Unmanned aerial vehicle adaptive energy efficiency prediction and optimization management method and system, and medium

Abstract

The invention discloses an unmanned aerial vehicle adaptive energy efficiency prediction and optimization management method and system and a medium, and the method comprises the steps: constructing a physical-data dual-drive AI prediction model according to a fusion electrochemical model or an empirical formula as a physical constraint before a task starts; bMS real-time data, historical flight logs and charging and discharging records serve as input of an AI prediction model, the current battery state is obtained according to the input, and an initial flight path with the optimal global energy consumption is initially predicted based on the current battery state and the environment situation; in the flight process, the current battery state and environment situation are continuously updated, the intelligent task and flight control module continuously monitors multiple trigger states, and when one trigger state meets a preset dynamic adjustment trigger condition, a dynamic adjustment decision is executed, and an optimization instruction is generated and issued to a flight controller for execution; according to the management method, through global and closed-loop energy efficiency optimization, unnecessary energy consumption is effectively reduced, and the cruising ability is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a UAV adaptive energy efficiency prediction and optimization management method and system and a medium. BACKGROUND

[0002] Unmanned aerial vehicles (UAVs) have been widely used in various industries, but their battery endurance has always been a core bottleneck limiting their operating range, duration, and efficiency. Especially for large and medium-sized UAVs, the problems of high power consumption, heavy battery weight, and slow charging are particularly prominent.

[0003] To improve the energy efficiency of UAVs, some explorations have been made in the prior art. For example, some schemes use multi-modal sensor data to generate an environment perception map, and use reinforcement learning (RL) and other methods for path planning and battery management. This kind of method learns the direct mapping from state to action through interaction with the environment, which improves the autonomy of the UAV to a certain extent. However, this kind of end-to-end reinforcement learning scheme also has some inherent limitations: 1. The battery model is too simplified and lacks physical insight: existing schemes often regard the battery as a simple energy pool, and its decision is mainly based on the current state of charge (SoC). This model ignores the nonlinear characteristics of the battery as a complex electrochemical system, such as temperature effects, aging effects, and the impact of different discharge rates on actual available energy, resulting in inaccurate energy consumption prediction, especially in complex flight tasks.

[0004] 2. The decision-making process lacks explainability and safety verification difficulty: the decision-making process based on reinforcement learning is usually like a "black box", and its decision logic is difficult for humans to understand and verify intuitively. This brings great challenges to the debugging, safety certification, and reliable deployment of the system in critical tasks.

[0005] 3. The decision-making level is shallow and the generalization ability is limited: its decision adjustment mainly focuses on tactical-level path and speed optimization, and lacks the ability to dynamically and autonomously adjust higher-level strategies that are oriented towards task goals. At the same time, for unexpected situations or abnormal battery states that have not appeared in the training data, the generalization ability and reliability of the coping strategy are questionable. SUMMARY

[0006] Based on the technical problems existing in the background technology, the present application proposes a UAV adaptive energy efficiency prediction and optimization management method and system and a medium, which effectively reduces unnecessary energy consumption through global and closed-loop energy efficiency optimization, and significantly improves the endurance.

[0007] The UAV adaptive energy efficiency prediction and optimization management method proposed by the present application comprises: Before the mission begins, the initial flight path with the optimal global energy consumption is initially predicted based on the current battery status and environmental conditions. During flight, the current battery status and environmental conditions are continuously updated. The intelligent mission and flight control module continuously monitors various trigger states. When one of the trigger states meets the preset dynamic adjustment trigger conditions, a dynamic adjustment decision is executed and an optimization command is generated and sent to the flight controller for execution.

[0008] Furthermore, the process for generating the current battery state is as follows: By integrating electrochemical models or empirical formulas as physical constraints, a physics-data dual-driven AI prediction model can be constructed. The current battery status is obtained by using real-time BMS data, historical flight logs, and charge / discharge records as inputs to the AI ​​prediction model.

[0009] Furthermore, the physical constraints of the AI ​​prediction model are specifically as follows: Energy conservation constraint: ; Voltage-current consistency: ; Piecewise monotonic constraints on internal resistance and temperature / aging; Discharge rate consistency constraint; in, For predicted future moments The remaining available energy, For the present The remaining available energy at any given moment. The current battery status is output by the battery status accurate sensing and prediction module. The discharge efficiency of the battery is a function of the current battery state. For at any time The total load power of the drone Let be the integral variable, representing the integral variable at the time of integration. At any instant within the time period, For an infinitesimal time increment, For load current, This is the load voltage. This is the equivalent internal resistance of the battery. This represents the current remaining battery percentage. It is an open-circuit voltage function. and temperature The function.

[0010] Furthermore, the AI ​​prediction model is an LSTM network-based model, and the loss function during the training process of the AI ​​prediction model... The settings are as follows: ; in, To predict the mean squared error loss between SoE and labeled SoE, To predict the loss between the voltage series and the measured series, To predict the loss between the equivalent internal resistance and the estimated equivalent internal resistance, This is a physical constraint penalty term, which corresponds to the physical constraints of the AI ​​prediction model. For time smoothing regularization, For weight decay regularization, For physical constraint weights, To smooth out the regularization weights, This is the weight decay factor.

[0011] Furthermore, the intelligent task and flight control module continuously monitors various trigger states, specifically: Using the current battery state and environmental conditions as input, the energy-efficient path planner employs energy expansion... Algorithm construction cost function Based on this, a local flight path with optimal energy consumption is generated; The autonomous mission decision engine continuously monitors the energy consumption deviation between the local flight path and the initially predicted initial flight path, and uses the energy consumption deviation exceeding the deviation threshold as the first trigger state to meet the preset dynamic adjustment trigger conditions. The autonomous mission decision engine continuously makes forward predictions about the state of the drone, and uses the prediction that the remaining energy cannot satisfy all the remaining tasks as the second trigger state to meet the preset dynamic adjustment trigger conditions. The autonomous task decision engine continuously monitors key environmental parameters and uses the change of the key environmental parameters exceeding the initially planned threshold range as the third trigger state to meet the preset dynamic adjustment trigger conditions. The key environmental parameters include wind speed and wind direction. The autonomous mission decision engine receives requests from ground stations or operators, requiring an immediate reassessment of external mission instructions as a fourth trigger state to meet preset dynamic adjustment trigger conditions.

[0012] Furthermore, in the step of using the energy consumption deviation exceeding the deviation threshold as the first trigger state to satisfy the preset dynamic adjustment trigger condition, wherein: Energy consumption deviation The formula is as follows: ; Deviation threshold The formula is as follows: ; in, This represents the cumulative energy consumption to date. The predicted cumulative energy consumption for the initial planning. Here is the numerical stability constant. This is the default setting; different values ​​are set depending on the level. This represents the risk aversion coefficient.

[0013] Furthermore, the cost function The formula is as follows: ; in, For the node To the node The corresponding number Each path segment For the flight path segment The prediction time function, For the speed, wind speed and distance of the drone, For the flight path segment The predicted energy consumption function, Due to high degree of variation, The current battery status is output by the battery status accurate sensing and prediction module. and As a weighting factor, satisfying .

[0014] Furthermore, the predicted energy consumption function Includes the following power components: Base power consumption : ; Aerodynamic losses : ; Climb / descent losses : ; Load power consumption : ; Battery loss : ; in, Based on the fundamental power consumption constant, For temperature coefficient, For reference temperature, The current temperature. This is the second-order drag coefficient. It is the third-order drag coefficient. Relative airflow velocity For takeoff mass, For the climb rate, To improve efficiency, The climbing drag coefficient, For load The on / off state, For load Rated power, For current, The equivalent internal resistance (Ω) of the battery prediction module. For thermal efficiency.

[0015] Furthermore, , , , , Offline calibration coefficient range: ; ; ; ; .

[0016] Furthermore, , , , , Online adaptive strategy: The coefficients are calibrated online using the sliding window least squares method. Calibration trigger condition: Prediction error greater than 5% for 10 consecutive times; Calibration frequency: Performed once every 100 predictions; Coefficient change limit: The adjustment range for a single adjustment is less than 10%.

[0017] Furthermore, the execution of dynamic adjustment decisions and generation of optimization instructions are sent to the flight controller for execution. The autonomous mission decision engine is based on the MPC framework. According to the preset "maximize mission coverage" strategy and mission point priority, it evaluates different mission-level adjustment schemes in combination with the utility function, and then executes dynamic adjustment decisions and generates optimization instructions accordingly. The utility function is: ; in, Adjust the strategy to achieve the optimal task. The set of remaining tasks. For the task Priority weights, In strategy and remaining available energy Next, complete the task. The task completion probability function This represents the risk aversion coefficient. This function represents the risk of task failure.

[0018] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0019] A computer-readable storage medium storing a plurality of classification programs, the plurality of classification programs being invoked by a processor to execute the method described above.

[0020] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as ROM, RAM, magnetic disk, or optical disk.

[0021] The advantages of the UAV adaptive energy efficiency prediction and optimization management method, system, and medium provided by this invention are as follows: Through global, closed-loop energy efficiency optimization, unnecessary energy consumption is effectively reduced. With the same battery capacity, the effective operating time and range of the UAV can be increased by 10%-30% or more, thus significantly improving endurance. Based on accurate energy prediction and intelligent autonomous decision-making, tasks can be dynamically adjusted to cope with emergencies, ensuring that the UAV has sufficient power to return safely under any circumstances. This greatly reduces the risk of mission failure or crash due to power depletion, thereby significantly improving mission success rate and safety. The UAV is transformed from a passive "power consumer" into an active "energy efficiency manager," with decision-making encompassing task-level adjustments, achieving a higher level of autonomy, thus realizing true intelligence and autonomy. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall architecture of the present invention; Figure 2 It is a top-level flowchart of management methods; Figure 3 This is a schematic diagram of the workflow of the battery status accurate sensing and prediction module; Figure 4 This is a schematic diagram of the internal interaction and decision-making process between the energy-sensing intelligent task and the flight control module; Figure 5 This is a schematic diagram illustrating the construction process of the energy-sensing path planning cost function. Detailed Implementation

[0023] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0024] like Figures 1 to 5 As shown, the UAV adaptive energy efficiency prediction and optimization management method proposed in this invention includes: Step 1: Before the mission begins, based on the current battery status and environmental conditions, initially predict the optimal initial flight path for global energy consumption. Step 2: During flight, the current battery status and environmental conditions are continuously updated. The intelligent mission and flight control module continuously monitors various trigger states. When one of the trigger states meets the preset dynamic adjustment trigger conditions, a dynamic adjustment decision is executed and an optimization command is generated and sent to the flight controller for execution.

[0025] This embodiment effectively reduces unnecessary energy consumption through global, closed-loop energy efficiency optimization. With the same battery capacity, it can increase the effective operating time and range of the drone by 10%-30% or more, thus significantly improving endurance. Based on accurate energy prediction and intelligent autonomous decision-making, it can dynamically adjust tasks to cope with emergencies and ensure that the drone has sufficient power to return safely under any circumstances. This greatly reduces the risk of mission failure or crash due to power depletion, thereby significantly improving mission success rate and safety. It transforms the drone from a passive "power consumer" into an active "energy efficiency manager," enabling its decision-making to cover mission-level adjustments and achieving a higher level of autonomy, thus achieving true intelligence and autonomy.

[0026] In this embodiment, the implementation of the above management method mainly relies on the management equipment, which includes a battery status accurate sensing and prediction module, a multi-source environment and task sensing module, an energy sensing intelligent task and flight control module, and an airborne equipment and payload power consumption management module.

[0027] I. Precise Battery Status Sensing and Prediction Module; (a1) Module functions; The battery status accurate sensing and prediction module is the foundation of the system's energy sensing. It obtains real-time data such as voltage, current, temperature, and SoC from the underlying battery management system (BMS) and integrates it with historical data (including drone flight logs and battery charge and discharge records) stored in non-volatile memory.

[0028] The battery state precision sensing and prediction module incorporates a physics-data dual-driven AI prediction model. This model is preferably based on machine learning (such as a lightweight recurrent neural network (RNN / LSTM)), and its core innovation lies in integrating electrochemical models or empirical formulas as physical constraints into its training and inference processes. This allows the model not only to learn data patterns but also ensures that its predictions (such as the trend of internal resistance changes with temperature and aging) conform to basic physical laws, thereby greatly improving the model's robustness, generalization ability, and interpretability. Based on the fused data, the model can predict the battery voltage sequence, accurate remaining energy (SoE), performance at different discharge rates, and battery state of health (SoH) in real-time and online over a future period.

[0029] (a2) Data Interface Specification: Input interfaces: BMS real-time data interface (CAN / SPI / I2C), historical data storage interface (eMMC / SD), environmental data interface (temperature sensor); Output interfaces: Prediction result publishing interface (ROS topic), model status interface (health status), calibration parameter interface (online learning); Data format: Standardized battery state data structure, including prediction results such as SoE, SoH, predicted voltage sequence, and internal resistance.

[0030] (a3) Inter-module communication mechanism; Data publishing: Battery status prediction results are published through the ROS topic mechanism, and real-time subscription is supported; State synchronization: Maintain state synchronization with the decision-making module to ensure the timeliness of prediction results; Parameter update: Receive feedback from the decision module and dynamically adjust the parameters of the prediction model.

[0031] (a4) AI prediction model; In this embodiment, the AI ​​prediction model is an LSTM network-based model, optimized by introducing physical constraints. Its mathematical form can be expressed as: Formula 1: ; in, For predicted future moments The remaining available energy, The voltage sequence is a predicted voltage series over a future period of time, specifically represented as a time series vector containing predicted voltage values ​​at multiple future discrete time points. The estimated equivalent internal resistance (Ω) of the battery prediction module reflects the battery's health status. This is an AI prediction model, whose internal structure is an LSTM network, used to capture dependencies in time series data. These are historical flight data and charge / discharge data, respectively. This provides real-time data (voltage, current, temperature) for the BMS. For environmental data (such as ambient temperature). These are the trained parameters of the AI ​​prediction model.

[0032] The input features for the AI ​​prediction model are: real-time BMS data, historical flight logs, charge / discharge records, and environmental data. BMS Real-time Data: Voltage Current ,temperature , Sampling frequency 10-20 Hz; Historical flight log: flight attitude, motor speed, power consumption, environmental conditions, time window 8-20s; Charge / discharge record: Number of cycles Depth of charge and discharge Temperature gradient ; Environmental data: Ambient temperature Humidity and air pressure (affect battery performance).

[0033] The output features of the AI ​​prediction model are: : Future Moments Remaining available energy, in Wh, with an accuracy requirement of ±2%; To predict voltage sequences, used to predict battery cutoff time and power limitations; It is the equivalent internal resistance, reflecting the battery's health status and temperature effects.

[0034] In this embodiment, during training, the loss function of the AI ​​prediction model includes not only the mean squared error between the predicted and true values, but also a physical constraint penalty term (corresponding to the physical constraints introduced in the AI ​​prediction model). This penalty term is used to constrain the model's output so that it does not violate basic electrochemical laws (e.g., the trend of internal resistance with temperature and aging should conform to physical laws), thereby improving the model's robustness and generalization ability.

[0035] To improve the interpretability and robustness of AI prediction models, the loss function during training... Defined as: ; in, To predict the mean squared error loss between SoE and labeled SoE, To predict the loss between the voltage series and the measured series, To predict the loss between the equivalent internal resistance and the estimated equivalent internal resistance (approximately using open-circuit-load transient or electrochemical impedance spectroscopy (EIS) for estimating parameters such as internal resistance). This is a physical constraint penalty term, which corresponds to the physical constraints of the AI ​​prediction model. Time smoothing regularization is used to suppress unreasonable high-frequency jitter. For weight decay regularization, For physical constraint weights, To smooth out the regularization weights, This is the weight decay factor.

[0036] As a preferred embodiment, the weighting suggestion is: [Physical constraint weights; it is recommended to start tuning from 0.1;] Smoothing regularization weight, recommended value is 0.005; Weight decay is recommended, 1e. 5.

[0037] The physical constraint penalty item in this embodiment This includes: energy conservation constraints, voltage-current consistency, piecewise monotonic constraints on internal resistance and temperature / aging, and discharge rate consistency. (a4-1) Energy conservation constraint: ; If the boundary is violated If so, a penalty will be imposed; Not lower than ; in, For the present The remaining available energy at any given moment. The current battery status is output by the battery status accurate sensing and prediction module. The discharge efficiency of the battery is a function of the current battery state. For at any time The total load power of the drone Let be the integral variable, representing the integral variable at the time of integration. At any instant within the time period, For an infinitesimal time increment, This is the battery's usable energy coefficient, with a default value of 0.95 (considering unusable areas). This refers to the battery's nominal capacity, measured in Wh. This is the discharge cutoff voltage, determined by the battery specifications. To predict voltage sequences.

[0038] (a4-2) Voltage-current consistency: ; in, It is an open-circuit voltage function. and temperature The function, This is the load voltage. The equivalent internal resistance of the battery is the current battery state. The function, This represents the current remaining battery percentage.

[0039] when A penalty is imposed when the voltage exceeds the tolerance range; for example, the tolerance range is ±0.05V (configurable as needed).

[0040] (a4-3) Piecewise monotonic constraints of internal resistance and temperature / aging; Under normal operating conditions of 10–40℃, ; In the low temperature (below 10℃) or high temperature (above 40℃) range, non-monotonic but It must not be lower than the physical lower limit; With the number of loops Increase, It exhibits a non-decreasing trend (allowing for small fluctuations).

[0041] Its physical boundary constraints: equivalent internal resistance The maximum value is 200% of the internal resistance of the new battery, and the minimum value is 80% of the internal resistance of the new battery.

[0042] (80% of the internal resistance of the new battery); (200% of the internal resistance of the new battery).

[0043] in, Equivalent internal resistance The minimum value, Equivalent internal resistance The maximum value, This refers to the internal resistance of the new battery.

[0044] It should be noted that the above physical boundary constraints can be extended and set as needed.

[0045] (a4-4) Discharge rate consistency: At a higher Below, predict voltage sequence The decrease is no less than the low The magnitude of the decline.

[0046] As a preferred embodiment, the training and inference configurations of the AI ​​prediction model are as follows: Data configuration: Sampling rate: 10–20Hz; Input window: 8–20s; Prediction step size Δt: 0.2–1s; Forward time domain H: 5–20s; Features: voltage / current / temperature / SoC, short-window and long-window statistics (mean, variance, gradient, discharge rate C-rate), ambient temperature, duration of the most recent full charge, etc.

[0047] Model architecture: 2-layer LSTM (64–128 hidden cells) + MLP head; quantization / distillation to meet onboard computing power requirements. Activation functions: tanh is used in LSTM layers, and ReLU is used in MLP layers; Dropout rate: 0.2-0.3 (to prevent overfitting), where the dropout rate represents the probability that each neuron is temporarily dropped during training.

[0048] Online calibration and adaptation: Sliding window residual calibration and fine-tuning switch are used (enabled when temperature / load is stable, small step size ≤1e). 4); Calibration trigger conditions: 5 consecutive prediction errors greater than 3% or voltage prediction error greater than 0.1V; Fine-tuning frequency: Performed once every 100 predictions to avoid overfitting.

[0049] Security Boundaries and Verification: Or SoE percentage ≥10–25%; ( (Configurable as needed) Prediction confidence: When the prediction variance is greater than the threshold, the system is downgraded to conservative mode.

[0050] in, For the energy required for the return flight, To reserve energy for minimum safety, The safety factor is a user-configurable parameter greater than 1, such as 1.2, 1.3, etc. For safety voltage margin, it is a configurable, small positive voltage value.

[0051] Model performance metrics: SoE prediction accuracy: Root mean square error (RMSE) less than 2%, mean absolute error (MAE) less than 1.5%; Voltage prediction accuracy: Root mean square error (RMSE) less than 0.05V; Internal resistance estimation accuracy: relative error less than 5%; Inference latency: less than 10ms (single prediction); Model size: less than 2MB (after quantization).

[0052] This embodiment adopts a white-box "prediction-optimization" paradigm, ensuring a traceable decision-making process that facilitates verification, debugging, and safety certification. The AI ​​prediction model incorporating physical constraints is also more robust than a purely data-driven model, exhibiting stronger generalization capabilities for unseen scenarios. Furthermore, through accurate SoH / RUL prediction and optimized charge / discharge management, batteries can be used and maintained more scientifically, extending their lifespan and reducing total lifespan costs.

[0053] II. Multi-source environment and task perception module; (b1) Module functions; The multi-source environment and mission awareness module is responsible for constructing a model of the external world in which the UAV operates. It integrates real-time data from onboard sensors (such as GPS, IMU, barometer, etc.) and information acquired from external data sources (such as online weather services and cloud-based digital elevation models, DEM) via wireless data links. Simultaneously, it analyzes user-predefined mission plans (including waypoints, mission priorities, and user-selected energy efficiency strategies) to construct a dynamic environment and mission situation map that includes the UAV's current status, dynamic wind fields, terrain undulations, and prioritized preset mission points.

[0054] (b2) Sensor data fusion strategy; Multi-source data fusion: Kalman filtering and particle filtering algorithms are used to fuse data from multiple sources such as GPS, IMU, and barometer; Environmental modeling: Constructing environmental representations such as dynamic wind field models, terrain elevation models, and atmospheric condition models; Task parsing: Parses user-defined task plans, including waypoint coordinates, task priority, time constraints, etc.

[0055] (b3) Energy efficiency strategy configuration; The energy efficiency strategy described here is not limited to preset modes such as time priority or maximum range, but can also be a set of multiple rules, for example: Safety margin strategy: Define the minimum remaining energy threshold required for a safe return (it can be a fixed value, such as 20%, or a dynamic value, such as "enough energy to withstand a Force 5 headwind return").

[0056] Task prioritization strategy: Clearly define the priority of different task points, and the abandonment strategy that the decision engine should follow when energy is insufficient (e.g., "abandoning from the lowest priority" or "abandoning the task point that is farthest away").

[0057] Risk preference strategy: By adjusting the utility function (such as the risk aversion coefficient) This can be used to define whether a system tends to be conservative (prioritizing a safe return) or aggressive (prioritizing attempting to complete more tasks) when faced with uncertainty.

[0058] (b4) Inter-module communication mechanism; Environmental data dissemination: Real-time release of environmental situation information, including wind fields, topography, atmospheric conditions, etc. Task status synchronization: Synchronize task execution status and progress information with the decision-making module; Strategy parameter transmission: Pass the user-configured energy efficiency strategy parameters to the decision-making module.

[0059] III. Intelligent mission and flight control module for energy sensing; The energy-sensing intelligent mission and flight control module includes a path planner, an autonomous mission decision engine, and a flight control command generator.

[0060] (c1) Functions of the intelligent mission and flight control module for energy perception; The energy-sensing intelligent mission and flight control module is the "brain" and decision-making core of the entire system. It receives the current battery status predicted by the battery status precision sensing and prediction module and the environmental situation from the multi-source environment and mission sensing module, and makes intelligent decisions. The decision results are issued to the airborne equipment and payload power management module in the form of power management commands, and to the UAV's underlying flight controller in the form of high-level flight control commands (including mission-level adjustment commands, such as skipping waypoints).

[0061] (c2) Path planner; Employ optimized path search algorithms (such as energy-expanded algorithms) The algorithm deeply integrates the dynamic energy consumption model of the battery state accurate perception and prediction module into its path cost function to generate the globally or locally energy-optimal flight path.

[0062] The path planner features real-time performance, multi-objective optimization, and dynamic adaptation. Real-time performance: Supports local path replanning (less than 100ms) and global path optimization (less than 500ms). Multi-objective optimization: balancing multiple objectives such as time efficiency, energy consumption, and safety margin. Dynamic adaptation: The path planning strategy is dynamically adjusted based on real-time environmental changes and battery status.

[0063] Using energy expansion When the algorithm searches for a path, it considers each possible path segment. (from node) arrive ), its cost The calculation is as follows: ; in, For the node To the node The corresponding number Each path segment For the flight path segment The flight time cost function is The function composed of The values ​​are the drone's speed (m / s), wind speed (m / s), and distance (m). For the flight path segment The predicted energy consumption function, Due to high degree of variation, The current battery status is output by the battery status accurate sensing and prediction module. and As a weighting factor, satisfying .

[0064] This embodiment is for energy expansion. Algorithm settings: Heuristic functions: ,That The historical average energy consumption rate The heuristic function value represents the path from the current node. The estimated cost to reach the destination, which is estimated here in terms of energy, is measured in watt-hours (Wh) or joules (J). The remaining distance refers to the path nodes found so far. The position is the straight-line distance from the final target point G.

[0065] Node expansion strategy: Prioritize expanding the most energy-efficient paths; Real-time performance: Local replanning is limited to completion within 100ms, and global planning is limited to completion within 500ms; Multi-resolution planning: Dynamically adjusts planning accuracy based on battery status (fine planning at high SoH, fast planning at low SoH).

[0066] In this embodiment, This is one of the key points of this embodiment. The function is a complex nonlinear function, and its inputs include flight parameters (speed). Altitude change ), environment (wind speed) (and the current battery status provided by the battery prediction module) (Including SoH, temperature, internal resistance, etc.). It can simulate the differences in energy consumption under different discharge rates and battery conditions.

[0067] To ensure real-time performance while also considering physical plausibility and learnability, a decomposition method of "semi-analytical + learnable coefficients" is adopted: Basic physical models: Let the relative airflow velocity be... (m / s) Norm operators are used to calculate the drone's velocity vector relative to the air; rate of climb. (m / s), where m is the takeoff mass (kg) and g is the gravitational acceleration (9.81 m / s²). For the drone's altitude, For time, climb rate Equal to its height Regarding time The derivative of .

[0068] Current-power relationship: ,in This represents the current battery voltage (V). Total power (W).

[0069] Discretization to Time step: ; ; in, The flight path was divided into tiny time steps. The number of For base power consumption, For aerodynamic losses, For climb / descent losses, For load power consumption, This is due to battery wear and tear.

[0070] Detailed breakdown of power components: Base power consumption : ; Aerodynamic losses : ; Climb / descent losses : ; Load power consumption : ; Battery loss : ; in, Based on the fundamental power consumption constant, a typical value is 15-25W. This is the temperature coefficient, typically 0.1-0.3 W / ℃. For reference temperature, typically 25℃. The current temperature. This is the secondary drag coefficient, with a typical value of 0.05-0.15 W / (m / s)². This is the third-order drag coefficient, with a typical value of 0.001-0.005 W / (m / s)³. Relative airflow velocity Takeoff mass (kg) For the climb rate, To improve efficiency, This is the climbing drag coefficient, typically 2-8 W / (m / s). For load The switch state (0 or 1). For load Rated power, For current, The equivalent internal resistance (Ω) of the battery prediction module. This represents thermal efficiency, typically 0.95-0.98.

[0071] Physical constraints: , , , , , ; Typical values: camera 10-20W, image transmission 8-15W, sensor 2-5W.

[0072] In this embodiment, , , , , The parameter calibration and adaptation include offline calibration coefficient range and online adaptive strategy.

[0073] Offline calibration coefficient range: ; ; ; ; .

[0074] Online adaptive strategy: The coefficients are calibrated online using the sliding window least squares method. Calibration trigger condition: Prediction error greater than 5% for 10 consecutive times; Calibration frequency: Performed once every 100 predictions; Coefficient change limit: The adjustment range for a single adjustment is less than 10%.

[0075] In this embodiment, the following settings are provided: Environmental parameter constraints: Wind speed limit: (When exceeding the limit, a conservative plan will be adopted); Temperature range: (If the error exceeds the limit, the model will be downgraded to the empirical model). Rate of change in height: (Exceeding the time segmentation planning).

[0076] In this embodiment, the following settings are provided: Performance metrics and verification: Prediction accuracy: The relative error is less than 8% (95% confidence interval); Calculation delay: single time The calculation takes less than 1ms; Memory usage: Model parameters less than 100KB; Real-time performance: Supports continuous prediction at 100Hz.

[0077] In this embodiment, a degradation strategy is set: Predicting energy consumption function when computing resources are limited Using a simplified model : ; In the simplified model, safety boundaries are set: Lower limit of energy prediction: , Experience value; Upper limit of power prediction: , This is the rated maximum power; Anomaly detection: Trigger path planner recalibration when three consecutive predictions exceed the boundary.

[0078] (c3) Autonomous task decision engine; This is key to achieving advanced intelligence. During flight, the engine continuously predicts the drone's state based on model-based predictive control (MPC) or a similar optimization framework. When it detects a significant deviation between actual and expected energy consumption, or when the predicted remaining energy is insufficient to meet all remaining tasks, it autonomously makes higher-level mission-level decisions based on user-defined energy efficiency strategies and mission priorities. These decisions are not limited to adjusting flight parameters, but also include adjustments to the mission itself, such as dynamically adjusting the mission execution sequence, or proactively deciding to skip secondary mission points to ensure the completion of core missions and safe return.

[0079] During flight, the autonomous decision engine continuously monitors various trigger states (deviation trigger, prediction trigger, environmental change trigger, external command trigger), and activates when at least one of the following preset dynamic adjustment trigger conditions is met: Deviation Trigger: The autonomous mission decision engine continuously monitors the energy consumption deviation between the local flight path and the initially predicted initial flight path, and uses the energy consumption deviation exceeding the deviation threshold as the first trigger state to meet the preset dynamic adjustment trigger conditions.

[0080] Assume the cumulative energy consumption up to now Predicted cumulative energy consumption based on initial planning : ; in, For energy consumption deviation, This is a numerical stability constant with a default value of 1.0Wh to prevent division by zero errors.

[0081] Dynamic threshold adjustment (based on risk appetite): ; in, The deviation threshold, This is the default setting. Different values ​​are used depending on the level: 0.10 for L1, 0.15 for L2, and 0.20 for L3. ; ; ; in, for Level deviation threshold, for Level deviation threshold, for Level deviation threshold, This represents the risk aversion coefficient, with a default value of 0.5 and a range of... , The larger the size, the more conservative the adjustment.

[0082] The triggering condition for deviation triggering is Specifically: This indicates that it has been triggered. class, This indicates that it has been triggered. class, This indicates that it has been triggered. For example, when γ=0.5, L1=0.085, L2=0.128, and L3=0.17.

[0083] Predictive Trigger: The autonomous mission decision engine continuously makes forward predictions about the state of the drone. Based on the latest prediction results from the battery state precision perception and prediction module, the prediction that the remaining energy cannot meet all remaining tasks is used as the second trigger state to satisfy the preset dynamic adjustment trigger conditions.

[0084] Assume the energy requirements of the remaining tasks. Energy requirements for return flight (Real-time estimation of wind field and altitude), safety reserve : ; in, Represents the predicted remaining energy The difference between the total number of tasks and all remaining tasks.

[0085] Safety Reserve The dynamic calculation formula is: ; Among them, coefficient ( hour, ), This refers to the battery's nominal capacity, measured in Wh, and is determined by the battery specifications.

[0086] Predicted trigger conditions: Trigger; Hierarchy: L1: → Slight energy deficiency; L2: → Moderate energy deficiency; L3: → Severe energy deficiency.

[0087] Environmental mutation trigger: The autonomous task decision engine continuously monitors key environmental parameters. When the change of the key environmental parameters exceeds the initially planned threshold range, it is used as the third trigger state to meet the preset dynamic adjustment trigger conditions. The key environmental parameters include wind speed and wind direction. For example, the multi-source environment and task perception module detected that key environmental parameters (such as wind speed and wind direction) changed drastically, and the magnitude of the change exceeded the tolerance range of the initial plan.

[0088] Wind speed change threshold (adjusted based on risk appetite): Headache changes : ; Total wind speed change : ; Wind shear / gust variance : , Represents a time window, for example ; For wind speed, This is the wind field variance threshold, with a default value of 2.0 m² / s², used to detect wind shear / gusts.

[0089] when Default value: ; ; .

[0090] In this embodiment, the formula for calculating the environmental change threshold in environment triggering is: ; in, For the new, dynamically adjusted environmental change thresholds, This is the default threshold for environmental changes.

[0091] The wind speed change threshold set above is compared to determine whether the environmental triggering conditions are met.

[0092] External command trigger: The system receives an external command from the ground station or operator that requires an immediate reassessment of the mission as the fourth trigger state, which satisfies the preset dynamic adjustment trigger conditions.

[0093] In one embodiment, the action priority and specific execution strategy of the above-mentioned various triggering states (deviation triggering, prediction triggering, environmental change triggering, external instruction triggering) are as follows: L1 (minor adjustment): Fine-tuning flight parameters: speed adjustment range ±15%, altitude adjustment range ±20m; Limit non-critical payloads: disable image transmission, reduce camera sampling rate, and limit other sensors; Execution time: Execute immediately, no confirmation required.

[0094] L2 (Local Replanning): Triggering local path replanning: Recalculate the optimal path from the current waypoint to the next waypoint; Evaluate task reordering alternatives: calculate the energy requirements of different task sequences and select the optimal solution; Execution time: Replanning completed within 5-10 seconds.

[0095] L3 (Mission-level adjustment): Perform task-level adjustments: Skip the lowest priority tasks and reorder the remaining tasks; Return to base or land on the spot if necessary: ​​when M is less than Forced return to home at 0.15E_nom; Execution time: Execute immediately, with a delay of less than 2 seconds.

[0096] In one embodiment, when multiple triggering conditions are met simultaneously, execution is performed according to the following priority: Priority 1: L3 trigger (highest priority, executes immediately); Priority 2: L2 trigger (medium priority, to be executed within 5 seconds); Priority 3: L1 trigger (lowest priority, to be executed within 10 seconds).

[0097] Additionally, handling conflicting trigger conditions: If L3 and L1 / L2 are triggered simultaneously, ignore L1 / L2 and execute the L3 action directly. If L2 and L1 are triggered simultaneously, L2 will be executed first, followed by L1 (if L1 is still triggered after L2 is executed).

[0098] Among the various triggering states mentioned above (deviation triggering, prediction triggering, environmental change triggering, and external command triggering), The range of values ​​is In this embodiment, The recommended value is: Aggressive strategies are suitable for short-distance, low-risk tasks. Balanced strategy, suitable for routine inspection tasks (recommended default value); A conservative strategy, suitable for long-distance, high-risk missions.

[0099] (c3) Flight control command generator; The decision output is transformed into specific instructions that the underlying flight controller can execute; the generated instructions are then verified for safety and feasibility.

[0100] Data subscription: Subscribe to the battery status prediction results of the battery status accurate perception and prediction module and the environmental situation information of the multi-source environment and task perception module; Command issuance: Issues power management commands to the airborne equipment and payload power management module, and flight control commands to the flight controller.

[0101] Status feedback: Receives execution status feedback from the underlying actuators, which is used for communication mechanisms between decision optimization modules to achieve closed-loop control optimization.

[0102] IV. Airborne Equipment and Payload Power Management Module; (d1) Module functions; The airborne equipment and payload power management module is one of the energy-saving execution units of the system. Based on the instructions of the energy-sensing intelligent mission and flight control module, it provides power to other airborne equipment or mission payloads (such as non-critical sensors, image transmission modules, etc.) on demand or switches them to low-power modes.

[0103] (d2) Power management strategy; Dynamic power allocation: Dynamically allocate the power budget for each load based on task priority and current energy status; Intelligent sleep management: Implement intelligent sleep strategies for non-critical loads to extend system battery life; Power limiting control: Automatically limits the power consumption of non-critical loads when energy is insufficient; Emergency mode switching: In an emergency, quickly switch to the minimum power consumption mode.

[0104] (d3) Load classification and management; Key payloads: Payloads that maintain basic flight functions, such as flight control, navigation and positioning, and communication links; Mission payload: Payloads such as cameras, sensors, and image transmission devices that perform specific tasks; Auxiliary loads: Loads that provide auxiliary functions such as lighting, heat dissipation, and entertainment.

[0105] (d4) Inter-module communication mechanism; Command reception: Receives power management commands from module C; Status feedback: Feedback the load power consumption status and execution results to module C; Load control: Directly controls the power management and operating mode of each load.

[0106] V. System Integration and Collaboration Mechanism; (e1) Data flow between modules: Real-time data stream: High-frequency sensor data (50-100Hz) is transmitted in real time between modules; Control command flow: Decision results and control commands are passed from the decision-making module to the execution module; Status feedback flow: Execution status and performance metrics are fed back from the execution module to the decision-making module; Configuration parameter stream: System configuration and policy parameters are updated synchronously across modules.

[0107] (e2) Collaborative working mechanism: Data synchronization: Each module keeps its data synchronized through a timestamp mechanism to ensure consistency in decision-making; State coordination: Modules coordinate their working states through a state machine mechanism to avoid conflicts; Fault isolation: A failure in a single module does not affect the normal operation of other modules; Performance monitoring: Monitor the performance metrics of each module in real time to ensure stable system operation.

[0108] (e3) System performance indicators: Response latency: The total latency from sensor data acquisition to control command execution is less than 200ms; Prediction accuracy: Battery state prediction accuracy is greater than 90%, and path energy consumption prediction accuracy is greater than 85%; Decision frequency: The autonomous task decision engine operates at a frequency of 10-20Hz, supporting real-time decision-making; System reliability: The system availability under normal conditions is greater than 99.5%.

[0109] These configurable strategies enable the system in this embodiment to flexibly adapt to the specific needs of different users and different task scenarios.

[0110] This embodiment constructs a system-level solution that deeply couples battery physics modeling, AI prediction, path planning, and task decision-making, possessing high technical barriers and comprehensive competitive advantages.

[0111] Example 1; This embodiment discloses an adaptive energy efficiency management method for unmanned aerial vehicles (UAVs), which is a dynamic, closed-loop optimization process, specifically including: S100 Before the task begins, the system first executes the battery status accurate perception and prediction module and the multi-source environment and task perception module. Battery Status Precision Sensing and Prediction Module: The system continuously collects real-time BMS data, flight logs, and historical battery data; using a built-in machine learning model, it predicts the battery's accurate remaining usable energy, voltage sequence, and performance online.

[0112] Multi-source environment and mission awareness module: The system integrates airborne sensor data and external data source information in real time to construct a dynamic environment and mission situation map that includes wind field, terrain and mission points (including priority).

[0113] S200: Based on the initial battery and environmental conditions, the system determines whether the mission has started. If it has started, it invokes the energy-optimal initial path planning to generate a global, energy-lower initial flight path. In this process, the current battery status and environmental conditions obtained in step S100 are used to plan the initial flight path with the optimal global energy consumption.

[0114] The S300 drone begins its flight mission. From this point on, the system enters a continuous in-flight closed-loop adaptive energy efficiency optimization cycle.

[0115] During flight, the S400 continuously updates battery status predictions and environmental conditions. The autonomous mission decision engine monitors the deviation between actual and predicted energy consumption in real time and makes forward predictions of future energy status.

[0116] S500: The system continuously determines whether the triggering conditions for dynamic adjustment are met; When the preset dynamic adjustment trigger conditions are met (such as deviation exceeding the threshold, insufficient predicted energy, strong headwinds leading to a surge in energy consumption, etc.), the autonomous mission decision engine executes dynamic adjustment decisions, including but not limited to fine-tuning flight parameters, triggering path replanning, or dynamically adjusting the mission sequence / target.

[0117] S600. When the triggering condition is met, the autonomous decision engine will execute dynamic adjustments and generate and issue optimized control commands, and proceed to step S700; when the triggering condition is not met, proceed directly to step S700. S700: The system checks whether the task has been completed or whether a return to base is required. If the mission is not completed and no return is required, return to step S300 to continue the flight mission and repeat the optimization loop. If the mission is completed or the flight has returned safely, the process ends.

[0118] Example 2; Typical scenario of autonomous task decision engine; like Figure 4 As shown, the autonomous decision-making engine is key to achieving advanced intelligence. This embodiment details the complete decision-making process of a UAV encountering an unforeseen strong headwind during a typical long-endurance inspection mission.

[0119] Scene setting: Mission type: Long-endurance inspection mission, total range 120km, with 8 pre-set inspection points; Initial conditions: Battery SoE = 5%, SoH = 92%, ambient temperature 25℃; Unexpected situation: After flying to the 3rd inspection point, the flight encountered a strong headwind, with the wind speed suddenly increasing from 5 m / s to 12 m / s; Time constraint: There are 3 hours left until sunset. The core mission must be completed and the team returned before dark.

[0120] (f1) Status update and deviation detection; Battery status anomaly detection: The battery state prediction module detected the actual discharge current. consistently higher than model predictions ; Five consecutive measurements showed: ; Deviation triggering conditions are met: ; The predicted SoE will drop sharply: from 85% to 72% (within 3 minutes).

[0121] Environmental mutation confirmed: The environmental perception module confirms a strong headwind using data from the IMU and pitot tube. The pitot tube, obtained by a sensor measuring relative airflow velocity, is used to estimate... and .

[0122] Wind speed changes: ,in, It is the set environmental trigger threshold; Wind shear detection: ; Trigger level: L3 (Severe environmental mutation).

[0123] (f2) Forward forecasting and risk assessment: The autonomous task decision engine is activated, and forward simulations are performed across multiple time scales.

[0124] Short-term forecast (5-15 minutes): Current path energy consumption : ; Predict the remaining SoE: , ;in, For the predicted remaining SoE for the next 5 minutes, The remaining SoE for the predicted next 15 minutes; Risk assessment: R_short=0.15 (medium risk).

[0125] Medium-term forecast (15-60 minutes): Energy required to complete the remaining 5 inspection points: ; Return energy requirements: (Consider strong headwinds); Safety Reserve: , ( hour); Total demand: ; risk assessment: (High risk).

[0126] Long-term forecast (60-180 minutes): Task completion probability: =0.12 (extremely low); Probability of safe return: =0.34 (low); risk assessment: =0.89 (extremely high risk).

[0127] (f3) Utility function optimization and decision generation; The autonomous task decision engine is based on the MPC framework. According to the preset "maximize task coverage" strategy and the task point priority defined below, it evaluates and selects the optimal task-level adjustment scheme by optimizing the utility function shown below.

[0128] Task priority weight definition: Inspection Points 1-2: (Completed, weights reset to zero); Inspection point 3: (Current point, high priority); Inspection points 4-5: (Core task point, highest priority); Inspection points 6-8: (Secondary task point, low priority).

[0129] The utility function is defined as: ; in, Adjust the strategy to achieve the optimal task. The set of remaining tasks. For the task Priority weights, satisfying , In strategy and remaining available energy Next, complete the task. The task completion probability function This represents the risk aversion coefficient. This function represents the risk of task failure.

[0130] The calculation formula is as follows: ; in, To complete the task Energy demand, This represents the influence coefficient of environmental factors (wind speed, temperature, etc.). This refers to the battery state influence coefficient (SoH, temperature, etc.).

[0131] Task failure risk function The formula is as follows: ; in, This is the critical energy threshold (the minimum energy required for a safe return). As a wind speed risk factor, Temperature risk factor This is the risk weighting coefficient.

[0132] Evaluation of decision-making options; Option 1: Continue with the original plan. Task completion probability: ; Probability of safe return: ; Utility value: .

[0133] Option 2: Skip inspection points 6-8 and complete the core task; Task completion probability: ; Probability of safe return: ; Utility value: .

[0134] Option 3: Return immediately and abandon the remaining mission; Task completion probability: ; Probability of safe return: ; Utility value: .

[0135] Option 4: Dynamic path replanning to bypass strong wind areas; Task completion probability: ; Probability of safe return: ; Utility value: .

[0136] Optimal decision: Option 2 (skip inspection points 6-8 and complete the core task).

[0137] (f4) Decision implementation and verification; Decision-making and implementation steps: Step 1: Task reordering; Retain 3-5 inspection points (core task); Skip inspection points 6-8 (minor tasks); Update the mission sequence: 3→4→5→Return.

[0138] Step 2: Path replanning; Trigger the path planner to recalculate the optimal path; Considering the impact of strong headwinds, a low-energy-consumption route should be selected. Planning time: less than 100ms (meets real-time requirements).

[0139] Step 3: Load Management; Shut down non-critical loads: image transmission module, auxiliary sensors; Reduce camera sampling rate: from 30fps to 10fps; Expected power savings: 12.5W.

[0140] Step 4: Flight parameter optimization; Speed ​​adjustment: Reduce from 18m / s to 15m / s (to reduce energy consumption); Altitude optimization: Select the optimal cruising altitude (to reduce wind resistance); Climb rate limit: .

[0141] Decision verification and monitoring: Real-time verification metrics: Energy consumption deviation: ,(Target); Task progress: Core task completion rate > 90%; Safety margin: (Before returning to port).

[0142] Exception handling mechanism: Three consecutive verification failures trigger a decision reassessment. Energy consumption deviation greater than 15%: downgrade to conservative mode; Sudden environmental change: Immediate replanning.

[0143] (f5) Decision results and performance evaluation: Execution result: Core task completion rate: 100% (inspection points 3-5 all completed); Overall task completion rate: 62.5% (5 / 8 inspection points); Safe return: Successful, final SoE = 23%; Total flight time: 18% shorter than originally planned.

[0144] Performance metrics: Decision response time: 45ms (meets the requirement of less than 100ms); Path replanning time: 78ms (meets the requirement of less than 100ms); Energy consumption prediction accuracy: 92.3% (meets the requirement of greater than 90%); Task success rate: 100% (core task).

[0145] (f6) System deployment architecture and collaboration mechanism: The system in this embodiment adopts a cloud-edge-device collaborative architecture to achieve optimal allocation of computing resources: Cloud (Ground Station / Cloud Server): Functions: Offline training and updating of AI models, global multi-drone mission collaborative planning; Computing power: GPU cluster, supporting large-scale parallel computing; Response time: seconds (non-real-time tasks); Data storage: PB-level historical data, model parameter library.

[0146] Edge computing platform (airborne computing platform): Functions: Online status prediction, real-time path planning, and task decision-making; Computing power: ARM A78 / Apple M series or NVIDIA Orin Nano; Response time: milliseconds (real-time decision-making); Data caching: GB-level real-time data and model parameters.

[0147] Terminal (underlying flight control / BMS): Functions: Instruction execution and data acquisition; Computing power: MCU level, real-time control; Response time: microseconds (control commands); Data acquisition: raw sensor data and execution status.

[0148] Collaboration mechanism: Data synchronization: Real-time transmission via 5G / WiFi 6 with a latency of less than 10ms; Task allocation: dynamic load balancing, complex optimization in the cloud, and real-time decision-making at the edge; Failover: When the edge device goes offline, the cloud takes over the decision-making; when the network is interrupted, the edge device operates autonomously. Model Update: After training in the cloud is completed, the model parameters at the edge are updated incrementally.

[0149] (f7) Security certification and compliance; Decision security verification: Formal verification: Using a model checker to verify the correctness of the decision logic; Simulation verification: Verify all decision-making scenarios in a digital twin environment; Flight verification: Verify system performance in a controlled environment.

[0150] Human-machine collaboration mechanism: Decision-making transparency: All decision-making processes are traceable and support human review; Manual intervention: Operators can take over control at any time; Decision confirmation: High-risk decisions require manual confirmation (such as abandoning the task).

[0151] Performance monitoring and logging: Decision log: Records all decision-making processes, input data, and output results; Performance metrics: Real-time monitoring of decision latency, accuracy, and success rate; Abnormal alarm: Abnormal situations will automatically trigger an alarm and be recorded.

[0152] The systems and methods described in the above embodiments can be implemented by a computer program. This computer program can be stored on a computer-readable storage medium. When the program is executed by a processor, the steps of the methods described in this invention can be implemented. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0153] Furthermore, this embodiment also provides a drone device or computer device, which includes a processor and a memory. The memory stores a computer program, and the processor is configured to execute the computer program to implement the steps of the method described in this invention. The drone device may be an autonomous drone, and the computer device may be an onboard computing module or a ground control station.

[0154] Example 3: Engineering Implementation and Real-Time Budgeting; This embodiment details the engineering implementation scheme of the UAV adaptive energy efficiency management system, including key technical details such as hardware and software platform specifications, real-time performance requirements, data management strategies, and system integration verification.

[0155] (g1) Hardware platform specifications and classifications, including (h1) to (h3): (h1) Small unmanned aerial vehicle platform (less than 5kg): Main control chip: multi-core ARM processor, clock speed 2.0-2.8GHz, 4-8 cores; Memory configuration: 8-16GB LPDDR4, high-bandwidth memory system; Storage system: Embedded storage + expandable memory card; Power consumption budget: Total system power consumption less than 50W, computing platform power consumption less than 15W; Cooling solution: passive cooling + small cooling fan, operating temperature -20°C to +60°C.

[0156] (h2) Medium-sized unmanned aerial vehicle platform (5-25kg): Main control chip: High-performance multi-core processor, clock speed 2.4-3.2GHz, 8-10 cores; Memory configuration: LPDDR5 16-32GB, high-bandwidth memory system; Storage system: High-speed solid-state storage + scalable storage; Power consumption budget: Total system power consumption less than 200W, computing platform power consumption less than 40W; Cooling solution: Active cooling system, liquid cooling optional, operating temperature -40°C to +70°C.

[0157] (h3) Large unmanned aerial vehicle (UAV) platform (greater than 25kg): Main control chip: High-performance multi-core processor, clock speed 2.8-4.0GHz, 8-16 cores; Memory configuration: LPDDR5 32-64GB, ultra-high bandwidth memory system; Storage system: High-speed solid-state storage + redundant storage configuration; Power consumption budget: Total system power consumption less than 500W, computing platform power consumption less than 100W; Cooling solution: Forced air cooling + liquid cooling system, operating temperature -50°C to +80°C.

[0158] (g2) Software architecture and middleware, including (i1) through (i3); (i1) Operating System and Real-Time Performance: Main operating system: Linux distribution or embedded operating system; Real-time kernel: A kernel that supports real-time scheduling with a maximum interrupt latency of less than 50μs; Containerization solution: Supports containerized deployment and GPU acceleration; Virtualization support: Supports virtualization technology (large platforms) and lightweight containers (small and medium-sized platforms).

[0159] (i2) Middleware configuration: Communication protocol: Distributed Data Service (DDS) + UAV communication protocol; QoS configuration: Real-time data is transmitted using best-effort transmission, while control commands are transmitted using reliable transmission; Message serialization: Supports efficient data serialization and zero-copy transmission.

[0160] (i3) Detailed specifications for topic configuration: High-frequency sensor data (50-100Hz): Battery management system status data, inertial measurement unit data, global positioning system data, and wind field estimation data.

[0161] Intermediate frequency control data (10-20Hz): Battery status prediction results, current path information, and task execution status.

[0162] Low-frequency command data (1-5Hz): Flight plan commands, mission adjustment commands, system health status.

[0163] (g3) Real-time performance metrics and constraints; Battery state prediction path: Data acquisition: BMS data reading takes less than 1ms; Feature extraction: Signal processing time less than 2ms; Model inference: The prediction algorithm executes in less than 10ms; Output results: Prediction results are published in less than 1ms; Total latency: less than 15ms from end to end.

[0164] Path planning path: Environmental perception: Sensor fusion in less than 5ms; Map update: Dynamic map building takes less than 10ms; Path search: The path search algorithm executes in less than 100ms (local replanning). Path optimization: Energy cost optimization in less than 50ms; Total latency: less than 165ms from end to end.

[0165] Task decision path: Status assessment: Multi-source data fusion time is less than 10ms; Risk calculation: Risk assessment takes less than 20ms; Decision optimization: Utility function optimization takes less than 200ms; Instruction generation: Control instruction generation takes less than 10ms; Total latency: less than 240ms end-to-end.

[0166] Performance distribution and worst-case analysis: Delay distribution requirements (P95 / P99): Predictive inference: P95 is less than 12ms, P99 is less than 15ms; Path planning: P95 less than 150ms, P99 less than 200ms; Task decision: P95 less than 300ms, P99 less than 400ms.

[0167] Jitter control: Predictive inference jitter: less than 2ms RMS; Path planning jitter: less than 10ms RMS; Task decision jitter: less than 20ms RMS.

[0168] (g4) Data management and communication infrastructure; Data compression and transmission: Sensor data: lossless compression algorithm, compression ratio 2:1-3:1; Image data: lossy compression algorithm, compression ratio 10:1-20:1; Prediction result: Incremental coding reduces redundant data transmission; Transmission protocol: A transmission protocol that supports packet loss recovery.

[0169] Cloud-edge-device communication specifications: Communication bandwidth: Supports high-bandwidth wireless communication; Latency requirements: End-to-end latency less than 10ms (edge ​​computing), less than 50ms (cloud computing); Reliability: High connectivity availability and data transmission success rate.

[0170] Data persistence strategy: Local storage: a circular buffer that retains data from the last 24 hours; Edge caching: Recently used strategy, retains data from the last 7 days; Cloud storage: Tiered storage strategy; Backup strategies: incremental backup and full backup strategies.

[0171] (g5) System integration and verification; Unit testing phase: Module functional testing: Coverage greater than 90%, boundary condition testing; Performance benchmarks: latency, throughput, resource usage; Interface compatibility testing: message format and protocol version compatibility.

[0172] Integration testing phase: Inter-module communication testing: topic publishing / subscription, service calls; End-to-end functional testing: verification of the complete workflow; Performance integration testing: concurrent performance testing of multiple modules.

[0173] System-level testing phase: Hardware-in-the-loop testing: System verification in a real hardware environment; Software-in-the-loop testing: system verification in a simulation environment; Semi-physical simulation testing: Partial real hardware + simulation environment.

[0174] Performance verification methods: Benchmarking: Performance measurement under standard workloads; Stress testing: Verification of system stability under extreme conditions; Long-term testing: Reliability verification through continuous operation for 24-72 hours; Regression testing: Performance regression verification after feature updates.

[0175] (g6) Certification and Compliance; Aviation safety certification: Airborne software airworthiness certification standards, airborne hardware airworthiness certification standards, system development process certification standards, and airworthiness requirements for unmanned aerial vehicle (UAV) systems.

[0176] Cybersecurity certifications include: information security management system certification, industrial automation and control system security certification, and cybersecurity framework compliance.

[0177] Environmental adaptability certification: Environmental testing methods and engineering guidelines, environmental conditions and testing procedures for airborne equipment, and guidelines for airworthiness certification of unmanned aerial vehicle systems.

[0178] (g7) Deployment and Operation; Deployment strategy: Blue-green deployment: Supports updates with zero downtime; Gray-scale release: Gradually expanding the usage of the new version; Rollback mechanism: Quickly revert to a stable version; Configuration management: Supports runtime configuration updates.

[0179] Monitoring and Alarms: Performance monitoring: Real-time monitoring of system performance metrics; Health check: Regularly check the status of system components; Abnormal alarm: Automatic alarm notification for abnormal situations; Log analysis: Structured log recording and analysis.

[0180] Maintenance and updates: Remote update: Supports over-the-air updates; Version management: Complete version release and rollback management; Patch management: Management of security patches and feature updates; Backup and recovery: System configuration and data backup and recovery.

[0181] (g8) System resource budget; Calculate resource budget: CPU resources: Supports multi-core parallel computing to meet real-time requirements; Memory resources: Supports large dataset processing and high-concurrency access; Storage resources: Supports high-speed data read and write and long-term data storage; Network resources: Supports high-bandwidth data transmission and low-latency communication.

[0182] Power consumption budget: Computing platform power consumption: Determine the power consumption budget based on the platform scale; Cooling system power consumption: The cooling requirements to support stable system operation; Communication system power consumption: The power consumption requirements to support data transmission and communication.

[0183] Performance budget: Real-time budget: To meet the real-time requirements of each functional module; Accuracy budget: To meet the requirements for forecasting and control accuracy; Reliability budget: To meet the requirements of system stability and reliability.

[0184] Example 4: Safety and Anomaly Handling; This embodiment details the safety mechanisms and anomaly handling strategies of the UAV adaptive energy efficiency management system to ensure that the system can operate safely and reliably under various abnormal conditions.

[0185] (j1) Sensor and model health monitoring, including (k1) to (k2); (k1) Residual gating mechanism: Voltage residual monitoring: Continuous monitoring of actual voltage With predicted voltage The deviation, when The degradation mechanism is triggered after K consecutive occurrences, where, To predict voltage sequences, This is the voltage deviation threshold.

[0186] Energy residual monitoring: Continuous monitoring of actual energy Energy predicted by the model The deviation, when The degradation mechanism is triggered after K consecutive occurrences, where, This is the energy deviation threshold.

[0187] Residual threshold setting: and Adjust dynamically according to battery specifications and system accuracy requirements.

[0188] (k2) Wind field credibility verification: Multi-sensor consistency verification: The reliability of wind field estimation is evaluated by verifying the consistency of data from multiple sources such as GNSS, IMU, and pitot tube. Inconsistency handling strategy: When data from multiple sensors is inconsistent, the safety margin is automatically increased and a conservative flight strategy is adopted; Credibility scoring mechanism: Establish a wind farm credibility scoring system and adjust decision-making strategies based on the scoring results.

[0189] (j2) System degradation strategy; Functionality degradation mechanism: Task-level adjustment disabled: When the system detects an anomaly, the task-level adjustment function is disabled, and only the basic return-to-home function is retained; Simplified path planning: The shortest energy consumption path planning is adopted, ignoring complex energy optimization strategies; Load management limits: Shut down non-critical loads to reduce system power consumption and extend battery life.

[0190] Flight parameter limitations: Speed ​​limit: Limit flight speed within a safe range to avoid excessive energy consumption fluctuations; Rate of ascent limit: Limits the rate of ascent and descent to reduce energy consumption; Altitude restriction: Limits changes in flight altitude to maintain a stable flight status.

[0191] (j3) Emergency response procedures; Low voltage emergency procedures: Hard threshold triggering: when If necessary, an emergency return to base or on-site landing should be triggered immediately. This is the discharge cutoff voltage. For the actual voltage being monitored, The voltage safety margin is set according to battery specifications and system requirements; Emergency route planning: Automatically plans the shortest and safest return route.

[0192] Emergency handling for low energy: Minimum return energy: when At that time, a forced return-to-base maneuver was executed, in which, For the energy required for the return flight, To ensure energy safety during the return journey, a minimum safety reserve of energy is required. The measured remaining available energy; Emergency Mode Activation: Activate emergency flight mode and adopt the most conservative flight strategy.

[0193] (j4) Anomaly detection and diagnosis; Anomaly pattern recognition: Sensor anomaly detection: Identifies hardware problems such as sensor malfunctions and data anomalies; Model anomaly detection: Identifying software problems such as prediction model bias and parameter anomalies; System anomaly detection: Identifies system problems such as performance degradation and insufficient resources.

[0194] Anomaly Classification and Handling: Minor anomalies: Adaptive adjustment strategies are employed to fine-tune system parameters; Moderate anomaly: Adopt a degradation strategy and disable some functions; Serious anomaly: Adopt emergency strategies and immediately return to base or land.

[0195] (j5) Log recording and tracing; Abnormal event log: Trigger condition recording: Records the specific conditions and thresholds for exception triggering; Decision-making process record: Record the decision-making logic and results during the exception handling process; Execution result recording: Records the execution result and system status after exception handling.

[0196] Data traceability and analysis: Historical data analysis: Analyze historical anomaly data to identify anomaly patterns and trends; Model update basis: Based on abnormal data, update and improve the prediction model; System optimization reference: Optimize system parameters and strategies based on anomaly analysis results.

[0197] Example 5: Experimental Design and Performance Verification; This embodiment details the experimental design method, performance verification indicators, and expected results of the UAV adaptive energy efficiency management system, providing a scientific basis for the practical application of the system.

[0198] (l1) Experimental data and baseline design; Experimental data source: Real flight data: Collects real drone flight logs, including various flight scenarios and environmental conditions; Experimental bench data: Simulate various scenarios such as temperature, wind field, and load through the experimental bench; Multi-scenario coverage: Ensure that experimental data covers a variety of real-world application scenarios.

[0199] Baseline system design: Baseline 1: A simple return-to-home strategy based solely on SoC thresholds, serving as the basis for performance comparison; Baseline 2: Energy-sensing path planning without battery state prediction, to verify the effectiveness of path planning; Baseline 3: End-to-end reinforcement learning strategy, compared with the performance of traditional AI methods.

[0200] (l2) Performance verification metrics; Predictive performance metrics: SoE Prediction Accuracy: RMSE and MAE metrics are used to assess the accuracy of energy predictions; Voltage series prediction: DTW distance index, used to assess the similarity of voltage series predictions; Internal resistance estimation accuracy: a relative error index used to assess the accuracy of internal resistance prediction.

[0201] Task execution metrics: Improved range: Percentage increase in range and time relative to baseline; Task completion rate: The success rate of task completion under various abnormal conditions; Safe return rate: The success rate of safe return under various abnormal conditions; Energy excess rate: δ_E distribution, used to assess the stability of energy prediction.

[0202] Real-time performance metrics: Prediction latency: The execution latency of the prediction algorithm, P50 / P95 quantile; Planning delay: Execution delay of the path planning algorithm, P50 / P95 quantile; Decision delay: The execution delay of the task decision algorithm, P50 / P95 quantile.

[0203] Here, P50 / P95 are quantile indicators, representing the median and 95th percentile of the delay / error distribution.

[0204] (l3) Experimental scenario design; Changes in environmental conditions: Temperature variation: Testing system performance under different ambient temperatures; Wind field variations: Testing system performance under different wind speeds and directions; Load variation: Test system performance under different load configurations.

[0205] Abnormal situation simulation: Sensor malfunction: Simulates abnormal or faulty sensor data; Sudden environmental changes: Simulate sudden environmental changes, such as strong winds or a sudden drop in temperature; System anomaly: Simulates a situation where the system performance degrades or resources are insufficient. (l4) Expected results and performance targets; Performance improvement goals: Battery life: Compared to baseline 1, battery life is improved by 10-30%; Task success rate: The task completion rate is significantly improved under various abnormal conditions; Safe return rate: Under various abnormal circumstances, the safe return rate is close to 100%.

[0206] System stability objective: Energy prediction stability: The δ_E distribution converges, and the prediction error is within an acceptable range; Real-time performance stability: All latency metrics meet real-time requirements; Anomaly handling capability: The system can operate stably and safely under various abnormal conditions.

[0207] (l5) Verification methods and procedures; Verification process design: Unit testing: Independent testing of each functional module; Integration testing: Testing the interfaces and interactions between modules; System testing: Perform end-to-end testing on the entire system; Performance testing: Verifying the system's performance metrics.

[0208] Verification environment requirements: Simulation environment: Initial verification is performed in the simulation environment; Semi-physical environment: Verification is performed in a partially real hardware environment; Real-world environment: Final verification is conducted in a real flight environment.

[0209] (l6) Results analysis and improvement; Data analysis methods: Statistical analysis: Perform statistical analysis on experimental data to identify performance trends; Comparative analysis: Comparative analysis with the baseline system quantifies performance improvements; Anomaly analysis: Analyze system performance under abnormal conditions and identify areas for improvement.

[0210] Continuous improvement mechanism: Model Updates: Based on experimental data, the prediction model is continuously updated and improved; Parameter optimization: Based on experimental results, optimize system parameters and configuration; Strategy Improvement: Based on experimental feedback, improve decision-making strategies and anomaly handling mechanisms.

[0211] Through the above embodiments, this embodiment demonstrates how to achieve in-depth, closed-loop, and adaptive management of UAV energy efficiency through an integrated intelligent system, thereby reliably completing tasks in complex and dynamic environments.

[0212] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive energy efficiency prediction and optimization management method for unmanned aerial vehicles (UAVs), characterized in that, include: Before the mission begins, the initial flight path with the optimal global energy consumption is initially predicted based on the current battery status and environmental conditions. During flight, the current battery status and environmental conditions are continuously updated. The intelligent mission and flight control module continuously monitors various trigger states. When one of the trigger states meets the preset dynamic adjustment trigger conditions, a dynamic adjustment decision is executed and an optimization command is generated and sent to the flight controller for execution.

2. The management method according to claim 1, characterized in that, The process of generating the current battery state is as follows: By integrating electrochemical models or empirical formulas as physical constraints, we can construct AI prediction models driven by both physics and data. The current battery status is obtained by using real-time BMS data, historical flight logs, and charge / discharge records as inputs to the AI ​​prediction model.

3. The management method according to claim 2, characterized in that, The physical constraints of the AI ​​prediction model are as follows: Energy conservation constraint: ; Voltage-current consistency: ; Piecewise monotonic constraints on internal resistance and temperature / aging; Discharge rate consistency constraint; in, For predicted future moments The remaining available energy, For the present The remaining available energy at any given moment. The current battery status is output by the battery status accurate sensing and prediction module. The discharge efficiency of the battery is a function of the current battery state. For at any time The total load power of the drone Let be the integral variable, representing the integral variable at the time of integration. At any instant within the time period, For an infinitesimal time increment, For load current, This is the load voltage. This is the equivalent internal resistance of the battery. This represents the current remaining battery percentage. It is an open-circuit voltage function. and temperature The function.

4. The management method according to claim 2, characterized in that, The AI ​​prediction model is an LSTM network-based model, and the loss function during the training process of the AI ​​prediction model is... The settings are as follows: ； in, To predict the mean squared error loss between SoE and labeled SoE, To predict the loss between the voltage series and the measured series, To predict the loss between the equivalent internal resistance and the estimated equivalent internal resistance, This is a physical constraint penalty term, which corresponds to the physical constraints of the AI ​​prediction model. For time smoothing regularization, For weight decay regularization, For physical constraint weights, To smooth out the regularization weights, This is the weight decay factor.

5. The management method according to claim 2, characterized in that, The intelligent mission and flight control module continuously monitors various trigger states, specifically: Using the current battery state and environmental conditions as input, the energy-efficient path planner employs energy expansion... Algorithm construction cost function Based on this, a local flight path with optimal energy consumption is generated; The autonomous mission decision engine continuously monitors the energy consumption deviation between the local flight path and the initially predicted initial flight path, and uses the energy consumption deviation exceeding the deviation threshold as the first trigger state to meet the preset dynamic adjustment trigger conditions. The autonomous mission decision engine continuously makes forward predictions about the state of the drone, and uses the prediction that the remaining energy cannot satisfy all the remaining tasks as the second trigger state to meet the preset dynamic adjustment trigger conditions. The autonomous task decision engine continuously monitors key environmental parameters and uses the change of the key environmental parameters exceeding the initially planned threshold range as the third trigger state to meet the preset dynamic adjustment trigger conditions. The key environmental parameters include wind speed and wind direction. The autonomous mission decision engine receives requests from ground stations or operators, requiring an immediate reassessment of external mission instructions as a fourth trigger state to meet preset dynamic adjustment trigger conditions.

6. The management method according to claim 5, characterized in that, In the phrase "using energy consumption deviation exceeding a deviation threshold as the first trigger state to satisfy a preset dynamic adjustment trigger condition", the following is included: Energy consumption deviation The formula is as follows: ； Deviation threshold The formula is as follows: ； in, This represents the cumulative energy consumption to date. The predicted cumulative energy consumption for the initial planning. Here is the numerical stability constant. This is the default setting; different values ​​are set depending on the level. This represents the risk aversion coefficient.

7. The management method according to claim 5, characterized in that, The cost function The formula is as follows: ； in, For the node To the node The corresponding number Each path segment For the flight path segment The prediction time function, For the speed, wind speed and distance of the drone, For the flight path segment The predicted energy consumption function, Due to high degree of variation, The current battery status is output by the battery status accurate sensing and prediction module. and As a weighting factor, satisfying .

8. The management method according to claim 7, characterized in that, The predicted energy consumption function Includes the following power components: Base power consumption : ; Aerodynamic losses : ; Climb / descent losses : ; Load power consumption : ; Battery loss : ; in, Based on the fundamental power consumption constant, For temperature coefficient, For reference temperature, The current temperature. This is the second-order drag coefficient. It is the third-order drag coefficient. Relative airflow velocity For takeoff mass, For the climb rate, To improve efficiency, The climbing drag coefficient, For load The on / off state, For load Rated power, For current, The equivalent internal resistance of the battery prediction module is given. For thermal efficiency.

9. The management method according to claim 8, characterized in that, , , , , Offline calibration coefficient range: ； ； ； ； 。 10. The management method according to claim 8, characterized in that, , , , , Online adaptive strategy: The coefficients are calibrated online using the sliding window least squares method. Calibration trigger condition: Prediction error greater than 5% for 10 consecutive times; Calibration frequency: Performed once every 100 predictions; Coefficient change limit: The adjustment range for a single adjustment is less than 10%.

11. The management method according to claim 5, characterized in that, The process of executing dynamic adjustment decisions and generating optimization instructions to be issued to the flight controller includes the autonomous mission decision engine, which is based on the MPC framework. According to the preset "maximize mission coverage" strategy and mission point priority, it evaluates different mission-level adjustment schemes in combination with utility functions, and then executes dynamic adjustment decisions and generates optimization instructions accordingly. The utility function is: ； in, Adjust the strategy to achieve the optimal task. The set of remaining tasks. For the task Priority weights, In strategy and remaining available energy Next, complete the task. The task completion probability function This represents the risk aversion coefficient. This function represents the risk of task failure.

12. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-11.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a plurality of classification programs, which are used by a processor to execute the method as described in any one of claims 1-11.

Images