A hydrogen energy unmanned aerial vehicle intelligent management method and system, an electronic device, and a storage medium

Abstract

The application discloses a hydrogen energy unmanned aerial vehicle intelligent management method and system, electronic equipment and a storage medium, aiming at solving the problems of inaccurate endurance prediction and rigid task planning. The method comprises the following steps: acquiring multi-source operation data in the flight process of the hydrogen energy unmanned aerial vehicle in real time; inputting the multi-source operation data into a pre-trained time series deep learning model to output a remaining flight time estimation value at the current moment and an uncertainty interval based on quantile regression; when a risk-adjusted endurance value determined based on the uncertainty interval indicates that the remaining endurance is insufficient to complete the remaining tasks in the original plan, determining an adjusted task waypoint sequence; based on the adjusted task waypoint sequence, and according to a safety path planning strategy comprising at least task adjustment, direct return and emergency standby landing level, an adjusted flight path for the hydrogen energy unmanned aerial vehicle to execute is generated. The application can improve the task execution efficiency and flight safety of the hydrogen energy unmanned aerial vehicle.

Description

Technical Field

[0001] This invention relates to the field of intelligent management technology for unmanned aerial vehicles (UAVs), specifically to an intelligent management method, system, electronic device, and storage medium for hydrogen-powered UAVs. Background Technology

[0002] Hydrogen-powered drones, powered by hydrogen fuel cells, have shown great potential in industrial applications such as power line inspection, logistics and transportation, and agricultural plant protection due to their advantages such as long flight time, high payload capacity and good low temperature adaptability.

[0003] However, predicting the remaining flight time of hydrogen-powered drones is a recognized technical challenge compared to the mature technology of lithium-ion battery drones. The remaining capacity of lithium-ion batteries can be estimated using a relatively simple voltage-capacity mapping relationship. But the flight range of hydrogen fuel cells is affected by a complex nonlinear coupling of various factors, such as: the pressure and temperature inside the hydrogen tank jointly determine the amount of usable hydrogen; the efficiency of the fuel cell stack dynamically decreases with operating time and conditions; external conditions such as ambient temperature, air pressure, and wind speed significantly affect stack performance and flight aerodynamic power consumption; and the power requirements of drones vary greatly under different flight conditions such as hovering, cruising, and climbing.

[0004] In existing technologies, the endurance estimation algorithm for lithium battery drones is often simply ported to hydrogen fuel cell drones. However, due to the fundamental difference in the underlying power principles, this approach typically results in prediction errors exceeding 15%. This leads to two extreme adverse consequences: First, to ensure safety, excessive endurance redundancy is left during mission planning, causing drones to frequently return to base prematurely, severely wasting their long endurance advantage and single-flight operational capabilities. Second, overly optimistic endurance estimations can cause drones to run out of fuel mid-mission and be unable to return, resulting in equipment crashes, cargo losses, or even secondary safety incidents.

[0005] Furthermore, when a drone discovers during flight that its remaining range is insufficient to complete its original mission plan, current technologies generally lack intelligent and automated mission replanning capabilities. Typically, this relies on ground station operators making manual judgments and adjusting flight paths, a method that is not only inefficient but also highly susceptible to decision-making errors due to operator stress or misjudgment in emergency situations. Therefore, there is an urgent need for a technological solution capable of accurately predicting remaining range and performing automated dynamic mission replanning. Summary of the Invention

[0006] To address the technical problems of inaccurate endurance prediction, rigid mission planning, and insufficient safety redundancy in hydrogen-powered drones mentioned above, this invention provides an intelligent management method, system, electronic device, and storage medium for hydrogen-powered drones.

[0007] This invention provides an intelligent management method for hydrogen-powered unmanned aerial vehicles (UAVs), comprising: acquiring multi-source operational data during the flight of the hydrogen-powered UAV in real time, wherein the multi-source operational data includes at least fuel status, fuel cell status, environmental status, and flight status; inputting the multi-source operational data into a pre-trained temporal deep learning model combining a temporal convolutional network and a multi-head attention mechanism, and outputting an estimate of the remaining flight time at the current moment and an uncertainty interval based on quantile regression; when the risk-adjusted endurance value determined based on the uncertainty interval indicates that the remaining endurance is insufficient to complete the remaining tasks originally planned, solving an optimization problem with maximizing the weighted value of the tasks to be completed as the optimization objective and constrained by the risk-adjusted endurance value, to determine an adjusted mission waypoint sequence; and generating an adjusted flight path for the hydrogen-powered UAV to execute based on the adjusted mission waypoint sequence and according to a safety path planning strategy that includes at least mission adjustment, direct return, and emergency diversion levels.

[0008] Furthermore, before inputting the multi-source operating data into the time-series deep learning model, the method further includes: constructing derived features based on the multi-source operating data, wherein the derived features include at least one of hydrogen consumption rate, stack efficiency decay rate, or power fluctuation index.

[0009] Furthermore, the safe path planning strategy includes: if the risk-adjusted endurance is sufficient to complete some waypoints, the adjusted flight path is generated at the mission adjustment level; if the risk-adjusted endurance is insufficient to complete any operational waypoints but sufficient for a return trip, the adjusted flight path is generated at the direct return trip level, and the adjusted flight path is a path directly to the return point; if the risk-adjusted endurance is insufficient for a return trip, the adjusted flight path is generated at the emergency diversion level, and the adjusted flight path is a path to the nearest safe diversion point.

[0010] Furthermore, the method also includes: when the deviation between the actual hydrogen consumption rate and the model prediction value exceeds a preset threshold, updating the parameters of the time-series deep learning model online using recent actual flight data.

[0011] Furthermore, the method also includes: visually presenting the adjusted flight path on a human-computer interaction interface.

[0012] Furthermore, the fuel status includes hydrogen cylinder pressure and temperature; the fuel cell stack status includes fuel cell stack voltage, current, and power; the environmental status includes ambient temperature, air pressure, and wind speed; and the flight status includes a sequence of historical flight conditions.

[0013] Furthermore, the temporal convolutional network is used to capture the dependence of hydrogen consumption patterns at different time scales; the multi-head attention mechanism is used to adaptively learn the dynamic weights of the impact of different features in the multi-source operational data on endurance at different flight stages.

[0014] This invention also provides a hydrogen-powered unmanned aerial vehicle (UAV) intelligent management system, comprising: an acquisition module for acquiring multi-source operational data during the flight of the hydrogen-powered UAV in real time, wherein the multi-source operational data includes at least fuel status, fuel cell stack status, environmental status, and flight status; a prediction module for inputting the multi-source operational data into a pre-trained temporal deep learning model combining a temporal convolutional network and a multi-head attention mechanism, and outputting an estimate of the remaining flight time at the current moment and an uncertainty interval based on quantile regression; a task optimization module for solving an optimization problem with maximizing the weighted value of the tasks to be completed as the optimization objective and constrained by the risk-adjusted flight time value when the risk-adjusted flight time value determined based on the uncertainty interval indicates that the remaining flight time is insufficient to complete the remaining tasks originally planned, thereby determining an adjusted task waypoint sequence; and a path generation module for generating an adjusted flight path for the hydrogen-powered UAV to execute based on the adjusted task waypoint sequence and according to a safe path planning strategy that includes at least task adjustment, direct return, and emergency diversion levels.

[0015] The present invention also provides an electronic device, including a processor and a memory, wherein the memory stores computer instructions, which, when executed by the processor, implement any of the methods described above.

[0016] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the methods described above.

[0017] Compared with existing technologies, this invention has the following advantages by constructing a closed-loop management process from data perception and intelligent prediction to decision optimization: 1. High prediction accuracy and strong operational reliability. The temporal deep learning model combined with temporal convolutional networks and multi-head attention mechanism adopted in this invention is specifically designed for the complex and nonlinear endurance characteristics of hydrogen-powered drones. It can reduce the endurance prediction error from more than 15% in existing technologies to less than 5%, significantly improving the accuracy and reliability of endurance prediction.

[0018] 2. Scientifically based decision-making and highly safe. This invention not only provides a point estimate of flight time but also outputs the uncertainty range, enabling the system to make decisions based on risk-adjusted flight time values. Combined with a three-tiered safe path planning strategy that includes mission adjustment, direct return to base, and emergency diversion, a robust multi-layered safety defense is constructed, fundamentally avoiding the risk of crashes due to flight exhaustion.

[0019] 3. Maximizes mission efficiency and improves economic benefits. When the flight range is insufficient, it can automatically replan the mission. Under the premise of ensuring a safe return, it intelligently selects the optimal subset of missions with the goal of maximizing weighted value, avoiding premature return and optimizing the effective workload of a single flight to the greatest extent, thereby improving economic benefits.

[0020] 4. Strong adaptability and wide applicability. Through online updates, the model can continuously learn and adapt to different individual drones, different fuel cell aging levels, and sudden environmental changes, ensuring the stability and accuracy of the model during long-term use. It achieves personalized adaptation of "one model per drone" and has good applicability. Attached Figure Description

[0021] Figure 1 The overall flowchart of the intelligent management method for hydrogen-powered drones provided by the present invention.

[0022] Figure 2 This is a schematic diagram of the architecture of the temporal deep learning model provided by the present invention.

[0023] Figure 3 This is a schematic diagram illustrating the remaining battery life prediction results provided by the present invention.

[0024] Figure 4 This is a comparative diagram of the waypoint sequences before and after dynamic replanning of the mission provided by the present invention.

[0025] Figure 5 A flowchart of the online update triggering logic provided by this invention.

[0026] Figure 6 A schematic diagram of the structure of the intelligent management system for hydrogen-powered drones provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] In the description of this invention, it should be understood that the terms "center", "longitudinal", "lateral", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] This invention provides a method, system, electronic device, and storage medium for intelligent management of hydrogen-powered drones. Hydrogen-powered drones have broad prospects in industrial applications due to their advantages such as long endurance, large payload capacity, and low-temperature resistance; however, endurance prediction remains a challenge due to the complexity of hydrogen fuel cell systems. This invention aims to solve the problems of inaccurate endurance prediction, rigid task planning, and insufficient safety redundancy in existing technologies by introducing advanced deep learning models and intelligent decision-making logic.

[0030] In embodiments of this invention, some core terms have specific meanings. "Temporal deep learning model" specifically refers to a deep learning model designed for processing time-series data. In one specific implementation of this invention, this model combines a temporal convolutional network (TCN) with a multi-head attention mechanism to learn complex temporal dependencies from multi-dimensional UAV operational data and predict future remaining flight time. "Risk-adjusted flight time value" is a conservative flight time value selected for subsequent decision-making based on the uncertainty range of the model-predicted remaining flight time. Typically, this value can be set as the lower limit of the uncertainty range, such as the lower bound of a 90% confidence interval, to provide sufficient redundancy for flight safety. "Safe path planning strategy" is a multi-level decision-making logic that determines, based on the comparison between the risk-adjusted flight time value and the time required to complete the mission, whether to continue executing the adjusted mission, directly command the UAV to return, or execute an emergency landing procedure in extreme circumstances.

[0031] Reference Figure 1 The overall process of the intelligent management method for hydrogen-powered unmanned aerial vehicles (UAVs) provided by this invention can include the following steps: S100: multi-source data acquisition and feature engineering, followed by online endurance prediction S200, task replanning decision-making based on the prediction results S300, and finally path generation and confirmation S400. The entire process forms a closed loop, capable of continuous operation during UAV flight, dynamically ensuring flight safety and mission efficiency. This method can be deployed on the UAV's onboard computing platform or on a ground station's computer system, communicating with the UAV in real time via a wireless data link.

[0032] In step S100, it is first necessary to acquire multi-source operational data during the flight of the hydrogen-powered drone in real time. This data is the foundation for all subsequent analysis and decision-making, and its comprehensiveness and accuracy directly affect the final result. Specifically, data acquisition is achieved through various sensors mounted on the drone, and the data is aggregated to the onboard computing unit via a CAN (Controller Area Network) bus or other internal buses.

[0033] The multi-source operational data can specifically include multiple dimensions. For example, fuel status can include the real-time pressure and temperature of the hydrogen cylinder, which directly relates to the quality of remaining usable hydrogen fuel. Pressure data can be measured by a high-pressure sensor installed at the valve of the hydrogen storage cylinder, and temperature data can be measured by a thermocouple or PT100 temperature sensor close to the cylinder. Acquiring both pressure and temperature simultaneously allows for a more accurate estimation of hydrogen quality based on the ideal gas law. Stack status can include the stack's output voltage, current, and total power, reflecting the core operating state of the power system. Voltage and current can be measured by Hall effect sensors or shunts built into the stack controller, and power is calculated by multiplying voltage and current. Environmental status can include ambient temperature, air pressure, and wind speed estimated using GPS (Global Positioning System) information; these factors significantly affect stack performance and flight aerodynamic power consumption. Ambient temperature, humidity, and air pressure can be acquired by airborne microelectromechanical system (MEMS) barometers and temperature and humidity sensors. Wind speed can be estimated by calculating the difference between the ground speed vector provided by GPS and the airspeed vector measured by the airspeed tube, thus obtaining a more accurate relative wind speed and direction. Flight status can include historical flight condition sequences, such as recording the power spectrum over a period of time to characterize whether the drone is hovering, cruising, climbing, or in other high-power operation modes. By fusing this multi-source heterogeneous data, the model can more comprehensively understand the complex factors affecting endurance. Before the data is input into the model, preprocessing is performed, such as median filtering for occasional sensor jumps, linear interpolation for brief signal loss, and normalization of all features (e.g., minimum or maximum normalization) to eliminate the influence of different physical dimensions and accelerate model convergence.

[0034] As a preferred implementation, preprocessing and feature engineering of the acquired multi-source operational data are key steps in improving model performance. Before inputting the data into the time-series deep learning model, the method further includes constructing derived features based on the multi-source operational data. These derived features make explicit the physical processes implicit in the original data, providing the model with higher-level, more informative input. For example, the "hydrogen consumption rate" can be calculated by combining the decrease in hydrogen cylinder pressure per unit time (e.g., 10 seconds) with temperature changes for compensation; it reflects the current fuel consumption rate more directly than a simple pressure value. The "pile efficiency decay rate" can be obtained by calculating the slope of the moving average of the ratio of pile output power to hydrogen consumption rate (i.e., pile efficiency) over a long period (e.g., 30 minutes); this feature characterizes the performance degradation trend of the pile due to aging or changes in operating conditions. The "power fluctuation index" can calculate the standard deviation or coefficient of variation of flight power within a short time window (e.g., 1 minute) to quantify flight stability; severe power fluctuations usually indicate higher energy loss. These derived features help the model more accurately capture changes in system state, thereby improving prediction accuracy.

[0035] In step S200, the acquired data is input into a pre-trained time-series deep learning model to predict the remaining flight time of the drone. The model is trained offline using a large amount of historical flight logs. The training data covers various aircraft types, different payloads, diverse weather conditions, and flight missions. Data augmentation techniques (such as random scaling, jittering, and adding noise to the time series) are used to improve the model's generalization ability.

[0036] Optionally, in one embodiment of the present invention, the temporal deep learning model employed has a specific internal structure, such as... Figure 2As shown, this model combines a Temporal Convolutional Network (TCN) with a multi-head attention mechanism. The input feature vector 210 contains the aforementioned multi-source operational data and its derived features, forming a multi-dimensional time series. The data first passes through the Temporal Convolutional Network (TCN) module 220, whose core consists of causal convolution and dilated convolution. Causal convolution ensures that the prediction at time point t depends only on data at and before t, conforming to temporal logic. Dilated convolution expands the receptive field of the convolution kernel with an exponentially increasing dilation rate, allowing the upper-layer network to see a very long time span, thereby effectively capturing the long-term dependencies of hydrogen consumption patterns at different time scales, such as power fluctuations at the second level to stack efficiency degradation at the minute or even hour level. The output of the TCN module is then fed to the multi-head attention module 230. The multi-head attention module 230 is used to adaptively learn the dynamic weights of different features in the input data on the impact of flight endurance at different flight stages. By mapping feature sequences to multiple independent query-key-value spaces (i.e., multi-head), each head learns a different dependency relationship. For example, during takeoff in extremely cold environments, one attention head might learn a strong correlation between ambient temperature and fuel cell voltage, and automatically increase the attention weight of the ambient temperature feature; during heavy-load climb missions, another head would focus more on features such as fuel cell power and hydrogen consumption rate. Finally, the feature information after multi-head attention weighted fusion is passed through one or more fully connected output layers 240 to output the final prediction result.

[0037] Reference Figure 3 The output of the time-series deep learning model is an estimate of the remaining flight time at the current moment and an uncertainty interval based on quantile regression. In the figure, the horizontal axis represents flight time, and the vertical axis represents remaining flight time. The time-series deep learning model, through quantile regression, outputs not only the remaining flight time estimate curve 310 (e.g., the 0.5 quantile), representing the most probable outcome, but also the upper bound of the uncertainty interval 320 (e.g., the 0.95 quantile) and the lower bound of the uncertainty interval 330 (e.g., the 0.05 quantile). This output format, compared to a single point estimate, quantifies the uncertainty of the prediction. During decision-making, the lower bound of the uncertainty interval 330 can be used as the risk-adjusted flight time value. This decision-making strategy based on the confidence lower bound ensures that even with some prediction error, the planned route still has a very high probability of being safe, effectively avoiding the risk of fuel depletion due to overestimation of flight time.

[0038] As a preferred implementation, to address potential unforeseen circumstances or discrepancies between the model and actual conditions during flight, this invention also introduces an online update mechanism. (See reference...) Figure 5The method also includes online model updates during flight. The system continuously monitors the deviation between the actual hydrogen consumption rate (calculated in real time by sensors) and the model's predicted value, as shown in step S510. Next, in step S520, it is determined whether the deviation exceeds a preset threshold (e.g., 10%) for N consecutive sampling periods (e.g., three consecutive 10-second periods). Setting the threshold for N consecutive exceedances is to filter out misjudgments caused by transient disturbances or sensor noise. If the determination is yes, an online update is triggered in step S530. The online update process uses recent flight data (e.g., the past 5 minutes) to perform lightweight fine-tuning of some parameters of the temporal deep learning model (typically the weights of the last few fully connected layers). This process uses gradient descent, iterating several times on recent data with a small learning rate, thereby quickly correcting the model to adapt to the current operating conditions without destroying the general knowledge already learned by the model. Through this online update, the model can quickly adapt to the aging state of a specific UAV fuel cell, individual differences, or environmental changes caused by factors not fully reflected in the training set, such as sudden winds and icing, thereby maintaining the long-term accuracy of predictions.

[0039] In step S300, the system makes a decision based on the prediction results. Specifically, when the risk-adjusted range value determined based on the uncertainty interval indicates that the remaining range is insufficient to complete the remaining tasks originally planned, a task replanning mechanism is triggered. The core of this mechanism is solving an optimization problem. This optimization problem is formalized as: maximizing the weighted value of the tasks to be completed as the optimization objective, i.e., Maximize(Σ(wi) Let xi)), where wi is the value weight of the i-th waypoint, and xi is a binary decision variable representing whether to visit that waypoint. This optimization problem is subject to a series of constraints, the core of which is the total flight time constraint: the total time required to complete the adjusted path must be less than or equal to the current risk-adjusted endurance value. Other constraints may include mandatory waypoints that must be visited. This design enables the UAV to intelligently "weigh the pros and cons" when endurance is limited, prioritizing high-value tasks to maximize the operational efficiency of a single flight. Different solvers can be selected based on the problem size and real-time requirements. For scenarios with fewer waypoints (e.g., less than 20), exact algorithms such as branch and bound can be used to find the optimal solution. For online scenarios with many waypoints and limited computation time, heuristic algorithms such as genetic algorithms and simulated annealing can be used to find high-quality suboptimal solutions within a limited time (e.g., within 5 seconds).

[0040] In step S400, based on the optimized and adjusted mission waypoint sequence, the system generates an adjusted flight path for the UAV to execute.

[0041] The safe path planning strategy is designed with a multi-level structure to address varying degrees of endurance constraints. In one specific implementation, the strategy includes three levels. The first level is task adjustment: if the calculated risk-adjusted endurance value is insufficient to complete all remaining tasks but is still sufficient to safely complete a subset of valuable waypoints and return, the system will perform task adjustment, generating a path that visits some optimized waypoints before returning. For example, if the risk-adjusted endurance value is 70 minutes, the original plan required 90 minutes for the remaining tasks, but the optimized subset of high-value tasks only requires 60 minutes (including the return trip), the system will generate a 60-minute task path. The second level is direct return: if the risk-adjusted endurance value is insufficient to complete any of the remaining operational waypoints (i.e., the time to visit the nearest waypoint and return exceeds the endurance value), but is still sufficient to allow the UAV to safely return from its current location to the takeoff point or a preset return point, the system will use the direct return level to generate a path directly to the return point. For example, if the risk-adjusted endurance is 30 minutes, a direct return trip would take 25 minutes, while visiting the nearest waypoint and then returning would take 45 minutes. In this case, the system will immediately abandon all tasks and generate a return route. The third level is emergency diversion: In the most extreme case, if the risk-adjusted endurance cannot even guarantee a safe return, the system will activate the emergency diversion level, immediately planning and generating a route to the nearest, pre-set safe diversion point. For example, if the risk-adjusted endurance is 15 minutes, a return trip would take 25 minutes, while flying to the nearest diversion point would only take 10 minutes, the system will choose emergency diversion. This tiered strategy ensures that the decision most conducive to aircraft safety is made under any circumstances.

[0042] As a preferred implementation, human-computer interaction is indispensable for operators to understand and trust the system's automated decisions. The method of this invention also includes visualizing the adjusted flight path on a human-computer interface. (See also...) Figure 4 This interface clearly displays a comparison of waypoint sequences before and after dynamic replanning of the mission. For example, on the map, the original planned waypoint sequence 410 is represented by a gray dashed line, the adjusted waypoint sequence 420 is represented by a green solid line, and the return point 430 is clearly marked. In addition to path comparison, the interface can also simultaneously display the current remaining range prediction confidence interval (e.g., ...). Figure 3 The system displays the specific reasons for this replanning (e.g., the text prompt: "Due to strong winds, the predicted flight time has been reduced by 12 minutes"), the estimated flight time of the new route, the expected value of the completed tasks, and the list of tasks to be discarded. Based on this information, operators can quickly make a judgment, choosing to click the "Confirm Execution" button to accept the system's suggestion, or, if necessary, click the "Manually Modify" button for manual intervention and fine-tuning. This transparent interactive design greatly improves the efficiency of human-machine collaboration and the operator's situational awareness.

[0043] Example 1 This embodiment describes the application of the method of the present invention in a basic scenario, in which an H25 hydrogen-powered drone performs a simple package delivery task from point A to point B. The task flow is as follows: take off from the base, fly to point A to pick up the goods, fly to point B to deliver the goods, and finally return to the base.

[0044] 1. Data Acquisition: After the UAV takes off, the system continuously acquires data on fuel status (hydrogen cylinder pressure, temperature), fuel cell status (voltage, current), environmental status (temperature, wind speed), and flight status (power sequence).

[0045] 2. Model Prediction: During the flight, the data is fed into a pre-trained temporal deep learning model. At a certain moment, the model outputs a prediction, indicating that the point estimate of the remaining flight time is 10 minutes, with a 90% confidence interval of [8 minutes, 12 minutes].

[0046] 3. Decision Trigger: The system sets a conservative risk appetite and uses the lower limit of the confidence interval as the risk-adjusted endurance value, i.e., 8 minutes. At this time, the flight control system calculates according to the original plan that it will take 9 minutes to fly from the current position to point B and return to the base. Since the required time (9 minutes) is greater than the risk-adjusted endurance value (8 minutes), the system triggers mission replanning.

[0047] 4. Task Optimization and Path Generation: In this scenario, the task waypoint sequence is fixed, with no subset available for optimization. The optimization problem is a safety decision. The system enters a safe path planning strategy for evaluation. Calculations show that returning directly to base from the current location would take 6 minutes. Since the risk-adjusted endurance (8 minutes) is sufficient to support a direct return (6 minutes), the system selects the "direct return" level and immediately generates a direct return path to base, replacing the original planned path to point B.

[0048] Example 2 This embodiment clearly demonstrates the core closed-loop process of the method of the present invention, which involves acquiring data, predicting uncertainties, making decisions based on risk values, and generating safe paths.

[0049] The scenario is set as an H100 Sky-Piercing UAV performing an inspection mission at a large photovoltaic power station. The mission includes 50 inspection points, of which 22 are core equipment such as inverters and transformers, which are assigned high value weights; the remaining 28 are ordinary photovoltaic panel arrays, which have lower value weights.

[0050] 1. Flight and Data Processing: The drone took off as planned. At the 40-minute mark of flight, the drone encountered strong crosswinds. The system detected a significant increase in wind speed data in real time. Simultaneously, the actual value of the derived characteristic, "hydrogen consumption rate," calculated based on hydrogen cylinder pressure and temperature data, also increased accordingly.

[0051] 2. Online Update: The system detected that the actual hydrogen consumption rate exceeded the model's predicted value by 15% for three consecutive sampling periods (30 seconds in total), surpassing the preset 10% threshold. The system automatically triggered the online update mechanism. Using actual flight data from the past 5 minutes, the weights of the last two fully connected layers of the time-series deep learning model were fine-tuned.

[0052] 3. Endurance Prediction and Replanning Trigger: After the online update, the model outputs a revised endurance prediction. The point estimate of remaining endurance time is lowered from 115 minutes to 105 minutes, and the risk-adjusted endurance value (lower confidence interval) is correspondingly lowered from 108 minutes to 98 minutes. The system calculates that completing all remaining waypoints will take 115 minutes, which is greater than the risk-adjusted endurance value of 98 minutes, thus officially triggering dynamic replanning of the mission.

[0053] 4. Task Optimization and Path Generation: The task optimization module is activated, solving an optimization problem aimed at maximizing the weighted value, with the constraint that the total flight time does not exceed 98 minutes. The solver (such as a genetic algorithm) provides an optimized sequence of waypoints within seconds. This sequence retains all 22 high-value core equipment inspection points and discards 28 low-value ordinary waypoints. The generated path belongs to the "task adjustment" level in the safe path planning strategy.

[0054] 5. Human-Machine Interaction: The replanning results are sent to the ground station's human-machine interface. On the interface, the original path is displayed in gray, and the new path is highlighted in green. A pop-up message appears: "Warning: Encountering strong winds, reduced endurance. The mission has been automatically optimized; prioritize inspecting core equipment." After reviewing the comparison of the old and new paths, estimated completion time, and remaining endurance, the operator clicks the "Confirm Execution" button. The drone then continues its mission along the adjusted path and safely returns after completing all high-value objectives.

[0055] This embodiment demonstrates how the present invention achieves a dynamic balance between mission efficiency and flight safety in response to sudden environmental changes through derived feature perception, online model updates, intelligent mission optimization, and multi-level safety strategies.

[0056] Example 3 This embodiment describes the application of the present invention under extreme working conditions. The scenario is set as an H25 UAV performing a power line inspection mission in the mountainous region of Northeast China during winter, with an ambient temperature of -35°C.

[0057] 1. Data Acquisition and Feature Weighting: In an extremely cold environment of -35℃, the system collected multi-source operational data, including fuel status (hydrogen cylinder pressure significantly lower than the initial pressure at room temperature), stack status (low output voltage), and environmental status (temperature -35℃). After the data was input into the model, the multi-head attention mechanism in the time-series deep learning model played a crucial role. Because the negative impact of low temperature on the stack chemical reaction rate and hydrogen availability was learned during the training phase, the attention module improved the values ​​for "ambient temperature" and the "stack efficiency decay rate" caused by low temperature.

[0058] 2. Accurate Prediction: Due to the attention mechanism's focus on key features, the model accurately captures the nonlinear effects under extremely cold conditions. The model ultimately predicts a flight time of 85 minutes, with a confidence interval of [75 minutes, 95 minutes].

[0059] 3. Decision-making and planning: Based on a 75-minute risk-adjusted flight endurance value, the system planned an inspection path for the drone that could be completed within its safe flight range. The drone ultimately flew for 82 minutes before safely returning home, with a prediction error of only 3.7%.

[0060] 4. Comparison of Beneficial Effects: If a traditional endurance estimation algorithm that does not consider the coupling effects of multiple factors, or a simple model without an attention mechanism, is used, it is likely to make an optimistic prediction of about 115 minutes based on normal temperature data. If the mission is carried out according to this erroneous prediction, the drone will fly a distance far exceeding its actual return-to-home capability, eventually crashing in a remote mountainous area due to fuel exhaustion, resulting in equipment damage and data loss. This invention, through an advanced model architecture, demonstrates high reliability and high safety under complex and extreme operating conditions.

[0061] The present invention also provides an intelligent management system for hydrogen-powered unmanned aerial vehicles. Figure 6 This is a schematic diagram of the structure of the intelligent management system for hydrogen-powered unmanned aerial vehicles provided by the present invention. (Refer to...) Figure 6The system includes an acquisition module 10, a prediction module 20, a task optimization module 30, and a path generation module 40. The acquisition module 10 is responsible for acquiring multi-source operational data during flight in real time through various sensors connected to the UAV, including fuel status, fuel cell status, environmental status, and flight status. The prediction module 20 internally stores a pre-trained time-series deep learning model. It receives data from the acquisition module and outputs an estimate of the remaining flight time at the current moment and an uncertainty interval based on quantile regression. The task optimization module 30 is the decision-making core of the system, continuously monitoring the output of the prediction module. When the risk-adjusted endurance value determined based on the uncertainty interval indicates that the remaining endurance is insufficient to complete the original planned task, the task optimization module 30 is activated. It solves a combinatorial optimization problem based on preset waypoint value weights, aiming to maximize the weighted value, with the risk-adjusted endurance value as the core constraint, ultimately determining an adjusted optimal sequence of waypoints under the current endurance capability. The path generation module 40, based on the adjusted mission waypoint sequence output by the mission optimization module 30 and following the aforementioned safe path planning strategies (mission adjustment, direct return, emergency landing), generates a specific adjusted flight path that can be executed by the UAV flight control system. These four modules work together to form a complete closed loop from data perception, state prediction, intelligent decision-making to final action execution, achieving intelligent and adaptive management of hydrogen-powered UAV flight missions.

[0062] The present invention also provides an electronic device, which can be an onboard computer of a UAV, an industrial control computer of a ground station, or any computing device capable of performing complex calculations. The electronic device includes a processor and a memory. The memory stores computer instructions designed to, when executed by the processor, implement the intelligent management method for hydrogen-powered UAVs as described in any of the above embodiments. For example, the processor can load and execute instructions to complete a series of computational tasks such as real-time data acquisition, model inference, optimization problem solving, and path planning.

[0063] Furthermore, this invention also provides a computer-readable storage medium. This medium can be non-volatile, such as a solid-state drive (SSD), SD card, USB flash drive, or optical disc. A computer program is stored on this storage medium. When this computer program is loaded and executed by a processor of a computing device (such as the aforementioned electronic device), the intelligent management method for hydrogen-powered drones described in any of the above embodiments can be implemented. This form allows the technical solution of this invention to be easily distributed, deployed, and upgraded in the form of a software product.

Claims

1. A method for intelligent management of hydrogen-powered unmanned aerial vehicles (UAVs), characterized in that, include: Real-time acquisition of multi-source operational data during the flight of hydrogen-powered drones, including at least fuel status, fuel cell stack status, environmental status, and flight status; The multi-source flight data is input into a pre-trained temporal deep learning model that combines a temporal convolutional network with a multi-head attention mechanism, and outputs the estimated remaining flight time at the current moment and the uncertainty interval based on quantile regression. When the risk-adjusted endurance value determined based on the uncertainty interval indicates that the remaining endurance is insufficient to complete the remaining tasks originally planned, solve an optimization problem with the objective of maximizing the weighted value of the tasks to be completed and constrained by the risk-adjusted endurance value, in order to determine the adjusted task waypoint sequence; Based on the adjusted mission waypoint sequence and in accordance with a safety path planning strategy that includes at least mission adjustment, direct return, and emergency diversion levels, an adjusted flight path is generated for the hydrogen-powered UAV to execute.

2. The method according to claim 1, characterized in that, Before inputting the multi-source runtime data into the time-series deep learning model, the method further includes: Based on the multi-source operating data, derived features are constructed, including at least one of hydrogen consumption rate, stack efficiency decay rate, or power fluctuation index.

3. The method according to claim 1, characterized in that, The safe path planning strategy includes: If the risk-adjusted endurance is sufficient to complete some waypoints, then the generation of the adjusted flight path adopts the mission adjustment level. If the risk-adjusted endurance is insufficient to complete any operational waypoint but sufficient to return, then the adjusted flight path is generated at the direct return level, and the adjusted flight path is a path that directly reaches the return point. If the risk-adjusted endurance value is insufficient for a return flight, the adjusted flight path is generated at the emergency diversion level, and the adjusted flight path is the path to the nearest safe diversion point.

4. The method according to claim 1, characterized in that, The method further includes: When the deviation between the actual hydrogen consumption rate and the model prediction exceeds a preset threshold, the parameters of the time-series deep learning model are updated online using recent actual flight data.

5. The method according to claim 1, characterized in that, The method further includes: The adjusted flight path is then visualized on the human-computer interaction interface.

6. The method according to claim 1, characterized in that, The fuel status includes hydrogen cylinder pressure and temperature; the fuel cell status includes fuel cell voltage, current and power; the environmental status includes ambient temperature, air pressure and wind speed; and the flight status includes a sequence of historical flight conditions.

7. The method according to claim 1, characterized in that, The temporal convolutional network is used to capture the dependence of hydrogen consumption patterns at different time scales; the multi-head attention mechanism is used to adaptively learn the dynamic weights of the impact of different features in the multi-source operational data on endurance at different flight stages.

8. A hydrogen-powered unmanned aerial vehicle (UAV) intelligent management system, characterized in that, include: The acquisition module is used to acquire multi-source operational data during the flight of the hydrogen-powered drone in real time. The multi-source operational data includes at least fuel status, fuel cell status, environmental status, and flight status. The prediction module is used to input the multi-source flight data into a pre-trained temporal deep learning model that combines a temporal convolutional network and a multi-head attention mechanism, and output the estimated remaining flight time at the current moment and the uncertainty interval based on quantile regression. The task optimization module is used to solve an optimization problem with the risk-adjusted endurance value as the optimization objective and the risk-adjusted endurance value as the constraint when the risk-adjusted endurance value determined based on the uncertainty interval indicates that the remaining endurance is insufficient to complete the remaining tasks originally planned, so as to determine the adjusted task waypoint sequence. The path generation module is used to generate an adjusted flight path for the hydrogen-powered UAV to execute, based on the adjusted mission waypoint sequence and in accordance with a safety path planning strategy that includes at least mission adjustment, direct return, and emergency diversion levels.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores computer instructions, which, when executed by the processor, implement the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-7.

Images