Unmanned aerial vehicle group low-energy-consumption flight path planning method based on distributed calculation

Abstract

The invention discloses an unmanned aerial vehicle group low-energy-consumption flight path planning method based on distributed computing, and the method comprises the steps: constructing a three-stage distributed computing node network which comprises a unique core edge node, region edge nodes configured according to the scale of an unmanned aerial vehicle group, and unmanned aerial vehicle airborne nodes consistent with the scale of the unmanned aerial vehicle group; the core edge nodes, the region edge nodes and the unmanned aerial vehicle airborne nodes are in communication connection, sub-regions are uniformly divided according to the area of a task region, each sub-region corresponds to one region edge node, and the boundaries of the sub-regions coincide with terrain obstacles; the computing power distribution is that the core edge node undertakes global aggregation, the area edge node undertakes local calculation, and the unmanned aerial vehicle airborne node undertakes preprocessing. The defects that in the prior art, a traditional distributed mode lacks a scenarized energy consumption model and global collaboration, and large-scale application of the unmanned aerial vehicle group is restricted are overcome, and the real-time performance of flight path planning and energy consumption control are improved.

Description

Technical Field

[0001] This invention relates to the field of UAV swarm collaborative control technology, specifically to a low-energy trajectory planning method for UAV swarms based on distributed computing. This method is applicable to multi-UAV collaborative operations in complex scenarios such as power line inspection in mountainous areas and environmental monitoring, and can significantly improve the real-time performance of trajectory planning and energy consumption control. Background Technology

[0002] In complex scenarios such as power line inspections and environmental monitoring in mountainous areas, drone swarms have gradually replaced traditional manual operations due to their high flexibility, wide coverage, and adaptability to complex terrain. Flight path planning, as a core component of collaborative operations, must meet three major requirements: adapting to environmental constraints such as obstacles, ensuring real-time collaborative collision avoidance among multiple drones, and controlling energy consumption to meet the requirements of long-term operations.

[0003] Current UAV swarm trajectory planning technologies are mainly divided into centralized and distributed approaches. Centralized planning relies on a central node (ground control station or lead UAV) to uniformly process data and generate trajectories, which has significant drawbacks in mountainous scenarios, such as real-time failures, poor reliability, and unreliable endurance. While traditional distributed planning solves the efficiency problem of centralized planning, it uses the same planning algorithm and frequency for all UAVs, failing to adapt to the varying loads of inspection and monitoring. Low-battery UAVs performing high-load computations further shorten their endurance, resulting in a lower task completion rate. Existing technologies face a core conflict between the "dynamic nature of complex environments, the specificity of tasks, and the need for long endurance" and the "insufficient real-time performance, poor reliability, and inefficient energy consumption." Centralized planning cannot adapt to the communication and data requirements of mountainous areas, while traditional distributed planning lacks scenario-based energy consumption models and global coordination, hindering the large-scale application of UAV swarms. Therefore, developing a distributed trajectory planning method adapted to mountainous areas that balances real-time performance and low energy consumption is crucial. Summary of the Invention

[0004] This invention proposes a low-energy trajectory planning method for UAV swarms based on distributed computing, aiming to solve the shortcomings of the existing technology, such as the lack of scenario-based energy consumption models and global coordination in traditional distributed systems, which restricts the large-scale application of UAV swarms, and improves the real-time performance and energy consumption control of trajectory planning.

[0005] A low-energy trajectory planning method for UAV swarms based on distributed computing includes constructing a three-level distributed computing node network. This network comprises a unique core edge node, regional edge nodes configured according to the UAV swarm size, and UAV-borne nodes of the same size as the swarm. The core edge node, regional edge nodes, and UAV-borne nodes are communicatively connected. The task area is evenly divided into sub-regions, with each sub-region corresponding to one regional edge node. The sub-region boundaries coincide with terrain obstacles. Computing power is allocated as follows: the core edge node handles global aggregation, the regional edge nodes handle local computation, and the UAV-borne nodes handle preprocessing. The specific method is as follows:

[0006] Step S1: Distributed data processing and environmental modeling to generate an environment-energy consumption coefficient mapping model;

[0007] Step S2: Planning layer trajectory optimization, including global pre-planning and local dynamic adjustment;

[0008] Step S3: Collaboration layer conflict management, resolving path conflicts through blockchain distributed consensus and distributed retrieval of time-space conflict matrix;

[0009] Step S4: Energy layer battery management and linkage, realizing energy and path linkage optimization based on the differences in UAV battery status;

[0010] Step S5: Two-way feedback optimization, correcting model parameters through two-way feedback from the planning layer to the perception layer and from the energy layer to the planning layer.

[0011] Furthermore, step S1 includes step S1.1 distributed fusion of multimodal sensor data and step S1.2 distributed construction of a two-layer environment model;

[0012] Step S1.1 includes:

[0013] The UAV-borne node collects lidar point cloud data, high-definition visual camera images, inertial measurement unit attitude data, and micro weather station environmental parameters, performs outlier detection using an improved adaptive Kalman filter, and uploads the compressed data to the regional edge node.

[0014] After decompressing the data at the edge nodes of the region, a sliding window noise estimation is performed, and the three-dimensional coordinates of obstacles, average wind speed, and ambient temperature within the region are output.

[0015] The core edge nodes receive the regional fusion results, convert them into a globally universal coordinate system and remove duplicates, and generate a global perception dataset.

[0016] Step S1.2 includes:

[0017] Static terrain layer: The core edge nodes will split and distribute the pre-loaded high-resolution digital elevation model by region, and fit the air drag coefficient at different altitudes through wind tunnel experiments;

[0018] Dynamic interference layer: The core edge nodes divide the task airspace into cubic grids. Each region's edge nodes update the parameters of their respective sub-matrixes, including the moving obstacle grid, wind power consumption coefficient grid, and battery temperature decay coefficient grid. The update priority is moving obstacle grid > wind power consumption coefficient grid > battery temperature decay coefficient grid.

[0019] Furthermore, the global pre-planning in step S2 includes:

[0020] Step S2.1.1: The core edge nodes generate multiple candidate path individuals with waypoints, and distribute them evenly according to the number of edge nodes in the region;

[0021] Step S2.1.2: The regional edge node calls the dynamic interference layer energy consumption coefficient, combines it with the UAV parameters, and calculates the total energy consumption of candidate individuals according to the objective function E=αL+βF+γC+δB, and selects the locally optimal individual; where L is the total path length, α is the basic energy consumption weight, F is the total wind energy consumption, β is the highest weight, C is the communication energy consumption, γ is the secondary energy consumption weight, B is the battery loss, and δ is the lowest weight.

[0022] Step S2.1.3: The core edge node receives the locally optimal individual, sorts them by total energy consumption, and assigns the individual with the lowest total energy consumption to the drone with the lowest remaining power, thereby controlling the difference in total energy consumption among drones in the group.

[0023] Furthermore, the local dynamic adjustment in step S2 includes:

[0024] Step S2.2.1: The core edge nodes split the historical flight data into sub-pools according to the scenario, and the regional edge nodes train the corresponding scenario-specific sub-models, including the rockfall obstacle avoidance sub-model, the strong wind height adjustment sub-model, and the low temperature energy saving sub-model.

[0025] Step S2.2.2: The core edge nodes fuse the sub-models by weighted average to generate a global model; the UAV onboard nodes quantize the global model into low precision and split it according to the scenario, load the corresponding sub-models according to the local perception data and output action commands; when communication is interrupted, the local pre-stored sub-model emergency library is activated; only the path within a specific range after the current waypoint is corrected to control the additional energy consumption increment.

[0026] Furthermore, the blockchain distributed consensus in step S3 includes: regional consensus nodes receiving the short-term future path prediction data of the drone, executing the Byzantine fault-tolerant consensus algorithm, and after verification through the request, preprocessing, preparation, commitment, and submission stages, packaging and generating blocks and uploading them to the core consensus node; the core consensus node verifying the integrity and hash consistency of the blocks, generating a global blockchain ledger and storing it by timestamp.

[0027] Further, the distributed retrieval of the time-space conflict matrix in step S3 includes: the core edge node splits the global time-space matrix by region, and the regional consensus node stores the sub-matrix; the UAV node sends a grid occupancy request to the regional consensus node, the regional consensus node retrieves the occupancy status, and after determining the conflict, calculates the priority according to the scoring formula Score=0.6×(1 - SOC / 100)+0.4×(2 - F), and retains the high-priority UAV path;

[0028] In the event of a cross-regional conflict, the consensus nodes in the affected regions report to the core edge nodes, which then coordinate based on global priorities.

[0029] Further, step S4 includes:

[0030] Step S4.1: The UAV onboard node calculates the actual usable battery capacity using the formula Ca=Cn×[1−k1×(0−T)−k2×(P / Pmax)], where Ca is the actual usable capacity, Cn is the nominal capacity, T is the battery temperature, P is the current load power, Pmax is the maximum load power, and k1 and k2 are fitting coefficients; and uploads the actual capacity, remaining power (SOC), and temperature to the regional consensus node periodically.

[0031] Step S4.2: Regional consensus nodes are controlled in layers according to SOC: SOC ≥ 60% executes remote tasks and high-load configuration; 30% ≤ SOC < 60% flies along the baseline path and is configured for medium load; SOC < 30% initiates energy-saving return and low-load configuration.

[0032] Step S4.3: The core edge node calculates the regional energy pressure index, issues energy-saving instructions to high-pressure areas, and issues task supplement instructions to low-pressure areas.

[0033] Further, step S5 includes:

[0034] The planning layer compares the "predicted energy consumption coefficient of the dynamic interference layer" with the "actual flight energy consumption coefficient". When the deviation exceeds the threshold, it feeds back to the perception layer to adjust the calculation parameters of the dynamic interference layer. The energy layer monitors the actual SOC decay rate. When it exceeds the expectation, it feeds back to the planning layer to adjust the weight of the battery loss parameter in the objective function. Feedback optimization is performed once every fixed number of steps to reduce the total energy consumption of the group and improve the success rate of path planning.

[0035] Furthermore, the high-speed backbone network is based on the latest communication standards, and its latency remains stable at a low level even in complex weather and obstructed mountainous environments; the "sun-blind" ultraviolet light communication conforms to general communication protocols and features low power consumption, medium communication distance, low bit error rate, and high resistance to electromagnetic interference; the power consumption of the UAV onboard node is a low percentage of the total power consumption of the UAV when it is fully loaded.

[0036] Furthermore, the core edge node is configured with dual power backup, and the regional edge node is equipped with a power replenishment module; all nodes of the three-level distributed computing node network are in a safe operating range when fully loaded, making it suitable for scenarios such as power inspection in mountainous areas, forest fire monitoring, and urban emergency rescue.

[0037] The beneficial effects of this invention are as follows: The low-energy trajectory planning method for UAV swarms based on distributed computing provided by this invention addresses the lack of scenario-based energy consumption models and global collaborative capabilities in existing distributed technologies. It innovatively constructs a two-layer energy consumption model that integrates multi-source environmental data, quantifying dynamic factors such as wind field changes, temperature distribution, and terrain features in complex environments into calculable energy consumption coefficients. Environmental data is collected in real time through a distributed sensing network, and an adaptive filtering algorithm is used for data fusion, effectively improving the accuracy of energy consumption prediction and environmental adaptability. A distributed parallel computing framework is adopted, decomposing the trajectory planning task into multiple sub-tasks for concurrent execution, significantly improving computational efficiency. Local preprocessing and regional optimization reduce data transmission volume and communication latency, ensuring real-time response capability for trajectory planning in complex environments. The innovatively designed multi-objective optimization algorithm comprehensively considers multiple factors such as path length, wind resistance, communication energy consumption, and battery consumption. Through a weighted optimization function, it achieves refined energy consumption management. This method can dynamically adjust the trajectory according to real-time environmental changes, ensuring that the UAV swarm always maintains optimal energy consumption.

[0038] Meanwhile, through a three-tiered distributed computing architecture, global energy consumption optimization is achieved while maintaining the robustness of the distributed system. Core edge nodes are responsible for global situational awareness and collaborative decision-making, regional edge nodes perform local optimization calculations, and UAV-borne nodes perform real-time data acquisition and processing. This hierarchical collaborative mechanism avoids the bottleneck of centralized computing and overcomes the shortcomings of traditional distributed systems that lack a global perspective. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart illustrating the low-energy trajectory planning method for unmanned aerial vehicle swarms in this invention.

[0041] Figure 2 This is a flowchart illustrating step S1, distributed data processing and environment modeling, in this invention.

[0042] Figure 3 This is a schematic diagram of the planning layer trajectory optimization process in step S2 of the present invention;

[0043] Figure 4 This is a schematic diagram of the collaboration layer conflict control process in step S3 of the present invention;

[0044] Figure 5This is a schematic diagram of the energy layer battery management process in step S4 of the present invention;

[0045] Figure 6 This is a schematic diagram of the bidirectional feedback optimization process in step S5 of the present invention. Detailed Implementation

[0046] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0047] A low-energy trajectory planning method for UAV swarms based on distributed computing, combined with... Figure 1 As shown, this technical solution addresses the problems of "high computational latency, unbalanced computing load, high communication energy consumption, and low energy management accuracy" in traditional centralized trajectory planning for large-scale UAV swarms in complex scenarios such as power line inspection in mountainous areas, forest fire monitoring, and urban emergency rescue. By constructing a three-level distributed computing node network, the traditional centralized task is decomposed into three stages of processing: "local preprocessing - regional optimization - global aggregation", which ensures the real-time performance, controllable energy consumption, and operational reliability of UAV swarm trajectory planning in complex environments.

[0048] First, the drone swarm is configured with core edge nodes, regional edge nodes, and drone-borne nodes. The core edge node is a single, dedicated node equipped with a high-performance server, a graphics processing unit, and ample memory and storage to support high-concurrency processing. Its core functions include global parameter distribution, cross-regional collaboration, final result aggregation, and global model fusion. Communication with the regional edge nodes utilizes a high-speed backbone network based on the latest communication standards, featuring high bandwidth, low latency, low packet loss rate, and high anti-interference capabilities. Through cross-regional transmission testing in mountainous areas, under complex weather and obstruction conditions, latency remained consistently low, meeting the real-time requirements of global collaboration.

[0049] The number of regional edge nodes is configured according to the scale of the drone swarm. For example, one regional edge node corresponds to every 10 drones. The regional edge nodes use medium-performance servers that support wide-temperature operation to adapt to complex outdoor environments. The core functions cover data processing, local optimization, conflict coordination, and sub-model training within the region. Communication with the drone's onboard nodes adopts "solar-blind" ultraviolet light communication. This communication method conforms to general communication protocols and has low power, medium communication distance, low bit error rate, and high electromagnetic interference resistance, meeting the communication needs of strong interference scenarios.

[0050] The number of onboard nodes for each drone is consistent with the size of the drone swarm. Each drone is equipped with an onboard node, which uses a lightweight processor with high energy efficiency, along with a storage module and a power management module that supports dynamic power adjustment. Its core functions include local sensor data preprocessing, model inference, energy consumption calculation, and flight parameter execution. Communication with edge nodes in the region uses ultraviolet light communication and integrates a miniature antenna. This antenna is small in size and adaptable to the needs of drones. After testing with multi-rotor drones, the power consumption of the onboard node when working at full load is a low percentage of the total power consumption of the drone, thus avoiding affecting the flight endurance.

[0051] The adaptation logic of the node network is as follows: The task area is evenly divided into sub-regions, with each sub-region corresponding to one edge sub-node. Geographic information map overlay analysis ensures that the communication distance between all drones and sub-nodes within the region is within an effective range, and that the sub-region boundaries coincide with terrain obstacles (such as mountains and rivers), reducing the frequency of cross-regional task scheduling. The computing power allocation principle is determined through computing power load testing of the drone swarm. Core nodes undertake global aggregation tasks, mainly including data alignment and model fusion; regional sub-nodes undertake local computing tasks, mainly including data fusion and sub-model training; and airborne nodes undertake preprocessing tasks, mainly including outlier detection and data compression, avoiding overload of a single node. Actual testing shows that each node operates within a safe operating range even at full load. Core edge nodes are configured with dual power backup, and regional sub-nodes are equipped with power replenishment modules to ensure continuous operation in complex scenarios.

[0052] Step S1: Distributed Data Processing and Environment Modeling

[0053] Combined with appendix Figure 1 , 2 As shown, this step, through an engineered process of distributed data cleaning, regional fusion, and global alignment, generates an "environment-energy consumption coefficient" mapping model, providing accurate energy consumption quantification for the planning layer. Parameters at each stage undergo multi-scenario testing and calibration to ensure data accuracy and real-time performance. Specifically, this includes:

[0054] Step S1.1: Distributed fusion of multimodal sensor data

[0055] Each drone's onboard nodes collect multi-source sensor data in real time and perform "outlier detection" using an improved adaptive Kalman filter (AKF). The sensor selection and preprocessing logic have been optimized through multi-scenario testing.

[0056] UAV-borne node data acquisition: LiDAR collects point cloud data at a fixed frequency. When the deviation of a point cloud from the average position of historical multi-frames exceeds a threshold, it is judged as a false point cloud caused by occlusion or reflection and is directly removed. After removing abnormal points, the point cloud efficiency is significantly improved. High-definition vision camera collects images at a fixed frame rate. When overexposure or low illumination causes the proportion of invalid pixels to exceed a threshold, it is judged as an invalid frame and does not participate in subsequent processing. After removing invalid frames, the image efficiency is improved. Inertial measurement unit (IMU) collects attitude data at a high frequency to help judge the validity of sensor data. For example, when a drastic attitude change is detected, LiDAR point cloud fusion is paused. Mini weather station collects environmental parameters (temperature, wind speed, wind direction) at a fixed frequency for subsequent energy consumption coefficient calculation. Finally, the effective data is compressed using an efficient compression algorithm and uploaded to the edge sub-node of the corresponding area via ultraviolet light communication at a fixed period. The measured data compression significantly reduces the communication bandwidth requirement and communication energy consumption.

[0057] Regional edge fusion computing: Regional edge sub-nodes receive compressed data from UAVs within the group, quickly decompress it using a hardware decompression module, and then perform AKF's "sliding window noise estimation." The fusion parameters have undergone multiple rounds of iterative testing and optimization. The length of the sliding window needs to balance noise estimation accuracy and real-time performance. Tests show that when the window length is at a specific period, the noise estimation error and latency are balanced. The noise covariance is dynamically adjusted, adjusting the measurement noise covariance based on the validity of sensor data and the process noise covariance based on environmental dynamics. The output results include the three-dimensional coordinates of obstacles within the region, average wind speed, and ambient temperature, updated at a fixed period. The fused data is then used for obstacle avoidance.

[0058] Global data alignment: Edge sub-nodes of each region upload the regional fusion results to the core edge nodes at fixed intervals through the high-speed backbone network. During the coordinate transformation process, the core nodes convert the local coordinate system of each region to the global universal coordinate system, and the transformation error is controlled at a low level. In the data deduplication process, duplicate obstacle data across regions are merged and the coordinates with the highest accuracy are retained. Finally, a global perception dataset is generated, stored in a unified format, and used by the planning layer. After the global data alignment is measured, the path conflict rate of UAVs across regions is reduced.

[0059] Step S1.2: Distributed Construction of a Two-Layer Environment Model

[0060] A two-layer model, including static and dynamic elements, was constructed to transform environmental factors into calculable energy consumption coefficients. The model parameters and update logic were experimentally verified to ensure quantification accuracy.

[0061] The static terrain layer is pre-loaded, requiring no real-time processing, thus reducing computational load during missions. Its data source is a high-resolution digital elevation model (DEM), acquired from a geographic information platform and calibrated through on-site surveying, containing information such as altitude, slope, and terrain type. During preprocessing and distribution, the core node splits the DEM into data blocks by region and pre-distributes them to sub-nodes in each region. These sub-nodes then distribute the data to UAVs within the group for storage, avoiding repeated loading during missions and reducing storage module power consumption. Key parameters are fitted using wind tunnel experiments to measure air drag coefficients at different altitudes. The air drag coefficient decreases with increasing altitude, showing minimal deviation from theoretical calculations and accurately reflecting the impact of altitude on flight energy consumption.

[0062] The dynamic interference layer needs to be updated in real time to adapt to environmental changes. The core node divides the task airspace into a cubic grid. Each grid stores parameters such as obstacle type, wind force coefficient, temperature coefficient, communication quality, and coverage status. Then, it is split into regional sub-matrices by region. Each regional sub-node is responsible for updating the parameters of its sub-matrix. The update logic has been verified in multiple scenarios. The mobile obstacle grid consists of moving targets detected by drones within the group and is updated at a fixed frequency. The marking of mobile obstacles provides a reliable basis for obstacle avoidance planning. The wind energy consumption coefficient grid is calculated using a formula with the angle between wind speed and flight direction as variables. The parameters are fitted through multiple flight experiments. The coefficient is greater than 1 when facing headwinds (energy consumption increases), less than 1 when facing tailwinds (energy consumption decreases), and close to 1 when facing crosswinds (no additional energy consumption). The battery temperature decay coefficient grid is also calculated using a formula with ambient temperature as a variable. The formula is fitted based on the performance test data of lithium batteries at different temperatures, resulting in a small capacity prediction error. The coefficient is less than 1 at low temperatures (capacity decay), and close to or slightly greater than 1 at suitable temperatures (capacity is stable or slightly improved). The update priority is set as mobile obstacle grid (highest priority) > wind energy consumption coefficient grid > battery temperature decay coefficient grid to ensure that high dynamic factors are updated first. The measured update delay of the dynamic interference layer meets the real-time requirements of the planning layer.

[0063] Step S2: Distributed trajectory optimization at the planning layer

[0064] Combined with appendix Figure 1 , 3 As shown, it specifically includes:

[0065] Step S2.1, Global Pre-planning

[0066] A baseline path covering the entire mission cycle is generated for each drone to ensure the lowest total energy consumption of the swarm. The calculation process has undergone multiple rounds of iterative optimization to ensure efficiency and accuracy in large-scale swarms.

[0067] S2.1.1 Candidate Individual Initialization and Allocation

[0068] Based on global perception data, the core node generates multiple candidate path individuals, each corresponding to a path with multiple waypoints. The spacing between waypoints is set according to task requirements. Then, the candidate individuals are evenly distributed according to the number of sub-nodes in the region to avoid centralized calculation of the core node. In actual tests, the time spent on core node allocation is negligible.

[0069] S2.1.2, Calculation of regional end fitness

[0070] Each regional sub-node calls the "dynamic interference layer energy consumption coefficient" of its region, and calculates the total energy consumption objective function for each candidate individual by combining the UAV's own parameters: E=αL+βF+γC+δB. The definition and weight values ​​of each parameter are determined through orthogonal experiments to ensure the accuracy and rationality of energy consumption quantification.

[0071] Specifically, parameter L represents the total path length, which is calculated by taking into account the Euclidean distance based on the waypoint coordinates of the static terrain layer and considering altitude changes. Its weight α is the basic energy consumption weight and the next most important. The contribution of path length to total energy consumption is moderate. Too high or too low a weight will affect the optimization effect. After the weight is set, the impact of path length error on total energy consumption is small, which is the optimal value.

[0072] The parameter F represents the total wind energy consumption, which is calculated by multiplying the sum of the wind coefficients of all grids covered by the path by the energy consumption per unit distance of the UAV. Its weight β is the most influential and has the highest weight. Wind contributes the most to the total energy consumption and is the most critical factor. After the weight is set, the wind energy consumption prediction error is small and the total energy consumption optimization effect is the best.

[0073] Parameter C represents communication energy consumption, which is calculated based on the average distance between the UAV and the regional sub-nodes, combined with communication power and time. Its weight γ is a secondary energy consumption weight and is relatively low. The proportion of communication energy consumption to total energy consumption is moderate. If the weight is too high, path efficiency will be excessively sacrificed. After the weight is set, the balance between communication energy consumption and path efficiency is optimal.

[0074] Parameter B represents battery loss, which is calculated by multiplying the sum of the temperature coefficients of the low-temperature grids covered by the path by the dwell time, combined with the battery capacity characteristics; its weight δ is the weight with the least impact and is the lowest. Experiments show that the contribution of low temperature to the total energy consumption is low and can be avoided by adjusting the path. After the weight is set, the battery loss prediction error is small and the impact on the total energy consumption is controllable.

[0075] The regional sub-nodes select the locally optimal individual with the minimum total energy consumption from the candidate individuals.

[0076] S2.1.3, Global Optimal Path Aggregation

[0077] The core node receives the locally optimal individuals from each regional sub-node, sorts them by total energy consumption from smallest to largest, and assigns the optimal individual as the baseline path to each drone. In terms of allocation strategy, the individual with the lowest total energy consumption is assigned to the drone with the lowest remaining battery power, ensuring that drones with low battery power consume the least energy. Balance control is achieved through energy consumption difference checks, ensuring that the total energy consumption difference among all drones in the swarm is small, avoiding excessive energy consumption by individual drones. Regarding latency control, the core node's aggregation calculation is short, and the globally pre-planned total latency meets real-time requirements. In testing, the baseline path generated by the globally pre-planned path significantly reduced the total energy consumption of the swarm compared to random paths.

[0078] Step S2.2, Local Dynamic Adjustment

[0079] This step involves quickly correcting the local path when the perception layer detects sudden environmental changes, avoiding energy consumption fluctuations caused by global replanning. Specifically, it includes:

[0080] Step S2.2.1: Distributed splitting of the experience pool and training of sub-models

[0081] The core node collects historical flight data from drones, dividing it into multiple sub-pools based on scenario type, covering scenarios such as rockfall, strong winds, and low temperatures. During the sub-model training phase, each region's edge sub-nodes train specialized sub-models for their corresponding scenarios. The training framework is adapted to hardware acceleration, and the training parameters are optimized through hyperparameter search. Specifically, the training objective of the rockfall obstacle avoidance sub-model is "minimum obstacle avoidance radius + minimum energy consumption increment," with the reward function relating obstacle avoidance radius and energy consumption increment. The training objective of the strong wind altitude adjustment sub-model is "finding the lowest drag altitude layer," with the reward function relating to the wind energy consumption coefficient. The training objective of the low-temperature energy-saving sub-model is "shortening the dwell time in low-temperature areas," with the reward function relating to the dwell time in low-temperature areas. In terms of training performance, each sub-node has a moderate number of training rounds, short training time, and high model accuracy after training. The model parameters are then uploaded to the core node.

[0082] Step S2.2.2: Core-end model fusion and airborne inference

[0083] During model fusion, the core node merges the sub-models using a weighted average method, with weights set based on the probability of scenario occurrence. The fused model generates a global model. For lightweight adaptation on the airborne end, the global model is quantized to low precision and split into independent sub-models according to the scenario to adapt to the storage and computing power of the airborne node. This includes model pruning (removing redundant features, reducing model dimensionality, and improving inference speed) and inference optimization (using instruction set acceleration for shorter inference latency). In the scenario adaptation inference stage, the UAV's airborne node automatically loads the corresponding sub-model based on local perception data and outputs action commands. During action execution, the UAV adjusts its flight parameters according to the commands, collecting energy consumption data in real time during the adjustment process and feeding it back to the reward function for optimization. Regarding emergency mechanisms, in the event of communication interruption, the locally pre-stored sub-model emergency library is activated, resulting in a high obstacle avoidance success rate and ensuring basic safety. Local adjustment constraints only correct the path within a specific range after the current waypoint, resulting in a small deviation from the original path and low additional energy consumption. In testing, local adjustments significantly improved the success rate of obstacle avoidance against sudden obstacles with minimal increase in additional energy consumption.

[0084] Step S3: Distributed Conflict Management at the Collaboration Layer

[0085] Combined with appendix Figure 1 , 4 As shown, this step addresses the drone swarm path conflict problem by combining distributed consensus and regional retrieval, avoiding excessive latency in centralized retrieval. It consists of two parts:

[0086] Part 1: Blockchain Distributed Consensus. A two-tier blockchain is constructed with core consensus nodes and regional consensus nodes to ensure the trusted sharing and tamper-proof nature of path prediction data. The consensus mechanism has been optimized through security and efficiency testing.

[0087] First, each regional consensus node receives short-term future path prediction data from drones within the group and executes the Byzantine Fault Tolerance consensus algorithm: the consensus process involves five stages: request, preprocessing, preparation, commitment, and submission; the number of verification nodes must ensure that a majority of drone nodes verify the validity of the data signature to avoid the injection of false path data; when packaging blocks, the verified data is packaged, generated at a fixed frequency, and uploaded to the core consensus node via a high-speed backbone network; security verification shows that under strong electromagnetic interference scenarios, the blockchain has a high tamper-proof success rate and a low data leakage rate, meeting the security requirements of sensitive scenarios.

[0088] Then, the core consensus node receives blocks from each region and performs lightweight verification, eliminating the need to repeatedly verify single-node signatures and only verifying block integrity and hash consistency. Integrity verification includes checking the number of block data and timestamp continuity; hash verification calculates the Merkle root of the block header and compares it with the hash value of the block body data; subsequently, a global blockchain ledger is generated, storing the verified regional blocks in timestamp order to support subsequent traceability and auditing; in terms of performance metrics, block generation time is short, communication bandwidth consumption is low, consensus success rate is high, and blockchain consensus significantly improves the credibility of path data and reduces the rate of fake data injection.

[0089] Part Two: Distributed Retrieval of Time-Space Conflict Matrix

[0090] First, the conflict matrix is ​​distributed and partitioned. The core node splits the global "time-space matrix" into regional sub-matrices based on the coverage of regional sub-nodes. The matrix parameters are optimized through conflict scenario testing. The time slice partitioning is consistent with the update frequency of the path prediction data to ensure time synchronization. The spatial grid partitioning is set based on the drone's fuselage size and obstacle avoidance safety distance to avoid missed or false collision detections. Sub-matrice storage is the responsibility of each regional consensus node, which stores its own sub-matrices and supports fast retrieval.

[0091] Then, local conflict detection and resolution are performed at the regional end. In the conflict detection stage, the drone node maps its own short-term path prediction data to the corresponding cell of its regional sub-matrix and sends a grid occupancy request to the regional consensus node. The regional node retrieves the occupancy status of the cell. If it is not occupied, it is marked as "occupied" and feedback indicates that the drone path is feasible. If it is occupied, it is determined to be a conflict and feedback indicates that the drone path is conflicted and needs to be adjusted. During conflict resolution, regional consensus nodes locally calculate the priority score of the conflicting drones. The scoring formula is designed based on energy consumption and safety requirements and has been verified through multiple rounds of conflict scenarios. The formula is: Score = 0.6×(1 - SOC / 100) + 0.4×(2 - F), where SOC is the drone's remaining battery power (%) and F is the wind energy consumption coefficient of the current path. The higher the score, the higher the path priority (prioritizing the retention of low-battery, low-energy-consumption paths). The adjustment strategy is to retain the paths of high-priority drones and issue adjustment instructions to low-priority drones. The adjustment constraint is that the deviation between the adjusted path and the original path is small, and the additional energy consumption is minimal. In the conflict verification stage, after adjustment, the drone remaps its path to the regional submatrix, and the regional nodes search again to ensure that no new conflicts are generated. The total latency of conflict detection and resolution within the measured area is short, and the conflict resolution success rate is high.

[0092] Finally, cross-regional conflict coordination involves the consensus nodes in the affected regions reporting conflict information, including the conflicting drone ID, conflict grid, and priority score, to the core node when a conflict involves drones in different regions. During coordination, the core node determines which drone's path to retain based on the global priority score and issues adjustment instructions to the regional sub-nodes of the other drone. Cross-regional conflict coordination significantly reduces the collision rate of drone swarms.

[0093] Step S4: Distributed Battery Management and Interconnection at the Energy Layer

[0094] Combined with appendix Figure 1 , 5 As shown, this step is based on the differences in battery status of individual drones, and achieves the linkage optimization of energy and path through local calculation, regional control, and global balancing. The parameters of each link have been tested by lithium battery performance and flight verification.

[0095] Specifically:

[0096] Step S4.1: Calculation of distributed battery energy consumption model

[0097] Each drone's onboard node calculates the actual usable battery capacity based on real-time data collected by the local battery management system (BMS), using the following formula:

[0098] Ca = Cn × [1 − k1 × (0 − T) − k2 × (P / Pmax )], the formula parameters are based on fitting from multiple charge-discharge tests, resulting in small capacity prediction error. Parameter definitions and test criteria are as follows: Cn is the nominal capacity of the lithium battery; T is the battery temperature, collected by the temperature sensor built into the BMS; tests show that capacity decreases as temperature decreases; P is the current load power, calculated by the BMS, including motor, sensor, and communication power; tests show that capacity decreases as load power increases; Pmax is the maximum load power of the drone.

[0099] The system uploads three key parameters—actual capacity, remaining power (SOC), and temperature—to the consensus node in its region via ultraviolet light communication at fixed intervals. The data volume is small, avoiding communication redundancy, and the energy consumption for data upload accounts for a very low proportion of the total energy consumption, which can be ignored.

[0100] Step S4.2, Regional SOC Hierarchical Control

[0101] Regional consensus nodes aggregate the actual capacity and SOC data of drones within the group, and divide them into three levels of control strategies according to SOC. The strategy parameters have been optimized through long-term flight testing to ensure a balance between energy consumption and mission efficiency under different battery states.

[0102] The first level is a high-battery state with SOC ≥ 60%. In terms of task allocation, it performs long-distance tasks and is allowed to pass through a small number of high-energy-consumption areas. The sensor configuration adopts high-frequency sampling and enables all necessary sensors. The load control adopts high load power and enables high-definition acquisition function. The hardware parameters are configured as follows: high motor speed, communication power adaptive according to distance, and the power management module works in high-performance mode. In the test, the task completion rate is high in high-battery mode, the energy consumption is higher than that in medium-battery mode, but the task efficiency is improved.

[0103] The second level is a medium-power state with 30% ≤ SOC < 60%. In terms of task allocation, the plane flies along the baseline path, avoiding high-energy-consumption areas and prioritizing low-energy-consumption areas. The sensor configuration uses medium-frequency sampling and disables some unnecessary acquisition functions. The load control uses medium-load power and only enables the regular acquisition function. The hardware parameters are configured as medium motor speed, low communication power priority, and the power management module operates in balanced mode. In the test, the energy consumption in the medium-power mode is lower than that in the high-power mode, the task completion rate is high, and the coverage of core tasks (such as transmission tower detection) is 100%.

[0104] The third level is the low battery state with SOC < 30%. In terms of task allocation, the current task is immediately terminated, and the energy-saving return mode is activated. Based on the dynamic interference layer data, a low-drag return path with tailwind is planned, prioritizing areas with tailwind > 2m / s and specific altitudes (suitable drag coefficient). The sensor configuration adopts low-frequency sampling, and unnecessary sensors are turned off. The load control adopts low load power, motor power is reduced by 15%, flight speed is reduced by 20%, all data acquisition functions are turned off, and only navigation and communication are retained. The hardware parameters are configured with low motor speed, low communication power, and the power management module working in energy-saving mode. In the test, the energy consumption of the return mode in low battery mode is lower than that of the conventional return mode, the safe landing rate is 100%, and there are no cases of loss of contact or crash.

[0105] Step S4.3, Global Energy Balance Scheduling

[0106] The core node receives the energy status of each region through the high-speed backbone network, including the regional average SOC, the number of drones with low power (SOC<30%), the number of drones with high power (SOC≥60%), and the regional energy consumption coefficient (average F value, B value). Based on this data, global energy balance scheduling is performed.

[0107] First, a regional energy consumption assessment is conducted to calculate the energy pressure index for each region. This index is related to the average SOC, the proportion of low-energy consumption, and the average energy consumption coefficient. The higher the index, the greater the energy pressure. Then, a scheduling strategy is implemented. Energy-saving instructions are issued to high-energy-pressure areas, including prioritizing low-energy consumption paths, reducing the workload, and having high-energy drones assist low-energy drones. Task supplement instructions are issued to low-energy-pressure areas, including increasing the workload and taking over some tasks from high-pressure areas, thereby improving the overall energy balance.

[0108] Step S5: Two-way feedback optimization

[0109] Combined with appendix Figure 1 , 6 As shown, this step is crucial to ensuring continuous optimization of the technology at each layer. Through bidirectional feedback from planning to sensing, and then to sensing and energy planning, the deviation of model parameters is corrected. The feedback cycle is set through stability testing to avoid system oscillation caused by excessively frequent feedback.

[0110] Regarding feedback from the planning layer to the perception layer, during path execution, the planning layer compares the "energy consumption coefficient predicted by the dynamic interference layer" with the "energy consumption coefficient of actual flight" in real time. When the deviation exceeds the threshold, the deviation data (including timestamp, grid coordinates, predicted value, and actual value) is fed back to the perception layer. The perception layer adjusts the energy consumption coefficient calculation parameters of the dynamic interference layer. The effect verification shows that after the adjustment, the subsequent prediction error is reduced to a low level.

[0111] At the feedback level from the energy layer to the planning layer, the energy layer monitors the SOC decay rate of the drone in real time. When the actual decay rate is faster than the expected decay rate (based on battery model prediction), the decay data (drone ID, flight time, temperature, and load power) is fed back to the planning layer. The planning layer adjusts the weight of the battery loss parameter in the objective function and prioritizes planning low-power paths for it. The effect verification shows that after the adjustment, the SOC decay rate of the drone tends to be as expected.

[0112] The trajectory planning is performed once every fixed number of steps, and after multiple rounds of feedback, the total energy consumption of the swarm is further reduced, the success rate of path planning is improved, and the adaptability of the model parameters is significantly enhanced. In the test, after closed-loop optimization, the drone swarm can still maintain stable operation under sudden environmental conditions (such as sudden increase in wind speed and sudden drop in temperature) with small energy consumption fluctuations.

[0113] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A low-energy trajectory planning method for unmanned aerial vehicle (UAV) swarms based on distributed computing, characterized in that, This includes constructing a three-level distributed computing node network, which includes a single core edge node, regional edge nodes configured according to the scale of the drone swarm, and drone-borne nodes consistent with the scale of the drone swarm. The core edge nodes, regional edge nodes, and UAV-borne nodes are interconnected. The task area is evenly divided into sub-regions, with each sub-region corresponding to one regional edge node. The boundaries of the sub-regions coincide with terrain obstacles. Computing power is allocated as follows: core edge nodes handle global aggregation, regional edge nodes handle local computation, and UAV-borne nodes handle preprocessing. The specific method is as follows: Step S1: Distributed data processing and environmental modeling to generate an environment-energy consumption coefficient mapping model; Step S2: Planning layer trajectory optimization, including global pre-planning and local dynamic adjustment; Step S3: Collaboration layer conflict management, resolving path conflicts through blockchain distributed consensus and distributed retrieval of time-space conflict matrix; Step S4: Energy layer battery management and linkage, realizing energy and path linkage optimization based on the differences in UAV battery status; Step S5: Two-way feedback optimization, correcting model parameters through two-way feedback from the planning layer to the perception layer and from the energy layer to the planning layer.

2. The method according to claim 1, characterized in that, Step S1 includes step S1.1 distributed fusion of multimodal sensor data and step S1.2 distributed construction of a two-layer environment model; Step S1.1 includes: The UAV-borne node collects lidar point cloud data, high-definition visual camera images, inertial measurement unit attitude data, and micro weather station environmental parameters, performs outlier detection using an improved adaptive Kalman filter, and uploads the compressed data to the regional edge node. After decompressing the data at the edge nodes of the region, a sliding window noise estimation is performed, and the three-dimensional coordinates of obstacles, average wind speed, and ambient temperature within the region are output. The core edge nodes receive the regional fusion results, convert them into a globally universal coordinate system and remove duplicates, and generate a global perception dataset. Step S1.2 includes: Static terrain layer: The core edge nodes will split and distribute the pre-loaded high-resolution digital elevation model by region, and fit the air drag coefficient at different altitudes through wind tunnel experiments; Dynamic interference layer: The core edge nodes divide the task airspace into cubic grids. Each region's edge nodes update the parameters of their respective sub-matrixes, including the moving obstacle grid, wind power consumption coefficient grid, and battery temperature decay coefficient grid. The update priority is moving obstacle grid > wind power consumption coefficient grid > battery temperature decay coefficient grid.

3. The method according to claim 1, characterized in that, The global pre-planning in step S2 includes: Step S2.1.1: The core edge nodes generate multiple candidate path individuals with waypoints, and distribute them evenly according to the number of edge nodes in the region; Step S2.1.2: The regional edge node calls the dynamic interference layer energy consumption coefficient, combines it with the UAV parameters, and calculates the total energy consumption of candidate individuals according to the objective function E=αL+βF+γC+δB, and selects the locally optimal individual; where L is the total path length, α is the basic energy consumption weight; F is the total wind energy consumption, β is the highest weight; C is the communication energy consumption, γ is the secondary energy consumption weight; B is the battery loss, and δ is the lowest weight; Step S2.1.3: The core edge node receives the locally optimal individual, sorts them by total energy consumption, and assigns the individual with the lowest total energy consumption to the drone with the lowest remaining power, thereby controlling the difference in total energy consumption among drones in the group.

4. The method according to claim 1, characterized in that, The local dynamic adjustment in step S2 includes: Step S2.2.1: The core edge nodes split the historical flight data into sub-pools according to the scenario, and the regional edge nodes train the corresponding scenario-specific sub-models, including the rockfall obstacle avoidance sub-model, the strong wind height adjustment sub-model, and the low temperature energy saving sub-model. Step S2.2.2: The core edge nodes fuse the sub-models by weighted average to generate a global model; the UAV onboard nodes quantize the global model into low precision and split it according to the scenario, load the corresponding sub-models according to the local perception data and output action commands; when communication is interrupted, the local pre-stored sub-model emergency library is activated; only the path within a specific range after the current waypoint is corrected to control the additional energy consumption increment.

5. The method according to claim 1, characterized in that, The blockchain distributed consensus in step S3 includes: regional consensus nodes receive the short-term future path prediction data of drones, execute the Byzantine fault-tolerant consensus algorithm, and after verification through the request, preprocessing, preparation, commitment, and submission stages, package and generate blocks and upload them to the core consensus nodes; the core consensus nodes verify the integrity and hash consistency of the blocks and generate a global blockchain ledger stored by timestamp.

6. The method according to claim 1, characterized in that, The distributed retrieval of the time-space conflict matrix in step S3 includes: the core edge node splits the global time-space matrix by region, and the regional consensus node stores the sub-matrix; the UAV node sends a grid occupancy request to the regional consensus node, the regional consensus node retrieves the occupancy status, and after determining the conflict, calculates the priority according to the scoring formula Score=0.6×(1 - SOC / 100)+0.4×(2 - F), and retains the high-priority UAV path; In the event of a cross-regional conflict, the consensus nodes in the affected regions report to the core edge nodes, which then coordinate based on global priorities.

7. The method according to claim 1, characterized in that, Step S4 includes: Step S4.1: The UAV onboard node calculates the actual usable battery capacity using the formula Ca=Cn×[1−k1×(0−T)−k2×(P / Pmax)], where Ca is the actual usable capacity, Cn is the nominal capacity, T is the battery temperature, P is the current load power, Pmax is the maximum load power, and k1 and k2 are fitting coefficients; and uploads the actual capacity, remaining power (SOC), and temperature to the regional consensus node periodically. Step S4.2: Regional consensus nodes are controlled in layers according to SOC: SOC ≥ 60% executes remote tasks and high-load configuration; 30% ≤ SOC < 60% flies along the baseline path and is configured for medium load; SOC < 30% initiates energy-saving return and low-load configuration. Step S4.3: The core edge node calculates the regional energy pressure index, issues energy-saving instructions to high-pressure areas, and issues task supplement instructions to low-pressure areas.

8. The method according to claim 1, characterized in that, Step S5 includes: The planning layer compares the "predicted energy consumption coefficient of the dynamic interference layer" with the "actual flight energy consumption coefficient". When the deviation exceeds the threshold, it feeds back to the perception layer to adjust the calculation parameters of the dynamic interference layer. The energy layer monitors the actual SOC decay rate. When it exceeds the expectation, it feeds back to the planning layer to adjust the weight of the battery loss parameter in the objective function. Feedback optimization is performed once every fixed number of steps to reduce the total energy consumption of the group and improve the success rate of path planning.

9. The method according to claim 1, characterized in that, The high-speed backbone network is based on the latest communication standards, and its latency remains stable at a low level even in complex weather and obstructed mountainous environments. The "sun-blind" ultraviolet light communication conforms to general communication protocols and features low power consumption, medium communication distance, low bit error rate, and high resistance to electromagnetic interference. The power consumption of the UAV onboard node is a low percentage of the total power consumption of the UAV when it is fully loaded.

10. The method according to claim 1, characterized in that, The core edge nodes are configured with dual power backup, and the regional edge nodes are equipped with power replenishment modules; all nodes in the three-level distributed computing node network are in a safe operating range when fully loaded, making it suitable for scenarios such as power inspection in mountainous areas, forest fire monitoring, and urban emergency rescue.

Images