Distributed system for dynamic perception and adaptation of unmanned aerial vehicle cluster environment

Abstract

The invention discloses a distributed system for dynamic perception and adaptation of an unmanned aerial vehicle cluster environment, and belongs to the technical field of computers. The distributed system comprises a sensing layer used for collecting state information of each node in an unmanned aerial vehicle cluster in real time, and a data processing layer used for preprocessing and analyzing the collected unmanned aerial vehicle node state information; the decision-making layer is used for generating an adjustment decision for an unmanned aerial vehicle cluster environment based on a preset unmanned aerial vehicle cluster adaptation strategy library and the dynamic feature data; the execution layer is used for performing corresponding adjustment operation on the unmanned aerial vehicle cluster nodes according to the adjustment decision; and the feedback layer is used for collecting new state information after the unmanned aerial vehicle node adjustment operation. Through innovative fusion and hierarchical collaboration of multiple technical modules and dynamic balance of a strategy library, cluster orderliness is maintained, and a self-growing closed loop of'perception-decision-execution-feedback 'is formed.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and more specifically, to a distributed system for dynamic perception and adaptation of a drone cluster environment. Background Art

[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 operation of multiple drones, enable efficient operations in complex environments. For example, they can quickly construct three-dimensional maps of disaster-stricken areas during emergency rescue operations and automatically monitor long-distance power transmission lines during power line inspections. However, existing drone swarm systems have numerous shortcomings in their dynamic environmental perception and adaptability. Traditional approaches often focus on single-dimensional data, such as monitoring the drone's location. This makes it difficult to comprehensively capture multi-source information about the drone's hardware resources (CPU, memory, disk I / O, etc.), software operating status (service processes, response time, etc.), and network communication status (latency, throughput, etc.). This results in an incomplete understanding of the swarm's overall operational status, hindering the timely detection of potential resource bottlenecks or potential faults. In terms of data processing, as drone swarms expand and mission complexity increases, the amount of collected data increases dramatically, becoming multi-source and heterogeneous. Existing data processing technologies struggle to efficiently clean, analyze, and integrate this massive amount of data. This is especially true in federated learning, where balancing data privacy and collaborative efficiency is difficult. Furthermore, there's a lack of effective methods for deep data mining to accurately extract dynamic environmental features, making it impossible to provide a reliable basis for decision-making. In the decision-making and execution stages, traditional decision-making algorithms suffer from low decision-making efficiency and poor optimization results when faced with complex NP-hard problems such as resource scheduling and task allocation in drone swarms, making them difficult to adapt to dynamically changing environmental demands. During execution, drone node management, service control, and network adjustments lack flexibility and intelligence, making it impossible to achieve precise resource allocation and efficient task coordination. Furthermore, most existing systems lack robust feedback mechanisms, making it difficult to adjust strategies based on execution results, making it difficult to form a closed-loop dynamic optimization system. In summary, the shortcomings of existing drone swarm systems in terms of dynamic environmental perception and adaptation seriously restrict their application effectiveness and promotion scope in complex scenarios. In view of this, we propose a distributed system for dynamic environmental perception and adaptation of drone swarms. Summary of the Invention

[0003] The purpose of the present invention is to provide a distributed system for dynamic perception and adaptation of drone cluster environments to solve the problems raised in the above background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions: A distributed system for dynamic perception and adaptation of drone swarm environments, including a perception layer, data processing layer, decision layer, execution layer, and feedback layer distributed across each node in the drone swarm, forming a closed-loop dynamic adaptation architecture. The perception layer, composed of sensor modules distributed across each drone node, collects node status information in real time and transmits it to the data processing layer via a distributed network. This 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, all of which are transmitted via a distributed network. The data processing layer is connected to the perception layer and is used to preprocess, protect privacy, fuse and analyze the original state information to 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; among them, the federated learning privacy protection module: the differential privacy unit dynamically adjusts the noise intensity according to the sensitivity of the data, using Laplace noise for high-sensitivity data and Gaussian noise for low-sensitivity data to meet the preset privacy budget; the homomorphic encryption unit encrypts the local model parameters and transmits them to the aggregation node, supporting ciphertext aggregation and encrypted distribution of global parameters; the dynamic node management unit allocates temporary keys, adjusts noise and encryption parameters when a new node joins, and cancels the key, clears the temporary parameters and triggers the incremental encryption update of the global model when the node exits; A decision layer, connected to the data processing layer, generates decisions on adjustments, service migration, or negative entropy injection based on a policy library and dynamic feature data; including a policy library for storing resource adjustment / service migration / negative entropy injection policies, and a quantum-inspired optimizer based on a quantum annealing algorithm for solving NP-hard problems; An execution layer, connected to the decision layer, for executing the adjustment decisions generated by the decision layer; including a hardware resource adjustment module, a service control module, and a network communication adjustment module; The feedback layer is connected to the execution layer and the data processing layer, and 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, it triggers the incremental training of the data processing layer model and the update of the strategy library, forming a self-growing closed loop.

[0005] 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 usage and memory occupancy; the software sensors are used to collect software operation status and network communication status information such as service response time and network delay.

[0006] Preferably, the data cleaning module adopts differentiated processing for different data types: median filtering is used to denoise continuous time series data, and interpolation and completion are used to repair discrete abnormal data, and finally unified into the Protocol Buffers standard format.

[0007] Preferably, the federated learning privacy protection module adds dynamic noise to the data through a differential privacy algorithm, combines homomorphic encryption to achieve secure transmission of ciphertext, and supports edge nodes to dynamically join / exit the training cluster.

[0008] Preferably, the digital twin fusion module is based on the LSTM-PPO hybrid model, which integrates the historical 1-hour load data and the real-time status data to generate a load trend forecast for the next 15 minutes and preview the strategic risks.

[0009] Preferably, the quantum-inspired optimizer solves the task allocation problem involving 100+ nodes based on the quantum parallelism of the quantum annealing algorithm and outputs fine-grained instructions containing resource allocation precision. The quantum-inspired optimizer is used to efficiently solve the NP-hard problems of cluster resource allocation, task scheduling, and service migration path planning. Specifically, it includes: a problem modeling module that converts the actual problem into an energy function containing decision variables, objective functions, and constraint penalty terms; The quantum annealing core module achieves global optimization through quantum state initialization, tunneling process simulation, and annealing scheduling; The classical-quantum interface module uses a hybrid architecture of quantum-inspired and classical computing to adapt to the limited computing power of drone nodes; The result verification and correction module performs feasibility verification and local correction on the output optimal solution; It also has lightweight, distributed collaboration and dynamic response capabilities, and can interrupt and restart the annealing process according to changes in cluster topology.

[0010] 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%, the incremental training of the LSTM-PPO model is triggered; if the strategy execution deviation exceeds the threshold, the strategy library parameters are automatically optimized.

[0011] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention builds a distributed system for the dynamic perception and adaptation of drone swarms. Through the collaborative collection of five categories of state information by multi-source sensors at the perception layer, combined with differential privacy protection at the data processing layer, LSTM-PPO digital twin modeling, and a decision layer driven by a quantum annealing algorithm, this system achieves full-process intelligence from data collection, privacy enhancement, trend prediction, to millisecond-level quantum optimization decision-making. The core breakthrough lies in the introduction of a system entropy change model to maintain cluster order, the use of quantum parallelism to shorten the decision-making time for NP-hard problems by more than 90%, and the use of a digital twin "trial and error sandbox" to avoid strategic risks in advance, forming a self-growing closed loop of "perception-decision-execution-feedback".

[0012] (2) The system of the present 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), reduces the frequency of manual intervention, improves emergency response speed, and achieves a high cluster task completion rate. Furthermore, sensitive data can be made “available but invisible” through differential privacy and homomorphic encryption. Its innovative architecture demonstrates unique value in scenarios such as disaster relief and smart inspection. It can not only meet the requirements of low latency and anti-interference, but also improve the efficiency of civil inspections through dynamic task allocation, providing an efficient and secure technical foundation for the large-scale application of distributed drone clusters. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Schematic diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Example

[0015] A distributed system for dynamic perception and adaptation of drone swarm environments, including a perception layer, data processing layer, decision layer, execution layer, and feedback layer distributed across each node in the drone swarm, forming a closed-loop dynamic adaptation architecture. The perception layer is composed of sensor modules distributed on each drone node, which is used to collect node status information in real time and transmit it to the data processing layer through a distributed network. The node status information includes hardware resource information (CPU utilization, memory occupancy, disk I / O rate, GPU computing power load), software operation status information (service process status, service response time, request error rate, container resource quota), network communication status information (network delay between nodes, throughput, packet loss rate, number of TCP connections), network adaptation parameters of drone nodes participating in federated learning (communication bandwidth threshold, encryption transmission delay tolerance), and multi-dimensional basic data required by the digital twin engine (hardware configuration fingerprint, software environment image hash value). Through the sensor modules distributed on each drone node, multi-dimensional status information is collected in real time to provide raw data support for system decision-making and transmitted through a distributed network.

[0016] Overall architecture design of federated learning This system adopts the FedAvg framework to build a "central node - edge node" two-level architecture to adapt to the dynamic and distributed characteristics of drone clusters. The specific architecture is as follows: 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; Edge nodes: These are drone nodes, divided into "training nodes" (20% computing 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). Communication link: Adopting "Mesh ad hoc network + 5G" dual link, the model parameter transmission between the training node and the central node is preferentially transmitted through 5G (bandwidth ≥ 100Mbps). When the 5G signal is weak, it automatically switches to the Mesh network (bandwidth ≥ 20Mbps) to ensure communication reliability.

[0017] Federated Learning Core Process (Complete Training Cycle) Taking the "UAV swarm load prediction model training" as an example, a 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: Table 1. UAV cluster load prediction model training ; UAV node hardware resource information includes CPU usage, memory occupancy, disk I / O rate, and GPU computing load. Software runtime status information includes the process status, service response time, request error rate, and container resource quota of the drone-mounted services. Network communication status information includes inter-UAV node network latency, throughput, packet loss rate, and number of TCP connections. Active monitoring mechanisms such as GPU computing load anomaly warnings and container resource quota overrun detection transform passive fault handling into proactive risk avoidance, improving cluster availability. This provides high-quality raw input to the data processing layer, directly impacting the resource scheduling accuracy of the decision-making layer.

[0018] Specifically, the sensor module in the perception layer includes hardware sensors and software sensors. The hardware sensors collect hardware resource information such as CPU utilization and memory usage, while the software sensors collect information on software operation and network communication status, such as service response time and network latency. Hardware sensors (such as the GPU temperature sensor) and software sensors (the computing load monitoring plug-in) utilize a spatiotemporal cross-validation mechanism. When the GPU temperature exceeds 85°C and the computing load remains above 90% for five consecutive minutes, an overload risk is identified. If a single sensor's data is abnormal (e.g., the temperature is normal but the load suddenly increases), a multi-node data comparison is triggered within 10 seconds to eliminate occasional noise interference.

[0019] Drone node software runtime status information, including service process status, service response time, request error rate, and container resource quota; drone node network communication status information, including inter-node network latency, throughput, packet loss rate, and number of TCP connections; network adaptation parameters for drone nodes participating in federated learning, including communication bandwidth threshold and encryption transmission delay tolerance; multi-dimensional basic data required by the digital twin engine, including drone hardware configuration fingerprints and software environment image hash values. Data collection frequency: hardware status 100Hz, network status 20Hz, federated learning parameters 5Hz; median filter window size: 5 sampling points, interpolation completion uses linear interpolation, isolation forest algorithm constructs 100 decision trees, and anomaly detection threshold is set to 3 times the standard deviation. The data processing layer, connected to the perception layer, preprocesses raw state information, protects privacy, and fuses and analyzes data to generate dynamic feature data. It includes a data cleaning module (denoising, error correction, and format unification), a federated learning privacy protection module (differential privacy perturbation + homomorphic encryption transmission), a digital twin fusion module (using an LSTM-PPO hybrid model to build a virtual image), and a data analysis module (extracting core parameters such as load balancing status and resource bottlenecks). This provides the federated learning privacy protection module with clean raw data to prevent noise from interfering with the differential privacy perturbation effect. It also provides the digital twin fusion module with time series data in a unified format, improving LSTM model training efficiency and ensuring state consistency between the virtual image and the physical cluster. Among them, the federated learning privacy protection module: the differential privacy unit dynamically adjusts the noise intensity according to the data sensitivity, using Laplace noise for highly sensitive data and Gaussian noise for low-sensitivity data to meet the preset privacy budget; the homomorphic encryption unit encrypts the local model parameters and transmits them to the aggregation node, supporting ciphertext aggregation and encrypted distribution of global parameters; the dynamic node management unit allocates temporary keys, adjusts noise and encryption parameters when a new node joins, and cancels the key, clears the temporary parameters and triggers the incremental encryption update of the global model when the node exits.

[0020] 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 usage and network latency), while interpolation and completion are used to correct discrete anomaly data (such as request error rates and container resource quotas), ultimately unifying the data into the standard Protocol Buffers format. Median filtering and interpolation algorithms are used to denoise and correct errors in the raw data from the perception layer, and the data format is unified into the standard Protocol Buffers format. Preprocessing such as cleaning and privacy enhancement ensures that the data input to the decision layer is accurate, secure, and available. Sliding window filtering and the isolation forest algorithm are used to identify and remove outliers to ensure data reliability.

[0021] The federated learning privacy protection module adds dynamic noise to the data through a differential privacy algorithm (the noise intensity is positively correlated with the data sensitivity), combines it with homomorphic encryption (such as the CKKS scheme) to achieve secure transmission of ciphertext, and supports edge nodes to dynamically join / exit the training cluster; allows cross-node collaborative training of global models while protecting user sensitive data; provides a secure communication foundation for the federated learning layer, supports edge nodes to dynamically join / exit the training cluster, and ensures the robustness of distributed training.

[0022] 1. Differential Privacy (Dynamic Noise Injection) (1) Algorithm principles and parameters Differential privacy adds noise (Laplace noise) to the data to ensure that deleting or adding one piece of data does not significantly change the data distribution. The core formula is: f'(x) = f(x) + Lap(Δf / ε), where: f(x): raw data (such as drone location coordinates, hardware usage); Lap(·): Laplace distribution with scale parameter Δf / ε (Δf is the data sensitivity and ε is the privacy budget); Sensitivity calculation: Dynamically set based on data type - drone location information (Δf=0.8, large impact on accuracy), service response time (Δf=0.3, small impact on accuracy); Privacy budget allocation: The total budget ε=1.0 is evenly divided according to the number of nodes (ε=0.01 per node when there are 100 nodes) to avoid data invalidation caused by excessive noise on a single node.

[0023] (2) Noise dynamic adjustment logic When the number of cluster nodes is ≥50, the noise scale is reduced by 15% (for example, the location information noise is reduced from 0.05 to 0.0425), because more nodes participate and the pressure of privacy protection is dispersed; When the data is used to train a federated learning model, the noise level decreases with each training round (5% every 10 rounds), ensuring higher data accuracy when the model converges later.

[0024] 2. CKKS Homomorphic Encryption (Secure Ciphertext Transmission) (1) Algorithm implementation steps CKKS (Cheon-Kim-Kim-Song) is an algorithm that supports near-homomorphic encryption. It can perform addition, subtraction, and multiplication operations in ciphertext, and is suitable for the "model parameter aggregation" requirement in federated learning. The implementation steps are as follows: Key generation: The federated learning center generates a 2048-bit public key (for encryption) and a private key (for decryption), and distributes the public key to each drone node; Data encryption: The node encrypts local data (such as model gradients) with the public key to generate ciphertext (about 4KB per record). The encryption time is ≤ 10ms. Ciphertext transmission and aggregation: Ciphertext is transmitted to the center through the Mesh network. The center calculates the gradient average (supports homomorphic addition) in the ciphertext state without decryption, thus preventing data leakage. Result decryption: The center uses the private key to decrypt the aggregated gradient, generate the global model parameters, and then encrypt and distribute them to each node to complete one round of training.

[0025] (2) Performance optimization Batch encryption: Packing 10 pieces of data into one batch unit reduces encryption time from 10ms per piece to 15ms per 10 pieces, increasing efficiency by 6.7 times. Key caching: The node caches the public key locally (valid for 1 hour) to avoid re-acquisition for each transmission, reducing network interaction time (from 50ms to 5ms).

[0026] The digital twin fusion module, based on a LSTM-PPO hybrid model, integrates historical one-hour load data with real-time status data (including node location and task type) to generate a 15-minute load trend forecast (with an error rate of ≤12%). It also provides a virtual "trial-and-error sandbox" for rehearsing strategic risks. By integrating historical time series data with real-time status data, the LSTM-PPO hybrid prediction model is trained to construct a virtual image of the drone cluster, generating dynamic feature data including load trend forecasts. This provides a "trial-and-error sandbox" for decision-makers: Strategies such as service migration and resource adjustment can be rehearsed in a virtual space (for example, simulating the migration of five microservices to a backup node). This allows for early exposure to risks such as network congestion and resource conflicts, helping to prevent real-world cluster volatility. Dynamic feature data including "15-minute load trends" is generated, enabling forward-looking decision-makers.

[0027] LSTM-PPO hybrid model 1. Model structure design LSTM (Long Short-Term Memory) is responsible for extracting time series data features, and PPO (Proximal Policy Optimization) is responsible for predictive policy optimization. The two work together to generate load trends for the next 15 minutes. The structure is shown in Table 2: Table 2 LSTM-PPO hybrid model structure design ; 2. Training and prediction logic Training data preparation: We used cluster load data from the past seven days (approximately 6 million records) and divided it into a training set (4.2 million records) and a validation set (1.8 million records) in a 7:3 ratio. Each record contained five types of features (CPU, GPU, memory, network latency, and number of tasks). LSTM feature extraction: Gating units (input gate, forget gate, and output gate) filter out redundant features, focusing on retaining load peaks (such as high CPU load at 10:00 and 16:00) and trend features (such as the linear increase in GPU load as tasks increase). PPO strategy optimization: Using the LSTM output features as input, the prediction parameters are adjusted through "policy gradient" to minimize the "MSE (mean square error)" between the predicted and actual loads. Each training round is iterated 100 times, and the MSE converges to ≤0.01 (corresponding to an error rate of ≤10%). Prediction execution: For real-time prediction, the load data for the last hour is input. The model outputs 90 prediction points within 500ms (10 seconds / point), with an error rate of ≤12%, and marks high-risk intervals (such as periods where the predicted GPU load is ≥85%).

[0028] 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 converts basic indicators such as CPU utilization and network latency into "decision factors" that can directly guide action (for example, a load balancing index > 0.8 triggers task rebalancing). This reduces the complexity of the decision-making algorithm and provides "lightweight feature vectors" for the quantum-inspired optimizer. By focusing on key dimensions such as resource bottlenecks and service quality, the quantum annealing algorithm reduces solution time by 50%, enabling more efficient handling of NP-hard scheduling problems.

[0029] The decision-making layer is connected to the data processing layer, and makes decisions on adjustments, service migration, or negative entropy injection based on the policy library and dynamic feature data. It includes a policy library for storage resource adjustment / service migration / negative entropy injection strategies, and a quantum-inspired optimizer that solves NP-hard problems based on the quantum annealing algorithm. The dynamic features output by the data processing layer are deeply integrated (including digital twin prediction results and federated learning node capability assessment) to form a "data-driven" decision-making mechanism; high-precision, fine-grained adjustment instructions are provided to the execution layer (such as CPU quotas for each container and bandwidth allocation for each network link) to ensure efficient and secure operations; thermodynamic entropy theory and quantum computing technology are introduced into cluster decision-making to solve the problems of "extensive decision-making, low efficiency, and poor scalability" in traditional distributed systems. The quantum-inspired optimizer is used to efficiently solve NP-hard problems in cluster resource allocation, task scheduling, and service migration path planning, specifically including: Problem modeling module, which transforms the actual problem into an energy function containing decision variables, objective functions and constraint penalty terms; The quantum annealing core module achieves global optimization through quantum state initialization, tunneling process simulation, and annealing scheduling; The classical-quantum interface module uses a hybrid architecture of quantum-inspired and classical computing to adapt to the limited computing power of drone nodes; The result verification and correction module performs feasibility verification and local correction on the output optimal solution; It also has lightweight, distributed collaboration and dynamic response capabilities, and can interrupt and restart the annealing process according to changes in cluster topology.

[0030] NP-hard problem solving algorithm (quantum annealing) Algorithm principles and adaptation scenarios Quantum annealing leverages the quantum tunneling effect to rapidly traverse the solution space for NP-hard problems such as task allocation and resource scheduling. Compared to traditional genetic algorithms, it can avoid falling into local optimal solutions. In this system, it is used for the "100+ node task allocation" scenario (goal: minimize total task latency and maximize resource utilization).

[0031] 2. Algorithm implementation steps Problem modeling: Convert task assignment into an integer programming problem and define variables: x_ij: 1 indicates that task i is assigned to node j, 0 indicates that it is not assigned; Objective function: min(ΣΣx_ij×t_ij) - max(ΣΣx_ij×r_ij) (t_ij is the execution delay of task i on node j, r_ij is the resource utilization); Constraints: Each task is assigned to only one node, and the node resource usage is ≤ the hardware limit (such as CPU ≤ 75%).

[0032] Quantum annealing process: Initialization: Set the initial temperature T0=100 (to simulate the energy of the quantum system) and randomly initialize the solution space (randomly assign tasks); Tunneling search: Through the quantum tunneling effect, jump to a new solution (adjust the x_ij value) with probability P = exp(-ΔE / T), where ΔE is the energy difference between the new solution and the current solution; Temperature decay: After every 100 iterations, the temperature decays according to T = T0 × 0.95, gradually reducing the tunneling probability and stabilizing to the optimal solution; Termination condition: The temperature drops to T=1 or the objective function does not improve after 20 consecutive iterations, and the optimal solution (task allocation plan) is output.

[0033] 3. Performance indicators Solution efficiency: In a 100-node, 50-task scenario, the solution time is reduced from 100ms for traditional genetic algorithms to less than 40ms, reducing decision latency by 60%. Solution quality: The output task allocation solution reduces total latency by 25% compared to traditional algorithms and reduces resource utilization variance from ±15% to ±8%; Fine-grained output: The plan includes resource allocation accuracy (CPU 0.1 core, memory 100MB), such as "Task 1 (image recognition) is assigned to node A: 0.8 CPU core + 2GB memory, execution latency ≤ 300ms."

[0034] Power inspection scenario: The strategy library prioritizes the use of the "photovoltaic 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 to shorten the task completion time.

[0035] Extreme weather response: When the perception layer detects a wind speed > 15m / s, the decision layer automatically triggers "wind resistance mode": Hardware layer: Reduce the drone's flight altitude to 50 meters and limit the maximum roll angle to 20°; Network layer: Switch to Mesh self-organizing network protocol, and adjust the communication frequency band to 5.8GHz (for stronger anti-interference ability); Mission layer: Pause high-precision shooting missions and switch to radar scanning mode.

[0036] The task instructions output by the decision generation module contain geo-fence information (such as "Do not enter the area with a radius of 100 meters of coordinates (x, y)"), which are synchronized to the command center GIS platform in real time through the Kafka message queue.

[0037] Specifically, the policy library stores drone node resource adjustment policies, service migration policies, and negative entropy injection policies designed based on the system entropy change model. This maintains cluster order by dynamically balancing system entropy. The policy library dynamically balances: pre-stores three types of policies: resource adjustment, service migration, and negative entropy injection. These policies maintain cluster order based on the system entropy change model and support custom policy plug-ins. These policies automatically match task types (such as inspection and rescue), reducing manual intervention costs.

[0038] The quantum-inspired optimizer uses a quantum annealing algorithm to solve task allocation problems and outputs fine-grained instructions that include resource allocation precision. The quantum-inspired optimizer: Using a quantum annealing algorithm to solve NP-hard problems such as multi-dimensional resource scheduling, it reduces decision latency from 100ms to under 40ms (a 60% reduction) compared to traditional genetic algorithms. It also generates refined instructions that include resource allocation precision (0.1 CPU core, 100MB of memory) (Decision Layer - Quantum-Inspired Optimizer).

[0039] The decision generation module combines the dynamic feature data output by the data processing layer with the adaptive policies in the policy library to generate specific instructions for adjusting drone node resource allocation, migrating services, and rescheduling tasks, providing operational basis for the execution layer. The module converts the feature parameters output by the data analysis module (such as the ID of the resource bottleneck node and the level of service degradation) into specific instructions (such as "migrate 3 GPU tasks from node A to node B" or "expand the database service replicas to 5") with an accuracy rate of ≥95%. This module supports fine-grained operations at the execution layer, generating refined instructions that include resource allocation accuracy (CPU 0.1 core, memory 100MB) and service migration order (migrating stateless services first, then stateful services), ensuring the efficiency and security of adjustment operations. The execution layer, connected to the decision layer, executes the adjustments it generates. It includes a hardware resource adjustment module (dynamically allocating CPU / GPU computing power), a service control module (process startup and shutdown, task migration), and a network communication adjustment module (bandwidth configuration, link optimization). By integrating multiple technologies (SDN / NFV network resiliency, TEE secure execution, and system-level optimization through negative entropy injection), the execution layer ensures efficient and secure decision implementation, enabling dynamic mapping between physical clusters and digital twins.

[0040] Specifically, the node management module dynamically adjusts node hardware resource allocation and performs task migration based on the negative entropy injection mechanism; the service control module verifies and executes key decisions through a trusted execution environment (such as Intel SGX) and supports dynamic switching of training modes for nodes participating in federated learning; the elastic network execution unit implements millisecond-level network topology switching based on SDN, and dynamically orchestrates service chains and optimizes communication paths through NFV technology.

[0041] Extreme environment execution mechanism: electromagnetic interference scenario: 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 the interference frequency band, and the network connectivity retention rate is ≥95%.

[0042] Hardware fault tolerance: When the node management module detects a CPU core failure in a drone, it triggers the "hot migration" mechanism: Migrate the node task to the adjacent node within 100ms; Send a hardware failure alarm to the perception layer and start the backup drone self-test process.

[0043] The system also includes a feedback layer, which connects to the execution layer and the data processing layer. This layer collects new state information after decision execution and feeds it back to the data processing layer. If the new state deviates from the prediction by more than a threshold, it triggers incremental model training and policy library updates in the data processing layer, forming a self-growing closed loop. The feedback layer scans for execution deviations in real time and drives policy iteration, enabling the cluster to autonomously adapt and self-repair in complex environments. Ultimately, this forms a self-growing closed loop of "data-driven decision making, execution verifying data." The feedback layer includes a feedback sensor module and a feedback data processing module. The feedback sensor module collects node state information 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. If the load prediction error exceeds 15%, it triggers incremental training of the LSTM-PPO model. If the policy execution deviation (e.g., an increase in error rate after service migration) exceeds a threshold, the policy library parameters are automatically optimized.

[0044] Model incremental training algorithm (LSTM-PPO incremental update) 1. Trigger conditions and data selection When the feedback layer detects a load forecast error > 15%, incremental training of the LSTM-PPO model is triggered. The core principle is to "update the model with the latest data without discarding historical parameters." The data selection rules are as follows: Training dataset: Actual load data for the past hour (approximately 36,000 records), including data from periods where the forecast error exceeded the standard (with a 20% increase in weight); Validation dataset: Nearly 10 minutes of real-time data (about 6,000 records), used to verify the accuracy of the updated model.

[0045] 2. Incremental training steps Parameter freezing and unfreezing: Freeze the parameters of the first two LSTM layers (preserving the ability to extract historical time series features), and unfreeze the parameters of the third LSTM layer and the PPO strategy layer; Local training: train the unfrozen layer with the new dataset, and set the learning rate to 1 / 10 of the initial training (0.001→0.0001) to avoid parameter oscillation; Accuracy verification: Test the model with the validation dataset. If the prediction error is ≤12%, update the model; otherwise, increase the number of training rounds (from 10 to 20) until the accuracy meets the target. Model replacement: The updated model is previewed in the digital twin environment for 10 minutes. After confirmation that there are no anomalies, the online model is replaced. The replacement takes ≤100ms and does not interrupt the prediction service.

[0046] Performance evaluation system: Test environment: simulates a 100-node cluster, including 20% ​​computing nodes (equipped with NVIDIA Jetson AGX Orin), 60% communication nodes, and 20% task nodes (equipped with multispectral cameras); Test indicators: Resource utilization: CPU / GPU average load ≤ 75%, memory utilization ≤ 80%; Task response delay: emergency tasks ≤ 500ms, routine tasks ≤ 2s; System reliability: MTBF ≥ 1000 hours, MTTR ≤ 30 minutes; Comparative data: Compared with traditional PID control-based cluster systems, this system improves task completion efficiency by 40% and reduces energy consumption by 20% (by reducing ineffective resource scheduling through negative entropy injection strategy).

[0047] This invention's multi-source data acquisition covers five major categories of drone status information: hardware resources (CPU / GPU usage, memory / disk I / O), software status (service processes, response time), and network performance (latency / throughput). This system combines hardware sensors (physical acquisition) with software sensors (program plug-ins) to achieve comprehensive monitoring of cluster node status. Data quality assurance: Raw data is denoised and corrected using algorithms such as median filtering and interpolation. Data acquisition frequency reaches microseconds, and position and posture 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 for ciphertext transmission. This ensures that local drone data is "available but not visible" in federated learning, meeting the needs of sensitive scenarios such as disaster relief and inspections.

[0048] The above shows and describes 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 above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A distributed system for dynamic perception and adaptation of drone swarm environments, characterized by: It includes the perception layer, data processing layer, decision layer, execution layer and feedback layer distributed in each node of the drone cluster, forming a closed-loop dynamic adaptive architecture; among which: The perception layer is composed of sensor modules at each node, which is used to collect node status information in real time and transmit it to the data processing layer through a distributed network; The data processing layer is connected to the perception layer and is responsible for the preprocessing, privacy protection, data fusion and analysis of the original state information, generating dynamic feature data, including data cleaning, federated learning privacy protection, digital twin fusion, and data analysis modules; among them, the federated learning privacy protection module: the differential privacy unit dynamically adjusts the noise intensity according to the sensitivity of the data, using Laplace noise for high-sensitivity data and Gaussian noise for low-sensitivity data to meet the preset privacy budget; the homomorphic encryption unit encrypts the local model parameters and transmits them to the aggregation node, supporting ciphertext aggregation and encrypted distribution of global parameters; the dynamic node management unit allocates temporary keys, adjusts noise and encryption parameters when a new node joins, and cancels the key, clears the temporary parameters and triggers the incremental encryption update of the global model when the node exits; The decision-making layer connects to the data processing layer and generates adjustments, service migration, or negative entropy injection decisions based on the policy library and dynamic feature data. It includes a policy library that stores relevant policies and a quantum-inspired optimizer based on the quantum annealing algorithm to solve NP-hard problems. The execution layer connects to the decision-making layer to execute adjustment decisions, including hardware resource adjustment, service control, and network communication adjustment modules; The feedback layer connects the execution layer and the data processing layer, collecting new state feedback after execution; if the deviation between the new state and the prediction exceeds the threshold, it triggers incremental training of the data processing layer model and update of the strategy library, forming a self-growing closed loop.

2. The distributed system for dynamic perception and adaptation of drone swarm environments according to claim 1 is characterized by: 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 usage and memory occupancy; the software sensors are used to collect software operation status and network communication status information such as service response time and network delay.

3. The distributed system for dynamic perception and adaptation of drone swarm environments according to claim 1, characterized in that: The data cleaning module adopts 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, and finally unified into the Protocol Buffers standard format.

4. The distributed system for dynamic perception and adaptation of drone swarm environments according to claim 1, characterized in that: The node status information includes hardware resource information, software operation 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.

5. The distributed system for dynamic perception and adaptation of drone swarm environments according to claim 1, characterized in that: 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 forecast for the next 15 minutes and preview strategic risks.

6. The distributed system for dynamic perception and adaptation of drone swarm environments according to claim 1, characterized in that: The quantum-inspired optimizer solves the task allocation problem involving 100+ nodes based on the quantum parallelism of the quantum annealing algorithm and outputs fine-grained instructions with resource allocation accuracy. The quantum-inspired optimizer is used to efficiently solve NP-hard problems in cluster resource allocation, task scheduling, and service migration path planning, specifically including: Problem modeling module, which transforms the actual problem into an energy function containing decision variables, objective functions and constraint penalty terms; The quantum annealing core module achieves global optimization through quantum state initialization, tunneling process simulation, and annealing scheduling; The classical-quantum interface module uses a hybrid architecture of quantum-inspired and classical computing to adapt to the limited computing power of drone nodes; The result verification and correction module performs feasibility verification and local correction on the output optimal solution; It also has lightweight, distributed collaboration and dynamic response capabilities, and can interrupt and restart the annealing process according to changes in cluster topology.

7. The distributed system for dynamic perception and adaptation of drone 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 greater than 15%, the incremental training of the LSTM-PPO model is triggered; if the strategy execution deviation exceeds the threshold, the strategy library parameters are automatically optimized.

Images