Computing resource allocation method for mobile detection task scheduling in extreme environment

Abstract

The invention discloses a computing resource allocation method for mobile detection task scheduling in an extreme environment, which comprises the following steps of: calculating a priority coefficient of a task to be scheduled based on the time sensitivity, importance weight and resource occupancy rate of the task; outputting the actual energy efficiency ratio of the calculation core based on the basic energy efficiency of the calculation core and the temperature attenuation coefficient and the radiation attenuation coefficient in the current environment; dynamic load balancing is executed, a two-dimensional scheduling model is constructed based on the priority coefficient and the actual energy efficiency ratio, high-priority tasks are directionally distributed to a calculation core with the high actual energy efficiency ratio, and low-priority tasks are distributed to a calculation core with the low actual energy efficiency ratio; and an energy consumption prediction module is introduced to predict the calculation load in a future set time period, and the working state of the calculation core is adjusted in advance according to a prediction result. The total power consumption of the calculation module can be reduced, the cruising ability of the detector can be prolonged, and collaborative optimization of the task real-time requirement and the system energy consumption can be realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and embedded computing technology, and more specifically to a computing resource allocation method for scheduling mobile detection tasks in extreme environments, applicable to low-power heterogeneous multi-core AI parallel computing resource allocation. Background Technology

[0002] Extreme environment mobile exploration is one of the core missions in deep space exploration, and its technological breakthroughs are directly related to the depth of human understanding of the geological evolution, resource distribution, and space environment of extreme environments such as the Moon. With the deep application of artificial intelligence (AI) technology in image processing, autonomous navigation, and environmental perception, extreme environment probes are evolving from the traditional "ground-based remote control with autonomous decision-making as a supplement" mode to a "highly autonomous and intelligent exploration" mode. This transformation places unprecedented demands on the onboard computing power of the probes.

[0003] However, extreme environments, such as diurnal temperature variations of up to 300°C, strong cosmic radiation, and periodic interruptions in sunlight, create a sharp contradiction with the probe's limited energy payload, which primarily relies on solar panels and batteries. On one hand, high-priority tasks (such as real-time obstacle avoidance and terrain recognition) require millisecond-level response times, relying on the parallel processing capabilities of high-performance computing cores (such as GPUs). On the other hand, low-priority tasks (such as environmental data storage and equipment status monitoring), while having lower real-time requirements, require continuous computing resources, and improper scheduling can easily lead to energy waste. The computing module accounts for 40%-60% of the probe's total power consumption. Traditional static resource allocation strategies, unable to dynamically match task load and computing power, often result in the dual problems of "delayed response for high-priority tasks" and "energy waste for low-priority tasks," severely restricting the duration and execution efficiency of the probe mission.

[0004] With heterogeneous multi-core computing architectures (CPU+GPU+FPGA) becoming mainstream, existing scheduling strategies have significant limitations:

[0005] (1) Most strategies allocate tasks based on only one dimension of task priority or resource load, ignoring the collaborative optimization of "priority-energy efficiency ratio"; for example, allocating high-priority tasks to general-purpose CPU cores leads to response delays, or allowing low-priority tasks to occupy high-energy-consuming GPU cores causes energy waste.

[0006] (2) Resource adjustment relies heavily on real-time load feedback and lacks forward-looking prediction of future task requirements. In extreme environments and complex terrains, such as crater mountains and gravel areas, the obstacle avoidance task may surge by 3-5 times in a few seconds. Temporarily activating the hibernation core will generate instantaneous power consumption peaks, which are 2-3 times higher than steady-state power consumption, further aggravating energy consumption.

[0007] (3) Existing models are mostly based on ground simulation environment verification, and do not fully consider the dynamic impact of environmental factors such as extreme environment, such as light and temperature, on computational energy consumption, resulting in a significant reduction in energy efficiency optimization effect in actual applications. Summary of the Invention

[0008] In view of this, the present invention provides a computational resource allocation method for scheduling mobile detection missions in extreme environments, which can at least partially solve the above-mentioned technical problems and achieve coordinated optimization of mission real-time requirements and system energy consumption.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] In a first aspect, embodiments of the present invention provide a computational resource allocation method for scheduling mobile detection missions in extreme environments, comprising the following steps:

[0011] S1. Construct a task priority quantification model, and calculate the priority coefficient P of the task to be scheduled based on the task's time sensitivity, importance weight, and resource utilization.

[0012] S2. Construct a computing core energy efficiency evaluation model. Based on the basic energy efficiency of the computing core and the temperature attenuation coefficient and radiation attenuation coefficient under the current environment, output the actual energy efficiency ratio E of the computing core.

[0013] S3. Perform dynamic load balancing. Based on the priority coefficient P and the task-core matching degree M, monitor the load rate of each computing core and maintain system load balance. The task-core matching degree M is: the normalized coefficient E of the priority coefficient P and the actual energy efficiency ratio. 归一化 The absolute value of the difference;

[0014] S4. Based on the priority coefficient P and the actual energy efficiency ratio M, a two-dimensional scheduling model is constructed to allocate high-priority tasks to computing cores with high actual energy efficiency ratios and low-priority tasks to computing cores with low actual energy efficiency ratios.

[0015] S5. An energy consumption prediction module is introduced. Based on historical environmental data and task load data, the computing load within a set period in the future is predicted, and the working status of the computing core is adjusted in advance according to the prediction results.

[0016] Furthermore, in step S1, the task priority quantification model integrates three dimensions of indicators—time sensitivity, importance weight, and resource utilization rate—through a weighted summation method, and outputs a normalized priority coefficient P.

[0017]

[0018] Where T represents time sensitivity, characterizing the task's tolerance for delay, and is quantified using the reciprocal of the delay threshold; W represents importance weight, assigned according to the task's contribution to the detector's safety and scientific objectives; and R represents resource utilization, calculated as the ratio of the computing power required by the task to the detector's total computing power. These represent the weighting coefficients.

[0019] Further, in step S2, the calculation of the core energy efficiency evaluation model obtains the actual energy efficiency ratio by multiplying the basic energy efficiency ratio by the environmental attenuation coefficient; the environmental attenuation coefficient is the product of the temperature attenuation coefficient and the radiation attenuation coefficient, and the temperature attenuation coefficient and the radiation attenuation coefficient are obtained by fitting measured data from extreme environments.

[0020] Furthermore, in step S3, the dynamic load balancing adopts priority queue management of tasks, sets load migration trigger threshold, priority coefficient of the same type of core, migration overhead tolerance threshold, overload core recovery threshold, idle core screening threshold, and low priority task interruption threshold, and periodically detects and migrates tasks to balance the load.

[0021] Furthermore, in step S4, the dual-dimensional scheduling model selects the computing core with the highest matching degree for task allocation by calculating the matching degree between the task priority coefficient and the normalized energy efficiency ratio of the computing core; the computing core includes low-power GPUs, FPGAs and general-purpose CPUs.

[0022] Furthermore, in step S5, the energy consumption prediction module adopts an LSTM-based neural network model. The input features include historical terrain data, environmental physical parameters, and historical computing load data. The output is the predicted value of the total computing load for the next 10-30 seconds.

[0023] Furthermore, an attention mechanism is introduced into the LSTM model to focus on key environmental features; and an energy consumption constraint layer is added to the model output to ensure that the predicted load does not exceed the current energy supply capacity.

[0024] In a second aspect, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method as described in any one of the first aspects.

[0025] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the method as described in any one of the first aspects.

[0026] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages:

[0027] Breaking through the traditional single-dimensional resource scheduling framework, a closed-loop scheduling paradigm of "prediction-matching-optimization" is proposed, providing a new theoretical model for energy efficiency optimization of heterogeneous multi-core computing resources in extreme environments. It can reduce the total power consumption of computing modules, extend the endurance of probes, and support lunar rovers to complete more scientific exploration tasks during periods without sunlight, such as up to 14 Earth days, providing key technical support for the collaborative scheduling of equipment in the construction of future lunar bases. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0029] Figure 1 A graph showing the impact of extreme environmental factors on the core energy efficiency of computation.

[0030] Figure 2 Line chart comparing the core energy efficiency ratios of different types;

[0031] Figure 3 The present invention provides a flowchart of a computational resource allocation method for scheduling mobile detection missions in extreme environments;

[0032] Figure 4 This is a schematic diagram of the LSTM prediction model structure provided by the present invention;

[0033] Figure 5 A curve showing the prediction accuracy of the LSTM prediction model provided by this invention compared with two other sets of models.

[0034] Figure 6 A curve comparing the error rates of the LSTM prediction model provided in this invention with those of two other models.

[0035] Figure 7 A bar chart comparing the total power consumption of different algorithms provided by this invention;

[0036] Figure 8 This is a comparison chart of task completion rates under different algorithms provided by this invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] The efficient execution of mobile reconnaissance missions in extreme environments relies on a deep understanding of computing hardware architecture, the characteristics of extreme environments, and mission attributes. Starting with the composition and characteristics of heterogeneous multi-core computing architectures, this paper systematically analyzes the constraints of extreme environments on computing resource scheduling. Based on task priority and computing requirements, it classifies extreme environment reconnaissance missions into different types, laying the technical foundation for the design of subsequent resource allocation strategies.

[0039] (1) Heterogeneous multi-core computing architecture:

[0040] Heterogeneous multi-core computing architecture is the hardware core for extreme environment probes to achieve efficient data processing and autonomous decision-making. It balances task processing efficiency and energy consumption by integrating computing cores of different types and performance levels. Compared to traditional homogeneous multi-core architectures, heterogeneous architectures are better suited to the diversity and dynamism of extreme environment probe missions. Typical components include the following core modules:

[0041] High-performance computing cores (GPUs / FPGAs): Primarily responsible for tasks requiring high parallelism and real-time performance. For example, low-power GPUs, such as the NVIDIA Jetson TX2, with a processing power of 3.2 trillion operations per second (TOPS) and an energy efficiency ratio of 3.2 TOPS / W, can quickly complete AI inference tasks such as terrain image segmentation and obstacle recognition. FPGAs, through hardware logic reconfiguration, support the acceleration of specific algorithms, such as LiDAR data filtering, achieving an energy efficiency ratio of up to 2.5 TOPS / W. Furthermore, their radiation resistance is superior to general-purpose processors, making them suitable for extreme environments with high radiation levels.

[0042] General-purpose computing CPUs: Represented by low-power ARM architecture processors, such as the Texas Instruments AM5728, these cores primarily handle serial tasks such as task scheduling, device control, and data storage. The energy efficiency ratio of these cores is typically 1.0-1.5 TOPS / W. Although their computing power is lower than that of GPUs, they feature rich instruction sets and strong compatibility, enabling them to efficiently handle low-priority periodic tasks, such as device status monitoring and environmental data backup.

[0043] Storage and communication modules include L2 / L3 cache, DRAM, and NAND flash memory, used for temporary caching and persistent storage of mission data. The storage system of extreme environment detectors must meet the requirements of radiation-hardened single-event upset fault tolerance and low power consumption. Radiation-hardened DRAM, such as Micron 2GB DDR4, is typically used, with data read / write latency controlled within 50ns to ensure rapid data exchange.

[0044] Energy Efficiency Management Unit: This unit controls the power consumption of the computing cores through Dynamic Voltage and Frequency Scaling (DVFS) technology. It can adjust the core operating frequency in real time according to the workload, such as dynamically switching the GPU frequency from 500MHz to 1.5GHz. Under light loads, the frequency is reduced to decrease power consumption, while under heavy loads, the frequency is increased to meet real-time requirements. For example, when the detector enters sleep mode, the CPU frequency can be reduced to 200MHz, decreasing the single-core power consumption from 1.2W to 0.3W.

[0045] The core advantage of heterogeneous multi-core architecture lies in "on-demand allocation of computing power": by allocating high-priority tasks to high-performance cores and low-priority tasks to general-purpose cores, the energy consumption per unit of computing power can be minimized while meeting the real-time requirements of tasks. However, this architecture also places higher demands on resource scheduling strategies—it needs to accurately match task characteristics with core performance to avoid wasting computing power or causing response delays due to improper scheduling.

[0046] (2) Characteristics of extreme environments and their impact on computational scheduling:

[0047] The extreme nature and unique characteristics of extreme environments are key constraints on computing resource scheduling, significantly impacting detector hardware stability, energy consumption control, and mission response speed. Key features include:

[0048] Extreme temperature fluctuations: In extreme environments, the diurnal temperature range can reach 300°C, with daytime highs reaching 127°C and nighttime lows dropping to -173°C. High temperatures increase the leakage rate of computing cores; for example, GPU power consumption at 127°C increases by 30%-50% compared to 25°C. Low temperatures, on the other hand, reduce the refresh rate of DRAM memory cells, potentially leading to data errors. This necessitates scheduling strategies that dynamically adjust core load based on temperature changes, such as limiting the continuous operation time of high-performance cores during high-temperature periods to prevent hardware failures caused by overheating.

[0049] Strong cosmic radiation: The extreme environment lacks atmospheric protection, and the detector is constantly exposed to galactic cosmic rays and high-energy solar particles, which can easily trigger single-event upsets (SEUs) and single-event locks (SELs) in the computing core. NASA estimates that the SEU rate of electronic equipment in extreme environments can reach 10⁻⁻⁶. 6This means that an unhardened GPU with 1GB of memory might experience approximately 8 errors per day. Therefore, scheduling strategies need to include fault tolerance mechanisms, such as allocating critical tasks like obstacle avoidance decisions to multi-core parallel processing and reducing the risk of errors caused by radiation through result comparison.

[0050] Reference Figure 1 The figure shows the key data visualization results of the "Computing Core Energy Efficiency Evaluation Model". The core is used to demonstrate the dynamic impact of temperature on the actual energy efficiency ratio (unit: TOPS / W) of the computing core in extreme environments, such as the moon, where the day-night temperature difference reaches 300℃. The horizontal axis covers a typical temperature range from -173℃ (extreme low temperature during the lunar night) to 127℃ (extreme high temperature during the lunar day), and the vertical axis represents the effective computing power of the core. The figure includes two energy efficiency curves for low-power GPUs and general-purpose CPUs. The data are all based on actual measurements and fitting in extreme environments. For example, the GPU has a basic energy efficiency ratio of 3.2 TOPS / W in a standard environment (25℃), but at a high temperature of 127℃, the temperature decay coefficient drops to 0.6, and the actual energy efficiency ratio is only 1.92 TOPS / W (a 40% decrease compared to the standard environment). The CPU has a basic energy efficiency ratio of 1.2 TOPS / W at 25℃ and an actual energy efficiency ratio of 0.8 TOPS / W at 127℃. This figure provides a quantitative basis for "directly allocating high-priority tasks to cores with high actual energy efficiency ratios" in the subsequent "dual-dimensional scheduling model".

[0051] Periodic interruptions in sunlight resources: The Moon's rotation period is 27.3 Earth days, and the probe experiences approximately 14 days of continuous sunlight followed by 14 days of complete darkness in this extreme environment. During the day, it relies on solar panels for power, consuming approximately 100-300W, while at night, it requires batteries with a capacity of approximately 500-1000Wh to maintain critical equipment operation. This energy constraint necessitates an energy-aware scheduling strategy: during the day, more high-performance cores can be activated to handle non-real-time tasks, such as terrain mapping; at night, redundant cores must be shut down, retaining only low-power CPUs for essential equipment monitoring to ensure endurance.

[0052] Reference Figure 2The image shows the core comparative visualization results of the "Computing Core Energy Efficiency Evaluation Model," used to intuitively present the differences in actual energy efficiency ratios of three types of computing cores—low-power GPUs, FPGAs, and general-purpose CPUs—under different temperatures in extreme environments. The horizontal axis covers the typical temperature range of extreme environments: -173℃, corresponding to the extreme low temperature of the lunar night; 25℃, standard environment; 127℃, the extreme high temperature of the lunar day. The vertical axis represents the actual energy efficiency ratio, in TOPS / W, reflecting the effective computing power per unit of power consumption. The data for the three curves in the figure are all based on the baseline energy efficiency ratios specified in the documentation (low-power GPU = 3.2 TOPS / W, FPGA = 2.5 TOPS / W, general-purpose CPU = 1.2 TOPS / W). The power efficiency ratio (PS / W) and temperature decay coefficient (0.7 at -173℃, 1.0 at 25℃, and 0.6 at 127℃) are calculated according to the formula "actual energy efficiency ratio E = base energy efficiency ratio E0 × temperature decay coefficient KT". The core characteristic is that the energy efficiency ratios of all three reach their peak values ​​under the standard environment of 25℃. Different degrees of decay occur at high temperatures of 127℃ and low temperatures of -173℃. Moreover, the energy efficiency ratio of low-power GPUs is always the highest, and that of general-purpose CPUs is the lowest. This figure provides a direct quantitative comparison basis for the subsequent "dual-dimensional scheduling model" of "directly allocating high-priority tasks to cores with high actual energy efficiency ratios, such as low-power GPUs and FPGAs, and allocating low-priority tasks to cores with low actual energy efficiency ratios, such as general-purpose CPUs".

[0053] Complex terrain distribution: Extreme environments feature diverse terrains such as craters, lunar maria, lunar highlands, and scree areas. The probe will frequently encounter terrain changes during its movement, such as a slope increasing from 0° to 30°. Terrain complexity is significantly positively correlated with task load—for example, in scree areas, obstacle avoidance tasks can be triggered 5-10 times more frequently than in flat areas, and the computational load can surge from 10% to 80% within seconds. This necessitates a scheduling strategy that can quickly respond to terrain changes, dynamically adjust the allocation of core resources, and prevent task congestion caused by load spikes.

[0054] (3) Classification of mobile detection missions in extreme environments:

[0055] The diversity of mobile reconnaissance missions in extreme environments determines their differentiated requirements for computing resources. Based on mission priority, real-time requirements, computational complexity, and energy sensitivity, they can be divided into the following four categories, and the core characteristics of each category are shown in Table 1:

[0056] Table 1

[0057]

[0058] High-priority real-time tasks: These tasks are directly related to the survival and safety of the probe and must be completed within strict time constraints. They mainly include:

[0059] Obstacle avoidance decision: Identify obstacles ahead, such as rocks with a diameter ≥ 0.5m, using visual sensors like stereo cameras, calculate detour paths, and require a response delay ≤ 50ms, otherwise it may cause the detector to collide;

[0060] Attitude adjustment: The vehicle's attitude is corrected in real time based on data from the gyroscope and accelerometer. For example, the center of gravity is adjusted when the slope is ≥15°. The calculation complexity is low, but the triggering frequency is high, about 100Hz.

[0061] These types of tasks need to be allocated to high-performance cores with an energy efficiency ratio of ≥3.2 TOPS / W, such as low-power GPUs, and 10%-20% of redundant computing power needs to be reserved to cope with sudden loads.

[0062] Medium-priority near-real-time tasks: Supporting the scientific exploration and autonomous navigation of the probe, allowing for some delay but ensuring continuous processing, including:

[0063] Topographic mapping: A 3D terrain model is generated by scanning with LiDAR. The data size of a single frame is about 10-50MB. The processing latency must be ≤1s, otherwise it will cause map stitching misalignment.

[0064] Path planning: Calculates the optimal movement path based on terrain data, such as the shortest distance or lowest energy consumption path. The computational complexity increases exponentially with the expansion of the detection range; for example, path planning for a 1km area requires 10... 6 This operation is performed once.

[0065] These tasks can be assigned to GPUs or FPGAs, consuming computing power when high-priority tasks are idle, with a latency tolerance of 500ms-1s.

[0066] Low-priority non-real-time tasks: mainly data management and equipment maintenance, with low real-time requirements but requiring continuous resource consumption, including:

[0067] Environmental data backup: Data such as temperature and radiation intensity are stored in flash memory. Each record is about 1KB, and 1-10GB of data is generated per day. Data can be processed in batches at night or during low-load periods.

[0068] Equipment status monitoring: periodic, such as collecting parameters like battery voltage and core temperature every 10 seconds, with low computational load, approximately 10³ operations per cycle, suitable for allocation to general-purpose CPU cores.

[0069] The latency tolerance for this type of task is ≥500ms, and it can be flexibly scheduled without affecting high-priority tasks.

[0070] Emergency tasks: Temporary tasks triggered by unusual events, with dynamically changing priorities, for example:

[0071] Equipment fault diagnosis: When a sensor, such as a camera, malfunctions, a self-test program must be initiated immediately, which takes approximately 10 seconds.5 Each operation requires a response latency of ≤1 second.

[0072] Ground command response: Receive emergency commands from the Earth control center, such as emergency hibernation, and complete command parsing and execution within 30 seconds.

[0073] Unexpected tasks require the use of reserved redundant computing power. If resources are insufficient, low-priority tasks must be interrupted to ensure timely response.

[0074] As can be seen from the above classification, the "priority-computational requirements-energy efficiency sensitivity" of extreme environment exploration missions exhibit significant differences, which provides a classification basis for the design of the "mission-resource" two-dimensional matching model of this invention.

[0075] This invention discloses a computational resource allocation method for scheduling mobile detection missions in extreme environments, referring to... Figure 3 As shown, it includes the following steps:

[0076] S1. Construct a task priority quantification model, and calculate the priority coefficient P of the task to be scheduled based on the task's time sensitivity, importance weight, and resource utilization.

[0077] S2. Construct a computing core energy efficiency evaluation model. Based on the basic energy efficiency of the computing core and the temperature attenuation coefficient and radiation attenuation coefficient under the current environment, output the actual energy efficiency ratio E of the computing core.

[0078] S3. Perform dynamic load balancing. Based on the priority coefficient P and the task-core matching degree M, monitor the load rate of each computing core and maintain system load balance. The task-core matching degree M is: the normalized coefficient E of the priority coefficient P and the actual energy efficiency ratio. 归一化 The absolute value of the difference;

[0079] S4. Based on the priority coefficient P and the actual energy efficiency ratio M, a two-dimensional scheduling model is constructed to allocate high-priority tasks to computing cores with high actual energy efficiency ratios and low-priority tasks to computing cores with low actual energy efficiency ratios.

[0080] S5. An energy consumption prediction module is introduced. Based on historical environmental data and task load data, the computing load within a set period in the future is predicted, and the working status of the computing core is adjusted in advance according to the prediction results.

[0081] This invention provides a dual-dimensional scheduling model based on "task priority and computing resource energy efficiency." By quantifying the real-time requirements of tasks and the energy efficiency ratio of computing cores, it achieves targeted matching of high-priority tasks with high-performance cores, while allocating low-priority tasks to general-purpose CPU cores. Combined with a dynamic load balancing algorithm, it controls the multi-core idle rate to a low threshold, reducing the overload probability to, for example, below 3%, thus reducing energy waste caused by resource mismatch at the source. Secondly, it introduces an extreme environment energy consumption prediction module. Based on historical terrain data and the temporal characteristics of light intensity, it constructs a prediction model using an LSTM neural network to accurately predict the computing load for the next 10-30 seconds. Based on the prediction results, it adjusts the number of cores and operating frequency in advance. For example, it wakes up dormant GPU cores 20 seconds before complex terrain appears, reducing peak startup power consumption by more than 40% and shortening the response time to less than 15ms, solving the energy consumption fluctuation problem of traditional "passive response" scheduling. Finally, this invention also improves the stability of the strategy in extreme environments by incorporating the impact of factors such as temperature and radiation on core energy efficiency into the scheduling decision.

[0082] The following is a detailed explanation of each of the above steps:

[0083] Step S1, Task Priority Quantification:

[0084] The priority of extreme environment exploration missions is influenced by multiple factors and needs to be quantified through multi-dimensional indicators to avoid subjective judgment bias. Based on the mission's impact on probe survival, real-time requirements, and computational resource needs, a three-dimensional quantitative model is constructed, which includes time sensitivity, importance weight, and resource utilization. The final output is a normalized priority coefficient, ranging from 0 to 1, with higher values ​​indicating higher priority.

[0085] Time sensitivity T: Characterizes the task's tolerance for latency, quantified as the reciprocal of the latency threshold. For high-priority real-time tasks such as obstacle avoidance decision-making, the latency threshold is ≤50ms, and the T value is calculated using the following formula:

[0086]

[0087] Importance weight W: Assigned based on the contribution of the mission to the safety and scientific objectives of the probe, determined using an expert scoring method combined with the Analytic Hierarchy Process (AHP). Specifically, tasks involving survival, such as obstacle avoidance decision-making and attitude adjustment, have W = 0.8~1.0; scientific exploration tasks, such as terrain mapping and path planning, have W = 0.5~0.7; and auxiliary tasks, such as equipment status monitoring and data backup, have W = 0.2~0.4.

[0088] Resource utilization rate R: Reflects the intensity of the task's demand for computing resources. It is calculated as the ratio of the task's required computing power (TOPS) to the detector's total computing power, using the following formula:

[0089]

[0090] Priority coefficient P: Calculated using a weighted summation formula, taking into account the above three-dimensional indicators. Each parameter in the formula corresponds to the influence weight of time sensitivity, importance weight, and resource utilization rate. .in, It is the weighting coefficient.

[0091] This quantification process can transform vague task priorities into calculable numerical indicators, providing an objective basis for subsequent resource matching.

[0092] Step S2: Calculate the core energy efficiency assessment;

[0093] The energy efficiency assessment of the computing core needs to combine the characteristics of extreme environments with the attributes of the core hardware to construct a two-dimensional assessment model of "basic energy efficiency - environmental degradation", and finally output the actual energy efficiency ratio (TOPS / W) of the core as a key basis for task allocation.

[0094] Base efficiency ratio E0: refers to the theoretical energy efficiency of the core under standard environmental conditions of 25°C and no radiation, determined by hardware parameters.

[0095] Low-power GPUs, such as the Jetson TX2: E0 = 3.2 TOPS / W;

[0096] FPGA, such as Xilinx Artix-7: E0 = 2.5 TOPS / W;

[0097] General purpose CPUs, such as AM5728: E0 = 1.2 TOPS / W.

[0098] Environmental attenuation coefficient K: Considering the impact of extreme environmental temperature and radiation on energy efficiency, it is obtained by fitting experimental data and is the product of the temperature attenuation coefficient and the radiation attenuation coefficient, as shown in the following formula:

[0099]

[0100] Among them, the temperature decay coefficient K T Modeling based on measured data: When the temperature rises from 25℃ to 127℃, the GPU's K... T A drop from 1.0 to 0.6 means a 40% decrease in energy efficiency; when the temperature drops from 25°C to -173°C, the CPU's K... T The radiation attenuation coefficient K decreases from 1.0 to 0.7. R Related to radiation dose, under extreme environmental daily average radiation dose, the K of the FPGA R =0.9, the best radiation resistance, the GPU's K R =0.8, CPU K R =0.75.

[0101] Actual energy efficiency ratio E: The core's effective energy efficiency under the current environment, which is the product of the base energy efficiency ratio and the environmental degradation coefficient, as shown in the following formula: .

[0102] For example, at a high temperature of 127℃, the actual energy efficiency of the GPU is E=3.2×0.6=1.92TOPS / W; at a low temperature of -173℃, the actual energy efficiency of the CPU is E=1.2×0.7=0.84TOPS / W.

[0103] To dynamically track changes in core energy efficiency, an energy efficiency monitoring module was designed to collect core temperature and radiation dose data every 100ms and update the E value in real time. Simultaneously, a core energy efficiency level classification standard was established.

[0104] High-performance core: E≥2.0TOPS / W, such as GPUs and FPGAs at room temperature;

[0105] Medium-performance cores: 1.0≤E<2.0TOPS / W, such as GPUs at high temperatures and CPUs at normal temperatures;

[0106] Low-performance cores: E < 1.0 TOPS / W, such as CPUs operating at low temperatures.

[0107] This evaluation model ensures that the impact of the environment on core performance is fully considered when allocating tasks, avoiding scheduling errors caused by deviations between theoretical and actual values.

[0108] Step S3, Dynamic Load Balancing Algorithm Design:

[0109] Dynamic load balancing algorithms are the core of achieving a two-dimensional matching of "tasks and resources." Their goal is to control the load difference between cores within a threshold range of ≤10%, while ensuring matching between task priority and core efficiency, thus avoiding idle or overloaded states. The algorithm process includes three stages: task queuing, core selection, and load adjustment.

[0110] As shown in Table 2, load migration trigger threshold, priority coefficient of the same type of core, migration overhead tolerance threshold, overload core recovery threshold, idle core filtering threshold, low priority task interruption threshold, and periodicity were designed.

[0111] Table 2

[0112]

[0113] Task queuing mechanism: A priority queue is used to store tasks to be processed, and the queue is arranged in descending order of priority coefficient P. High-priority tasks, such as those with P ≥ 0.8, directly enter the "emergency scheduling area" and must be allocated resources within 5ms; medium- and low-priority tasks, such as those with P < 0.8, enter the "buffer scheduling area" and wait for allocation when the core is idle. At the same time, 15% of the buffer queue space is reserved for sudden tasks to ensure that they can interrupt low-priority tasks with P ≤ 0.3.

[0114] Core selection rules: Target cores are selected based on the matching degree M between task priority and core efficiency. The matching degree is calculated using the following formula: the smaller the M value, the higher the degree of matching between the task and the core.

[0115]

[0116] Among them, E 归一化 This involves standardizing the actual energy efficiency ratio (E) within a range of 0-1, where, for example, high-efficiency core...

[0117] E 归一化 =0.8~1.0, medium efficiency =0.4~0.7, low efficiency =0.1~0.3. For example, for high-priority tasks, P=0.9 and high-efficiency core E 归一化 =0.9, M=0, is the optimal match; if all high-performance cores are at full load, then the medium-performance core with the smallest M is selected for allocation.

[0118] Load balancing strategy: When the core load factor (the ratio of the current task's computing power to its maximum computing power) exceeds 80%, load migration is triggered.

[0119] Identify low-priority tasks on overloaded cores, such as those with P≤0.5, and calculate their migration overhead, which is the sum of data transfer time and reload time.

[0120] Filter idle cores with a load rate of <50%, and prioritize cores of the same type. For example, migrate GPU tasks to other GPUs to reduce adaptation overhead.

[0121] If the migration reduces the overloaded core load rate to ≤70% and the target core load rate to ≤75%, then perform the migration.

[0122] Simulation results show that the algorithm can reduce the standard deviation of multi-core load from 25% in static allocation to less than 8%, control the core idle rate to ≤5%, reduce the overload probability to ≤2%, and reduce the total power consumption by 22%-28% compared with the traditional round-robin scheduling algorithm.

[0123] Step S4, the overall process of the two-dimensional scheduling model:

[0124] The model's execution process is divided into three stages: initialization, task access, and dynamic scheduling.

[0125] Initialization: Read heterogeneous multi-core hardware parameters, E0 of each core, and maximum computing power; initialize the environmental monitoring module: temperature and radiation data acquisition; set load balancing thresholds: idle rate ≤5%, overload rate ≤2%.

[0126] Task access: Receive task requests, extract task parameters: latency threshold, computing power requirement, importance, calculate priority coefficient P, and store them in the corresponding priority queue.

[0127] Dynamic scheduling:

[0128] A scheduling cycle is triggered every 5ms to update the actual energy efficiency ratio E of each core, combined with environmental data;

[0129] High-priority tasks are extracted from the "emergency dispatch zone" and allocated to high-performance cores according to the principle of minimum M value; for the "buffer zone"...

[0130] "Scheduling area" tasks are allocated to cores with matching energy efficiency in descending order of P value when the cores are idle;

[0131] If core overload or idleness is detected, initiate the load balancing adjustment strategy in step S3 above to migrate low-priority tasks.

[0132] This model achieves precise matching between extreme environment detection tasks and computing resources through a closed-loop design of "priority quantification - energy efficiency assessment - dynamic equilibrium", laying the foundation for the subsequent introduction of an energy consumption prediction module.

[0133] Step S5: Implementation of the extreme environment energy consumption prediction module:

[0134] To achieve proactive resource scheduling for extreme environment exploration missions, this embodiment designs an extreme environment energy consumption prediction module. By collecting correlation data between extreme environments and computational load, a prediction model based on an LSTM (Long Short-Term Memory) network is constructed to accurately predict the computational load for the next 10-30 seconds. The core objective of this module is to provide advance notice for the "task-resource" dual-dimensional scheduling model, avoiding power consumption peaks caused by sudden load spikes. Its implementation process encompasses three stages: data acquisition and preprocessing, LSTM model construction, and model training and validation.

[0135] 5.1) Data Acquisition and Preprocessing:

[0136] The accuracy of energy consumption prediction in extreme environments depends on high-quality input data. This requires collecting extreme environment parameters, task load characteristics, and core energy consumption data, and preprocessing to eliminate noise and redundancy, providing reliable input for model training. Data collection dimensions and sources:

[0137] Environmental parameters include extreme ambient temperature (°C), solar irradiance (W / m²), terrain slope (°), and radiation dose (rad / h). Data are collected by a sensor array onboard the detector. The temperature sensor samples at 1 Hz; the irradiance sensor samples at 0.1 Hz due to the relatively slow change in light intensity; the terrain slope is measured in real time using lidar at a sampling frequency of 5 Hz; and the radiation dose is recorded every 10 seconds by a radiation detector.

[0138] Task load characteristics: including the trigger frequency (times / s) of each type of task, the TOPS (topics per second) of computing power required per task, and the task priority coefficient P. The data is output in real time by the task scheduling module, with a sampling frequency consistent with the scheduling cycle, such as 5ms / time, to ensure the capture of instantaneous load fluctuations.

[0139] Calculate core power consumption data: including real-time power consumption (W), operating frequency (MHz), and load rate (%) for each core. This data is collected by the energy efficiency management unit via a hardware interface at a sampling frequency of 10Hz, simultaneously recording the core type (GPU / CPU / FPGA).

[0140] The input data can be found in Table 3:

[0141] Table 3

[0142]

[0143] Data preprocessing steps:

[0144] Data alignment and interpolation: Due to differences in sampling frequencies among different sensors, all data needs to be standardized to a 1Hz time scale. Missing values, such as those caused by temporary sensor malfunctions, are filled using linear interpolation. For example, if light intensity data has not been updated within 10 seconds, the intermediate value is fitted using values ​​from previous and subsequent time points.

[0145] Outlier handling: Abnormal data is identified using the 3σ criterion, such as a sudden increase in radiation dose exceeding the normal range by 3 times. Extreme outliers are replaced with the median of adjacent time points to avoid noise affecting model training.

[0146] Feature normalization: Map environmental parameters and load features to the [0,1] interval, using the min-max normalization formula:

[0147]

[0148] Where x is the original data, x min x max These are the minimum and maximum values ​​of the feature, determined statistically based on historical detection data, such as a temperature range of -173℃ to 127℃.

[0149] Sequence Construction: The preprocessed data is used to construct input and target sequences according to time steps. The input sequence consists of the environmental and load features of the previous N time steps, for example, N=30, corresponding to 30 seconds of historical data; the target sequence is the total computational load TOPS for the next 10-30 seconds, forming datasets for three prediction scenarios: "historical 30s → future 10s", "historical 30s → future 20s", and "historical 30s → future 30s".

[0150] After preprocessing, the dataset contains over 100,000 samples, generated based on 1,000 hours of data from a simulated extreme environment exploration mission. The training set accounts for 70%, the validation set for 20%, and the test set for 10%.

[0151] 5.2) Construction of LSTM prediction model:

[0152] Extreme environments have strong temporal correlations in their computational loads. For example, increased terrain complexity can continuously trigger obstacle avoidance tasks. Traditional time series prediction models such as ARIMA struggle to capture long-term dependencies, while LSTM, through its gating mechanism, can effectively memorize historical features, making it suitable for prediction tasks in such scenarios.

[0153] Overall model structure:

[0154] Input layer: Receives feature vectors from 30 time steps. Each time step contains 12-dimensional features: 4-dimensional environmental parameters + 5-dimensional task load features + 3-dimensional core state features. The input dimension is (30, 12).

[0155] LSTM layer: Contains two stacked LSTM units. The first layer has 64 neurons, and the second layer has 32 neurons. Both use the ReLU activation function, and the output of the first layer is subjected to Dropout with a ratio of 0.2 to suppress overfitting. The LSTM unit controls the flow of information through forget gates, input gates, and output gates. For example, the forget gate decides to discard irrelevant terrain data from the historical sequence, while the input gate retains the impact of the current light intensity on energy consumption.

[0156] Fully connected layer: The feature vector output by the LSTM layer is flattened and mapped to the predicted value through two fully connected layers. The first layer has 16 neurons with ReLU activation; the second layer has 1 neuron with linear activation, and outputs the total computational load TOPS for the next 10-30 seconds.

[0157] The overall structure of LSTM is as follows Figure 4 As shown.

[0158] Key parameter design:

[0159] Time step: The length of the input sequence is set to 30 (corresponding to 30s). Based on experimental verification, a time step shorter than 20s will lose long-term terrain change information, a time step longer than 40s will introduce redundant noise, and 30s is the balance point between prediction accuracy and computational efficiency.

[0160] Prediction step size: Supports three outputs: 10s, 20s, and 30s, which are achieved by adjusting the time dimension of the output layer. The 10s prediction is used to quickly respond to load fluctuations, while the 30s prediction is used to plan core wake-up strategies in advance.

[0161] Optimizer and Loss Function: The Adam optimizer is used with a learning rate of 0.001, β1=0.9, and β2=0.999 to accelerate convergence. The mean squared error (MSE) loss function is selected, and the formula is as follows: .

[0162] in, For actual load, For load prediction, n is the number of samples. MSE can effectively penalize large prediction errors, meeting the need to avoid load peaks in extreme environments.

[0163] Model adaptation to extreme environmental scenarios:

[0164] For extreme environmental terrain change scenarios, such as suddenly entering a crater, an attention mechanism is added after the LSTM layer to enable the model to automatically focus on key features such as terrain slope and obstacle density, thereby improving the prediction sensitivity of sudden load.

[0165] Considering energy constraints in extreme environments, such as limited computing power at night, an energy consumption constraint layer is added to the model output to ensure that the predicted load does not exceed the current energy supply capacity. For example, when powered by batteries at night, the load limit is 50% of the daytime load to avoid invalid predictions.

[0166] 5.3) Model Training and Validation:

[0167] The model training adopts a "pre-training + fine-tuning" strategy. First, the basic model is trained based on simulated data, and then the parameters are fine-tuned by combining actual extreme environment detection data. Finally, the prediction performance is verified through multi-dimensional indicators.

[0168] Training process:

[0169] Pre-training: The model was trained using 80,000 simulated extreme environment data points, covering scenarios such as flat lunar maria, complex craters, and extreme temperatures. The training process involved 50 iterations, with the validation set MSE calculated at the end of each iteration. The training stopped when the MSE showed no decrease for five consecutive iterations, resulting in the base model. The simulated data was generated using an extreme environment simulator to ensure that parameters such as terrain and temperature conformed to the physical laws of real extreme environments.

[0170] Fine-tuning: Incorporate 2,000 actual exploration data from Chang'e-4 and Yutu-2, as well as publicly available environment and mission logs; freeze the first 50% of the parameters of the LSTM layer, and only fine-tune the parameters of the fully connected layer and attention mechanism, iterating 20 times to adapt the model to the characteristics of real extreme environments.

[0171] Performance verification metrics:

[0172] Forecast accuracy: Calculated by the relative error between the predicted load and the actual load, using the following formula:

[0173]

[0174] The test set results show that the prediction accuracy is ≥90% for 10s, ≥87% for 20s, and ≥85% for 30s, which meets the scheduling strategy's requirements for lead time.

[0175] Peak capture rate: The statistical model has the ability to predict peak loads that exceed twice the average load. The results show that the capture rate for peak scenarios such as a surge in obstacle avoidance tasks is ≥82%, which can effectively wake up the GPU cores in advance.

[0176] Computational overhead: The inference time of the model on the detector embedded platform Jetson TX2 is ≤50ms, which is much smaller than the prediction step size of 10s and will not add any additional latency.

[0177] Comparative experiment:

[0178] Compared with traditional time series forecasting models ARIMA and BP, as shown in Table 4, the proposed LSTM model outperforms in all metrics, especially in long-term forecasting of 30s and peak scenarios, demonstrating its adaptability to complex loads in extreme environments.

[0179] Table 4

[0180]

[0181] The prediction accuracy comparison curves of the LSTM prediction model provided by this invention with the other two sets of models are shown below. Figure 5 As shown, the error rate is compared to, for example Figure 6 As shown.

[0182] Model Deployment and Updates:

[0183] The model is deployed to the detector computing unit in a lightweight ONNX format. It undergoes five rounds of incremental training every 24 hours based on newly acquired environmental and load data, dynamically updating the model parameters to adapt to long-term changes in extreme environments, such as seasonal fluctuations in light intensity.

[0184] Through the above implementation, the extreme environment energy consumption prediction module can accurately predict the computing load in the next 10-30 seconds, providing a basis for decision-making on dynamically adjusting the number and frequency of multi-core parallel cores. In conjunction with the "task-resource" dual-dimensional scheduling model, it realizes the upgrade of energy consumption management from "passive response" to "active optimization".

[0185] To verify the comprehensive performance of the proposed dual-dimensional scheduling strategy of "task priority-computing resource energy efficiency" combined with the energy consumption prediction module, an extreme environment mobile detection simulation environment was constructed, and multiple sets of comparative experiments were designed. Quantitative analysis was carried out from the dimensions of power consumption control, task response speed, and load balancing capability, and a comparison was made with existing mainstream scheduling algorithms to clarify the advantages and applicable scenarios of the proposed strategy.

[0186] 1. Experimental environment and parameter settings:

[0187] 1.1 Simulation Platform Setup:

[0188] An experimental platform was built based on NASA's open-source extreme environment probing simulation toolkit and heterogeneous multi-core computing simulation framework. This platform has the following capabilities:

[0189] Extreme environment simulation: It can reproduce extreme environmental temperature fluctuations, from -173℃ to 127℃, radiation dose, daily average of 100mSv, and terrain complexity including features such as flat areas, gravel areas, and ring mountain areas. It supports light cycles, with 14-day day / 14-day night switching.

[0190] Computational hardware simulation: Simulates a heterogeneous multi-core architecture, including two low-power GPUs, such as Jetson TX2, with a basic energy efficiency of 3.2 TOPS / W; four general-purpose CPUs, such as AM5728, with a basic energy efficiency of 1.2 TOPS / W; and two FPGAs, such as Artix-7, with a basic energy efficiency of 2.5 TOPS / W, supporting dynamic voltage and frequency adjustment and core sleep / wake-up control;

[0191] Task generator: Randomly generates workloads according to the defined task types: high-priority real-time tasks, medium-priority near-real-time tasks, low-priority non-real-time tasks, and sudden tasks. The trigger frequency of obstacle avoidance decision tasks is positively correlated with the terrain complexity, such as 0.1Hz in flat areas, 1Hz in gravel areas, and 2Hz in ring-shaped mountainous areas.

[0192] 1.2 Experimental parameter configuration:

[0193] Experimental period: Simulates 1 lunar day (27.3 Earth days), with 14 days of daytime powered by solar energy (300W) and 14 days of nighttime powered by a battery (800Wh).

[0194] Comparison Algorithms: Seven mainstream scheduling strategies were selected as benchmarks, including:

[0195] Static Priority Scheduling (SPS): allocates cores according to a fixed priority without dynamic adjustment;

[0196] Round-Robin Scheduling (RRS): Tasks take turns using cores, regardless of priority.

[0197] Shortest Job First (SJF): Prioritizes tasks with lower computational complexity;

[0198] Load balancing scheduling (LBS): only pursues load balancing across cores;

[0199] Feedback Dynamic Scheduling (FDS): Adjusts resources based on real-time load feedback;

[0200] Energy-Efficiency First (EEF) scheduling: prioritizes allocation to low-power cores;

[0201] Two-Dimensional Scheduling without Prediction (TDS-WP): The version of the strategy in this invention after removing the energy consumption prediction module;

[0202] Performance metrics:

[0203] Total power consumption (Wh): The cumulative energy consumption of the computing module during the experimental period;

[0204] Task response latency (ms): The average time from task generation to completion, with a focus on the latency exceeding the standard rate of high-priority tasks, such as the proportion of tasks with a latency >50ms;

[0205] Core load balancing: The standard deviation of the load rate of each core; the smaller the deviation, the more balanced the load.

[0206] Battery life: The duration (in hours) during which the battery can operate in darkness.

[0207] 2. Experimental Results and Analysis

[0208] 2.1 Comparison of Power Consumption Control Performance

[0209] Figure 7 The total power consumption distribution of different algorithms during the experimental period is shown. The results indicate that:

[0210] The strategy proposed in this invention (including the prediction module) has the lowest total power consumption of 1250Wh, which is 27.3% lower than static priority scheduling (SPS, 1720Wh) and 15.5% lower than two-dimensional scheduling without prediction (TDS-WP, 1480Wh).

[0211] While Energy Efficiency First Scheduler (EEF) can reduce some power consumption (1580Wh), it leads to increased latency for high-priority tasks due to excessive use of low-performance cores, resulting in poor overall performance.

[0212] The power consumption of Load Balancing (LBS) and Feedback Dynamic Scheduling (FDS) is 1650Wh and 1590Wh, respectively. Due to the lack of coordinated optimization of energy efficiency and priority, their power consumption is higher than that of the strategy of this invention.

[0213] Figure 8 The results show the task completion rates of different algorithms within the experimental period. The results indicate that the strategy proposed in this invention (including the prediction module) has the highest task completion rate, reaching 96%.

[0214] 2.2 Comparison of Task Response Speed

[0215] Table 5 shows the average response latency and latency exceedance rate for high-priority tasks such as obstacle avoidance decision-making. The results show that the average latency of the strategy proposed in this invention is 32ms, with an exceedance rate of only 1.2%, which is superior to all compared algorithms. Specifically, the average latency of the obstacle avoidance task in the gravel area decreased from 48ms for TDS-WP to 29ms, because the prediction module wakes up the GPU core earlier, shortening the core startup time (from 15ms to 8ms).

[0216] The out-of-target rates for Static Priority Scheduling (SPS) and Round Robin Scheduling (RRS) reached 18.5% and 25.3%, respectively, due to the inability to dynamically match high-priority tasks with high-performance cores.

[0217] While Shortest Job First (SJF) can reduce latency for short tasks, it has an overshoot rate of 12.7% for high-priority long tasks such as path planning, making it unsuitable for extreme environment tasks.

[0218] Table 5

[0219]

[0220] 2.3 Load Balancing and Stability Analysis

[0221] The standard deviation of the strategy of this invention is 7.8%, which is significantly lower than SPS (22.3%) and RRS (18.5%), and slightly better than TDS-WP (9.5%), indicating that the dynamic load balancing algorithm effectively balances the load of each core.

[0222] In extreme temperature scenarios, such as daytime temperatures of 127°C, the strategy of this invention adjusts the core energy efficiency evaluation value to maintain the actual load rate of the GPU core at 60%-70%, thus avoiding overload problems at high temperatures. Compared with the FDS algorithm, which has an overload rate of 8.3%, the strategy of this invention has an overload rate of only 2.1%.

[0223] 2.4 Impact of Terrain Complexity on Performance

[0224] To verify the adaptability of the strategy in complex terrain, three scenarios—flat area, scree area, and ring mountain area—were selected for comparison. The results show:

[0225] In the ring-shaped mountainous area, the obstacle avoidance task frequency is 2Hz. The total power consumption of the strategy of this invention is 420Wh, which is 17.6% lower than that of TDS-WP (510Wh). This is because the prediction module increases the number of GPU cores 20s in advance, avoiding power fluctuations caused by load surges.

[0226] The delay rate of high-priority tasks in the ring-shaped mountainous area is 2.3%, which is much lower than the EEF 15.7% and LBS 11.2%, demonstrating the advantage of the "priority-energy efficiency" dual-dimensional matching.

[0227] 3. Experimental conclusions:

[0228] The scheduling strategy proposed in this invention, through the collaborative design of "dual-dimensional matching + energy consumption prediction", reduces total power consumption by 22%-28% while keeping the latency exceedance rate of high-priority tasks within 1.5%, and its overall performance is superior to existing mainstream algorithms.

[0229] The energy consumption prediction module is particularly effective in complex terrain and energy-constrained scenarios such as nighttime and mountainous areas, reducing peak power consumption by more than 40% and extending battery life by 18%-25%.

[0230] The dynamic load balancing algorithm effectively balances the multi-core load, ensuring that the load standard deviation is ≤8%, thus improving system stability under extreme conditions.

[0231] The above results verify the feasibility and superiority of the proposed strategy in mobile detection missions in extreme environments, and provide a quantitative reference for the scheduling of computational resources for actual detectors.

[0232] This invention addresses the low-power consumption requirements of computing resource scheduling in mobile detection missions in extreme environments. It proposes a heterogeneous multi-core AI parallel computing resource allocation strategy that integrates a two-dimensional matching of "task priority and computing resource energy efficiency" with energy consumption prediction. Through theoretical modeling, algorithm design, and simulation experiments, the following core achievements were obtained: At the theoretical level, a two-dimensional scheduling model of tasks and resources was constructed. By quantifying task priorities and the actual energy efficiency of computing cores, targeted matching of high-priority tasks with high-efficiency cores was achieved, solving the problem of "difficulty in balancing energy consumption and real-time performance" in traditional single-dimensional scheduling. An extreme environment energy consumption prediction module was introduced, using an LSTM neural network to predict the computing load for the next 10-30 seconds (accuracy ≥85%), filling the application gap of predictive resource allocation in the field of extreme environment detection. At the algorithm level, a dynamic load balancing algorithm was designed. Through a task priority queue and core matching degree screening mechanism, the standard deviation of multi-core load was controlled within 8%. Simultaneously, the core state was adjusted in advance based on the prediction results, reducing the peak power consumption of temporary startups by more than 40%. Simulation results show that this strategy reduces total power consumption by 27.3% compared to static priority scheduling, controls the latency exceedance rate of high-priority tasks to 1.2%, and extends battery life by 25% during nighttime. Its overall performance is superior to existing mainstream algorithms.

[0233] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a computational resource allocation method for scheduling mobile detection missions in extreme environments as described in any of the above embodiments.

[0234] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the computational resource allocation method for scheduling mobile exploration missions in extreme environments as described in any of the above embodiments.

[0235] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0236] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for computing resource allocation for mobile exploration task scheduling in extreme environment, characterized in that, The method comprises the following steps: S1, constructing a task priority quantification model, calculating the priority coefficient P of the to-be-scheduled task based on the time sensitivity, importance weight and resource occupancy rate of the task; S2, constructing a computing core energy efficiency evaluation model, outputting the actual energy efficiency ratio E of the computing core based on the basic energy efficiency of the computing core and the temperature attenuation coefficient and radiation attenuation coefficient under the current environment; S3, performing dynamic load balancing, monitoring the load rate of each computing core according to the priority coefficient P and the task-core matching degree M, and maintaining system load balancing; the task-core matching degree M is a normalized coefficient E of the priority coefficient P and the actual energy efficiency ratio 归一化 the absolute value of the difference S4, based on the priority coefficient P and the actual energy efficiency ratio M, constructing a two-dimensional scheduling model, directing the allocation of high-priority tasks to computing cores with high actual energy efficiency ratio, and allocating low-priority tasks to computing cores with low actual energy efficiency ratio; S5, introducing an energy consumption prediction module, predicting the computing load in a future set period based on historical environmental data and task load data, and adjusting the working state of the computing core in advance according to the prediction result.

2. The method of claim 1, wherein, In step S1, the task priority quantification model fuses the three-dimensional indexes of time sensitivity, importance weight and resource occupancy rate by weighted summation, and outputs the normalized priority coefficient P; ； Wherein, T represents time sensitivity, representing the tolerance degree of the task to delay, and the reciprocal of the delay threshold is quantified; W represents the importance weight, which is valued according to the contribution of the task to the safety of the probe and the scientific goal; R represents the resource occupancy rate, which is calculated by the ratio of the required computing power of the task to the total computing power of the probe; respectively represent weight coefficients.

3. The method of claim 1, wherein, In step S2, the computing core energy efficiency evaluation model obtains the actual energy efficiency ratio by multiplying the basic energy efficiency ratio by the environmental attenuation coefficient; the environmental attenuation coefficient is the product of the temperature attenuation coefficient and the radiation attenuation coefficient, and the temperature attenuation coefficient and the radiation attenuation coefficient are obtained based on extreme environment measurement data.

4. The method of claim 1, wherein, In step S3, the dynamic load balancing manages tasks by priority queue, sets load migration trigger threshold, same type core priority coefficient, migration overhead tolerance threshold, overload core recovery threshold, idle core screening threshold, and low-priority task interruption threshold, and periodically detects and migrates tasks to balance the load.

5. The method of claim 1, wherein, In step S4, the two-dimensional scheduling model selects the computing core with the highest matching degree of the task priority coefficient and the normalized energy efficiency ratio of the computing core for task allocation; the computing core includes low-power GPU, FPGA and general-purpose CPU.

6. The method of claim 1, wherein, In step S5, the energy consumption prediction module adopts an LSTM-based neural network model, the input features include historical terrain data, environmental physical parameters and computing load history data, and the output is the total computing load prediction value in the future 10-30 seconds.

7. The method of claim 6, wherein, The LSTM model is followed by an attention mechanism for focusing on key environmental features, and an energy consumption constraint layer is added at the output end of the model to ensure that the predicted load does not exceed the current energy supply capacity.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the method of any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-7.

Images