CPU / GPU computing power), a service control module (process start / stop, task migration), and a network communication adjustment module (bandwidth configuration, link optimization). Through the integration of multiple technologies (network resilience of SDN / NFV, secure execution of TEE, and system-level optimization through negative entropy injection), the execution layer ensures efficient and secure implementation of decisions, achieving dynamic mapping between physical clusters and digital twins.

[0112] Specifically, the node management module dynamically adjusts the allocation of node hardware resources and performs task migration based on the negative entropy injection mechanism; the service control module verifies key decisions through a trusted execution environment (such as Intel SGX) and supports the dynamic switching of training modes for nodes participating in federated learning; and the elastic network execution unit achieves millisecond-level network topology switching based on SDN and dynamically orchestrates service chains and optimizes communication paths through NFV technology.

[0113] Extreme environment execution mechanism: Electromagnetic interference scenario:

[0114] The elastic network execution unit automatically switches to the frequency hopping communication protocol, changing the communication frequency every 50ms; the SDN controller reroutes data to bypass interference frequency bands, and the network connectivity retention rate is ≥95%.

[0115] Hardware fault tolerance:

[0116] When the node management module detects a CPU core failure in a drone, it triggers a "hot migration" mechanism:

[0117] Migrate the task on this node to a neighboring node within 100ms;

[0118] Send a hardware fault alarm to the perception layer and initiate the self-test process of the backup drone.

[0119] It also includes a feedback layer, which connects to the execution layer and the data processing layer. This feedback layer collects new state information after decision-making and feeds it back to the data processing layer. If the deviation between the new state and the prediction exceeds a threshold, it triggers incremental training of the model and updates the policy library in the data processing layer, forming a self-growing closed loop. The feedback layer scans execution deviations in real time and drives policy iteration, enabling the cluster to autonomously adapt and self-repair in complex environments, ultimately forming a self-growing closed loop of "data-driven decision-making and execution-verified data." The feedback layer includes a feedback sensor module and a feedback data processing module. The feedback sensor module collects the state information of nodes after adjustment operations, while the feedback data processing module processes and analyzes the feedback state information and feeds the results back to the data processing layer. The feedback data processing module compares the new state information with the digital twin prediction results. If the load prediction error > 15%, it triggers incremental training of the LSTM-PPO model; if the policy execution deviation (such as an increase in error rate after service migration) exceeds a threshold, it automatically optimizes the policy library parameters.

[0120] Model incremental training algorithm (LSTM-PPO incremental update)

[0121] 1. Triggering conditions and data selection

[0122] When the feedback layer detects "load prediction error > 15%", it triggers incremental training of the LSTM-PPO model. The core principle is "updating the model with the latest data without discarding historical parameters." Data selection rules are as follows:

[0123] Training dataset: Nearly 1 hour of actual load data (approximately 36,000 records), including data from periods where prediction errors exceeded the standard (weighted by 20%).

[0124] Validation dataset: Real-time data from the past 10 minutes (approximately 6,000 records) used to validate the accuracy of the updated model.

[0125] 2. Incremental Training Steps

[0126] Parameter freezing and unfreezing: Freeze the parameters of the first two layers of LSTM (to preserve the ability to extract historical time-series features), and unfreeze the parameters of the third LSTM layer and the PPO strategy layer;

[0127] Local training: Train the thawed layer with the new dataset, setting the learning rate to 1 / 10 of the initial training rate (0.001→0.0001) to avoid parameter oscillations;

[0128] Accuracy validation: Test the model with a validation dataset. If the prediction error is ≤12%, update the model; otherwise, increase the number of training epochs (from 10 epochs to 20 epochs) until the accuracy meets the target.

[0129] Model replacement: The updated model is previewed in the digital twin environment for 10 minutes. After confirming that there are no abnormalities, the online model is replaced. The replacement time is ≤100ms, and the prediction service is not interrupted.

[0130] Performance evaluation system:

[0131] Test environment: Simulate a 100-node cluster, including 20% ​​computing nodes (equipped with NVIDIA Jetson AGXOrin), 60% communication nodes, and 20% task nodes (equipped with multispectral cameras);

[0132] Test metrics:

[0133] Resource utilization: CPU / GPU average load ≤75%, memory utilization ≤80%;

[0134] Task response latency: Emergency tasks ≤ 500ms, routine tasks ≤ 2s;

[0135] System reliability: MTBF ≥ 1000 hours, MTTR ≤ 30 minutes;

[0136] Comparative data: Compared with traditional cluster systems based on PID control, this system improves task completion efficiency by 40% and reduces energy consumption by 20% (by reducing ineffective resource scheduling through negative entropy injection strategy).

[0137] This invention employs multi-source data acquisition, covering five major categories of status information for UAVs: hardware resources (CPU / GPU utilization, memory / disk I / O), software status (service processes, response time), and network performance (latency / throughput). It combines hardware sensors (physical acquisition) and software sensors (program plugins) to achieve comprehensive monitoring of cluster node status. Data quality is guaranteed through algorithms such as median filtering and interpolation completion to denoise and correct raw data. Data acquisition frequency reaches microsecond levels, and position and attitude accuracy reaches centimeter levels, providing a high-precision data foundation for decision-making. Differential privacy algorithms are used to add noise perturbations to the data, combined with homomorphic encryption technology to achieve encrypted transmission. In federated learning, this ensures that UAV local data is "usable but not visible," meeting the needs of sensitive scenarios such as disaster relief and inspection.

[0138] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed system for dynamic perception and adaptation in an unmanned aerial vehicle (UAV) swarm environment, characterized in that, This includes a perception layer, data processing layer, decision-making layer, execution layer, and feedback layer distributed across each node of the drone swarm, forming a closed-loop dynamic adaptive architecture; among which: The perception layer consists of sensor modules at each node, used to collect node status information in real time and transmit it to the data processing layer through a distributed network. The sensor modules of the perception layer include hardware sensors and software sensors. The hardware sensors are used to collect hardware resource information such as CPU utilization and memory usage. The software sensors are used to collect information such as service response time, network latency, software running status, and network communication status. The perception layer adopts a spatiotemporal cross-verification mechanism: when the hardware sensor detects that the GPU temperature is > 85℃ and the software sensor detects that the GPU computing load is > 90% for 5 consecutive minutes, it is determined that there is a risk of node overload; if the data of a single sensor is abnormal, it triggers a comparison of data from multiple nodes within 10 seconds to eliminate occasional noise interference, and the collected data is transmitted to the data processing layer through a distributed network. The data processing layer connects to the perception layer and is responsible for the preprocessing, privacy protection, data fusion and analysis of raw state information, generating dynamic feature data, including data cleaning, federated learning privacy protection, digital twin fusion, and data analysis modules; The data cleaning module employs differentiated processing for different data types: median filtering is used to denoise continuous time series data with a filtering window of 5 sampling points, and interpolation is used to complete and repair discrete outlier data, ultimately unifying it into the Protocol Buffers standard format; Federated learning privacy protection module: Based on a two-level architecture of "central node-edge node", the central node is equipped with an Intel Xeon 8375C CPU and an NVIDIA A100 GPU, and the edge nodes are divided into 20% training nodes and 80% data nodes, which transmit data through Mesh self-organizing network + 5G dual link; The differential privacy unit dynamically adjusts the noise intensity based on data sensitivity, using Laplace noise for highly sensitive data and Gaussian noise for low-sensitivity data, while meeting the preset privacy budget. For drone location information with a sensitivity Δf=0.8, Laplace noise is added; for service response time with a sensitivity Δf=0.3, Gaussian noise is added. The total privacy budget ε=1.0 and is evenly distributed across the number of nodes. Furthermore, when the number of cluster nodes is ≥50, the noise scale is reduced by 15%, and during federated learning, the noise scale is reduced by 5% every 10 training rounds. The homomorphic encryption unit encrypts local model parameters and transmits them to the aggregation node, supporting ciphertext aggregation and encrypted distribution of global parameters. The homomorphic encryption unit uses the CKKS algorithm: the central node generates a 2048-bit public / private key, and the training node uses the public key to encrypt local model parameters. The encryption time is ≤10ms / parameter, and 15ms / 10parameters in batch processing. After the ciphertext is transmitted to the central node, it supports ciphertext aggregation. After aggregation, the global parameters are encrypted and distributed. The dynamic node management unit allocates temporary keys, adjusts noise and encryption parameters when a new node is added, and cancels the key, clears temporary parameters and triggers global model incremental encryption update when a node leaves. The digital twin fusion module is based on the LSTM-PPO hybrid model, which integrates historical 1-hour load data with real-time status data from the perception layer to generate load trend predictions for the next 15 minutes and simulate strategy risks. Data analysis module: Extracts dynamic feature data such as cluster load balancing index, resource bottleneck node ID, and service quality degradation level, providing lightweight feature vectors for the decision-making level; The decision-making layer connects to the data processing layer, generating and adjusting, migrating services, or injecting negative entropy decisions based on the strategy library and dynamic feature data. It includes a strategy library containing storage-related strategies and a quantum-inspired optimizer that solves NP-hard problems based on the quantum annealing algorithm. The strategy library includes storage resource adjustment strategies, service migration strategies, and negative entropy injection strategies. When the system entropy value exceeds a preset threshold, negative entropy is injected to maintain the order of the cluster. It can automatically match exclusive strategies according to task type. For example, in the inspection scenario, the photovoltaic panel identification task strategy is called first and charging nodes are planned. Quantum-inspired optimizer: Based on the quantum parallelism of the quantum annealing algorithm, it adopts a "classical-quantum hybrid architecture" to adapt to the limited computing power of UAVs, solving NP-hard problems of cluster resource allocation, task scheduling, and service migration path planning. Problem modeling module: Transforms task assignment into an integer programming problem, defining decision variables x_ij, where 1 indicates that task i is assigned to node j, and the objective function is: min(ΣΣx_ij×t_ij)-max(ΣΣx_ij×r_ij), where t_ij is the execution latency and r_ij is the resource utilization rate. The constraints are that node resource usage ≤ CPU 75% and memory 80%. The core module of quantum annealing: The initial temperature T0=100. Every 100 iterations, the temperature decreases by T=T0×0.

95. Through the quantum tunneling effect, the jump probability P=exp (-ΔE / T) achieves global optimization. The termination condition is that the temperature drops to 1 or the objective function does not improve after 20 consecutive iterations. ΔE is the difference between the calculated energy function value corresponding to the new solution and the current solution during the quantum annealing process. The result verification and correction module performs feasibility verification on the output optimal solution and outputs fine-grained instructions with resource allocation precision of 0.1 CPU cores and 100MB memory, such as "image recognition task is assigned to node A: 0.8 CPU cores + 2GB memory, execution latency ≤300ms". It can also interrupt and restart the annealing process according to changes in cluster topology. The execution layer connects to the decision-making layer and executes adjustment decisions, including modules for hardware resource adjustment, service control, and network communication adjustment. Hardware resource adjustment module: dynamically allocates CPU / GPU computing power, and triggers task hot migration to a nearby node within 100ms when a CPU core failure is detected; Service control module: Based on the Intel SGX Trusted Execution Environment, it verifies key decisions and supports dynamic switching between training nodes and data nodes; Network communication adjustment module: Based on SDN, it realizes millisecond-level network topology switching, orchestrates service chains through NFV technology, and changes the frequency hopping communication protocol every 50ms in electromagnetic interference scenarios to bypass interference frequency bands; The feedback layer connects the execution layer and the data processing layer, collecting feedback on the new state after execution; if the deviation between the new state and the prediction exceeds the threshold, it triggers incremental training of the model and updating of the policy library in the data processing layer, forming a self-growing closed loop.

2. The distributed system for dynamic perception and adaptation of unmanned aerial vehicle (UAV) swarm environments according to claim 1, characterized in that: The node status information includes hardware resource information, software running status information, network communication status information, network adaptation parameters of drone nodes participating in federated learning, and multi-dimensional basic data required by the digital twin engine.

3. The distributed system for dynamic perception and adaptation of unmanned aerial vehicle (UAV) swarm environments according to claim 1, characterized in that: The feedback data processing module of the feedback layer compares the new state information with the digital twin prediction results. If the load prediction error is > 15%, it triggers incremental training of the LSTM-PPO model; if the policy execution deviation exceeds the threshold, it automatically optimizes the policy library parameters.

Images