Unmanned all-in-one machine dynamic computing power distribution system and method based on AI edge computing

Abstract

The invention discloses an unmanned all-in-one machine dynamic computing power distribution system and method based on AI edge computing, and relates to the technical field of unmanned all-in-one machine dynamic computing power distribution.The system collects environment, computing power, equipment and heat pipe data through an intelligent sensor to generate a feature data set; constructing and optimizing an AI model by using a convolutional neural network to serve as an adaptive neural model, and automatically switching a quantization mode to adapt to different power consumption states; monitoring and analyzing loads of the CPU, the NPU and the FPGA, calculating a task unloading ratio coefficient, and judging whether local computing power is qualified or not; monitoring a resource use condition in real time, calculating a task scheduling weight coefficient, and adjusting and calculating resource allocation according to a preset threshold value; monitoring a task unloading state, calculating a task unloading stability coefficient, and ensuring that a task unloading proportion meets requirements; according to the system, dynamic and intelligent computing power scheduling and task allocation are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic computing power allocation for unmanned integrated machines, and specifically to a dynamic computing power allocation system and method for unmanned integrated machines based on AI edge computing. Background Art

[0002] With the continuous advancement of technology, the demand for automation and intelligence is growing across various industries, particularly in the field of unmanned aerial vehicles (UAVs). UAVs are widely used in logistics and transportation, agricultural monitoring, environmental protection, and other fields, requiring powerful computing power to process real-time data and execute tasks. However, UAVs have limited computing resources, and traditional computing resource allocation methods often fail to dynamically adjust to task requirements and device status, resulting in uneven resource utilization and wasted energy.

[0003] Currently, most UAVs use a fixed computing power allocation model. This model cannot flexibly adjust computing resources based on task complexity, the real-time status of the device, or environmental changes, thus affecting overall efficiency and energy management. This is especially true when multiple computing units, such as CPUs, NPUs, and FPGAs, operate simultaneously. Traditional approaches often suffer from overload, inefficiency, and irrational computing resource allocation. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention provides a dynamic computing power allocation system and method for unmanned integrated machines based on AI edge computing to solve the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a dynamic computing power allocation system for unmanned integrated machines based on AI edge computing, a dynamic computing power allocation system for unmanned integrated machines based on AI edge computing, including a data acquisition module, an adaptive neural model module, a heterogeneous computing management module, a computing power dynamic allocation module and a collaborative computing module; The data acquisition module is used to deploy intelligent sensors on the unmanned integrated machine to collect environmental perception data, computing power status data, equipment status data and heat pipe data, and preprocess the collected data, finally extract key features and establish a feature data set; The adaptive neural model module is used to build an initial AI model using a convolutional neural network (CNN), and is trained and optimized through task data and resource status to adapt to different computing power environments. The module optimizes the model through distillation, quantization, and pruning, deploys multiple versions of the AI ​​model, and automatically switches quantization modes under different power consumption states. The trained model is then run as an adaptive neural model to optimize the scheduling of computing power and the allocation of computing tasks. The heterogeneous computing management module is used to monitor and analyze the load and resource utilization of the computing resources CPU, NPU and FPGA in the unmanned integrated machine in real time, obtain the task offloading ratio coefficient Roff by calculation, and compare it with the first threshold Q1 to determine whether the local computing power is qualified. If it is unqualified, a corresponding strategy is generated; The computing power dynamic allocation module is used to monitor the usage of computing resources in real time after receiving the first warning instruction, and calculate the task scheduling weight coefficient WQZ based on the data of the feature data set, and compare it with the preset second threshold Q2 to determine whether the computing resource allocation of the task is normal. If the computing resource allocation is abnormal, a corresponding strategy is generated; The collaborative computing module is used to monitor the status of task unloading in real time after receiving the second warning instruction, and calculate the task unloading stability coefficient in combination with the feature data set. , and compare and analyze it with the third threshold Q3 to determine whether the task offloading ratio is within the qualified range. If it is not within the qualified range, a corresponding strategy is generated.

[0006] Preferably, the data acquisition module includes an acquisition unit, a data fusion unit, a data processing unit and a feature data extraction unit; The acquisition unit is used to install an RGB camera on the front of the unmanned aerial vehicle to collect image information for machine vision tasks; install a laser radar on the bottom of the unmanned aerial vehicle to collect three-dimensional contour information of objects; install a temperature sensor inside the unmanned aerial vehicle to collect the temperature values ​​of the CPU, NPU and FPGA chips; collect computing power status data by real-time monitoring of the internal usage of the CPU, NPU and FPGA; and collect device status data by real-time monitoring of battery status and communication status. The data fusion unit is used to fuse the collected data, including time alignment: aligning the collected data by timestamp to ensure that they are collected at the same time; spatial calibration: using the depth information of the laser radar to match each pixel in the image captured by the camera to the position of the object in the actual three-dimensional space; The data processing unit is used to filter the noise of the data, including image data denoising: using Gaussian filtering to remove random noise in the image; LiDAR point cloud filtering: removing abnormal points and deviation points caused by sensor errors; GPS data de-jittering: applying Kalman filtering to smooth the GPS trajectory and reduce errors caused by signal jitter; The feature data extraction unit is used to extract features from the data, including image feature extraction: using edge detection and target detection algorithms to identify the location and category of the target object; computing load feature extraction: extracting the current computing power occupancy and load trend from the CPU, NPU and FPGA load data; battery status feature extraction: extracting the remaining power and discharge rate; chip temperature feature extraction: extracting the temperature change trend of the CPU, NPU and FPGA, and establishing a feature data set.

[0007] Preferably, the adaptive neural model module is used to construct an initial AI model using a convolutional neural network (CNN), and to train and test it using task input data and computing resource status data, so that it has the ability to adapt to different computing power conditions. Based on environmental perception data, the initial CNN model is trained so that it can perform core computing tasks of target recognition and path planning; the model is distilled, quantized, and pruned to reduce computational complexity while retaining key feature extraction capabilities; multiple versions of AI models are deployed to adapt to different computing power environments, and model selection is dynamically adjusted based on computing power status data, device status data, and heat pipe data; during actual operation, the feature vectors used to extract different task requirements and computing power occupancy are used to further optimize the model selection logic, and the quantization mode is automatically switched under different power consumption states; and the trained model is run as an adaptive neural model to optimize the scheduling of computing power and the allocation of computing tasks.

[0008] Preferably, the heterogeneous computing management module includes a first computing unit and a first analyzing unit; The first computing unit is used to monitor the load and resource usage of the CPU, NPU, and FPGA in real time, and calculate the task offloading ratio coefficient Roff after dimensionless processing based on the characteristic data set data. The formula is as follows: ; Where, Indicates the CPU load, Indicates the load of the NPU, represents the load of the FPGA, Indicates the model computation load value, obtained by model training.

[0009] Preferably, the first analysis unit is configured to preset a first threshold Q1 in advance, and compare and analyze the task offloading ratio coefficient Roff with the first threshold Q1, and obtaining the first evaluation result includes: When the task offloading ratio coefficient Roff ≤ the first threshold Q1, it indicates that the local computing power is qualified and there is no risk of overload in response to the task computing requirements, and continuous monitoring is required; When the task offloading ratio coefficient Roff is greater than the first threshold Q1, it indicates that the local computing power is unqualified and there is a risk of overload in response to the task computing demand. The first warning instruction is triggered and the first strategy is generated: adjust the task priority according to the task scheduling weight.

[0010] Preferably, the computing power dynamic allocation module includes a second computing unit and a second analyzing unit; The second computing unit is used to monitor the computing resource usage in real time when receiving the first warning instruction, and calculate the task scheduling weight coefficient WQZ after dimensionless processing based on the characteristic data set data. The formula is as follows: ; Where, represents the priority weight coefficient of task j, represents the computational load of task j, represents the maximum computing power of task j, represents the power consumption requirement value of task j, Indicates the remaining battery power of the drone. represents the task power consumption scheduling impact coefficient, represents the execution delay of task j, Indicates the maximum delay, represents the task execution delay scheduling impact coefficient, represents the chip temperature value when processing task j, Indicates the maximum chip temperature, Indicates the scheduling impact coefficient of chip temperature.

[0011] Preferably, the second analysis unit is used to preset a second threshold Q2 in advance, and compare and analyze the task scheduling weight coefficient WQZ with the second threshold Q2, and obtain the second evaluation result including: When the task scheduling weight coefficient WQZ ≤ the second threshold Q2, it indicates that the task computing resource allocation is normal, the task execution meets the computing scheduling requirements, and there is no risk of computing resource anomalies and scheduling policy instability, so the task execution is postponed; When the task scheduling weight coefficient WQZ is greater than the second threshold Q2, it indicates that the task computing resource allocation is abnormal, the task execution does not meet the computing scheduling requirements, and there is a risk of computing resource abnormality and unstable scheduling strategy. The second warning instruction is triggered and the second strategy is generated: offloading the computing task to the edge computing node for processing.

[0012] Preferably, the collaborative computing module includes a third computing unit and a third analyzing unit; The third calculation unit is used to monitor the status of task unloading in real time when receiving the second warning instruction, and calculate the task unloading stability coefficient after dimensionless processing based on the characteristic data set data. , the formula is as follows: ;

[0013] Where, Represents a collection of all tasks that need to be uninstalled. represents the number of computation cycles of task j, represents the communication cost of transmitting task j to the edge computing node, Indicates the available computing power of the edge computing node, Indicates the proportion of local computing load, Indicates the maximum load value that the local computing resources can bear. Indicates the influence coefficient of local computing resource load on task offloading, Indicates the real-time network delay, represents the maximum network delay that the task can accept, a2 represents the impact coefficient of network delay on offloading, and w1 and w2 represent weight coefficients.

[0014] Preferably, the third analysis unit is used to preset a third threshold Q3 in advance and set the task offloading stability coefficient Comparing and analyzing with the third threshold value Q3 to obtain a third evaluation result includes: When task offloading stability factor When the value is less than or equal to the third threshold Q3, it indicates that the task offloading ratio is within the acceptable range and there is no risk of overloading edge computing nodes or causing unstable network transmission. Continuous monitoring is required. When task offloading stability factor When the value exceeds the third threshold Q3, it indicates that the task offloading ratio is in an unqualified range, which may lead to overload of edge computing nodes and unstable network transmission. The third warning instruction is triggered, and the third strategy is generated: reduce the task offloading ratio by 20%, reduce the data compression ratio by 10%, and increase the communication cost by 10%; split the task into multiple subtasks to reduce the network load.

[0015] Preferably, a method for allocating dynamic computing power of an unmanned integrated machine based on AI edge computing comprises the following steps: Step 1: Deploy smart sensors on the UAV to collect environmental perception data, computing power status data, equipment status data, and heat pipe data. Preprocess the collected data, extract key features, and establish a feature dataset. Step 2: Build an initial AI model using a convolutional neural network (CNN). This model is trained and optimized using task data and resource status to adapt to different computing environments. The module optimizes the model through distillation, quantization, and pruning, deploys multiple versions of the AI ​​model, and automatically switches quantization modes under different power consumption conditions. The trained model is then run as an adaptive neural model to optimize computing power scheduling and computing task allocation. Step 3: Monitor and analyze the load and resource utilization of the CPU, NPU, and FPGA computing resources in the UAV in real time. Calculate the task offload ratio coefficient Roff and compare it with the first threshold Q1 to determine whether the local computing power is qualified. If it is unqualified, generate a corresponding strategy. Step 4: Monitor the usage of computing resources in real time, and calculate the task scheduling weight coefficient WQZ based on the data of the feature dataset. This coefficient is then compared with the preset second threshold Q2 to determine whether the computing resource allocation of the task is normal. If the computing resource allocation is abnormal, a corresponding strategy is generated. Step 5: Monitor the status of task offloading in real time and calculate the task offloading stability coefficient based on the feature data set. , and compare and analyze it with the third threshold Q3 to determine whether the task offloading ratio is within the qualified range. If it is not within the qualified range, a corresponding strategy is generated.

[0016] The present invention provides a dynamic computing power allocation system and method for unmanned integrated machines based on AI edge computing. It has the following beneficial effects: (1) This is a dynamic computing power allocation system and method for unmanned aerial vehicles based on AI edge computing. Through real-time dynamic monitoring and intelligent scheduling, the system can automatically adjust the computing resource allocation according to the actual task requirements and equipment status of the unmanned aerial vehicle, ensure the load balance of the CPU, NPU and FPGA, avoid resource overload or waste, and thus significantly improve the utilization of computing resources.

[0017] (2) This is a dynamic computing power allocation system and method for unmanned integrated machines based on AI edge computing. It optimizes computing power allocation and task scheduling through an adaptive neural model, combines environmental perception data, computing power status and device status to achieve low-power operation, and automatically switches quantization modes under different power consumption states, reducing energy consumption during the calculation process, thereby significantly improving the energy efficiency of the system.

[0018] (3) This unmanned aerial vehicle dynamic computing power allocation system and method based on AI edge computing can identify potential system overload risks in real time through a three-stage early warning mechanism of computing power status, task scheduling, and task offloading, and generate corresponding strategies based on the task offloading ratio and computing resource status to ensure the stability of the system in a high-load environment and reduce equipment failures and unstable operations.

[0019] (4) This is a dynamic computing power allocation system and method for unmanned integrated machines based on AI edge computing. Through priority scheduling and task offloading strategies for computing tasks, the system can dynamically adjust the computing resource allocation and offloading strategies of tasks according to task type, computing load and resource status, ensuring that tasks can be completed in the shortest time and improving the overall task execution efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flowchart of a dynamic computing power allocation system for unmanned integrated machines based on AI edge computing in the present invention; Figure 2 This is a schematic diagram of the steps of a method for dynamic computing power allocation of an unmanned integrated machine based on AI edge computing in the present invention. DETAILED DESCRIPTION

[0021] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0022] Example 1 See also Figure 1 , the present invention provides an unmanned integrated machine dynamic computing power allocation system based on AI edge computing, including a data acquisition module, an adaptive neural model module, a heterogeneous computing management module, a computing power dynamic allocation module and a collaborative computing module; The data acquisition module is used to deploy intelligent sensors on the unmanned integrated machine to collect environmental perception data, computing power status data, equipment status data and heat pipe data, and preprocess the collected data, finally extract key features and establish a feature data set; The adaptive neural model module is used to build an initial AI model using a convolutional neural network (CNN), and is trained and optimized through task data and resource status to adapt to different computing power environments. The module optimizes the model through distillation, quantization, and pruning, deploys multiple versions of the AI ​​model, and automatically switches quantization modes under different power consumption states. The trained model is then run as an adaptive neural model to optimize the scheduling of computing power and the allocation of computing tasks. The heterogeneous computing management module is used to monitor and analyze the load and resource utilization of the computing resources CPU, NPU and FPGA in the unmanned integrated machine in real time, obtain the task offloading ratio coefficient Roff by calculation, and compare it with the first threshold Q1 to determine whether the local computing power is qualified. If it is unqualified, a corresponding strategy is generated; The computing power dynamic allocation module is used to monitor the usage of computing resources in real time after receiving the first warning instruction, and calculate the task scheduling weight coefficient WQZ based on the data of the feature data set, and compare it with the preset second threshold Q2 to determine whether the computing resource allocation of the task is normal. If the computing resource allocation is abnormal, a corresponding strategy is generated; The collaborative computing module is used to monitor the status of task unloading in real time after receiving the second warning instruction, and calculate the task unloading stability coefficient in combination with the feature data set. , and compare and analyze it with the third threshold Q3 to determine whether the task offloading ratio is within the qualified range. If it is not within the qualified range, a corresponding strategy is generated.

[0023] This embodiment utilizes a multi-layered monitoring and adaptive adjustment mechanism, including real-time monitoring of computing resource usage, calculation of task scheduling weights, and assessment of task offloading stability factors, to dynamically optimize the UAV's computing resource allocation and task offloading strategies. Under varying load conditions, the system automatically adjusts computing resource allocation and offloading ratios, effectively avoiding resource overload and uneven offloading, thereby improving task execution efficiency and ensuring stable and efficient system operation in various environments.

[0024] Example 2: This example is an explanation of Example 1. Specifically, the data acquisition module includes an acquisition unit, a data fusion unit, a data processing unit, and a feature data extraction unit. The acquisition unit is used to install an RGB camera on the front of the unmanned aerial vehicle to collect image information for machine vision tasks; install a laser radar on the bottom of the unmanned aerial vehicle to collect three-dimensional contour information of objects; install a temperature sensor inside the unmanned aerial vehicle to collect the temperature values ​​of the CPU, NPU and FPGA chips; collect computing power status data by real-time monitoring of the internal usage of the CPU, NPU and FPGA; and collect device status data by real-time monitoring of battery status and communication status. The data fusion unit is used to fuse the collected data, including time alignment: aligning the collected data by timestamp to ensure that they are collected at the same time; spatial calibration: using the depth information of the laser radar to match each pixel in the image captured by the camera to the position of the object in the actual three-dimensional space; The data processing unit is used to filter the noise of the data, including image data denoising: using Gaussian filtering to remove random noise in the image; LiDAR point cloud filtering: removing abnormal points and deviation points caused by sensor errors; GPS data de-jittering: applying Kalman filtering to smooth the GPS trajectory and reduce errors caused by signal jitter; The feature data extraction unit is used to extract features from the data, including image feature extraction: using edge detection and target detection algorithms to identify the location and category of the target object; computing load feature extraction: extracting the current computing power occupancy and load trend from the CPU, NPU and FPGA load data; battery status feature extraction: extracting the remaining power and discharge rate; chip temperature feature extraction: extracting the temperature change trend of the CPU, NPU and FPGA, and establishing a feature data set.

[0025] In this embodiment, by integrating multiple intelligent sensors such as RGB cameras, lidar, and temperature sensors, the data acquisition module can comprehensively and real-timely acquire the UAV's environmental perception, computing power status, equipment status, and heat pipe data. The data fusion unit performs time alignment and spatial calibration to ensure the accuracy and consistency of the collected data. The data processing unit further improves data quality through noise filtering, and the feature data extraction unit extracts key features from multiple dimensions to provide the system with high-precision real-time data support, thereby enhancing the system's adaptability to environmental changes and the reliability of task execution.

[0026] Example 3. This example is an explanation of Example 2. Specifically, the adaptive neural model module is used to construct an initial AI model using a convolutional neural network (CNN), and to train and test it using task input data and computing resource status data, so that it has the ability to adapt to different computing power conditions. Based on environmental perception data, the initial CNN model is trained so that it can perform core computing tasks of target recognition and path planning; the model is distilled, quantized, and pruned to reduce computational complexity while retaining key feature extraction capabilities; multiple versions of AI models are deployed to adapt to different computing power environments, and model selection is dynamically adjusted based on computing power status data, device status data, and heat pipe data; in actual operation, the feature vectors based on different extraction task requirements and computing power occupancy are used to further optimize the model selection logic, and the quantization mode is automatically switched under different power consumption states; and the trained model is run as an adaptive neural model to optimize the scheduling of computing power and the allocation of computing tasks.

[0027] In this embodiment, by optimizing the convolutional neural network (CNN) of the adaptive neural model module and combining task input data with computing resource status data, the system can dynamically adjust computing task allocation and model selection based on real-time computing power conditions. This dynamic optimization capability based on environmental perception enables the model to run efficiently under different computing power environments. It also reduces computational complexity through distillation, quantization, and pruning techniques while retaining key feature extraction capabilities, greatly improving the system's computing power utilization and task execution efficiency, ensuring that tasks can be executed smoothly and efficiently under different power consumption states.

[0028] Example 4: This example is an explanation of Example 3. Specifically, the heterogeneous computing management module includes a first computing unit and a first analyzing unit. The first computing unit is used to monitor the load and resource usage of the CPU, NPU, and FPGA in real time, and calculate the task offloading ratio coefficient Roff after dimensionless processing based on the characteristic data set data. The formula is as follows: ; Where, Indicates the CPU load, Indicates the load of the NPU, represents the load of the FPGA, Indicates the model computation load value, obtained by model training.

[0029] In this embodiment, the first computing unit of the heterogeneous computing management module monitors the CPU, NPU, and FPGA load in real time, performs dimensionless processing on the characteristic dataset, and calculates the task offload ratio coefficient Roff. This method dynamically assesses the load of each computing unit and adjusts the task offload strategy in real time, effectively optimizing computing resource allocation, avoiding overload and inefficiency, thereby improving computing resource utilization and ensuring that the unmanned aerial vehicle maintains optimal load balancing in various computing environments.

[0030] Example 5: This example is explained in Example 4. Specifically, the first analysis unit is used to preset a first threshold Q1 in advance, and compare and analyze the task offloading ratio coefficient Roff with the first threshold Q1. Obtaining a first evaluation result includes: When the task offloading ratio coefficient Roff ≤ the first threshold Q1, it indicates that the local computing power is qualified and there is no risk of overload in response to the task computing requirements, and continuous monitoring is required; When the task offloading ratio coefficient Roff is greater than the first threshold Q1, it indicates that the local computing power is unqualified and there is a risk of overload in response to the task computing demand. The first warning instruction is triggered and the first strategy is generated: adjust the task priority according to the task scheduling weight.

[0031] In this embodiment, the first analysis unit compares the task offload ratio coefficient Roff with a preset first threshold Q1, accurately assessing the status of local computing power. When Roff exceeds the threshold, the system promptly identifies the risk of insufficient computing power, triggers an alert, and generates an optimization strategy to adjust task priorities. This mechanism effectively prevents computing resource overload and ensures the stability and efficiency of task scheduling.

[0032] Example 6: This example is an explanation of Example 5. Specifically, the computing power dynamic allocation module includes a second computing unit and a second analyzing unit. The second computing unit is used to monitor the computing resource usage in real time when receiving the first warning instruction, and calculate the task scheduling weight coefficient WQZ after dimensionless processing based on the characteristic data set data. The formula is as follows: ; Where, represents the priority weight coefficient of task j, represents the computational load of task j, represents the maximum computing power of task j, represents the power consumption requirement value of task j, Indicates the remaining battery power of the drone. represents the task power consumption scheduling impact coefficient, represents the execution delay of task j, Indicates the maximum delay, represents the task execution delay scheduling impact coefficient, represents the chip temperature value when processing task j, Indicates the maximum chip temperature, represents the scheduling impact coefficient of chip temperature;

[0033] Table 1.

[0034] In this embodiment, after receiving the first warning instruction, the second computing unit monitors computing resource usage in real time and calculates the task scheduling weight coefficient WQZ based on the characteristic data set. This factor, combined with multiple factors such as the task's computational load, power consumption requirements, execution latency, and chip temperature, allows for precise allocation of computing resources to each task. This mechanism effectively balances the priorities and resource requirements of different tasks, ensuring that the system can properly schedule tasks when computing power is insufficient, avoiding overload or excessive power consumption.

[0035] Example 7: This example is an explanation of Example 6. Specifically, the second analysis unit is used to preset a second threshold Q2 in advance, and compare and analyze the task scheduling weight coefficient WQZ with the second threshold Q2. Obtaining the second evaluation result includes: When the task scheduling weight coefficient WQZ ≤ the second threshold Q2, it indicates that the task computing resource allocation is normal, the task execution meets the computing scheduling requirements, and there is no risk of computing resource anomalies and scheduling policy instability, so the task execution is postponed; When the task scheduling weight coefficient WQZ is greater than the second threshold Q2, it indicates that the task computing resource allocation is abnormal, the task execution does not meet the computing scheduling requirements, and there is a risk of computing resource abnormality and unstable scheduling strategy. The second warning instruction is triggered and the second strategy is generated: offloading the computing task to the edge computing node for processing.

[0036] In this embodiment, a second analysis unit compares and analyzes the task scheduling weight coefficient WQZ with a second threshold Q2. When the task scheduling weight coefficient WQZ exceeds the second threshold Q2, it promptly detects abnormal task computing resource allocation and triggers a second warning instruction. This mechanism automatically generates an offload strategy, offloading computing tasks from local computing resources to edge computing nodes for processing, thereby avoiding local resource overload and scheduling instability, ensuring that tasks can be stably executed in a more suitable computing environment, and improving the overall computing efficiency and resource utilization of the system.

[0037] Example 8: This example is an explanation of Example 7. Specifically, the collaborative computing module includes a third computing unit and a third analyzing unit. The third calculation unit is used to monitor the status of task unloading in real time when receiving the second warning instruction, and calculate the task unloading stability coefficient after dimensionless processing based on the characteristic data set data. , the formula is as follows: ;

[0038] Where, Represents a collection of all tasks that need to be uninstalled. represents the number of computation cycles of task j, represents the communication cost of transmitting task j to the edge computing node, Indicates the available computing power of the edge computing node, Indicates the proportion of local computing load, Indicates the maximum load value that the local computing resources can bear. Indicates the influence coefficient of local computing resource load on task offloading, Indicates the real-time network delay, represents the maximum network delay that the task can accept, a2 represents the influence coefficient of network delay on offloading, w1 and w2 represent weight coefficients, , ,and ;

[0039] Table 2.

[0040] In this embodiment, the task offloading status is monitored in real time by the third computing unit in the collaborative computing module. The task offloading stability coefficient is calculated in combination with the characteristic data set, which effectively evaluates the stability and resource usage during the task offloading process. When the task is offloaded to the edge computing node, the system can adjust the offloading strategy in real time based on factors such as the task's computing cycle, communication cost, and network latency. This mechanism ensures stable offloading of tasks under different loads and network conditions, thereby improving the computing efficiency of the overall system and the stability of task execution, and avoiding task failures or delays under high load or unstable network conditions.

[0041] Example 9: This example is explained in Example 8. Specifically, the third analysis unit is used to preset the third threshold Q3 in advance and set the task offloading stability coefficient Comparing and analyzing with the third threshold value Q3 to obtain a third evaluation result includes: When task offloading stability factor When the value is less than or equal to the third threshold Q3, it indicates that the task offloading ratio is within the acceptable range and there is no risk of overloading edge computing nodes or causing unstable network transmission. Continuous monitoring is required. When task offloading stability factor When the value exceeds the third threshold Q3, it indicates that the task offloading ratio is in an unqualified range, which may lead to overload of edge computing nodes and unstable network transmission. The third warning instruction is triggered, and the third strategy is generated: reduce the task offloading ratio by 20%, reduce the data compression ratio by 10%, and increase the communication cost by 10%; split the task into multiple subtasks to reduce the network load.

[0042] In this embodiment, the third analysis unit compares and analyzes the task offloading stability coefficient with the preset third threshold value Q3, which can accurately assess the stability of task offloading and the load of edge computing nodes and network transmission. When the task offloading ratio exceeds the qualified range, the system can automatically trigger the third warning instruction and adopt an optimization strategy, including reducing the task offloading ratio, adjusting the data compression ratio, and increasing the communication cost, thereby effectively avoiding edge computing node overload and network instability. This mechanism improves the system's adaptability in high-load environments, ensuring the smooth operation of computing resources and the efficiency of task execution.

[0043] Example 10, a method for allocating dynamic computing power of unmanned integrated machines based on AI edge computing, please refer to Figure 2 , including the following steps: Step 1: Deploy smart sensors on the UAV to collect environmental perception data, computing power status data, equipment status data, and heat pipe data. Preprocess the collected data, extract key features, and establish a feature dataset. Step 2: Build an initial AI model using a convolutional neural network (CNN). This model is trained and optimized using task data and resource status to adapt to different computing environments. The module optimizes the model through distillation, quantization, and pruning, deploys multiple versions of the AI ​​model, and automatically switches quantization modes under different power consumption conditions. The trained model is then run as an adaptive neural model to optimize computing power scheduling and computing task allocation. Step 3: Monitor and analyze the load and resource utilization of the CPU, NPU, and FPGA computing resources in the UAV in real time. Calculate the task offload ratio coefficient Roff and compare it with the first threshold Q1 to determine whether the local computing power is qualified. If it is unqualified, generate a corresponding strategy. Step 4: Monitor the usage of computing resources in real time, and calculate the task scheduling weight coefficient WQZ based on the data of the feature dataset. This coefficient is then compared with the preset second threshold Q2 to determine whether the computing resource allocation of the task is normal. If the computing resource allocation is abnormal, a corresponding strategy is generated. Step 5: Monitor the status of task offloading in real time and calculate the task offloading stability coefficient based on the feature data set. , and compare and analyze it with the third threshold Q3 to determine whether the task offloading ratio is within the qualified range. If it is not within the qualified range, a corresponding strategy is generated.

[0044] In this embodiment, through the comprehensive execution of the above steps, the unmanned all-in-one machine can realize comprehensive monitoring and analysis of environmental perception data, computing power status data, equipment status data and heat pipe data, and then optimize computing task scheduling and resource allocation through adaptive neural models. The system can dynamically adjust the use of computing resources according to the real-time monitored computing power load, task scheduling weight and unloading stability to ensure that tasks are executed efficiently according to priority and avoid computing power overload or resource waste. When the load is too high or task resources are unevenly distributed, the system automatically generates an optimization strategy and adjusts the task offloading ratio, significantly improving the system's computing power and resource utilization, thereby achieving efficient and stable task scheduling in a changing computing power environment.

[0045] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by technicians in this field for each set of sample data; as long as it does not affect the proportional relationship between the parameter and the quantized value.

[0046] The above formulas are obtained by collecting a large amount of data and performing software simulation, and a formula close to the actual value is selected. The coefficients in the formula are set by those skilled in the art according to actual conditions. The above is only a preferred specific implementation method of the present invention, but the protection scope of the present invention is not limited to this. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, can make equivalent replacements or changes based on the technical solution and inventive concept of the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A dynamic computing power allocation system for unmanned integrated machines based on AI edge computing, characterized in that: It includes data acquisition module, adaptive neural model module, heterogeneous computing management module, computing power dynamic allocation module and collaborative computing module; The data acquisition module is used to deploy intelligent sensors on the unmanned integrated machine to collect environmental perception data, computing power status data, equipment status data and heat pipe data, and preprocess the collected data, finally extract key features and establish a feature data set; The adaptive neural model module is used to build an initial AI model using a convolutional neural network (CNN), and is trained and optimized through task data and resource status to adapt to different computing power environments. The module optimizes the model through distillation, quantization, and pruning, deploys multiple versions of the AI ​​model, and automatically switches quantization modes under different power consumption states. The trained model is then run as an adaptive neural model to optimize the scheduling of computing power and the allocation of computing tasks. The heterogeneous computing management module is used to monitor and analyze the load and resource utilization of the computing resources CPU, NPU and FPGA in the unmanned integrated machine in real time, obtain the task offloading ratio coefficient Roff by calculation, and compare it with the first threshold Q1 to determine whether the local computing power is qualified. If it is unqualified, a corresponding strategy is generated; The computing power dynamic allocation module is used to monitor the usage of computing resources in real time after receiving the first warning instruction, and calculate the task scheduling weight coefficient WQZ based on the data of the feature data set, and compare it with the preset second threshold Q2 to determine whether the computing resource allocation of the task is normal. If the computing resource allocation is abnormal, a corresponding strategy is generated; The collaborative computing module is used to monitor the status of task unloading in real time after receiving the second warning instruction, and calculate the task unloading stability coefficient in combination with the feature data set. , and compare and analyze it with the third threshold Q3 to determine whether the task offloading ratio is within the qualified range. If it is not within the qualified range, a corresponding strategy is generated.

2. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 1 is characterized in that: The data acquisition module includes an acquisition unit, a data fusion unit, a data processing unit and a feature data extraction unit; The acquisition unit is used to install an RGB camera on the front of the unmanned aerial vehicle to collect image information for machine vision tasks; install a laser radar on the bottom of the unmanned aerial vehicle to collect three-dimensional contour information of objects; install a temperature sensor inside the unmanned aerial vehicle to collect the temperature values ​​of the CPU, NPU and FPGA chips; collect computing power status data by real-time monitoring of the internal usage of the CPU, NPU and FPGA; and collect device status data by real-time monitoring of battery status and communication status. The data fusion unit is used to perform data fusion on the collected data, including time alignment: aligning the collected data by time stamps to ensure that they are collected at the same time; space Calibration: Using the depth information from the LiDAR, each pixel in the image captured by the camera is matched to the position of the object in the actual 3D space. The data processing unit is used to filter the noise of the data, including image data denoising: using Gaussian filtering to remove random noise in the image; LiDAR point cloud filtering: removing abnormal points and deviation points caused by sensor errors; GPS data de-jittering: applying Kalman filtering to smooth the GPS trajectory and reduce errors caused by signal jitter; The feature data extraction unit is used to extract features from the data, including image feature extraction: using edge detection and target detection algorithms to identify the location and category of the target object; computing load feature extraction: extracting the current computing power occupancy and load trend from the CPU, NPU and FPGA load data; battery status feature extraction: extracting the remaining power and discharge rate; chip temperature feature extraction: extracting the temperature change trend of the CPU, NPU and FPGA, and establishing a feature data set.

3. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 2 is characterized in that: The adaptive neural model module is used to build an initial AI model using a convolutional neural network (CNN). It uses task input data and computing resource status data for training and testing, enabling it to adapt to different computing power conditions. Based on environmental perception data, the initial CNN model is trained to perform core computing tasks such as target recognition and path planning. The model is then distilled, quantized, and pruned to reduce computational complexity while retaining key feature extraction capabilities. Deploy multiple versions of AI models to adapt to different computing power environments, and dynamically adjust model selection based on computing power status data, device status data, and heat pipe data; during actual operation, the feature vectors based on different extraction task requirements and computing power occupancy are used to further optimize the model selection logic, and automatically switch quantization modes under different power consumption states; and run the trained model as an adaptive neural model to optimize computing power scheduling and computing task allocation.

4. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 3 is characterized in that: The heterogeneous computing management module includes a first computing unit and a first analyzing unit; The first computing unit is used to monitor the load and resource usage of the CPU, NPU, and FPGA in real time, and calculate the task offloading ratio coefficient Roff after dimensionless processing based on the characteristic data set data. The formula is as follows: ； Where, Indicates the CPU load, Indicates the load of the NPU, represents the load of the FPGA, Indicates the model computation load value, obtained by model training.

5. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 4 is characterized in that: The first analysis unit is configured to preset a first threshold Q1 in advance, and compare and analyze the task offloading ratio coefficient Roff with the first threshold Q1, and obtain a first evaluation result including: When the task offloading ratio coefficient Roff ≤ the first threshold Q1, it indicates that the local computing power is qualified and there is no risk of overload in response to the task computing requirements, and continuous monitoring is required; When the task offloading ratio coefficient Roff is greater than the first threshold Q1, it indicates that the local computing power is unqualified and there is a risk of overload in response to the task computing demand. The first warning instruction is triggered and the first strategy is generated: adjust the task priority according to the task scheduling weight.

6. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 5 is characterized in that: The computing power dynamic allocation module includes a second computing unit and a second analyzing unit; The second computing unit is used to monitor the computing resource usage in real time when receiving the first warning instruction, and calculate the task scheduling weight coefficient WQZ after dimensionless processing based on the characteristic data set data. The formula is as follows: ； Where, represents the priority weight coefficient of task j, represents the computational load of task j, represents the maximum computing power of task j, represents the power consumption requirement value of task j, Indicates the remaining battery power of the drone. represents the task power consumption scheduling impact coefficient, represents the execution delay of task j, Indicates the maximum delay, represents the task execution delay scheduling impact coefficient, represents the chip temperature value when processing task j, Indicates the maximum chip temperature, Indicates the scheduling impact coefficient of chip temperature.

7. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 6 is characterized in that: The second analysis unit is configured to preset a second threshold Q2 in advance, and compare and analyze the task scheduling weight coefficient WQZ with the second threshold Q2 to obtain a second evaluation result, including: When the task scheduling weight coefficient WQZ ≤ the second threshold Q2, it indicates that the task computing resource allocation is normal, the task execution meets the computing scheduling requirements, and there is no risk of computing resource anomalies and scheduling policy instability, so the task execution is postponed; When the task scheduling weight coefficient WQZ is greater than the second threshold Q2, it indicates that the task computing resource allocation is abnormal, the task execution does not meet the computing scheduling requirements, and there is a risk of computing resource abnormality and unstable scheduling strategy. The second warning instruction is triggered and the second strategy is generated: offloading the computing task to the edge computing node for processing.

8. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 7 is characterized in that: The collaborative computing module includes a third computing unit and a third analyzing unit; The third calculation unit is used to monitor the status of task unloading in real time when receiving the second warning instruction, and calculate the task unloading stability coefficient after dimensionless processing based on the characteristic data set data. , the formula is as follows: ； Where, Represents a collection of all tasks that need to be uninstalled. represents the number of computation cycles of task j, represents the communication cost of transmitting task j to the edge computing node, Indicates the available computing power of the edge computing node, Indicates the proportion of local computing load, Indicates the maximum load value that the local computing resources can bear. Indicates the influence coefficient of local computing resource load on task offloading, Indicates the real-time network delay, represents the maximum network delay that the task can accept, a2 represents the impact coefficient of network delay on offloading, and w1 and w2 represent weight coefficients.

9. The AI ​​edge computing-based unmanned integrated machine dynamic computing power allocation system according to claim 8, characterized in that: The third analysis unit is used to preset a third threshold Q3 in advance and set the task offloading stability coefficient Comparing and analyzing with the third threshold value Q3 to obtain a third evaluation result includes: When task offloading stability factor When the value is less than or equal to the third threshold Q3, it indicates that the task offloading ratio is within the acceptable range and there is no risk of overloading edge computing nodes or causing unstable network transmission. Continuous monitoring is required. When task offloading stability factor When the value exceeds the third threshold Q3, it indicates that the task offloading ratio is in an unqualified range, which may lead to overload of edge computing nodes and unstable network transmission. The third warning instruction is triggered, and the third strategy is generated: reduce the task offloading ratio by 20%, reduce the data compression ratio by 10%, and increase the communication cost by 10%; split the task into multiple subtasks to reduce the network load.

10. A method for dynamic computing power allocation of an unmanned integrated machine based on AI edge computing, comprising the dynamic computing power allocation system for an unmanned integrated machine based on AI edge computing as described in any one of claims 1 to 9, comprising the following steps: Step 1: Deploy smart sensors on the UAV to collect environmental perception data, computing power status data, equipment status data, and heat pipe data. Preprocess the collected data, extract key features, and establish a feature dataset. Step 2: Build an initial AI model using a convolutional neural network (CNN). This model is trained and optimized using task data and resource status to adapt to different computing environments. The module optimizes the model through distillation, quantization, and pruning, deploys multiple versions of the AI ​​model, and automatically switches quantization modes under different power consumption conditions. The trained model is then run as an adaptive neural model to optimize computing power scheduling and computing task allocation. Step 3: Monitor and analyze the load and resource utilization of the CPU, NPU, and FPGA computing resources in the UAV in real time. Calculate the task offload ratio coefficient Roff and compare it with the first threshold Q1 to determine whether the local computing power is qualified. If it is unqualified, generate a corresponding strategy. Step 4: Monitor the usage of computing resources in real time, and calculate the task scheduling weight coefficient WQZ based on the data of the feature dataset. This coefficient is then compared with the preset second threshold Q2 to determine whether the computing resource allocation of the task is normal. If the computing resource allocation is abnormal, a corresponding strategy is generated. Step 5: Monitor the status of task offloading in real time and calculate the task offloading stability coefficient based on the feature data set. , and compare and analyze it with the third threshold Q3 to determine whether the task offloading ratio is within the qualified range. If it is not within the qualified range, a corresponding strategy is generated.

Images