Resource optimization management method based on reconfigurable meter

Abstract

The invention discloses a resource optimization management method based on a reconfigurable meter, and the method remarkably improves the calculation efficiency, cruising ability and cooperation safety of an unmanned aerial vehicle cluster through the dynamic task distribution of a heterogeneous multi-core architecture, intelligent scheduling driven by reinforcement learning, a low-power-consumption model compression technology and an ultra-low delay communication technology. Efficient computing power distribution, energy consumption control and real-time cooperation of the unmanned aerial vehicle cluster are achieved, and the cruising ability, safety and task execution efficiency of the unmanned aerial vehicle cluster in a complex task scene are improved. The method is suitable for scenes of security patrol, disaster search and rescue, agricultural monitoring and the like, solves the problems of rigid resource allocation, high energy consumption and large communication delay in the prior art, and has wide industrial application prospects.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) clusters, and in particular, to a resource optimization management method based on reconfigurable computing. Background Art

[0002] With the increasing requirements for the autonomy, intelligence, multi-tasking, etc. of UAV operations, UAVs have evolved from single-vehicle operations to cluster operations, posing higher requirements for multi-vehicle cluster communication technology. By adopting a collaborative method of multiple UAVs, information sharing is achieved through communication between UAVs, expanding the perception of the environmental situation, and enabling collaborative task allocation, collaborative search, collaborative reconnaissance and attack, which can effectively improve the survival ability and overall combat effectiveness of the UAV cluster. The UAV cluster can autonomously drive intelligent machines to complete tasks that cannot be accomplished by a single UAV without human intervention.

[0003] Currently, task allocation in UAV clusters is a key technology for UAV clusters to achieve efficient combat missions. With the development of UAV cluster technology and the transformation of combat styles, the combat mission areas of UAV clusters are constantly expanding, the scope covered by task allocation is constantly expanding, and the scale and complexity of task allocation problems are constantly increasing, which pose new challenges to UAV cluster task allocation technology.

[0004] Existing UAV clusters face the following problems in complex task scenarios:

[0005] 1. Rigid allocation of computing resources: Traditional fixed computing architectures cannot adapt to dynamic task requirements, resulting in wasted computing power or insufficient performance.

[0006] 2. Insufficient endurance due to high power consumption: Intensive computing tasks lead to a sharp increase in energy consumption and limited endurance.

[0007] 3. High communication latency in multi-vehicle collaboration: Multi-vehicle collaboration relies on high-bandwidth and low-latency communication, and existing technologies are difficult to meet the millisecond-level interaction requirements, affecting task real-time performance.

[0008] 4. Single traditional scheduling algorithm: Static scheduling strategies cannot take into account multi-objective optimization (such as energy consumption, latency, accuracy).

[0009] Therefore, the present invention proposes a resource optimization management method based on reconfigurable computing to improve the computing efficiency, endurance, and collaborative operation performance of UAV clusters. Summary of the Invention

[0010] The objective of the present invention is to address the drawbacks existing in the prior art as presented in the background art, and to propose a resource optimization management method based on reconfigurable computing. The aim is to solve the problems of computing power allocation, energy consumption control, and collaborative efficiency of UAV swarms in high-concurrency task scenarios through heterogeneous multi-core architectures, dynamic task scheduling, low-power model optimization, and ultra-low latency communication technologies, so as to improve the computing efficiency, endurance, and collaborative operation performance of UAV swarms.

[0011] To achieve the above objective, the present invention adopts the following technical solutions:

[0012] A resource optimization management method based on reconfigurable computing includes the following steps:

[0013] S1. Construct a heterogeneous multi-core computing architecture, including a flexibly combined general-purpose computing (CPU) and dedicated computing (NPU unit / DSP unit / FPGA unit / ASIC unit);

[0014] S2. According to the real-time task load of the UAV swarm, dynamically allocate tasks to the general-purpose computing unit and the dedicated computing unit through a hardware task segmentation strategy;

[0015] S3. Adopt an intelligent task scheduling unit based on reinforcement learning to monitor the utilization rate of computing resources in real time and dynamically adjust the voltage frequency and computing power allocation of each computing unit;

[0016] S4. Optimize the deep learning model in combination with a multi-dimensional model compression algorithm to enhance the autonomous perception and decision-making ability of the UAV swarm;

[0017] S5. Deploy an adaptive dynamic power management mechanism to balance the computing power demand and energy consumption through dynamic voltage and frequency scaling (DVFS) and an AI-predicted power consumption model;

[0018] S6. Implement real-time data interaction of the UAV swarm through low-latency wireless communication technologies using adaptive beamforming and MIMO multi-antenna technologies, including UWB, millimeter-wave communication, and 5G / 6G URLLC protocols, reducing the communication packet loss rate by 30% and compressing the end-to-end delay to less than 5 ms.

[0019] As a further step in the method of the present invention, in S1 and S2, a hardware task segmentation strategy is adopted to flexibly combine general-purpose computing and dedicated computing, which respectively undertake functions such as neural network inference, signal processing, custom acceleration, and execution of customized algorithm cores;

[0020] According to the real-time load characteristics and operation requirements of the UAV (such as image recognition, collaborative navigation), different computing units can be dynamically selected and allocated to achieve the optimal energy efficiency ratio.

[0021] As a further step in the method of the present invention, between S1 and S2, the heterogeneous multi-core computing architecture reduces data transfer latency and bandwidth bottlenecks through on-chip storage and 3D stacked storage technologies, thereby improving the efficiency of multi-sensor data processing during deep learning inference or multi-sensor data processing, achieving an overall computational efficiency improvement of approximately 20%.

[0022] As a further step in the method of the present invention, in S3-1, the intelligent task scheduling unit based on reinforcement learning adopts PPO or DDPG algorithms to monitor the utilization rate of multi-core resources inside the chip, and dynamically allocate the voltage frequency and computing power of computing units such as NPU, DSP, and FPGA according to task priorities and environmental feedback.

[0023] As a further step in the method of the present invention, in S4-1, the multi-dimensional model compression algorithm is based on MDCA, and performs multiple compressions and optimizations on deep learning models from the parameter dimension, structure dimension, and operation dimension, and is applicable to mainstream network structures such as CNN, YOLO, and ResNet.

[0024] As a further step in the method of the present invention, in S4-2, the multi-dimensional model compression algorithm includes depthwise separable convolution and cross-layer fusion technologies. The depthwise separable convolution first performs depthwise convolution and then realizes information fusion between channels through 1×1 convolution, thereby significantly compressing the convolution kernel parameters and computational volume, reducing the computational amount by 70%-90%, and reducing the weights by 50%-80%;

[0025] On this basis, the cross-layer fusion (Cross-LayerFusion) technology can further merge operators such as convolution, BN, and activation functions into larger operators, reducing the repeated access to intermediate feature maps;

[0026] It is found through deployment and testing on mobile or embedded platforms that the combination of depthwise separable convolution and cross-layer fusion can increase the average inference speed by more than 1.5 times, reduce the model size by approximately 35%, and reduce the overall energy consumption by approximately 40%.

[0027] As a further step in the method of the present invention, in S5-1, the adaptive dynamic power management mechanism predicts the computing power requirements in future time periods through a recurrent neural network (RNN) or long short-term memory network (LSTM), and allocates power resources in advance to ensure the priority execution of obstacle avoidance and path planning tasks, with a prediction accuracy of not less than 95%;

[0028] S5-2, the adaptive dynamic power management includes DVFS (Dynamic Voltage and Frequency Scaling) and an AI prediction power consumption model. DVFS automatically adjusts the working voltage and frequency of each computing unit in the chip according to flight tasks and load conditions;

[0029] The AI prediction power consumption model uses deep learning (such as RNN, LSTM, etc.) to predict the computing power demand and energy status in the next time period, and adjusts the power distribution in advance to ensure the priority of computing power in critical stages (such as safety obstacle avoidance or emergency tasks).

[0030] As a further step in the method of the present invention, S5-1, UWB (Ultra Wide Band) + millimeter wave (mmWave): When the UAV cluster performs short-range high-bandwidth communication, UWB can be used to compress the delay of a single data exchange to less than 5 ms;

[0031] S5-2, 5G / 6G URLLC: In scenarios such as formation flight, emergency obstacle avoidance, or real-time collaboration that are extremely sensitive to latency, URLLC can reduce the end-to-end delay to the millisecond level, greatly enhancing the synchronization and safety of the UAV swarm.

[0032] As a further step in the method of the present invention, S7, it also includes autonomous path planning based on MARL (Multi-Agent Reinforcement Learning). The autonomous path planning based on MARL (Multi-Agent Reinforcement Learning) includes a hybrid strategy of RL + genetic algorithm (GA) and a hybrid search of A* and Dijkstra;

[0033] S7-1, RL + genetic algorithm (GA) hybrid strategy, by integrating multi-agent reinforcement learning (MARL) and genetic algorithm (GA), realizes dynamic optimization and autonomous decision-making of global and local paths in an environment with multiple objectives (such as energy consumption, shortest flight range, obstacle avoidance, etc.) and multiple constraints (such as no-fly zones, communication ranges, flight altitude limits, etc.);

[0034] S7-2, A* and Dijkstra hybrid search, uses the A* algorithm for rapid obstacle avoidance in local path search, and obtains a high-precision feasible solution through the Dijkstra algorithm for global path design, and combines GA for iterative update, taking into account both the global optimal solution and local real-time performance.

[0035] As a further step in the method of the present invention, S8, it also includes a lightweight security encryption mechanism, and uses elliptic curve encryption (ECC), hash chain technology, and AI anomaly detection algorithms;

[0036] S8-1, elliptic curve encryption (ECC) + hash chain (HashChain), can provide high-strength encryption security with a small key volume, and further prevent data tampering and replay in combination with the hash chain, which is suitable for the UAV cluster environment with limited computing resources;

[0037] When the CPU occupancy rate does not exceed 5%, it can maintain a data integrity check accuracy rate of more than 98%, and effectively resist man-in-the-middle attacks and data replay attacks, providing a solid security guarantee for cluster operations;

[0038] S8-2. The AI anomaly detection algorithm uses autoencoders (AE) or graph neural networks (GNN) to detect anomalies and identify attacks in real-time communication data, accurately screening for suspicious data packets or malicious instructions.

[0039] Compared with the prior art, the present invention integrates multiple computing units and simultaneously provides efficient communication, energy management, and security modules, providing strong technical support for the autonomous flight, path planning, task allocation, and information interaction of UAV swarms in complex environments. Combining the latest wireless communication methods such as 5G / 6G, UWB, and LoRa, LPCC can effectively meet the large bandwidth and low latency requirements for multi-aircraft collaboration, laying a solid foundation for the large-scale application of UAV swarms in multiple industrial scenarios.

[0040] The main core components include:

[0041] ① Heterogeneous multi-core architecture collaboration: Through the flexible combination of CPU, NPU, DSP, FPGA, and ASIC, it supports neural network inference, signal processing, and customized algorithm acceleration.

[0042] ② Dynamic task scheduling: The intelligent scheduling unit based on reinforcement learning optimizes resource allocation in real-time, improving the energy efficiency ratio.

[0043] ③ Low-power AI model optimization: Adopting multi-dimensional compression algorithms and depthwise separable convolution techniques to reduce the model's computational complexity and energy consumption.

[0044] ④ Adaptive power management: Combining DVFS and AI prediction models to dynamically adjust computing power and power consumption, extending the battery life.

[0045] ⑤ Ultra-low latency communication: Integrating UWB, millimeter wave, and 5G / 6G URLLC technologies to ensure the real-time performance of multi-aircraft collaboration.

[0046] ⑥ Lightweight security mechanism: Through ECC encryption and AI anomaly detection, ensure data security and system stability.

[0047] The beneficial effects are as follows:

[0048] The heterogeneous multi-core architecture increases the system's computing throughput by 25%, shortens the task response time by 20%, and reduces energy consumption by 30%.

[0049] The model compression technology reduces the inference power consumption by 40%, the dynamic power management extends the battery life by 20%, and the model inference efficiency is improved by 30%.

[0050] The end-to-end communication delay is compressed to less than 5 ms, and the packet loss rate is reduced by 30%.

[0051] The lightweight encryption mechanism achieves data integrity of over 98% and resists common network attacks.

[0052] The AI anomaly detection algorithm uses autoencoders or graph neural networks to perform anomaly detection and attack recognition on real-time communication data, accurately screening for suspicious data packets or malicious instructions. Specific implementation manners

[0053] The content of the present invention can be more easily understood by referring to the following detailed description of the preferred implementation methods of the present invention and the included embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. In case of conflict, the definitions in this specification shall prevail.

[0054] When an equivalent, concentration, or other value or parameter is expressed as a range, a preferred range, or a range defined by a series of upper preferred values and lower preferred values, it should be understood that all ranges formed by any pairing of any range upper limit or preferred value with any range lower limit or preferred value are specifically disclosed, regardless of whether the ranges are separately disclosed. For example, when the range "1 to 5" is disclosed, the described range should be interpreted as including the ranges "1 to 4", "1 to 3", "1 to 2", "1 to 2 and 4 to 5", "1 to 3 and 5", etc. When a numerical range is described herein, unless otherwise stated, the range is intended to include its end values and all integers and fractions within the range.

[0055] The singular form includes plural discussion objects, unless otherwise clearly specified in the context. "Optional" or "any one" means that the matter or event described thereafter may or may not occur, and this description includes the case where the event occurs and the case where the event does not occur.

[0056] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments.

[0057] A resource optimization management method based on reconfigurable computing includes the following steps:

[0058] S1. Construct a heterogeneous multi-core computing architecture, including a flexibly combined general-purpose computing (CPU) and dedicated computing (NPU unit / DSP unit / FPGA unit / ASIC unit);

[0059] S2. According to the real-time task load of the drone cluster, dynamically allocate tasks of the general-purpose computing unit and the dedicated computing unit through a hardware task segmentation strategy;

[0060] S3. Adopt an intelligent task scheduling unit based on reinforcement learning to monitor the computing resource utilization rate in real time and dynamically adjust the voltage frequency and computing power allocation of each computing unit;

[0061] S4. Optimize deep learning models by combining multi-dimensional model compression algorithms to enhance the autonomous perception and decision-making capabilities of drone swarms.

[0062] S5. Deploy an adaptive dynamic power management mechanism to balance computing power requirements and energy consumption through dynamic voltage and frequency scaling (DVFS) and AI-based predictive power consumption models.

[0063] S6. Real-time data interaction among drone clusters is achieved through low-latency wireless communication technology using adaptive beamforming and MIMO multi-antenna technology, including UWB, millimeter wave communication and 5G / 6G URLLC protocol. The communication packet loss rate is reduced by 30%, and the end-to-end delay is compressed to less than 5ms.

[0064] As a further step in the method of the present invention, S1 and S2 adopt a hardware task partitioning strategy to flexibly combine general computing and special computing, respectively undertaking functions such as neural network reasoning, signal processing, custom acceleration, and customized algorithm core execution;

[0065] Based on the real-time load characteristics and operational requirements of the UAV (such as image recognition and collaborative navigation), different computing units can be dynamically selected and allocated to achieve the optimal energy efficiency ratio;

[0066] This paragraph states that in multi-task parallel scenarios such as image recognition + path planning, this heterogeneous multi-core architecture can increase the overall system processing throughput by about 25% and reduce energy consumption by about 30%, and can still maintain efficient and stable performance in high-concurrency computing scenarios with multi-machine collaboration.

[0067] In the method of the present invention, S1 and S2, the heterogeneous multi-core computing architecture, uses on-chip storage and 3D stacked storage technology to reduce data transfer delays and bandwidth bottlenecks, thereby improving the efficiency of multi-sensor data processing during deep learning reasoning or multi-sensor data processing, and achieving an overall computing efficiency improvement of approximately 20%;

[0068] This storage architecture can significantly reduce the frequency of off-chip memory accesses and reduce the occupancy of the external bus, providing faster and more stable computing power support for drones to perform high-intensity perception and decision-making tasks in multi-machine collaborative scenarios.

[0069] In the method of the present invention, in S3-1, the intelligent task scheduling unit based on reinforcement learning adopts the PPO or DDPG algorithm to monitor the utilization rate of multi-core resources inside the chip, and dynamically allocate the voltage frequency and computing power of computing units such as NPU, DSP, and FPGA according to task priorities and environmental feedback. Different algorithm cores can be flexibly switched according to the flight state of the drone, environmental complexity, and communication conditions. Practical application tests show that this scheduling unit can increase the task execution efficiency by about 25%, and still maintain good adaptability under the constraints of multiple objectives (energy consumption, latency, accuracy). Compared with traditional static allocation, the average response time is shortened by about 20%, and the energy efficiency ratio (TOPS / W) is increased by about 15%, which is particularly significant when the drone cluster executes rapid movement or needs to suddenly respond to multiple task instructions.

[0070] In the method of the present invention, in S4-1, the multi-dimensional model compression algorithm is based on MDCA, and performs multiple compressions and optimizations on the deep learning model from the parameter dimension, structure dimension, and operation dimension, and is applicable to mainstream network structures such as CNN, YOLO, and ResNet;

[0071] Through means such as model pruning, quantization, and knowledge distillation, the inference efficiency can be increased by more than 30% while maintaining the accuracy, and the chip inference power consumption can be reduced by about 40%, providing assistance for long-term endurance and real-time decision-making in the drone formation scenario;

[0072] Tests on hardware platforms with limited resources such as NPU and FPGA show that the model optimized by MDCA can significantly shorten the inference latency and reduce the overall energy consumption under the same computing resources, and is particularly suitable for high-real-time tasks such as continuous target detection and environmental perception required by the drone cluster.

[0073] In the method of the present invention, in S4-2, the multi-dimensional model compression algorithm includes depthwise separable convolution and cross-layer fusion technology. Traditional convolution has a large number of multiply-accumulate operations in both the spatial and channel dimensions, while depthwise separable convolution first performs depthwise convolution (DepthwiseConv), and then uses 1×1 convolution (PointwiseConv) to achieve information fusion between channels, thereby significantly compressing the convolution kernel parameters and the amount of computation. The amount of computation is reduced by 70%-90%, and the weights are reduced by 50%-80%. On this basis, the cross-layer fusion (Cross-LayerFusion) technology can further combine operators such as convolution, BN, and activation functions into larger operators, reducing the repeated access to intermediate feature maps;

[0074] After deployment and testing on mobile or embedded platforms, it is found that the combination of depthwise separable convolution and cross-layer fusion can increase the average inference speed by more than 1.5 times, reduce the model size by about 35%, and reduce the overall energy consumption by about 40%. In the lossless or slightly lossy accuracy mode, it can effectively meet the collaborative operation scenarios with extremely high real-time requirements such as path planning and target detection for drone swarms.

[0075] In the method of the present invention, in S5-1, the adaptive dynamic power management mechanism predicts the computing power requirements in future time periods through a recurrent neural network (RNN) or a long short-term memory network (LSTM), and allocates power resources in advance to ensure the priority execution of obstacle avoidance and path planning tasks, and the prediction accuracy is not less than 95%.

[0076] In S5-2, the adaptive dynamic power management includes DVFS (Dynamic Voltage and Frequency Scaling) and an AI prediction power consumption model. DVFS automatically adjusts the operating voltage and frequency of each computing unit in the chip according to the flight mission and load conditions, briefly increases the computing power in high-load scenarios (such as large target detection or path search) to meet real-time requirements, and reduces power consumption in the idle or low-load stage to balance performance and endurance.

[0077] This mechanism can reduce the overall power consumption by about 30%, and increase the endurance time of the drone by about 20% while ensuring the timeliness of the task, which is crucial for drone swarms performing cruise monitoring or continuous operations.

[0078] The AI prediction power consumption model uses deep learning (RNN, LSTM, etc.) to predict the computing power requirements and energy status in the next time period, and adjusts the power distribution in advance to ensure the priority of computing power in critical stages (such as safe obstacle avoidance or emergency tasks).

[0079] The accuracy of the prediction model can reach 95%, which can give priority to flight control and obstacle avoidance tasks in emergency scenarios, thus significantly reducing the failure risk caused by insufficient energy consumption.

[0080] In the method of the present invention, in S5-1, UWB (Ultra-Wideband) + mmWave (millimeter wave): When the drone swarm communicates at close range with high bandwidth, UWB can be used to compress the delay of a single data exchange to less than 5 ms; in medium and long-distance, high-traffic (such as high-definition video stream) communication scenarios, millimeter wave can further improve data throughput and communication stability.

[0081] After adopting adaptive beamforming and MIMO multi-antenna technology, the overall communication efficiency is increased by about 25%, and the packet loss rate is reduced by about 30%, which is particularly crucial for multi-aircraft synchronization and information interaction of large drone formations in complex environments (such as urban canyons, mountains, etc.).

[0082] S5-2, 5G / 6G URLLC: In scenarios such as formation flight, emergency obstacle avoidance, or real-time collaboration that are extremely sensitive to latency, URLLC can reduce the end-to-end latency to the millisecond level, significantly enhancing the synchronization and safety of the UAV swarm;

[0083] For UAVs with high-speed maneuverability or those requiring precise formation, URLLC provides a highly reliable communication link, enabling timely issuance of control commands or receipt of sensor feedback at critical moments, avoiding delays or safety risks caused by network fluctuations.

[0084] In the method of the present invention, S7 also includes autonomous path planning based on MARL (Multi-Agent Reinforcement Learning) to collaboratively optimize the UAV swarm to improve task execution efficiency; the autonomous path planning based on MARL (Multi-Agent Reinforcement Learning) includes an RL + Genetic Algorithm (GA) hybrid strategy and a hybrid search of A* and Dijkstra;

[0085] S7-1, RL + Genetic Algorithm (GA) hybrid strategy, by integrating multi-agent reinforcement learning (MARL) and genetic algorithm (GA), realizes dynamic optimization and autonomous decision-making of global and local paths in multi-objective (energy consumption, shortest flight path, obstacle avoidance, etc.) and multi-constraint (no-fly zone, communication range, flight altitude limit, etc.) environments.

[0086] In high-dimensional search scenarios, the overall navigation efficiency is increased by about 20%, and the convergence speed is accelerated by about 30%. When facing sudden obstacles or regional changes, the replanning time is reduced by an average of 10%, ensuring the coordination and real-time performance of the UAV swarm in complex environments;

[0087] S7-2, A* and Dijkstra hybrid search, uses the A* algorithm for fast obstacle avoidance in local path search, and the Dijkstra algorithm to obtain a high-precision feasible solution for global path design, and combines GA for iterative update, taking into account both the global optimal solution and local real-time performance;

[0088] This method can reduce redundant or ineffective searches, has good adaptability and stability in maps with large scales or complex terrains, and significantly improves the efficiency when multi-UAV formations perform tasks such as patrol monitoring and disaster search and rescue.

[0089] In the method of the present invention, S8 also includes a lightweight security encryption mechanism, and uses elliptic curve encryption (ECC), hash chain technology, and AI anomaly detection algorithms;

[0090] S8-1, Elliptic Curve Encryption (ECC) + Hash Chain, can provide high-strength encryption security with a small key volume, and further prevents data from being tampered with and replayed in combination with the hash chain, which is suitable for the UAV swarm environment with limited computing resources;

[0091] When the CPU occupancy rate does not exceed 5%, it can maintain a data integrity verification accuracy rate of more than 98%, effectively resist man-in-the-middle attacks and data replay attacks, and provide a solid security guarantee for cluster jobs;

[0092] S8-2. The AI anomaly detection algorithm uses an autoencoder (Autoencoder) or a graph neural network (GNN) to perform anomaly detection and attack recognition on real-time communication data, and accurately screen suspicious data packets or malicious instructions;

[0093] It can quickly isolate potential risks in the early stage, prevent malicious behaviors from spreading in the UAV formation, and achieve a high level of security and robustness for the overall cooperation system.

[0094] Based on the above, the present invention proposes the following:

[0095] Embodiment 1, urban security patrol:

[0096] 1. Allocate image recognition (NPU) and path search (DSP) tasks through a heterogeneous multi-core architecture. The NPU processes real-time image recognition (such as the YOLO optimized model), and the DSP executes path planning;

[0097] 2. The intelligent scheduling unit dynamically adjusts the FPGA computing power according to the environmental complexity (for example, dynamically according to the pedestrian flow density), and the response time is shortened by 18%;

[0098] 3. Use the compressed YOLO model for real-time target detection, and the inference delay is reduced by 40%;

[0099] 4. Achieve millisecond-level interaction of formation instructions through UWB communication, and transmit high-definition monitoring video streams through millimeter waves;

[0100] 5. The DVFS mechanism reduces the CPU frequency during low-load periods, and the overall power consumption is reduced by 28%.

[0101] Embodiment 2, disaster search and rescue mission:

[0102] 1. Use a hybrid strategy of MARL and genetic algorithm to optimize the global path, and the global path convergence speed is increased by 30%;

[0103] 2. The AI prediction model pre-allocates the computing power required for obstacle avoidance tasks;

[0104] 3. Transmit search and rescue coordinates through ECC encryption, and use a hash chain to verify data integrity;

[0105] 4. The 3D stacked storage technology reduces data transfer latency, and the multi-sensor data processing efficiency is increased by 25%-30%;

[0106] 5. The AI algorithm based on an autoencoder screens communication data in real time to isolate malicious instructions;

[0107] 6. Elliptic curve encryption ensures the security of communication data.

[0108] The examples involved in this article are only illustrative and are used to explain some features of the method described in the present invention. The appended claims are intended to claim the broadest scope conceivable, and the embodiments presented herein are only illustrative of the selected embodiments according to the combinations of all possible embodiments. Therefore, the intention of the applicant is that the appended claims are not limited by the selection of examples illustrating the features of the present invention. Some of the numerical ranges used in the claims also include sub-ranges within them, and variations within these ranges should also be construed as being covered by the appended claims whenever possible.

Claims

1. A resource optimization management method based on reconfigurable computing, characterized in that S1. Construct a heterogeneous multi-core computing architecture, including a general-purpose computing (CPU) and a dedicated computing (NPU unit / DSP unit / FPGA unit / ASIC unit) that are flexibly combined; S2. According to the real-time task load of the UAV cluster, dynamically allocate tasks of the general-purpose computing unit and the dedicated computing unit through a hardware task segmentation strategy; S3. Adopt an intelligent task scheduling unit based on reinforcement learning to monitor the computing resource utilization rate in real time and dynamically adjust the voltage frequency and computing power allocation of each computing unit; S4. Optimize the deep learning model by combining a multi-dimensional model compression algorithm to improve the autonomous perception and decision-making ability of the UAV cluster; S5. Deploy an adaptive dynamic power management mechanism to balance the computing power demand and energy consumption through dynamic voltage and frequency regulation (DVFS) and an AI prediction power consumption model; S6. Implement real-time data interaction of the UAV cluster through low-latency wireless communication technologies using adaptive beamforming and MIMO multi-antenna technologies, including UWB, millimeter-wave communication, and 5G / 6G URLLC protocols, reducing the communication packet loss rate by 30% and compressing the end-to-end delay to less than 5 ms.

2. The resource optimization management method based on reconfigurable computing according to claim 1, wherein For S1 and S2, adopt a hardware task segmentation strategy to flexibly combine general-purpose computing and dedicated computing, and respectively undertake functions such as neural network inference, signal processing, custom acceleration, and execution of customized algorithm cores; According to the real-time load characteristics and operation requirements of the UAV, different computing units can be dynamically selected and allocated to achieve the optimal energy efficiency ratio.

3. The resource optimization management method based on reconfigurable computing according to claim 2, characterized in that, For S1 and S2, the heterogeneous multi-core computing architecture reduces data transfer latency and bandwidth bottlenecks through on-chip storage and 3D stacking storage technologies, thereby improving the efficiency of multi-sensor data processing during deep learning inference or multi-sensor data processing, and obtaining an overall computing efficiency improvement of about 20%.

4. The resource optimization management method based on reconfigurable computing according to claim 1, wherein S3-1. The intelligent task scheduling unit based on reinforcement learning adopts the PPO or DDPG algorithm to monitor the utilization rate of multi-core resources inside the chip, and dynamically allocate the voltage frequency and computing power of computing units such as NPU, DSP, and FPGA according to task priorities and environmental feedback.

5. A resource optimization management method based on a reconfigurable meter according to claim 1, characterized in that, S4-1. The multi-dimensional model compression algorithm is based on MDCA, and performs multiple compressions and optimizations on the deep learning model from the parameter dimension, structure dimension, and operation dimension, and is applicable to mainstream network structures such as CNN, YOLO, and ResNet.

6. The resource optimization management method based on reconfigurable computing according to claim 1, characterized in that, S4-2. The multi-dimensional model compression algorithm includes depthwise separable convolution and cross-layer fusion technologies. Depthwise separable convolution first performs channel-wise convolution and then uses 1×1 convolution to achieve information fusion between channels.

7. A resource optimization management method based on a reconfigurable meter according to claim 1, characterized in that, S5-1. The adaptive dynamic power management mechanism predicts the computing power demand in the future period through a recurrent neural network or a long short-term memory network, and allocates power resources in advance to ensure the priority execution of obstacle avoidance and path planning tasks; S5-2. The adaptive dynamic power management includes DVFS and an AI prediction power consumption model. DVFS automatically adjusts the working voltage and frequency of each computing unit in the chip according to the flight task and load conditions. The AI prediction power consumption model uses deep learning to predict the computing power demand and energy status in the next time period, and adjusts the power distribution in advance to ensure the priority of computing power in critical stages.

8. A resource optimization management method based on a reconfigurable meter according to claim 1, characterized in that S5-1, Ultra-wideband + Millimeter-wave: When the UAV cluster conducts short-range high-bandwidth communication, it can adopt UWB to compress the delay of a single data exchange to less than 5 ms. S5-2, 5G / 6G URLLC: In scenarios such as formation flight, emergency obstacle avoidance, or real-time collaboration that are extremely sensitive to latency, URLLC can reduce the end-to-end delay to the millisecond level, greatly enhancing the synchronization and security of the UAV swarm.

9. A resource optimization management method based on a reconfigurable meter according to claim 1, characterized in that, It also includes S7, MARL-based autonomous path planning, and MARL-based autonomous path planning includes a hybrid strategy of RL + genetic algorithm and a hybrid search of A* and Dijkstra. S7-1, RL + genetic algorithm hybrid strategy, by fusing multi-agent reinforcement learning and genetic algorithm, realizes dynamic optimization and autonomous decision-making of global and local paths in a multi-objective and multi-constraint environment. S7-2, A* and Dijkstra hybrid search, uses the A* algorithm for rapid obstacle avoidance in local path search, and obtains a high-precision feasible solution through the Dijkstra algorithm for global path design, and combines GA for iterative update to balance the global optimal solution and local real-time performance.

10. The resource optimization management method based on a reconfigurable meter according to claim 1, characterized in that, It also includes S8, a lightweight security encryption mechanism, and adopts elliptic curve encryption and hash chain technology as well as an AI anomaly detection algorithm. S8-1, Elliptic curve encryption + Hash chain, can provide high-strength encryption security with a small key volume. Combining with the hash chain further prevents data from being tampered with and replayed. It is suitable for the UAV cluster environment with limited computing resources. When the CPU occupancy rate does not exceed 5%, it can maintain a data integrity verification accuracy rate of more than 98% and effectively resist man-in-the-middle attacks and data replay attacks, providing a solid security guarantee for cluster operations. S8-2, The AI anomaly detection algorithm uses an autoencoder or a graph neural network to detect anomalies and identify attacks in real-time communication data, and accurately screens suspicious data packets or malicious instructions.