Adaptive UAV charging method and system based on multi-source perception and artificial intelligence decision-making

Abstract

This invention discloses an adaptive UAV charging method and system based on multi-source perception and artificial intelligence decision-making, belonging to the field of UAV ground support facilities and automatic control technology. The method includes: a spatialized sensor array collecting micro-environment data and fusing it with weather forecasts, constructing a micro-environment field through dynamic weights; constructing a GRU network with temporal features as input, outputting comprehensive risk probability and hidden state features, and simultaneously outputting weighted fusion values ​​of sub-risks such as thermal runaway, overvoltage, and communication anomalies; generating charging current, pulse frequency, and heat dissipation power actions based on a state space containing comprehensive risk probability and hidden states using deep reinforcement learning; arbitrating actions for safety through a three-level rule base and fuzzy logic conflict resolution; and reinforcing online learning with safety experience playback and elastic weights. This invention achieves a paradigm shift from passive response to active prediction, significantly improving charging efficiency and battery life while ensuring safety, and exhibits high adaptability and robustness.

Description

Technical Field

[0001] This invention relates to the field of UAV ground support facilities and automatic control technology, specifically to an adaptive UAV charging method and system based on multi-source perception and artificial intelligence decision-making, which is applicable to achieving safe, efficient, and adaptive intelligent charging management in UAV automated airports or fixed hangars. Background Technology

[0002] With the large-scale application of drones in logistics delivery, power line inspection, agricultural monitoring and other fields, unattended drone automatic charging stations have become key facilities for maintaining continuous operation.

[0003] A search revealed that the existing patent document CN117360839A (Adaptive Omnidirectional Anti-Offset Wireless Charging System for Drones) primarily achieves anti-offset charging through the structural design of the transmitting coil and the judging coil, as well as the switching of efficiency thresholds. However, its decision-making is based solely on the charging efficiency threshold, without involving multi-source environmental perception or dynamic optimization using artificial intelligence. The safety strategy remains passive and reactive. Furthermore, existing charging solutions generally suffer from the following deficiencies: First, the safety protection mechanism is passive and lagging, often employing a "perception-threshold-action" model, only executing an emergency shutdown after detecting that the battery temperature exceeds a fixed threshold (e.g., 60°C), failing to predict and prevent thermal runaway. Second, the charging strategy is rigid, mostly using preset constant current-constant voltage curves or making only limited fine-tuning based on battery temperature, ignoring differences in external meteorological conditions (temperature, humidity, wind speed) and individual battery health status. Third, the perception method has blind spots; a single sensor or simple superposition cannot accurately construct the micro-environmental field within the charging chamber. Fourth, intelligent decision-making is mostly based on a single model or simple superposition, lacking deep collaboration between rules and learning mechanisms.

[0004] Therefore, there is an urgent need for a new charging solution that can achieve precise spatial perception, dynamic fusion of multi-source data, and collaborative decision-making through multiple mechanisms. Summary of the Invention

[0005] This invention aims to provide an adaptive drone charging method and system based on multi-source perception and artificial intelligence decision-making, which can achieve a fundamental shift from passive response-based safety to proactive prediction and dynamic optimization-based protection.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows: On the one hand, an adaptive drone charging method based on multi-source perception and artificial intelligence decision-making is provided, including the following steps: Step 1, collecting micro-environmental data through multiple temperature and humidity sensors deployed inside and outside the charging chamber, and acquiring weather forecast data and real-time drone battery status data; determining the current charging stage based on the changes in induction coil voltage, battery state of charge change rate, and real-time risk probability during the charging process, wherein the charging stage includes landing stage, stable charging stage, and risk mode; using a dynamic weighted fusion algorithm based on charging stage and real-time risk probability to fuse macro-weather forecast data and micro-sensor data to generate high-precision micro-environmental field data; Step 2, constructing a temporal feature vector for drone charging safety, inputting it into a pre-trained gated recurrent unit network, and outputting a comprehensive risk probability value and a gated recurrent unit hidden state feature vector; Step 3, constructing a system containing the comprehensive risk probability value and the gated recurrent unit hidden state feature vector. The hidden state feature vector of the unit and the state space vector of the conventional physical quantities are used to generate adaptive charging strategy actions using a deep reinforcement learning model. The actions include at least the charging current rate, charging pulse frequency, and active cooling fan power. The reward function R of the reinforcement learning is designed as a multi-objective weighted sum, and the weights of each sub-objective are dynamically adjusted. Step 4: A three-level rule base including hard limit protection rules, arbitration rules, and suggestion rules is established to perform mandatory verification and correction on the preliminary strategy output by reinforcement learning. When multiple rules conflict, a conflict resolution strategy based on fuzzy logic is used to calculate the comprehensive confidence level and output the final control command. Step 5: Historical execution experience is stored according to safety level, high-risk samples are oversampled, model updates are dynamically triggered based on performance indicators, and an elastic weight consolidation algorithm is used to prevent catastrophic forgetting during online learning. The final command is executed and the experience data is stored in the safety experience replay pool.

[0007] On the other hand, an adaptive drone charging system based on multi-source perception and artificial intelligence decision-making is provided, including a spatial perception module, a real-time risk assessment module, a dynamic strategy optimization module, a hierarchical safety arbitration module, and a closed-loop learning module, to implement the above-mentioned method.

[0008] In the comprehensive risk assessment based on gated cyclic units, this invention further discloses that: the GRU network not only outputs a single risk probability, but also simultaneously outputs estimated values ​​and corresponding adaptive weights for multiple risk components such as thermal runaway risk, overvoltage risk, and communication anomaly risk, and finally obtains the results through weighted fusion. P risk This improves risk interpretability and decision robustness. Specifically: Thermal runaway risk component R thermal Based on battery temperature change rate dt / dt and absolute temperature T battGaussian membership function is used. Calculate, where c A Based on typical thermal runaway critical values; overpressure risk component. R overvoltage Based on voltage V and safety upper limit V max The ratio, combined with the voltage change rate dV / dt, is quantized using a logistic function; communication anomaly risk component. R comm The bit error rate (BER) is calculated based on the proportion of time the latency exceeds a threshold in the most recent 100 frames of the CAN bus. The weights of the three risk components mentioned above are... α k The adaptive output from the fully connected layer at the end of the GRU network ensures that the risk assessment can dynamically change with operating conditions. Those skilled in the art can implement the above-mentioned sub-risk estimation based on the description in this specification.

[0009] Compared with the prior art, the present invention has the following significant advantages: 1. It has achieved a revolution in safety paradigms, enabling early warning of multiple risks such as thermal runaway, transforming "post-event loss mitigation" into "pre-event disaster prevention"; 2. It breaks through the traditional contradiction between safety and efficiency, and improves the overall charging efficiency by more than 20% under the same safety level through prediction-decision joint optimization and dynamic multi-objective reward function; 3. The multi-mechanism collaboration (prediction + reinforcement learning + rule arbitration) produced unexpected technical effects: pure rule systems have a high failure probability in high-complexity scenarios, and pure reinforcement learning systems have security risks in the early stages of training. However, the hybrid architecture of this invention is significantly better than the pure rule scheme in terms of charging efficiency and much lower than the pure reinforcement learning scheme in terms of the occurrence rate of security events, achieving a synergistic effect of 1+1>2. 4. Spatialized sensor layout and dynamic weight fusion algorithm solve the problem of perception blind spots, and the constructed high-precision "micro-environment field" provides a reliable data foundation for decision-making; 5. In particular, this invention introduces the sub-risk components and adaptive weights of the GRU output into the weighted fusion, so that the risk prediction results have both physical interpretability and the nonlinear fitting ability of deep learning. Compared with simple weighted or single output models, the early warning time of thermal runaway is extended by an average of 35%, and the false alarm rate of overpressure is reduced by 42%. This combination is not a simple superposition of various features, but a deep neural network structure optimization based on domain knowledge, which produces prediction accuracy and robustness that cannot be achieved by a single physical model or a pure black box model. It is a typical example of synergistic effect.

[0010] 6. Safety experience replay and EWC mechanism solve the problem of catastrophic forgetting in online learning, and significantly improve the ability to identify dangerous boundaries through risk sample oversampling. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating an adaptive drone charging method based on multi-source perception and artificial intelligence decision-making according to an exemplary embodiment; Figure 2 This is a schematic diagram of an adaptive drone charging system structure based on multi-source perception and artificial intelligence decision-making, according to an exemplary embodiment. Figure 3 This is a schematic diagram of the overall hardware and data flow architecture of the system of the present invention according to an exemplary embodiment; Figure 4 This is a flowchart illustrating the rule matching and conflict resolution logic of a three-level security arbitrator according to an exemplary embodiment; Figure 5 This is a flowchart illustrating a security experience playback and online learning process according to an exemplary embodiment. Detailed Implementation

[0012] To more clearly illustrate the technical features of the present invention, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the present invention. To simplify the disclosure of the present invention, the components and arrangements of specific examples are described below. Of course, these are merely examples and are not intended to limit the invention.

[0013] Example 1 like Figure 1 As shown in the figure, an adaptive drone charging method based on multi-source perception and artificial intelligence decision-making provided by an embodiment of the present invention includes the following steps: Step 1: Collect micro-environmental data through multiple temperature and humidity sensors deployed inside and outside the charging chamber, and acquire weather forecast data and real-time battery status data of the drone; determine the current charging stage based on the changes in induction coil voltage, battery state of charge change rate, and real-time risk probability during the charging process. The charging stage includes landing stage, stable charging stage, and risk mode; use a dynamic weighted fusion algorithm based on charging stage and real-time risk probability to fuse macro-weather forecast data and micro-sensor data to generate high-precision micro-environmental field data.

[0014] In step 1, the temperature and humidity sensor deployment points include at least the air inlet, the exhaust outlet, the battery contact surface, two diagonal dead corners inside the cabin, and the reference point outside the cabin, totaling 8 sensors. Each sensor collects data at an adjustable frequency of 1~5Hz and transmits it to the edge computing unit via the CAN bus.

[0015] The dynamic weighted fusion algorithm is based on the current charging stage. Adjusting the weights: Setting the initial weights for the sensors Initial weights of weather forecasts Adjustment amount and Satisfy: When (During the landing phase) , ;when During the (stable charging phase), , ;when or real-time risk probability hour, , The fusion temperature was obtained after normalization. ,in The average temperature collected by the spatialized sensor array. For weather forecast temperature data, Similarly, the fusion humidity is obtained. ,in The average humidity value is collected by a spatialized sensor array. This is humidity data for weather forecasts.

[0016] The rules for determining the charging phase are as follows: when a sudden change in the induced voltage of the receiving coil is detected and lasts for more than 2 seconds, it is determined to be a descent phase; when the battery state of charge increases by 2% without any abnormal signal, it is determined to be a stable charging phase; when the real-time risk probability... >0.4 or temperature change rate Enter risk mode at this time.

[0017] Step 1 further includes reconstructing the three-dimensional temperature and humidity distribution field inside the charging compartment using inverse distance weighted interpolation: for any spatial point (x, y, z), its predicted temperature value The weight , Let be the Euclidean distance from the i-th sensor to this point, with a sampling period of 1 second.

[0018] The weather forecast data is obtained every 5 minutes via a 4G / 5G communication module, which acquires future time period temperature, humidity, wind speed, and rainfall probability data provided by the meteorological bureau's API. The real-time status data of the drone battery is collected at a frequency of 1Hz through the battery management system interface, and includes at least battery temperature, state of charge, health, voltage, and current.

[0019] The microenvironmental field data is transferred between step 1 and step 2 via shared memory. The trigger condition is that step 2 is started immediately when the rate of change of the fused microenvironmental field data exceeds a preset threshold or when the charging phase is switched.

[0020] Step 2: Construct a temporal feature vector for drone charging safety, input it into a pre-trained gated recurrent unit network, and output a comprehensive risk probability value. And the hidden state feature vector of the gated recurrent unit.

[0021] The time-series feature vector includes at least 12 dimensions: battery temperature, ambient temperature, humidity, temperature change rate dT / dt, voltage change rate dV / dt, historical risk event markers, battery health status SOH, state of charge SOC, charging current, ambient air pressure, wind speed, and number of cycles.

[0022] The gated loop unit network adopts a two-layer structure with a hidden layer dimension of 64; its core calculation process is: resetting the gate. , update door Candidate hidden state Ultimately hidden state , for Activation function For element-wise multiplication, and These are the trainable parameters of the model; the output layer is obtained by the sigmoid function. At the same time, extract the last time step. As the hidden state feature vector of the gated recurrent unit.

[0023] The pre-training adopts a transfer learning and adversarial verification strategy: first, the network is pre-trained on a public battery aging dataset, and then fine-tuned using small sample data collected for the target drone model; adversarial verification distinguishes source domain and target domain samples by training a binary classifier, and uses a gradient inversion layer to confuse the domain labels of the feature extractor. When the classifier accuracy drops below 52%, the domain is considered to be successfully adapted.

[0024] Step 3: Construct a comprehensive risk probability value. The hidden state feature vectors of the gated recurrent unit and the state space vectors of conventional physical quantities are used to generate adaptive charging strategy actions using a deep reinforcement learning model. These actions include at least the charging current rate, charging pulse frequency, and active cooling fan power. The reward function R for reinforcement learning is designed as a multi-objective weighted sum, with the weights of each sub-objective adjusted according to... Make dynamic adjustments.

[0025] The comprehensive risk probability value It is generated by weighted fusion of at least three risk components: thermal runaway risk, overvoltage risk, and communication anomaly risk; wherein, the thermal runaway risk component adopts a Gaussian membership function. Based on battery temperature change rate or battery temperature calculation; overvoltage risk component is quantified using an S-shaped function based on the ratio of battery voltage to the upper limit of safe voltage and voltage change rate; communication anomaly risk component is weighted based on CAN bus bit error rate and data packet delay exceeding threshold frequency; adaptive weights are assigned to each sub-risk. The gated recurrent unit network is synchronously output through a fully connected layer, ultimately forming... ,and ,in For the first Individual risk value, The total number of sub-risks.

[0026] The logical relationship between step 2 and step 3 is as follows: The GRU hidden state feature vectors serve as key inputs to the state space of the reinforcement learning agent, and the risk is recalculated every decision cycle. When the priority exceeds 0.7, the priority of the security arbitration module will be forcibly increased.

[0027] The deep reinforcement learning model employs a policy optimization-based deep Q-network (DQN) or proximal policy optimization (PPO) structure; the state space vector includes fused ambient temperature, fused ambient humidity, battery temperature, battery temperature change rate, battery health (SOH), and state of charge (SOC). The values, GRU hidden state feature vectors, current charging duration, previous command current, and previous fan duty cycle are included.

[0028] The action space is a three-dimensional continuous / discrete hybrid form: charging current ratio Ic∈[0.2C,1.5C], pulse frequency fp∈[0,100Hz], and cooling fan duty cycle D∈[0,1]. The policy network adopts a 3-layer fully connected structure, and the output layer corresponds to the mean and variance of the actions.

[0029] The reward function ,in, to According to Dynamically adjusted weighting coefficients; For charging efficiency, For safety indicators, For battery health indicators, β is the motion smoothness index; β is the balance coefficient, ranging from 0.5 to 1.0.

[0030] The weights of each sub-objective are based on Dynamic adjustment: , w3 and w4 remain fixed; satisfying Typical parameters are =0.55, =0.15, w3=0.2, w4=0.1, k=0.4, α=15.

[0031] Step 3 also includes guided initialization: in the early stage of training, demonstration data {St,at} is generated using an expert rule base, and the policy network is pre-trained through imitation learning by behavior cloning. Then, reinforcement learning is used for fine-tuning to accelerate convergence and avoid early dangerous exploration.

[0032] Step 4: Establish a three-level rule base including hard limit protection rules, arbitration rules, and suggestion rules to perform mandatory verification and correction on the initial strategy output by reinforcement learning; when multiple rules conflict, a conflict resolution strategy based on fuzzy logic is used to calculate the comprehensive confidence level and output the final control command.

[0033] The steps 3 and 4 adopt a serial logic of "action candidate-verification-correction": after the reinforcement learning model outputs an action, it is immediately sent to the security arbitration step for rule verification. If the verification passes, it is executed directly; otherwise, the arbitration step outputs the corrected action and feeds the corrected experience back to the reinforcement learning replay buffer.

[0034] The hard limit protection rules are directly bound to the hardware protection circuit, including but not limited to: immediately disconnecting the charging circuit and issuing an audible and visual alarm when the battery temperature exceeds 60°C; hardware current limiting when the charging current exceeds 1.8C; this level of rule has the highest priority and does not depend on software judgment.

[0035] The arbitration rules include at least: when When the current is greater than 0.6, the charging current is forcibly limited to below 0.5C and the duty cycle of the cooling fan is set to 100%; when the battery health SOH is less than 75% and the current output by reinforcement learning is greater than 0.8C, the clamping current is forcibly reduced to 0.5C; the priority of this rule is between hard limit and recommendation.

[0036] The recommended rules include: when the ambient humidity is >80%, it is recommended to increase the charging pulse frequency to above 50Hz to enhance the sensitivity of insulation monitoring; when historical risk events are frequent, it is recommended to reduce the exploration noise figure of reinforcement learning.

[0037] The conflict resolution strategy based on fuzzy logic specifically refers to: the comprehensive confidence level of each triggered rule Ri. ,in Based on the priority of the rules, For the degree of matching of rule conditions, The maximum conflict coefficient, ,in and Rules respectively and A suggested set of action parameters is provided; the rule with the highest confidence level is selected as the final execution rule. If multiple rules have the same and highest confidence level, a weighted average method is used to generate a compromise instruction. The basic priority of the rules can be dynamically adjusted according to the scenario: in a high humidity environment, the priority of rules related to voltage limitation is automatically increased by 10%; when the battery is severely aged, the priority of rules related to temperature rise is increased by 15%.

[0038] Step 5: Store historical execution experience according to security level, oversample high-risk samples, dynamically trigger model updates based on performance indicators, and use elastic weight consolidation algorithm to prevent catastrophic forgetting during online learning; execute the final instruction and store the experience data in the safe experience replay pool.

[0039] The specific storage classification based on safety level is as follows: the successful queue has a capacity of 10,000 and stores normal charging experience; the general queue has a capacity of 5,000; the high-risk queue has a capacity of 2,000 and stores experience close to the danger boundary; the sampling weight of high-risk samples is 5 times that of ordinary samples.

[0040] The conditions for dynamically triggering model updates based on performance metrics include: the average reward of the most recent 50 decisions has decreased by more than 12% compared to the historical average reward; or The absolute value of the deviation between the predicted value and the actual occurrence rate of risk events is greater than 18%; or a routine assessment update is triggered every 100 charging cycles.

[0041] The loss function of the elastic weight consolidation algorithm is: ,in , The old ones are model parameters. For the new model parameters, The loss function for the new task. The regularization coefficient is . The diagonal elements of the Fisher information matrix represent parameters. The importance of λ is given, with λ ranging from 0.5 to 2.

[0042] Step 5 also includes a model version control mechanism: the system maintains three versions: the current production model, the candidate model, and the historical best model. When the average reward of the new model on the validation set is higher than that of the current production model and the occurrence rate of security events is not higher than that of the old model, the production model is automatically switched.

[0043] The experience replay pool in step 5 forms a data closed loop with step 3: after each execution of the final instruction, the state, action, reward, next state, risk probability and whether arbitration is triggered are combined into a five-tuple and stored in the replay pool. Step 3 periodically samples batch data from the replay pool for online policy updates.

[0044] Example 2 like Figure 2 As shown in the figure, an adaptive drone charging system based on multi-source perception and artificial intelligence decision-making provided by an embodiment of the present invention includes: The spatial perception module is used to collect micro-environmental data through multiple temperature and humidity sensors deployed inside and outside the charging cabin, and to acquire weather forecast data and real-time battery status data of the drone; it determines the current charging stage based on the changes in induction coil voltage, battery state of charge change rate, and real-time risk probability during the charging process, including the landing stage, stable charging stage, and risk mode; it uses a dynamic weighted fusion algorithm based on the charging stage and real-time risk probability to fuse macro-weather forecast data and micro-sensor data to generate high-precision micro-environmental field data; The real-time risk assessment module is used to construct a temporal feature vector for drone charging safety, input it into a pre-trained gated recurrent unit network, and output a comprehensive risk probability value. And the hidden state feature vector of the gated recurrent unit; The dynamic strategy optimization module is used to construct a comprehensive risk probability value. The hidden state feature vectors of the gated recurrent unit and the state space vectors of conventional physical quantities are used to generate adaptive charging strategy actions using a deep reinforcement learning model. These actions include at least the charging current rate, charging pulse frequency, and active cooling fan power. The reward function R for reinforcement learning is designed as a multi-objective weighted sum, with the weights of each sub-objective adjusted according to... Make dynamic adjustments; The hierarchical security arbitration module is used to establish a three-level rule base including hard limit protection rules, arbitration rules, and recommendation rules. It performs mandatory verification and correction on the initial strategy output by reinforcement learning. When multiple rules conflict, a conflict resolution strategy based on fuzzy logic is used to calculate the comprehensive confidence level and output the final control command. The closed-loop learning module is used to classify and store historical execution experience according to security level, oversample high-risk samples, dynamically trigger model updates based on performance indicators, and use an elastic weight consolidation algorithm to prevent catastrophic forgetting during the online learning process; it executes the final instruction and stores the experience data in the safe experience replay pool.

[0045] The key technologies of adaptive drone charging systems are as follows.

[0046] I. System Hardware Architecture: like Figure 3As shown, the system hardware adopts a layered modular design. Sensors and communication units in the perception layer are connected to the core computing layer via industrial buses (such as CAN, RS485) or Ethernet. The core computing layer is handled by embedded edge computing devices with AI acceleration capabilities (such as the NVIDIA Jetson series), responsible for running all algorithms. The programmable charging power supply and temperature control system in the execution layer receive digital instructions from the computing layer. All data streams are synchronously recorded and encrypted locally by a "data black box" module.

[0047] This embodiment deploys eight temperature and humidity sensors within the charging compartment, specifically arranged as follows: one on each side of the air inlet (positions P1 and P2), one on each side of the exhaust outlet (positions P3 and P4), one directly below the battery contact surface (position P5), one in each of the two diagonal blind spots inside the compartment (positions P6 and P7), and one external reference point (position P8). Sensor data is collected at a frequency of 1Hz and transmitted to the edge computing unit via the CAN bus. Simultaneously, the system obtains forecast data from the meteorological bureau's API every 5 minutes via a 4G network. The spatial layout of the sensor array ensures the spatial representativeness of the collected data, providing a foundation for subsequent data fusion.

[0048] II. Dynamic Weight Fusion Algorithm: The dynamic weighted fusion algorithm adopts an adaptive weighted fusion based on the charging stage and risk state, and its core steps are as follows: Step 1: Initialize basic weights: Set initial values ​​for sensor data weights Initial weights of weather forecast data ,satisfy .

[0049] Step 2: Weight adjustment based on charging stage: According to the current charging stage Dynamically adjust the weights: , , Among the adjustment amounts and satisfy: when (During the landing phase) , , when During the (stable charging phase), , , when or real-time risk probability hour, , .

[0050] Step 3: Weight Normalization: The adjusted weights are then normalized. .

[0051] Step 4: Data Fusion Calculation The fused microenvironmental field data were calculated using a weighted average method: , , in and The average values ​​of temperature and humidity collected by the spatialized sensor array. and This is weather forecast data.

[0052] III. GRU Risk Prediction Model: The time-series risk prediction model adopts a gated recurrent unit (GRU) neural network architecture, and its core calculation process is as follows: Step 1: Input feature vector: The preprocessed temporal feature vector is denoted as... ,in For feature dimension, This serves as the time step index. The feature vector must contain at least: the original sensor values ​​and the temperature change rate. Voltage change rate Historical risk event markers, battery cycle counts, etc.

[0053] Step 2: Forward computation of GRU cells: The core computation of the GRU unit includes two gating mechanisms: reset gate and update gate. Reset door Controlling the degree to which historical information is forgotten: , Update Gate Control the ratio of new information to historical information: , Candidate hidden state Integrate current input with historical information after reset: , Final hidden state The update gate controls the weighted combination of historical states and candidate states: , in for Activation function For element-wise multiplication, and These are the trainable parameters of the model.

[0054] Step 3: Fuzzy logic sub-risk calculation: Hiding the output state of GRU Mapping to multiple sub-risk dimensions, fuzzy logic rules are used to calculate the value of each sub-risk. Taking thermal runaway sub-risk as an example... For example, its fuzzy rule can be expressed as:

[0055] The membership function is a Gaussian function. .

[0056] Step 4: Calculate the overall risk probability: The final comprehensive risk probability value is obtained by weighting and fusing multiple sub-risks. : , in For the first Individual risk value, For the corresponding weight coefficients, satisfying , This represents the total number of sub-risks. Final output. .

[0057] To evaluate the distributional differences between the source domain (public dataset) and the target domain (target device data), this invention employs an adversarial verification method. Specifically, a binary classifier is trained to distinguish whether a sample comes from the source domain or the target domain, while a feature extractor is trained to confuse the classifier. When the classifier accuracy approaches 50%, it indicates that the feature distributions of the two domains have aligned. This invention uses the degree of confusion from adversarial verification as an evaluation metric for transfer learning effectiveness, triggering retraining when the classification accuracy exceeds 65%. Experiments show that this mechanism reduces cross-device prediction error by 42%. Specifically, the adversarial verification network includes a gradient inversion layer, where the feature extractor and domain classifier align their feature distributions by inverting gradients. When the domain classifier accuracy stabilizes below 52%, it is considered that the feature distributions of the two domains are no longer significantly different.

[0058] IV. Dynamic Multi-Objective Reward Function: The reward function of the deep reinforcement learning model adopts a dynamic multi-objective weighted design, and its expression is: , The sub-objective functions are defined as follows: Charging efficiency function , representing the increment of the state of charge per unit time. Safety function Punishment for excessively rapid temperature rise and high-risk conditions. Battery health function Punishment for high-current charging of aging batteries Action stability function The punishment current fluctuated violently; Dynamic weight adjustment mechanism: Weighting coefficients With real-time risk probability Dynamic changes, with safety weights For example: , in , Based on the weights, For adjustment coefficients, This is the sensitivity parameter.

[0059] Similarly, efficiency weight Follow It rises and then automatically decreases: , Each weight satisfies the normalization constraint: , This dynamic weighting mechanism enables the agent to operate in low-risk scenarios ( (Smaller) Prioritize charging efficiency, in high-risk scenarios ( (For larger cases) it automatically switches to a conservative strategy without manual switching. Simulation verification shows that when When the weight increases from 0.2 to 0.8, the safety weight w2 automatically increases by 2.3 times (from 0.15 to 0.55), while the efficiency weight w1 decreases accordingly to 0.4 times its original value (from 0.55 to 0.15). The battery health weight w3 and the action stability weight w4 remain unchanged. This dynamic adjustment enables the agent to achieve a smooth policy transition under different risk levels.

[0060] V. Fuzzy Logic Conflict Resolution Algorithm: like Figure 4 As shown, when multiple rules are triggered simultaneously and conflict with each other, this invention employs a conflict resolution strategy based on fuzzy logic. The core calculation steps are as follows: Step 1: Quantify rule priority: Each rule Basic priority The priority is determined by the rule level: Level 1 rules have a priority of 100, Level 2 rules have a priority of 60-80, and Level 3 rules have a priority of 30-50.

[0061] Step 2: Calculate the condition matching degree: rule The condition part is related to the current context. Matching degree Calculated using fuzzy membership functions: , in For rules No. Membership function for each condition, This is the value of the variable in the current context.

[0062] Step 3: Calculate the conflict degree between rules: rule With rules The degree of conflict between Defined as: , in and Rules respectively and Recommended set of action parameters; A larger value indicates a more severe conflict between the two rules. rule The overall conflict penalty coefficient is: .

[0063] Step 4: Calculate the overall confidence level: rule Overall confidence level It is determined by the basic priority, condition matching degree, and conflict penalty coefficient: .

[0064] Step 5: Optimal rule selection: The rule with the highest overall confidence level will be selected as the final execution rule. .

[0065] in This represents the set of all triggered rules. If multiple rules have the same and highest confidence level, a weighted average method is used to generate a compromise instruction.

[0066] VI. Safety Experience Review and Online Learning: like Figure 5 As shown, the security experience classification and storage mechanism is as follows: 1) Store historical execution experience in three categories according to security level: Success queue :capacity Store normal charging experience. General queue :capacity Store common experience. High-risk queue :capacity It stores experience near dangerous boundaries, with a sampling weight 5 times that of ordinary samples; 2) The empirical sampling probability is defined as: ; 3) Elastic Weight Consolidation Algorithm: To prevent catastrophic forgetting during online learning, the Elastic Weight Consolidation (EWC) algorithm is employed. Let the parameters of the old model be... The new model parameters are The EWC loss function is then: , in The loss function for the new task. The regularization coefficient is . The diagonal elements of the Fisher information matrix represent parameters. Importance: , The Fisher information matrix is ​​calculated and stored after training on the old task is completed, and is used to protect important parameters in subsequent learning.

[0067] 4) Smart update trigger conditions: Model updates are not simply executed periodically, but are dynamically triggered based on the following performance metrics: 1. Average Reward Decline Indicator: When the most recent Average reward of each decision Compared to historical average rewards The drop exceeded the threshold Time-triggered update: ; 2. Risk prediction deviation index: when The deviation between the predicted value and the actual incidence rate of risk events exceeds the threshold. Time-triggered update: ; 3. Periodic triggering: every time it is completed Each charge triggers a routine evaluation update.

[0068] In an outdoor environment (temperature 35℃, humidity 80%RH), the method of this invention, the method in prior art 1 (based on efficiency threshold switching), and the traditional constant current and constant voltage method were compared: the method of this invention improved charging efficiency by 22%, reduced battery peak temperature by 8℃, and did not trigger any safety alarms during charging; in the ablation experiment, the safety event incidence rate increased by 120% after removing the risk prediction module, and an overcurrent alarm occurred after removing the arbitration module, proving the necessity of the hybrid architecture. In cross-model testing, adversarial verification reduced prediction error by 42%. The above data fully demonstrate that the various technical means of this invention work together to produce an unexpectedly superior overall technical effect of "active prediction - dynamic decision-making - safety arbitration".

[0069] VII. GRU Risk Prediction Based on Sub-Risk Weighted Fusion: Three branches are connected in parallel at the output layer of the GRU network, corresponding to the estimated values ​​of thermal runaway risk, overvoltage risk, and communication anomaly risk, respectively. Weights α1, α2, and α3 are generated by an attention mechanism. For a typical charging process, when the battery temperature rises at a rate of 2.5℃ / s, the thermal runaway risk branch outputs 0.92, the overvoltage branch outputs 0.12, and the communication branch outputs 0.05. The network adaptively outputs weights (0.78, 0.15, 0.07), ultimately... =0.78×0.92+0.15×0.12+0.07×0.05≈0.73. Compared to outputting only a single probability, this method can provide the safety arbitration module with specific risk sources, facilitating accurate responses from the rule base (such as immediately limiting current and increasing fan speed in response to thermal runaway). This embodiment has been verified on a simulation test bench, reducing the false trigger rate by 26%.

[0070] The hardware of this invention is based on mature industrial-grade components, and the software algorithm can be deployed on a general-purpose edge computing platform, enabling large-scale production and deployment. It can be widely applied to automated drone airports in scenarios such as logistics hubs, power line inspection networks, and security patrol bases, effectively addressing the core pain points in energy replenishment regarding safety, efficiency, and adaptability, and has broad market application prospects.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. An adaptive drone charging method based on multi-source perception and artificial intelligence decision-making, characterized in that, Includes the following steps: Step 1: Collect micro-environmental data by deploying multiple temperature and humidity sensors inside and outside the charging chamber, and acquire weather forecast data and real-time status data of the drone battery; determine the current charging stage based on the changes in induction coil voltage, battery state of charge change rate, and real-time risk probability during the charging process. The charging stage includes landing stage, stable charging stage, and risk mode; use a dynamic weighted fusion algorithm based on charging stage and real-time risk probability to fuse macro-weather forecast data and micro-sensor data to generate high-precision micro-environmental field data. Step 2: Construct a temporal feature vector for drone charging safety, input it into a pre-trained gated recurrent unit network, and output the comprehensive risk probability value and the hidden state feature vector of the gated recurrent unit. Step 3: Construct a state space vector containing the comprehensive risk probability value, the hidden state feature vector of the gated recurrent unit, and conventional physical quantities. Use a deep reinforcement learning model to generate adaptive charging strategy actions. The actions include at least the charging current rate, the charging pulse frequency, and the power of the active cooling fan. The reward function R of the reinforcement learning is designed as a multi-objective weighted sum, and the weights of each sub-objective are dynamically adjusted according to the comprehensive risk probability value. Step 4: Establish a three-level rule base including hard limit protection rules, arbitration rules, and suggestion rules to perform mandatory verification and correction on the initial strategy output by reinforcement learning; when multiple rules conflict, a conflict resolution strategy based on fuzzy logic is used to calculate the comprehensive confidence level and output the final control command. Step 5: Store historical execution experience according to security level, oversample high-risk samples, dynamically trigger model updates based on performance indicators, and use elastic weight consolidation algorithm to prevent catastrophic forgetting during online learning; execute the final instruction and store the experience data in the safe experience replay pool.

2. The method according to claim 1, characterized in that, The temperature and humidity sensors are deployed at least at the air inlet, exhaust outlet, battery contact surface, two diagonal dead corners inside the cabin, and an external reference point, totaling eight sensors. Each sensor collects data at an adjustable frequency of 1~5Hz and transmits it to the edge computing unit via the CAN bus.

3. The method according to claim 1, characterized in that, The dynamic weighted fusion algorithm is based on the current charging stage. Adjusting the weights: Setting the initial weights for the sensors Initial weights of weather forecasts Adjustment amount Satisfy: When hour, ;when hour, ;when or real-time risk probability hour, ; The fusion temperature was obtained after normalization. ,in The average temperature collected by the spatialized sensor array. For weather forecast temperature data, ; Similarly, the fused humidity is obtained. ,in The average humidity value is collected by a spatialized sensor array. This is humidity data for weather forecasts.

4. The method according to claim 1, characterized in that, The gated loop unit network adopts a two-layer structure with a hidden layer dimension of 64; its core calculation process is: resetting the gate. , update door Candidate hidden state Ultimately hidden state , for Activation function For element-wise multiplication, and These are the trainable parameters of the model; the output layer is obtained by the sigmoid function. At the same time, extract the last time step. As the hidden state feature vector of the gated recurrent unit.

5. The method according to claim 1, characterized in that, The comprehensive risk probability value It is generated by weighted fusion of at least three risk components: thermal runaway risk, overvoltage risk, and communication anomaly risk; wherein, the thermal runaway risk component adopts a Gaussian membership function. Based on battery temperature change rate or battery temperature calculation; overvoltage risk component is quantified using an S-shaped function based on the ratio of battery voltage to the upper limit of safe voltage and voltage change rate; communication anomaly risk component is weighted based on CAN bus bit error rate and data packet delay exceeding threshold frequency; adaptive weights are assigned to each sub-risk. The gated recurrent unit network is synchronously output through a fully connected layer, ultimately forming... ,and ,in For the first Individual risk value, The total number of sub-risks.

6. The method according to claim 1, characterized in that, The deep reinforcement learning model is a deep Q-network or a near-end policy optimization network based on policy optimization; the state space vector includes fused ambient temperature, fused ambient humidity, battery temperature, battery temperature change rate, battery health, and state of charge. The values, GRU hidden state feature vectors, current charging duration, previous command current, and previous fan duty cycle are included.

7. The method according to claim 1, characterized in that, The reward function ,in, to According to Dynamically adjusted weighting coefficients; For charging efficiency, For safety indicators, For battery health indicators, β is an indicator of motion smoothness; β is the balance coefficient.

8. The method according to claim 1, characterized in that, The conflict resolution strategy based on fuzzy logic specifically refers to: the comprehensive confidence level of each triggered rule Ri. ,in Based on the priority of the rules, For the degree of matching of rule conditions, The maximum conflict coefficient, ,in and Rules respectively and The recommended set of action parameters is used; the rule with the highest confidence is selected as the final execution rule. If multiple rules have the same and the highest confidence, a weighted average method is used to generate a compromise instruction.

9. The method according to any one of claims 1-8, characterized in that, The loss function of the elastic weight consolidation algorithm is: ,in , The old ones are model parameters. For the new model parameters, The loss function for the new task. The regularization coefficient is . The diagonal elements of the Fisher information matrix represent parameters. The importance of.

10. An adaptive drone charging system based on multi-source perception and artificial intelligence decision-making, characterized in that, include: The spatial perception module is used to collect micro-environmental data through multiple temperature and humidity sensors deployed inside and outside the charging cabin, and to acquire weather forecast data and real-time battery status data of the drone. Based on the changes in induction coil voltage, battery state of charge change rate, and real-time risk probability during the charging process, the current charging stage is determined, which includes the landing stage, stable charging stage, and risk mode. A dynamic weighted fusion algorithm based on the charging stage and real-time risk probability is used to fuse macro-weather forecast data and micro-sensor data to generate high-precision micro-environmental field data. The real-time risk assessment module is used to construct a temporal feature vector for drone charging safety, input it into a pre-trained gated recurrent unit network, and output a comprehensive risk probability value. And the hidden state feature vector of the gated recurrent unit; The dynamic strategy optimization module is used to construct a comprehensive risk probability value. The hidden state feature vectors of the gated recurrent unit and the state space vectors of conventional physical quantities are used to generate adaptive charging strategy actions using a deep reinforcement learning model. These actions include at least the charging current rate, charging pulse frequency, and active cooling fan power. The reward function R for reinforcement learning is designed as a multi-objective weighted sum, with the weights of each sub-objective adjusted according to... Make dynamic adjustments; The hierarchical security arbitration module is used to establish a three-level rule base including hard limit protection rules, arbitration rules, and recommendation rules. It performs mandatory verification and correction on the initial strategy output by reinforcement learning. When multiple rules conflict, a conflict resolution strategy based on fuzzy logic is used to calculate the comprehensive confidence level and output the final control command. The closed-loop learning module is used to classify and store historical execution experience according to security level, oversample high-risk samples, dynamically trigger model updates based on performance indicators, and use an elastic weight consolidation algorithm to prevent catastrophic forgetting during the online learning process; it executes the final instruction and stores the experience data in the safe experience replay pool.

Images