A method and system for controlling charging of a drone

Abstract

The application relates to the technical field of unmanned aerial vehicle (UAV) charging control, and discloses a UAV charging control method and a control system. The method comprises the following steps: after responding to a UAV charging request, starting a charging control process comprising an environment parameter collection and battery state initialization stage. After establishing a charging connection, a self-adaptive observer performs system state estimation and external disturbance estimation in parallel based on the collected environment parameters and the initialized battery state. According to the real-time results of the above estimation, the charging strategy is dynamically adjusted, including refreshing the power control signal and activating the compensation mechanism, and the energy output of the charging equipment is cooperatively controlled accordingly until the battery reaches the predetermined charging target. Through the double consideration of the environment and the battery state, combined with the parallel estimation and dynamic adjustment mechanism, the charging process can accurately match the battery demand of the UAV, effectively cope with various interference factors, improve the reliability and adaptability of the charging process, and provide a new solution for efficient charging of the UAV.

Description

Technical Field

[0001] This invention relates to the field of drone charging control technology, specifically to a drone charging control method and control system. Background Technology

[0002] With the rapid development of drone technology, its applications have broadly covered multiple fields such as logistics delivery, power line inspection, environmental monitoring, and emergency rescue. The continuous and stable operation of drones highly depends on battery power supply, and charging technology, as a key component ensuring drone endurance, directly impacts the drone's operational efficiency and application value. Currently, drone charging technology has made significant progress, evolving from early manual charging to automated and intelligent charging, resulting in various charging modes including fixed charging stations, mobile charging platforms, and wireless charging.

[0003] However, existing drone charging technologies still face many unresolved issues in practical applications. Regarding environmental adaptability, most charging solutions fail to adequately consider the impact of environmental parameters on the charging process. Drone operations are complex and varied, potentially encountering harsh environmental conditions such as high temperatures, low temperatures, high humidity, and strong electromagnetic interference. For example, in low-temperature environments, battery activity significantly decreases; using conventional charging strategies not only results in low charging efficiency but may also cause irreversible damage to the battery. In high-temperature environments, continuous high-current charging can easily lead to battery overheating and even safety hazards. While some existing charging systems are equipped with simple environmental sensing modules, they can only achieve basic parameter acquisition and display, failing to deeply integrate environmental parameters with charging control strategies, resulting in weak adaptability to environmental changes during the charging process.

[0004] Existing solutions also have significant shortcomings in the coordination between battery state sensing and charging control. Battery state is the core basis for formulating charging strategies, including key parameters such as remaining capacity, health status, internal resistance, and temperature. Currently, most charging systems perform rudimentary initial detection of battery state, only acquiring basic information such as remaining capacity, lacking sufficient precision in sensing deeper parameters such as battery health status and aging degree. More critically, during charging, system state estimation and external disturbance estimation often employ a serial processing method, resulting in high data processing latency. When faced with disturbances such as sudden changes in battery internal state or external load fluctuations, the charging system cannot respond promptly, and the adjustment of power control signals exhibits lag, thus affecting charging stability. Furthermore, most charging systems lack a robust compensation mechanism; when problems such as voltage fluctuations and current surges occur, compensation measures cannot be quickly activated, easily leading to charging interruptions or battery damage.

[0005] Drone batteries come in a variety of types, and different models and brands exhibit significant differences in charging characteristics. Existing charging systems lack versatility, often requiring customized charging strategies for specific battery models, increasing user costs and operational complexity. These problems are even more pronounced in automated charging scenarios, where the reliability and efficiency of the charging process directly determine the continuity of drone operations. Current technologies are insufficient to meet the high demands of large-scale, routine drone applications. Therefore, developing a drone charging control method capable of accurately sensing environmental and battery status, rapidly responding to various disturbances, and dynamically adapting to charging needs has become an urgent requirement in this field. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for controlling the charging of unmanned aerial vehicles (UAVs) to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a method for controlling the charging of a drone, the method comprising:

[0008] In response to receiving a charging request from the drone, a charging control process is initiated, which includes an environmental parameter acquisition phase and a battery status initialization phase.

[0009] After the charging connection is established, based on the collected environmental parameters and the initialized battery state, the adaptive observer performs system state estimation and external disturbance estimation in parallel.

[0010] Based on the real-time results of the system state estimation and external disturbance estimation, the charging strategy is dynamically adjusted, including the refresh of the power control signal and the activation of the compensation mechanism.

[0011] Based on the refreshed power control signal and the activated compensation mechanism, the energy output of the charging equipment is controlled in a coordinated manner until the battery reaches the predetermined charging target.

[0012] Preferably, the specific implementation of the environmental parameter acquisition stage and the battery state initialization stage includes: synchronously capturing airflow velocity, ambient light intensity, and electromagnetic interference background values ​​in the charging compartment through a multi-source sensor array; sampling the battery polarization voltage using a high-precision analog-to-digital converter to establish a voltage-time rate of change curve; spatiotemporally aligning the airflow velocity, ambient light intensity, electromagnetic interference background values, and voltage-time rate of change curve to form a multimodal environmental feature vector; and constructing a dynamic weight matrix for the battery initial state based on the multimodal environmental feature vector.

[0013] Preferably, the parallel execution of the system state estimation and external disturbance estimation is achieved through the following steps: configuring a dual-channel Kalman filter bank, with the first channel processing the battery electrochemical impedance spectroscopy data and the second channel analyzing the thermal conduction gradient distribution; inputting the dynamic weight matrix into the dual-channel filter bank to output the battery internal resistance change trajectory and the predicted thermal accumulation value, respectively; identifying abnormal fluctuation patterns in the environmental feature vector through a convolutional neural network and calculating the disturbance intensity coefficient; and fusing the internal resistance change trajectory, the predicted thermal accumulation value, and the disturbance intensity coefficient to generate a system stability index.

[0014] Preferably, the steps of dynamically adjusting the charging strategy include: establishing a fuzzy control rule base based on system stability indicators, wherein the fuzzy control rule base contains a nonlinear mapping relationship between charging rate and temperature threshold; using a particle swarm optimization algorithm to search for the optimal charging trajectory in the fuzzy control rule base, wherein the optimal charging trajectory satisfies battery life constraints; monitoring the charge distribution state on the battery surface in real time, and triggering an adaptive recalibration procedure when uneven charge is detected; verifying the feasibility of the charging trajectory through a digital twin model, and outputting the verified power modulation waveform.

[0015] Preferably, the activation process of the compensation mechanism specifically includes: constructing a disturbance propagation model, which includes analysis of the coupling path of electromagnetic interference through the charging interface; designing a multi-order notch filter bank, the center frequency of which is dynamically adjusted according to the real-time acquired electromagnetic interference spectrum; when the system stability index exceeds the critical value, activating the phase compensation circuit to pre-distort the power modulation waveform; and judging the convergence of the compensation system through the Lyapunov function to ensure that the disturbance suppression effect meets the design requirements.

[0016] Preferably, the implementation steps of the energy output of the collaborative control charging device include: distributing the charging current using a multi-phase interleaved parallel topology, dynamically adjusting the switching sequence of each phase according to the predicted value of heat accumulation; collecting current ripple samples in each switching cycle and extracting harmonic component characteristics through fast Fourier transform; adjusting the dead time of the gate drive signal according to the harmonic component characteristics to achieve zero-voltage switching conversion; and using an optical fiber isolator to synchronize the control signals of each phase to ensure that the power synthesis waveform meets the battery charging characteristic requirements.

[0017] Preferably, the method further includes a continuous optimization process, which is achieved by: embedding a micron-level temperature sensor array on the surface of the battery electrode to collect three-dimensional thermal distribution data; establishing an electro-thermal-mechanical multiphysics coupling model to invert the internal stress distribution state of the battery in real time; when a local hot spot is detected, activating a directional air cooling system to implement precise thermal management; and optimizing the parameters of the electro-thermal-mechanical multiphysics coupling model through a reinforcement learning algorithm to continuously improve the prediction accuracy with the number of charging cycles.

[0018] Preferably, the method further includes an abnormal situation handling strategy, the execution steps of which include: deploying an acoustic emission detection device to capture the ultrasonic signal characteristics of lithium plating phenomenon inside the battery; when a specific frequency ultrasonic harmonic is detected, immediately switching to pulse charging mode and reducing the average current; simultaneously starting an X-ray diffraction analyzer to detect changes in the lattice structure of the electrode material; if the lattice distortion rate exceeds the safety threshold, triggering an emergency circuit breaker mechanism and saving the failure analysis data.

[0019] Preferably, the method further includes accurate determination of charging termination, and the control steps include: measuring the AC internal resistance of the battery using the four-terminal method and plotting the internal resistance-charge change curve; starting the relaxation measurement program when the second derivative of the internal resistance-charge change curve shows an inflection point and recording the open-circuit voltage decay characteristics; predicting the duration of the voltage plateau period using the Gaussian process regression algorithm and calculating the optimal charging termination point; maintaining microcurrent maintenance after disconnecting the main charging circuit until the battery voltage stabilizes at the rated floating value.

[0020] Preferably, the present invention also includes a drone charging control system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor, when executing the computer program, implements the steps of the drone charging control method described above.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] Comprehensive environmental parameter collection accurately captures key factors in the charging scenario, such as temperature, humidity, and electromagnetic interference intensity, providing a complete environmental basis for subsequent charging strategy formulation. Battery state initialization, through precise detection of multi-dimensional parameters such as remaining battery capacity, health status, and internal resistance, establishes an initial state model of the battery, avoiding deviations in charging strategies caused by ambiguous initial state perception. This dual pre-preparation mechanism ensures that charging control is based on accurate perception from the outset, laying the foundation for the stability of the entire charging process.

[0023] By simultaneously performing two estimation tasks using an adaptive observer, data processing time is significantly reduced, enabling the system to acquire accurate information in real time regarding changes in the battery's internal state and external environmental disturbances. The adaptive observer possesses excellent self-adjustment capabilities, automatically adjusting estimation parameters based on the characteristics of different battery models and the disturbance features under varying environments, ensuring high accuracy in estimation results even under complex conditions. This real-time and accurate dual estimation capability allows the charging system to promptly capture subtle changes during the charging process, providing reliable support for the dynamic adjustment of the charging strategy.

[0024] The adjustment of the charging strategy includes not only refreshing the power control signal but also activating the compensation mechanism, forming a dual control system of "active adjustment + passive compensation." Real-time refreshing of the power control signal allows for precise adjustment of the charging power based on changes in battery status and the impact of environmental disturbances, ensuring the charging process always meets the battery's current needs and avoiding problems such as overcharging, undercharging, or low charging efficiency that may occur under conventional fixed-power charging modes. Timely activation of the compensation mechanism, on the other hand, enables rapid initiation of voltage compensation and current stabilization measures in the event of sudden events such as voltage fluctuations or current surges, effectively offsetting the negative impact of disturbances on the charging process and ensuring its continuity and stability. This dynamic adjustment mechanism gives the charging system excellent robustness, enabling it to adapt to the charging characteristics of different drone battery models and complex and ever-changing charging environments.

[0025] This method does not simply adjust energy output based on the power control signal, but rather integrates the refreshed power control signal with the activated compensation mechanism. This ensures that the charging device's energy output meets both the battery's current charging needs and can cope with various disturbances. This collaborative control mode avoids control deviations that may result from a single control signal, ensuring that charging energy is stably and efficiently input into the battery. Whether it's protecting battery activity by reducing initial charging power in low-temperature environments or ensuring charging stability by activating anti-interference compensation mechanisms in electromagnetic interference environments, this method achieves optimal charging performance through collaborative control.

[0026] In practical applications, the advantages of this method become even more pronounced. For logistics delivery drones, rapid charging is often required in different environments during operation. This method can quickly adapt to changes in the site environment and differences in battery status, ensuring charging efficiency and battery safety. For power line inspection drones, their operating environment is mostly in complex outdoor scenarios with prominent issues such as electromagnetic interference and temperature fluctuations. The disturbance estimation and compensation mechanism of this method can effectively address these issues and ensure the stability of the charging process. For scenarios involving multiple drone models operating collaboratively, the adaptive capability of this method can adapt to different battery characteristics without human intervention, reducing operational complexity and usage costs. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the working principle of the drone charging control method described in this invention.

[0028] Figure 2 A flowchart for environmental parameter acquisition and battery status initialization;

[0029] Figure 3 A flowchart for parallel estimation of system state and external disturbances;

[0030] Figure 4This is a battery state diagram;

[0031] Figure 5 Comparison chart of abnormal situations occurring and the handling effects of drone charging systems. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] Please see Figure 1 This invention provides a charging control method for unmanned aerial vehicles (UAVs). The method includes: an environmental parameter acquisition phase where external environmental data is acquired in real time through a sensor network deployed within the charging chamber; and a battery state initialization phase where the initial electrochemical characteristics of the battery are recorded and analyzed using high-precision measuring equipment. After the charging connection is established, an adaptive observer performs system state estimation and external disturbance estimation in parallel based on the acquired environmental parameters and the initialized battery state. The system state estimation focuses on changes in internal battery parameters such as internal resistance and temperature, while the external disturbance estimation performs quantitative analysis of environmental factors such as electromagnetic interference and thermal fluctuations. Based on the estimation results, the charging strategy is dynamically adjusted. The adjustment process includes refreshing the power control signal and activating the compensation mechanism. The power control signal is generated in real time by a digital signal processor, and the compensation mechanism is selectively activated based on the disturbance intensity. Finally, based on the refreshed power control signal and the activated compensation mechanism, the energy output of the charging device is coordinated and controlled. The energy output uses multiphase modulation technology to ensure stable current until the battery voltage or capacity reaches the predetermined charging target.

[0034] Example 1: See Figure 2In practical implementation, the environmental parameter acquisition phase and the battery state initialization phase involve the deployment and data processing of a multi-source sensor array. The multi-source sensor array consists of a miniature anemometer, a photosensitive sensor, and an electromagnetic field probe. These components are precisely installed at specific locations on the top and side walls of the charging compartment to ensure comprehensive coverage of the monitoring area. The miniature anemometer uses the principle of thermal anemometers to capture airflow speed data in the charging compartment at a sampling frequency of 1000 times per second. The airflow speed data is transmitted in real time to the central processing unit for preliminary filtering. The photosensitive sensor is based on a calibrated photodiode design, and its spectral response range covers the visible to near-infrared band. It is used to continuously monitor ambient light intensity, and the ambient light intensity value is recorded in lux. The electromagnetic field probe is a wideband radio frequency receiver with a scanning frequency extending from 10MHz to 1GHz. The electromagnetic interference background value is quantified in decibels and microvolts. All sensors in the multi-source sensor array are synchronously triggered through a high-speed serial interface to ensure the time consistency of data acquisition. A high-precision analog-to-digital converter operates at 24-bit resolution with a sampling interval of 1 microsecond, continuously sampling the battery polarization voltage for 1000 data points. A least-squares fitting algorithm is then used to generate a voltage-time variation curve. The voltage-time variation rate curve is derived by calculating the first derivative of the curve, using the central difference method to reduce error. The voltage-time variation rate curve reflects the battery's transient response characteristics and is used for subsequent state initialization. Spatiotemporal alignment is achieved through a hardware timestamp synchronization module. Based on the GPS clock signal, this module aligns airflow velocity, ambient light intensity, electromagnetic interference background values, and the voltage-time variation rate curve to a unified time reference. The aligned dataset is then integrated into a multimodal environmental feature vector. This vector is four-dimensional, corresponding to the average and variance values ​​of the four parameters. The average and variance values ​​are calculated using a sliding window method with a window size of 100 sampling points. Based on the multimodal environmental feature vector, the dynamic weight matrix is ​​constructed using principal component analysis. Principal component analysis extracts the main components of the multimodal environmental feature vector and assigns weight coefficients. The weight coefficients are adaptively adjusted according to the fluctuation range of the environmental parameters. The fluctuation range is determined by calculating the standard deviation of the parameters. The size of the dynamic weight matrix is ​​fixed at 4x4. The matrix elements represent the relative importance of each environmental parameter to the initial state of the battery. The dynamic weight matrix is ​​used to weight the initial state data of the battery in the subsequent estimation stage.

[0035] In some embodiments, the installation location of the multi-source sensor array can be adjusted according to the geometry of the charging compartment. For example, in a large charging compartment, the array may be deployed in multiple quadrants to enhance spatial resolution. Airflow velocity measurement is not limited to thermal anemometers; ultrasonic anemometers can also be used to improve accuracy, but the core principle remains the same: quantifying airflow through fluid dynamics. The calibration process for photosensitive sensors includes baseline measurements under standard illumination conditions to ensure data accuracy. The frequency range of electromagnetic field probes may be extended to higher frequency bands to address the electromagnetic interference characteristics of specific environments, but the basic function remains capturing background interference values. The sampling interval of the high-precision analog-to-digital converter can be configured to be shorter or longer to accommodate the polarization characteristics of different battery types. The establishment of voltage-time rate of change curves always relies on mathematical derivation to ensure repeatability. The time synchronization accuracy of the spatiotemporal alignment module may be improved to the nanosecond level, but the core objective remains data fusion. The method for constructing multimodal environmental feature vectors may introduce more statistical features, such as skewness and kurtosis, to enrich the vector information. The algorithm for constructing the dynamic weight matrix can be replaced by factor analysis or independent component analysis, but principal component analysis is the standard method, and its purpose is dimensionality reduction and weight allocation.

[0036] In practical implementation, the data acquisition process of the multi-source sensor array includes an initial calibration phase, which is automatically executed before each charging request. During calibration, the sensors are exposed to a known reference environment, consisting of a standard airflow generator, a standard light source, and a standard electromagnetic field source. The calibration data is used to correct sensor drift and ensure the accuracy and reliability of the measurements. Airflow velocity data is low-pass filtered with a cutoff frequency set to 50Hz to remove high-frequency noise. Ambient light intensity data undergoes logarithmic transformation to linearize the response curve. Electromagnetic interference background values ​​are converted to a frequency domain representation using a fast Fourier transform, highlighting interference peaks. The sampling circuit of the high-precision analog-to-digital converter includes an anti-aliasing filter, which is a Butterworth type fourth-order filter to prevent high-frequency distortion from affecting voltage sampling. During battery polarization voltage sampling, the battery is in an open-circuit state to avoid errors introduced by the load. The derivative calculation of the voltage-time rate of change curve uses the five-point Stokes formula to improve numerical stability. The spatiotemporal alignment module relies on field-programmable gate array (FPGA) hardware. The FPGA processes multiple data streams in parallel, achieving microsecond-level timestamp synchronization accuracy. The generation of multimodal environmental feature vectors includes a data normalization step, which scales each parameter value to zero mean and unit variance, eliminating the influence of dimensions. During the construction of the dynamic weight matrix, principal component analysis (PCA) calculations are performed by a digital signal processor (DSP). The DSP uses the Jacobi iteration method to solve for eigenvalues, assigning the highest weight to the component with the largest eigenvalue. The dynamic weight matrix is ​​ultimately stored in non-volatile memory for later use.

[0037] Optionally, the multi-source sensor array can integrate a temperature compensation circuit. This circuit adjusts the sensor output based on thermocouple readings to compensate for errors caused by changes in ambient temperature. However, temperature compensation is not a mandatory step and depends on the thermal stability of the charging compartment. A high-precision analog-to-digital converter may be equipped with a multiplexer, which allows alternating sampling of multiple battery cells; therefore, multiplexing is an optional extension. The spatiotemporal alignment module supports external clock input from a high-precision oscillator to improve synchronization reliability, although a GPS clock already meets basic requirements. The construction of the dynamic weight matrix can be optimized using an artificial neural network trained on historical data to improve weight allocation; however, principal component analysis (PCA), as a lightweight approach, is preferred for real-time applications.

[0038] In practical implementation, the coordination between the environmental parameter acquisition phase and the battery state initialization phase is achieved through an interrupt-driven mechanism. This mechanism triggers processing tasks when sensor data is ready. Data from the multi-source sensor array is transferred to the buffer via direct memory access, reducing processor load. Airflow speed data is displayed in real-time on the monitoring interface for operator visualization and verification. Ambient light intensity data is correlated with timestamps to generate a historical illumination log. An alarm is triggered when the electromagnetic interference background value exceeds a threshold set according to international electromagnetic compatibility standards. The sampling sequence of the high-precision analog-to-digital converter is precisely controlled by a timer using a quartz crystal oscillator with frequency stability down to one part per million. Before sampling the battery polarization voltage, the battery connector undergoes a contact resistance check, which is ensured to be below 1 milliohm using a four-wire measurement method. Data points on the voltage-time rate of change curve undergo outlier removal using the Grubbs test with a confidence level set to 95%. The spatiotemporal alignment module's algorithm includes interpolation processing, specifically linear interpolation, to fill potential data loss gaps. The multimodal environmental feature vector is stored as a structure containing timestamps and parameter arrays for easy subsequent indexing. The update strategy for the dynamic weight matrix is ​​incremental. The incremental update fine-tunes the weights after each charging cycle, based on the latest environmental data, but the matrix size remains unchanged.

[0039] It is understandable that the synchronous acquisition of multi-source sensor arrays relies on a common clock source. This common clock source ensures that all sensor samples are instantaneously consistent, avoiding data distortion caused by phase differences. The resolution selection of the high-precision analog-to-digital converter is based on the dynamic range of the battery voltage; 24-bit resolution covers millivolt-level variations, meeting high-precision requirements. The establishment of the voltage-time rate of change curve is essentially a differential process, revealing the rate of chemical reactions within the battery and providing input for state estimation. The necessity of spatiotemporal alignment stems from the temporal correlation of multimodal data, ensuring that the feature vector reflects the environmental state at the same moment. The construction of the dynamic weight matrix reflects parameter importance weighting, which improves the accuracy of state initialization.

[0040] In some embodiments, the components of the multi-source sensor array may employ a redundant design, including a backup sensor array that automatically switches in case of a main array failure. Airflow velocity measurement may be combined with a pressure sensor, which provides additional airflow dynamics parameters; however, the core parameter remains the velocity value. Ambient light intensity monitoring can be extended to the ultraviolet band, which affects certain battery materials, but standard implementations are limited to visible light. Acquisition of electromagnetic interference background values ​​may include time-domain analysis, which captures transient pulses, but frequency-domain scanning is sufficient for feature extraction. The sampling interval of the high-precision analog-to-digital converter can be dynamically adjusted based on the degree of battery aging, which is estimated through cycle counting; however, a fixed interval simplifies implementation. The mathematical model of the voltage-time rate of change curve may employ polynomial fitting, with the order of polynomial fitting adjustable, but linear fitting is the default choice. The synchronization signal for the spatiotemporal alignment module may come from a network time protocol, which is suitable for distributed systems, but a local clock is sufficient. The dimensionality of the multimodal environmental feature vector may be increased, with the added dimension including derived statistics such as moving averages; however, a four-dimensional vector balances complexity and efficiency. The algorithm for constructing the dynamic weight matrix can integrate machine learning techniques and machine learning adaptive learning environment modes, but principal component analysis ensures real-time performance.

[0041] Optionally, a humidity sensor can be added during the environmental parameter acquisition phase. This sensor monitors the relative humidity of the charging compartment, which affects battery performance. However, this parameter is not included in Example 1 and is therefore optional. AC impedance spectroscopy may be introduced during the battery state initialization phase, as it provides richer electrochemical information; however, DC polarization voltage sampling is the fundamental method. The alignment accuracy of multimodal environmental feature vectors can be improved to sub-microsecond levels. Sub-microsecond alignment requires higher-speed hardware, but microsecond-level alignment is sufficient for most applications. The dynamic weight matrix can be persisted to a database, which records historical weight trends for long-term analysis; however, real-time processing is preferably stored in memory.

[0042] Example 2: See Figure 3In specific implementation, the parallel execution of system state estimation and external disturbance estimation is achieved by configuring a dual-channel Kalman filter bank. The dual-channel Kalman filter bank consists of two independent and structurally identical Kalman filters, which are deployed on dedicated digital signal processor cores. The first channel Kalman filter processes the battery electrochemical impedance spectroscopy data, which is acquired by a frequency scanner. The frequency scanner applies small AC excitation signals at logarithmic intervals in the range of 1Hz to 10kHz and measures the battery response. The measurement data includes the real and imaginary impedance values. The second channel Kalman filter analyzes the thermal conduction gradient distribution, which is captured by an infrared thermal imager. The infrared thermal imager scans the battery surface at a rate of 30 frames per second to generate a temperature field matrix. The dynamic weight matrix is ​​fed as input into a dual-channel Kalman filter bank. The first-channel Kalman filter outputs the trajectory of the battery's internal resistance change, which is obtained by updating the battery's equivalent circuit model parameters online using the recursive least squares method. The second-channel Kalman filter outputs the predicted value of thermal accumulation, which is based on the discretized heat diffusion equation for state prediction. The prediction step size is consistent with the frame period of the infrared thermal imager. The convolutional neural network is implemented using a five-layer structure, including an input layer, two convolutional layers, one pooling layer, and a fully connected output layer. The input layer receives multimodal environmental feature vectors from Example 1. The convolutional layers identify abnormal fluctuation patterns in the environmental feature vectors using preset convolutional kernels. Abnormal fluctuation patterns include sudden peaks in electromagnetic interference or turbulent changes in airflow. The fully connected output layer calculates the disturbance intensity coefficient, which is a scalar value between 0 and 1. The battery internal resistance change trajectory, thermal accumulation prediction value, and disturbance intensity coefficient are fused using a weighted average method. The weights of the weighted average method are dynamically allocated based on the estimated variance of each input quantity. After fusion, a system stability index is generated. The system stability index is a continuous value between 0 and 1. A value below 0.5 indicates that the system is in a stable state, and a value above 0.8 triggers a high-level alarm. The entire parallel execution process is hardware accelerated on a field-programmable gate array (FPGA). The parallel architecture of the FPGA ensures that the two Kalman filter channels and the convolutional neural network can run synchronously.

[0043] In some embodiments, the dual-channel Kalman filter bank can be implemented using extended Kalman filters or lossless Kalman filter variants to accommodate the nonlinear characteristics of the battery system, but the basic Kalman filter structure is the standard configuration due to its computational efficiency and stability. The frequency range for acquiring battery electrochemical impedance spectroscopy data can be extended to lower or higher frequencies, such as from 0.1 Hz to 100 kHz, to capture a more complete battery dynamic response, but the range of 1 Hz to 10 kHz already covers the main relaxation processes. The acquisition device for the thermal conductivity gradient distribution is not limited to an infrared thermal imager; a contact thermocouple array can also be used, but an infrared thermal imager offers the advantage of non-contact measurement. The number of layers in the convolutional neural network can be increased or decreased, for example, by adding batch normalization layers or dropping layers to improve generalization ability, but a five-layer structure strikes a balance between model complexity and real-time performance. The output range of the perturbation intensity coefficient can be adjusted to other intervals, such as -1 to 1, to distinguish perturbation polarity, but a range of 0 to 1 is sufficient to represent the perturbation amplitude. The threshold values ​​for system stability indicators can be adjusted according to specific application scenarios. For example, in scenarios with more stringent requirements, the alarm threshold can be set to 0.7, but 0.5 and 0.8 are used as general reference values.

[0044] In practical implementation, the state vector of the first-channel Kalman filter includes the battery open-circuit voltage, ohmic internal resistance, and polarization internal resistance, with a 3-dimensional state vector. The process noise covariance matrix is ​​initialized based on historical battery data, and the observation matrix is ​​constructed based on the principle of electrochemical impedance spectroscopy (EIS). The battery EIS data is updated every 100 milliseconds, and the Kalman gain matrix is ​​recursively calculated to ensure the smoothness of the battery's internal resistance change trajectory. The state vector of the second-channel Kalman filter includes the battery core temperature, surface temperature gradient, and heat flux density, with a 4-dimensional state vector. The heat conduction gradient distribution data is obtained by dimensionality reduction of the temperature field matrix from an infrared thermal imager. The dimensionality reduction method is principal component analysis, retaining the first three principal components. The predicted heat accumulation value is output through a linear combination of the state vectors, and the combination coefficients are determined by the material's thermophysical parameters. The training process of the convolutional neural network uses historical environmental data, which includes labeled anomalous events. The training algorithm is stochastic gradient descent with momentum, the convolutional kernel size is 3x3, and the activation function is ReLU. The sensitivity of identifying anomalous fluctuation patterns is controlled by adjusting the network weights. The weighted average method used in the data fusion stage employs inverse variance weighting. The variance of the battery internal resistance change trajectory is extracted from the error covariance matrix of the Kalman filter, the variance of the thermal accumulation prediction value is estimated using the Monte Carlo method, and the variance of the disturbance intensity coefficient is set to a fixed value. The fusion weights are renormalized to ensure a sum of 1. The system stability index is calculated as a weighted linear combination of all input quantities, with the combination coefficients set based on expert experience. The index value is refreshed every 50 milliseconds, with the refresh frequency synchronized with the control system cycle.

[0045] Optionally, a third channel can be added to the dual-channel Kalman filter bank for estimating the battery's state of charge, but the description in Example 2 is limited to a dual-channel structure. Measurement of the battery's electrochemical impedance spectroscopy data can employ multi-frequency simultaneous excitation techniques, which improve measurement speed, but sweep frequency measurement provides a higher signal-to-noise ratio. Analysis of the thermal conduction gradient distribution can be combined with computational fluid dynamics simulations, which provide information on the internal temperature field, but significantly increase implementation complexity. The input to the convolutional neural network can be extended to time-series data containing historical environmental feature vectors to identify dynamic patterns, but the current implementation uses instantaneous vectors. The data fusion method can be replaced with Kalman filter fusion or DS evidence theory to handle uncertainties, but the weighted average method is simple to implement and meets accuracy requirements.

[0046] In practical implementation, the parallel execution hardware adopts a pipelined architecture, which allows data to flow continuously between different processing units. Matrix operations of the dual-channel Kalman filter bank use a dedicated floating-point arithmetic unit supporting single-precision floating-point numbers. Convolution operations of the convolutional neural network are accelerated through a hardware multiplier array containing 256 parallel multiplication units. The output interface for system stability indicators is a shared memory area, employing a dual-port random access memory design. This allows the charging strategy module to directly read the latest indicator values, which are accompanied by timestamps and validity flags. The validity flags are set by the data integrity check module. The storage format for the battery internal resistance change trajectory is a circular buffer with a depth of 1000 points, supporting historical data backtracking. The thermal accumulation prediction value includes a confidence interval with a 95% probability range. An emergency power reduction command is triggered when the disturbance intensity coefficient exceeds 0.9. The timing of the entire system is managed by a high-precision timer with a resolution of 1 microsecond. Synchronization signals are distributed to each processing unit to ensure consistent estimation steps.

[0047] It is understandable that the parallel design of the dual-channel Kalman filter bank fully utilizes the independence of battery electrochemical impedance spectroscopy data and thermal conduction gradient distribution data. This independence allows the two channels to be optimized independently without interference. The introduction of convolutional neural networks compensates for the shortcomings of traditional filtering methods in environmental disturbance identification, which requires pattern recognition capabilities. The choice of the weighted average method is based on the quantifiability of the estimation errors of each input quantity. Quantifiability provides a clear mathematical basis for weighted fusion. The system stability index unifies the multi-source estimation results into a single scalar, which simplifies the decision-making logic of subsequent strategy modules. The implementation of the field-programmable gate array ensures that the complex algorithm can be completed within a finite time budget, which is crucial for real-time control systems.

[0048] In some embodiments, cross-coupling terms can be introduced between the channels of the Kalman filter bank to model the battery electrothermal effect. However, Embodiment 2 assumes weak coupling between the two channels, so independent processing is used. Preprocessing of the battery electrochemical impedance spectroscopy data can include impedance spectrum fitting, which uses an equivalent circuit model to extract parameters. However, the raw impedance data can be directly input into the filter. The acquisition frequency of the thermal conduction gradient distribution can be increased to 100 frames / second to capture rapid thermal changes, but this increases the data processing burden. The number of convolutional kernels in the convolutional neural network can be optimized, as it affects feature extraction capability, but is limited by hardware resources. The calculation of the perturbation intensity coefficient can introduce frequency weights, which emphasize the harmfulness of perturbations in specific frequency bands. However, the current implementation treats it as uniform weighting. The generation of system stability indices can use nonlinear fusion functions, such as fuzzy logic, but linear weighting is already sufficiently effective.

[0049] Optionally, system state estimation can be enhanced by adding a state of charge estimator based on voltage-current integration, which serves as an auxiliary verification mechanism. External disturbance estimation can incorporate a mechanical vibration sensor to detect structural disturbances. Convolutional neural network training can be performed online, adapting to environmental changes while offline training ensures initial performance. Data fusion weights can be adaptively learned using recursive least squares, but fixed weights simplify implementation. The system stability index output can incorporate hysteresis characteristics to prevent fluctuations near the threshold, but the current design uses a deterministic output.

[0050] See Figure 4 This figure is a core visualization result of battery state analysis in a UAV charging control system. The horizontal axis represents charging current, and the vertical axis represents battery voltage. The scatter plot colors are mapped to the state of charge (SOC) through a color gradient on the right, clearly showing the battery's electrical characteristics at different SOC stages through a gradient from deep purple to bright yellow. The trend in the figure shows that as the SOC gradually increases, the charging current first rises and then falls, while the battery voltage continuously decreases, exhibiting a significant negative correlation. This relationship intuitively reflects the dynamic coupling characteristics of voltage, current, and SOC during battery charging, and is a key input for dual-channel Kalman filter bank analysis: the first channel processes the battery's electrochemical impedance spectrum based on this voltage-current data, and the second channel combines thermal information to analyze the thermal conduction gradient, ultimately outputting the battery's internal resistance change trajectory and predicted heat accumulation value. This figure provides a visual basis for system state estimation, assisting in the subsequent fusion of environmental disturbance information to generate system stability indicators. It is an important data support for the closed loop of battery state perception-strategy adjustment-energy output in UAV charging control, demonstrating the core value of accurate battery perception in intelligent charging technology.

[0051] Example 3: In specific implementation, the operation steps for dynamically adjusting the charging strategy are based on a fuzzy control rule base established according to the system stability index. The fuzzy control rule base contains a nonlinear mapping relationship between charging rate and temperature threshold. The nonlinear mapping relationship is defined by a set of "if-then" rules, such as "if the system stability index is high and the battery temperature is low, then the charging rate is high". The input variables of the rule base are the system stability index and the real-time battery temperature, and the output variables are the charging rate setpoint and the temperature alarm threshold. The membership functions of the input and output variables adopt trigonometric functions. The fuzzy inference method is the Mamdani minimum-maximum method, and the defuzzification method adopts the centroid method. The particle swarm optimization algorithm searches for the optimal charging trajectory in the fuzzy control rule base. The particle position in the particle swarm optimization algorithm represents the charging current sequence, the particle velocity represents the charging current adjustment step size, the population size is set to 50 particles, and the number of iterations is set to 100 generations. The optimal charging trajectory needs to meet the battery life constraint condition. The battery life constraint condition is defined based on the Arrhenius model to define the maximum allowable temperature. The objective function is the weighted sum of charging time and temperature penalty terms. Real-time monitoring of the battery surface charge distribution is achieved through a surface potential sensor array. This array contains 16 sensing points evenly distributed on the battery surface. When charge non-uniformity is detected, an adaptive recalibration procedure is triggered. This procedure is implemented by applying a balancing current, the magnitude of which is calculated based on the charge non-uniformity. The digital twin model runs on a cloud server. Its inputs include historical battery data, real-time parameters, and candidate solutions for the optimal charging trajectory. The output is a validated power modulation waveform, a PWM signal with an adjustable frequency range from 1kHz to 10kHz. Validation is performed by comparing the battery response predicted by the digital twin model with the safety boundary.

[0052] The activation process of the compensation mechanism involves constructing a disturbance propagation model. This model includes analysis of the coupling path of electromagnetic interference through the charging interface. The coupling path is determined by measuring S-parameters using a vector network analyzer, and the S-parameters are stored as a frequency response matrix. A multi-order notch filter bank is designed, consisting of three digital filters. The transfer function of each filter is:

[0053]

[0054] in: Let represent the transfer function of the i-th filter. Let Z be the variable for transformation. Let be the digital angular frequency corresponding to the center frequency of the i-th filter. The radius of the pole of the i-th filter (which determines the bandwidth) is given. The value range is 1 to 3. The center frequency is dynamically adjusted based on the real-time acquired electromagnetic interference spectrum. The adjustment algorithm is based on the gradient descent method to find the peak frequency of the interference.

[0055] In some embodiments, the number of rules in the fuzzy control rule base can be expanded according to the control accuracy requirements, for example, from a simple 9 rules to 25 rules, but the core principles remain unchanged. The parameter settings of the particle swarm optimization algorithm can be adjusted, for example, the inertia weight can be linearly reduced from 0.9 to 0.4 to balance global and local searches, but the basic algorithm framework remains consistent. The density of the surface potential sensor array can be increased, for example, by using a 64-point array to improve spatial resolution, but a 16-point array already meets basic requirements. The complexity of the digital twin model can be increased, for example, by introducing an electrochemical-thermal coupling model, but current implementations are based on equivalent circuit models. The order of the multi-order notch filter bank can be varied, for example, using second-order or fourth-order filters, but a third-order design strikes a balance between selectivity and computational cost. The phase compensation circuit can be implemented using analog circuits or digital signal processors, but digital implementations offer greater flexibility.

[0056] In practical implementation, the establishment process of the fuzzy control rule base includes a knowledge acquisition stage, in which initial rules are determined through interviews with domain experts and analysis of historical data. The rule base is stored in non-volatile memory and supports online updates. The particle swarm optimization algorithm executes once per second. In each optimization iteration, the particle position update formula includes cognitive and social components. The constraint handling adopts the penalty function method. The optimal charging trajectory output is a time-current sequence with a sequence length of the predicted time domain for the next 30 seconds. The sampling frequency of the surface potential sensor array is 100Hz. The charge non-uniformity is quantified by calculating the standard deviation of the voltage at each sensing point. When the standard deviation exceeds the threshold of 5mV, an adaptive recalibration procedure is triggered. The balancing current is applied for 1 second, and the current magnitude is proportional to the non-uniformity. The digital twin model runs in a cloud virtual machine environment. The model parameters are updated periodically via a wireless network. The verification process includes checking safety indicators such as peak temperature and voltage overshoot. The duty cycle resolution of the power modulation waveform reaches 0.1%. The disturbance propagation model is constructed using transmission line theory, modeling the charging interface as a multi-port network. The electromagnetic interference coupling path is described by an S-parameter matrix with dimensions of 2x2xN (N being the number of frequency points). The multi-order notch filter bank is implemented based on a direct type II second-order section structure, with a real-time coefficient update period of 1 millisecond. The gradient descent step size is set to 0.001, and the center frequency tracking range covers 150kHz to 30MHz. The phase compensation circuit consists of a digital filter with a zero-pole configuration transfer function. Predistortion processing is applied before the power modulation waveform is generated, and the group delay compensation is calculated based on the interference frequency.

[0057] In some embodiments, an ambient temperature factor can be added as an input variable to the fuzzy control rule base, serving as a third input variable; however, the description in Embodiment 3 is limited to two input variables. The objective function of the particle swarm optimization algorithm can incorporate a cycle lifetime decay term, estimated based on rainflow counting, but the current implementation focuses on optimizing the charging process. The layout of the surface potential sensor array can be optimized, for example, by using a non-uniform distribution to focus on monitoring edge regions, but a uniform distribution is the standard approach. The bandwidth of the multi-order notch filter bank can be adaptively adjusted, correlated with interference intensity, but a fixed bandwidth simplifies implementation. The dynamically adjusted charging strategy can integrate a rule-based emergency strategy, which overrides the optimization results when an anomaly is detected; however, the description in Embodiment 3 prioritizes optimization. The activation threshold of the compensation mechanism can be set as an adjustable parameter, set according to the battery type, but 0.8 is used as a general critical value. The phase compensation circuit can add an adaptive filter feedforward path, which further suppresses periodic interference, but the basic compensation structure already includes the main functions.

[0058] In practical implementation, the fuzzy control rule base operates on an embedded microcontroller platform. The microcontroller performs 100 fuzzy inferences per second, with input variables acquired via ADC sampling and output variables output via DAC. The particle swarm optimization algorithm executes on the application processor, with each optimization taking no more than 500 milliseconds. The optimal charging trajectory is sent to the power converter via a CAN bus. The signal from the surface potential sensor array is transmitted through an isolation amplifier. The adaptive recalibration program is controlled by dedicated logic circuitry. The balancing current source is a programmable current source with a maximum output current of 2A. The digital twin model uses Ethernet as its communication interface, and the verification results are verified using CRC checksums to ensure integrity. The power modulation waveform generation module includes a waveform memory storing various predefined waveform templates. The parameters of the disturbance propagation model are calibrated using a vector network analyzer every 24 hours, and the S-parameter matrix is ​​stored in flash memory. The coefficient updates of the multi-order notch filter bank are processed by a coprocessor, and a stochastic gradient descent variant is used to avoid local minima. The predistortion parameter table of the phase compensation circuit is experimentally determined, covering the entire temperature range, and the group delay compensation is calculated using interpolation.

[0059] It is understandable that the fuzzy control rule base's ability to handle uncertainty makes it suitable for dealing with nonlinear problems in battery charging, including electrochemical relaxation and thermal hysteresis effects. The swarm intelligence of particle swarm optimization algorithms enables them to effectively explore high-dimensional search spaces, which correspond to multi-stage charging current optimization. Monitoring the surface charge distribution directly reflects the uniformity of ion migration within the battery, and this uniformity affects the battery's aging rate. Virtual verification using digital twin models reduces the risks of physical verification, which can damage the battery. The frequency-selective suppression characteristics of multi-order notch filter banks are particularly effective against narrowband interference, which is commonly found in switching power supply harmonics. The stability theory of Lyapunov functions provides a rigorous mathematical foundation for the compensation system, ensuring predictable performance.

[0060] In some embodiments, the defuzzification method for the fuzzy control rule base can be replaced with the area center method or the maximum value averaging method, but the centroid method provides a smoother output response. The social component of the particle swarm optimization algorithm can be calculated based on dynamic topology, such as ring or star connections, but the globally optimal topology is the standard choice. The update strategy for the digital twin model can be based on digital threading technology, which enables full lifecycle data synchronization, but periodic updates are sufficient. The implementation of multi-order notch filter banks can employ a lattice structure to improve numerical stability, but a direct structure is computationally more efficient. The predistortion of the phase compensation circuit can be combined with a memory polynomial model, which handles power amplifier nonlinearity, but current linear predistortion is sufficient.

[0061] Example 4: In specific implementation, referring to Table 1, the implementation steps of the energy output of the collaborative control charging equipment adopt a multi-phase interleaved parallel topology to distribute the charging current. The multi-phase interleaved parallel topology includes four independent Buck converters as power stages. The input terminals of the four Buck converters are connected in parallel to the DC bus, and the output terminals are connected to the battery terminals. The switching timing of each phase Buck converter is dynamically adjusted according to the thermal accumulation prediction value. The dynamic adjustment logic is implemented based on a lookup table stored in the microcontroller. The index of the lookup table is the thermal accumulation prediction value, and the output is the phase delay between each phase. The phase delay ensures that the conduction time of the four Buck converters is evenly distributed within a single switching cycle. Current ripple samples are acquired during each switching cycle using Hall effect current sensors installed after the inductor of each Buck converter. The sampling point is selected at the midpoint of the switch's off-state to capture the average current value. The current ripple samples are digitized using a 12-bit analog-to-digital converter at a rate of 1 trillion samples per second. The digitized samples are then sent to a digital signal processor (DSP) for Fast Fourier Transform (FFT) analysis. The FFT points are set to 1024, and the extracted harmonic component characteristics include the fundamental amplitude, second harmonic amplitude, and third harmonic amplitude. The dead time of the gate drive signal is adjusted based on the harmonic component characteristics using a programmable delay line (PLL) chip. The PLL chip receives the optimal dead time value calculated by the DSP. The optimal dead time value is inversely proportional to the third harmonic amplitude, and the adjustment step size is set to 10 nanoseconds. The condition for achieving zero-voltage switching is verified by monitoring the zero-crossing point of the drain-source voltage of the switch. The control signals of each phase are synchronized using fiber optic isolators. The fiber optic isolators use multimode fiber to transmit PWM signals. The transmission delay is precisely calibrated and compensated to ensure that the power synthesis waveform meets the battery charging characteristics requirements. The battery charging characteristics requirements include a current ripple factor of less than 5% in the constant current stage and a voltage overshoot of less than 2% in the constant voltage stage.

[0062] Table 1: Correspondence between predicted thermal accumulation values ​​and phase delay

[0063] Predicted range of heat accumulation (normalized) Phase delay (degrees) 0.0-0.2 90 0.2-0.5 75 0.5-0.8 60 0.8-1.0 45

[0064] In the specific implementation, the power devices in the multiphase interleaved parallel topology are selected as MOSFETs with a withstand voltage of 60V and an on-resistance of 10 milliohms. The switching frequency of each Buck converter is set to 100kHz, the output inductance is 10 microhenries, and the output capacitance is 100 microfarads. The thermal accumulation prediction value is obtained from the system stability estimation module of Example 2 through the serial peripheral interface bus, with an update frequency of 100Hz. The Hall effect current sensor has a bandwidth of 1MHz and a sensitivity of 100mV / A. The sampling window of the current ripple sample covers at least 16 complete switching cycles to ensure the frequency resolution of the Fast Fourier Transform. The results of the Fast Fourier Transform are stored in a dual-port random access memory for the harmonic analysis module to read. The generation of the gate drive signal is based on a dedicated PWM controller. The dead time register of the PWM controller is connected to the programmable delay line chip through a parallel bus. The verification circuit for zero-voltage switching conversion includes a high-speed comparator. The high-speed comparator generates an acknowledgment signal when the voltage across the switching transistor drops below 1V. The light source of the fiber optic isolator is an 850 nm vertical cavity surface-emitting laser, and the optical receiver is a PIN photodiode. The synchronization signal uses Manchester encoding to improve anti-interference capability. The quality of the power synthesized waveform is monitored online using an eight-bit high-precision oscilloscope, and the monitoring data is fed back to the central controller for closed-loop adjustment.

[0065] In some embodiments, the number of phases in a multiphase interleaved parallel topology can be extended to six or eight phases to support higher power applications, but a four-phase design strikes a good balance between complexity and performance. The sampling time for current ripple samples can be selected at different points during the switching process (either during the on-time or off-time), but the midpoint of the off-time period more accurately reflects the average current value. The number of analysis points for the Fast Fourier Transform can be adjusted to 512 or 2048 points to accommodate different real-time requirements, with 1024 points providing moderate frequency resolution and computational overhead. The selection of harmonic component characteristics can include higher harmonics such as the fifth harmonic, but the fundamental to third harmonics already contain the main energy components of the current ripple. The type of fiber optic isolator can be replaced by a magnetic isolator or a capacitive isolator, but fiber optic isolators offer the highest voltage isolation level and electromagnetic interference immunity.

[0066] Optionally, a current sharing control loop can be added to the multiphase interleaved parallel topology. This loop balances the current by adjusting the PWM duty cycle of each phase. However, the description in Example 4 focuses on achieving natural current sharing through interleaved timing. Analysis of current ripple samples can be combined with wavelet transform to extract transient features, but fast Fourier transform is more suitable for periodic harmonic analysis. Dead time adjustment can be based on direct measurement of the switching transistor junction temperature, but the method in this embodiment uses harmonic characteristics as an indirect and effective observation indicator. Verification of the power synthesized waveform can incorporate artificial intelligence algorithms for pattern recognition, but traditional monitoring methods based on oscilloscopes are sufficiently reliable.

[0067] In practical implementation, the control core of the multiphase interleaved parallel topology is a 32-bit microcontroller. The microcontroller runs a real-time operating system with a task scheduling cycle of 10 microseconds. The phase delay lookup table is stored in flash memory and supports online updates via a debug interface. The signal conditioning circuit of the Hall effect current sensor includes a second-order anti-aliasing filter with a cutoff frequency set to 500kHz. The temperature drift coefficient of the reference voltage source for the analog-to-digital converter is less than 10ppm / °C. The Fast Fourier Transform (FFT) is performed by a hardware accelerator, with a 1024-point transformation taking less than 50 microseconds. Harmonic component characteristics are extracted by comparing the amplitudes of each harmonic with preset thresholds. The programmable delay line chip has a 16-bit configuration word and a delay resolution of 5 picoseconds. The verification signal for zero-voltage switching is ANDed with the PWM generation logic to ensure that the switch is only allowed to conduct when the zero-voltage condition is met. The transmission channel of the fiber optic isolator adopts a redundant design, with automatic switching between primary and backup channels and a synchronization accuracy error of less than 100 picoseconds. The compliance check for battery charging characteristics is performed once per second, and the check results are recorded in non-volatile memory.

[0068] It is understandable that the multiphase interleaved parallel topology effectively cancels the current ripple of each phase through phase interleaving. This current ripple cancellation reduces the ripple coefficient of the total output current, thereby reducing battery thermal stress. Current ripple samples are collected and real-time feedback is provided during each switching cycle, enabling the system to dynamically respond to load changes and component parameter drift. Adjusting the dead time of the gate drive signal based on harmonic component characteristics directly optimizes switching losses, thus improving the overall efficiency of the charging system. Fiber optic isolators are used to synchronize the control signals of each phase to overcome timing deviations in long-distance transmission, ensuring the accuracy of multiphase power combining.

[0069] In practical implementation, the heat dissipation management of the multiphase interleaved parallel topology adopts direct water cooling of the substrate. The water-cooled plate is in direct contact with the power MOSFETs, and the predicted heat accumulation value is used to control the speed of the coolant pump. The installation position of the Hall effect current sensor is optimized through electromagnetic field simulation to minimize the measurement error caused by proximity effect. The digital data of the current ripple sample is timestamped to ensure synchronization with the switching cycle. The spectral leakage problem of the Fast Fourier Transform is handled by adding a Hanning window, and the Hanning window function is pre-stored in the hardware accelerator. The dead time adjustment of the gate drive signal follows the monotonicity principle, that is, the dead time decreases as the harmonic amplitude increases, but a minimum dead time limit is set to prevent bridge arm shoot-through. The fiber length of the fiber isolator is customized according to the rack layout, and the length error is calibrated by the delay compensation table in the software. The threshold parameters required by the battery charging characteristics can be configured by the host computer software to adapt to different battery models.

[0070] Example 5: In specific implementation, the continuous optimization process is achieved through a micron-scale temperature sensor array embedded in the surface of the battery electrodes. This array contains 100 platinum resistance thermometer-based sensing units, distributed in a 10x10 matrix on the electrode active material. Each sensing unit is 50 μm x 50 μm in size. The array is fabricated using photolithography and flexible circuit board leads. The sampling frequency for acquiring three-dimensional thermal distribution data is 10 Hz, and the data is transmitted to the signal conditioning circuit via a serial peripheral interface. An electro-thermal-mechanical multiphysics coupling model is established and runs on a real-time processor. The inputs to the electro-thermal-mechanical multiphysics coupling model include the three-dimensional thermal distribution data from the micron-scale temperature sensor array, the charging current value, and the battery voltage value. The algorithm for real-time inversion of the internal stress distribution state of the battery is based on the finite element method. The finite element mesh divides the battery interior into 100,000 hexahedral elements, and the coupled solution of the temperature field, electric field, and stress field is calculated for each element. When a localized hotspot is detected, a directional air-cooling system is activated for precise thermal management. This system consists of sixteen miniature brushless fans, mounted 5 mm above the battery pack. Each fan's speed is independently adjustable, with the speed control signal generated based on the location and temperature deviation of the localized hotspot. The parameters of the electro-thermal-mechanical multiphysics coupling model are optimized using a reinforcement learning algorithm. This algorithm employs a deep Q-network structure, with historical charging data and real-time 3D thermal distribution data as inputs, and outputs the model parameter adjustments. A learning rate of 0.01 ensures that prediction accuracy continuously improves with the number of charging cycles.

[0071] The abnormal situation handling strategy deploys an acoustic emission detection device to capture the ultrasonic signal characteristics of lithium plating inside the battery. The device includes four broadband piezoelectric sensors mounted on the four sides of the battery casing via coupling agent, with a frequency response range of 20kHz to 200kHz and a dynamic range exceeding 100 dB. Upon detection of specific frequency ultrasonic harmonics, the system immediately switches to pulse charging mode and reduces the average current. These specific frequencies are identified as characteristic peaks around 150kHz and 220kHz. The identification algorithm is based on Fast Fourier Transform and mode matching. The duty cycle of the pulse charging mode is linearly reduced from 100% to 50%, and the average current decreases proportionally to the ultrasonic signal amplitude. Simultaneously, an X-ray diffraction analyzer is activated to detect changes in the lattice structure of the electrode material. The X-ray diffraction analyzer uses a copper target Kα-ray source with a wavelength of approximately 0.154 nm, a diffraction angle scanning range of 10 degrees to 80 degrees, and a scanning speed of 5 degrees per second. If the lattice distortion rate exceeds the safety threshold of 1%, the emergency circuit breaker mechanism is triggered and the failure analysis data is saved. The lattice distortion rate is calculated by comparing the offset of the current diffraction peak position with the standard card peak position. The emergency circuit breaker mechanism disconnects the main contactor within 100 microseconds. The failure analysis data includes all sensor readings in the 60 seconds prior to the fault.

[0072] In some embodiments, the material for the micrometer-scale temperature sensor array can be a nickel-chromium alloy or a semiconductor PN junction, but platinum resistance thermometers offer better linearity and stability. The mesh generation for the electro-thermal-mechanical multiphysics coupling model can use tetrahedral elements or adaptive meshes, but hexahedral elements strike a balance between computational accuracy and efficiency. The implementation of the directional air-cooling system can be replaced with liquid-cooled microchannels or Peltier effect cooling, but air-cooling systems are simpler in structure and have a faster response. The network structure of the reinforcement learning algorithm can be replaced with the policy gradient method or the actor-commentator algorithm, but deep Q-networks are suitable for such discrete decision problems. The number of sensors in the acoustic emission detection device can be increased to eight to improve positioning accuracy, but four sensors are sufficient for basic detection requirements. The X-ray source for the X-ray diffraction analyzer can be a molybdenum target or a cobalt target to change the detection depth, but a copper target is a commonly used laboratory configuration.

[0073] In practical implementation, each sensing unit of the micron-level temperature sensor array is individually calibrated at 0°C and 50°C. Calibration coefficients are stored in the sensor's built-in memory. Three-dimensional thermal distribution data is reconstructed into a continuous temperature field using a three-dimensional interpolation algorithm. The solver for the electro-thermal-mechanical multiphysics coupling model is based on an explicit time integration method. At each time step, the coupling effect of the conductivity matrix, thermal conductivity matrix, and elasticity matrix is ​​calculated. The real-time inversion of the internal stress distribution of the battery is output as an equivalent stress cloud map. The directional air-cooling system is controlled using a fuzzy PID controller, with a fan speed adjustment range of 0 to 12000 rpm and a response time of less than 200 milliseconds. The reinforcement learning algorithm is trained on a battery cycling test bench. The reward function comprehensively considers temperature uniformity, charging speed, and stress level. Model parameter optimization is performed once after every 10 complete charges. The acoustic emission detection device's data acquisition card has a sampling rate of 1 MHz. Ultrasonic signal feature extraction includes waveform duration, rise time, and ring count. Ultrasonic harmonics of specific frequencies are initially screened using a bandpass filter bank. The switching of pulse charging modes is executed by a solid-state relay, and the switching process is completed within 10 microseconds. The average current is adjusted through PWM duty cycle control. The sample chamber of the X-ray diffraction analyzer is a radiation-proof sealed structure. The diffraction pattern is acquired by a two-dimensional detector, and the lattice distortion rate is calculated using a full-spectrum fitting method. The trigger circuit of the emergency circuit breaker mechanism is designed with hardware redundancy, including two independent comparator circuits. Failure analysis data is stored in a ferroelectric memory with power-down protection.

[0074] Optionally, the continuous optimization process could introduce a battery expansion deformation sensor as an additional input to the electro-thermal-mechanical multiphysics coupling model, but the implementation of Example 5 focuses on the temperature field as the primary observation object. Anomaly handling strategies could include adding gas chromatography to monitor electrolyte decomposition products, but the description in Example 5 is limited to acoustic emission and X-ray diffraction analysis. The layout of the directional air-cooling system could be optimized to a surround structure to improve heat dissipation uniformity, but this embodiment uses a top-mounted layout. The reward function of the reinforcement learning algorithm could incorporate economic indicators such as energy consumption costs, but the current implementation focuses on technical performance optimization.

[0075] In practical implementation, the leads of the micron-scale temperature sensor array are made of silver paste-printed wires with a wire width of 30 micrometers. Impedance matching networks ensure signal transmission quality, and the three-dimensional thermal distribution data is updated every 100 milliseconds. The material parameter database of the electro-thermal-mechanical multiphysics coupling model contains the elastic modulus and thermal expansion coefficient under different charging states, and the refresh cycle for real-time inversion of the internal stress distribution of the battery is 1 second. The fan drive circuit of the directional air cooling system uses an H-bridge topology, and wind speed feedback is obtained through a built-in Hall sensor. The deep Q-network of the reinforcement learning algorithm contains three hidden layers, each with 128 neurons, and the experience replay buffer has a capacity of 10,000 records. The acoustic emission detection device has a sensor sensitivity of 80 dB, a preamplifier gain of 40 dB, and the ultrasonic signal feature extraction process includes wavelet denoising. The parameters of the pulse charging mode are set through a lookup table, with the lookup table index being the ultrasonic signal amplitude and battery temperature. The average current reduction rate is limited to 0.1C per second to avoid secondary impact. The X-ray diffraction analyzer has a beam diameter of 0.5 mm and a scanning step size of 0.02 degrees. The lattice distortion rate is calculated using the least squares method to fit the diffraction peak positions. The emergency circuit breaker has a contactor breaking capacity of 100 amperes, and the arc suppression circuit uses an RC absorption network. The failure analysis data includes detailed waveform records for the 10 cycles prior to the fault.

[0076] Understandably, micron-scale temperature sensor arrays offer unprecedented spatial resolution, making it possible to detect localized thermal anomalies. The multiphysics coupling characteristics of the electro-thermal-mechanical multiphysics coupling model can more realistically simulate the complex interactions within a battery. Precise thermal management of the directional air-cooling system achieves a balance between energy efficiency and heat dissipation. The self-learning capability of reinforcement learning algorithms enables the system to continuously evolve. The acoustic emission detection device provides early warning of lithium plating, buying valuable time for safety intervention. X-ray diffraction analysis provides direct observation of the crystal structure, offering material-level evidence of failure. The rapid response of the emergency circuit breaker mechanism effectively prevents the escalation of faults.

[0077] In some embodiments, the distribution pattern of the micrometer-scale temperature sensor array can be optimized to a hexagonal tessellation to improve space utilization, but this embodiment uses a simple rectangular grid. The solution algorithm for the electro-thermal-mechanical multiphysics coupling model can be changed to an implicit method to improve numerical stability, but the explicit method is faster. The control algorithm for the directional air-cooling system can introduce predictive control to adjust the wind speed in advance, but embodiment 5 uses feedback control. Optionally, the abnormal situation handling strategy can add infrared thermal imaging as a supplementary detection method to the X-ray diffraction analyzer, but the implementation of embodiment 5 maintains the independence of the two technologies. The action time of the emergency circuit breaker mechanism can be further shortened by optimizing the contactor structure, but 100 microseconds already meets the safety standard.

[0078] In practical implementation, continuous optimization processes and anomaly handling strategies are interconnected via a shared data bus. For example, high-stress region information obtained from the inversion of the electro-thermal-mechanical multiphysics coupling model can be used to adjust the monitoring focus areas of the acoustic emission detection device. Long-term drift of the micron-level temperature sensor array is compensated for through periodic recalibration, with a recalibration cycle set every 30 days. The fan lifespan of the directional air-cooling system is predicted based on cumulative operating time, and the prediction results are used for planned maintenance. The sensor coupling status of the acoustic emission detection device is monitored in real time using the guided wave method to ensure good acoustic contact. Radiation safety of the X-ray diffraction analyzer is ensured by an interlocking device, allowing radiation emission only when the sample chamber is sealed. Periodic testing of the emergency circuit breaker mechanism is conducted by injecting simulated fault signals, with a testing cycle of every 24 hours.

[0079] See Figure 5 This chart, presented as a grouped bar chart, visually illustrates the frequency and handling effectiveness of five types of anomalies during drone charging. The horizontal axis represents the anomaly type, and the vertical axis represents the number of occurrences. Red bars represent the number of occurrences, and green bars represent the number of successful handlings. The data shows that voltage fluctuations are the most frequent anomaly, with a high success rate in handling, demonstrating the effectiveness of compensation mechanisms such as multi-order notch filter banks and phase compensation circuits. Electromagnetic interference exceeding limits occurs second most frequently, with the number of successful handlings approaching the number of occurrences, reflecting the suppression capability of the disturbance propagation model and notch filter bank for electromagnetic interference. The occurrence and handling frequency of abnormal temperature increases and uneven charge distribution are relatively balanced, indicating that the directional air cooling system and adaptive recalibration program can respond promptly to these anomalies. Current surges occur the least frequently, with a lower number of successful handlings, suggesting that there is still room for optimization in the emergency circuit breaker mechanism for these anomalies. This figure is a quantitative representation of the abnormal situation handling strategy. By comparing the occurrence and handling data of different abnormalities, the effectiveness of the fault handling mechanism in each link can be evaluated. It provides data support for subsequent reinforcement learning to optimize the electro-thermal-mechanical multi-physics coupling model and improve the emergency circuit breaking mechanism. It is an important visualization tool for the robustness analysis of the UAV charging control system.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for controlling the charging of a drone, characterized in that, The method includes: In response to receiving a charging request from the drone, a charging control process is initiated, which includes an environmental parameter acquisition phase and a battery status initialization phase. After the charging connection is established, based on the collected environmental parameters and the initialized battery state, the adaptive observer performs system state estimation and external disturbance estimation in parallel. Based on the real-time results of the system state estimation and external disturbance estimation, the charging strategy is dynamically adjusted, including the refresh of the power control signal and the activation of the compensation mechanism. Based on the refreshed power control signal and the activated compensation mechanism, the energy output of the charging equipment is controlled in a coordinated manner until the battery reaches the predetermined charging target. The specific implementation of the environmental parameter acquisition phase and the battery state initialization phase includes: synchronously capturing airflow velocity, ambient light intensity, and electromagnetic interference background values ​​in the charging compartment through a multi-source sensor array; sampling the battery polarization voltage using a high-precision analog-to-digital converter to establish a voltage-time rate of change curve; spatiotemporally aligning the airflow velocity, ambient light intensity, electromagnetic interference background values, and voltage-time rate of change curve to form a multimodal environmental feature vector; and constructing a dynamic weight matrix for the battery initial state based on the multimodal environmental feature vector.

2. The UAV charging control method according to claim 1, characterized in that, The parallel execution of system state estimation and external disturbance estimation is achieved through the following steps: configuring a dual-channel Kalman filter bank, with the first channel processing battery electrochemical impedance spectroscopy data and the second channel analyzing the thermal conduction gradient distribution; inputting the dynamic weight matrix into the dual-channel filter bank to output the battery internal resistance change trajectory and thermal accumulation prediction value, respectively; identifying abnormal fluctuation patterns in environmental feature vectors through a convolutional neural network and calculating the disturbance intensity coefficient; and fusing the internal resistance change trajectory, thermal accumulation prediction value, and disturbance intensity coefficient to generate a system stability index.

3. The UAV charging control method according to claim 2, characterized in that, The steps of the dynamic adjustment charging strategy include: establishing a fuzzy control rule base based on system stability indicators, wherein the fuzzy control rule base contains a nonlinear mapping relationship between charging rate and temperature threshold; using a particle swarm optimization algorithm to search for the optimal charging trajectory in the fuzzy control rule base, wherein the optimal charging trajectory satisfies battery life constraints; monitoring the charge distribution state on the battery surface in real time, and triggering an adaptive recalibration procedure when uneven charge is detected; verifying the feasibility of the charging trajectory through a digital twin model, and outputting the verified power modulation waveform.

4. The UAV charging control method according to claim 3, characterized in that, The activation process of the compensation mechanism specifically includes: constructing a disturbance propagation model, which includes analysis of the coupling path of electromagnetic interference through the charging interface; designing a multi-order notch filter bank, the center frequency of which is dynamically adjusted according to the real-time acquired electromagnetic interference spectrum; when the system stability index exceeds the critical value, activating the phase compensation circuit to pre-distort the power modulation waveform; and judging the convergence of the compensation system through the Lyapunov function to ensure that the disturbance suppression effect meets the design requirements.

5. The UAV charging control method according to claim 4, characterized in that, The implementation steps of the energy output of the coordinated control charging device include: distributing the charging current using a multi-phase interleaved parallel topology, and dynamically adjusting the switching sequence of each phase according to the predicted value of heat accumulation; collecting current ripple samples in each switching cycle and extracting harmonic component characteristics through fast Fourier transform; adjusting the dead time of the gate drive signal according to the harmonic component characteristics to achieve zero-voltage switching conversion; and using an optical fiber isolator to synchronize the control signals of each phase to ensure that the power synthesis waveform meets the battery charging characteristic requirements.

6. The UAV charging control method according to claim 1, characterized in that, The method also includes a continuous optimization process, which is achieved by: embedding a micron-level temperature sensor array on the surface of the battery electrode to collect three-dimensional thermal distribution data; establishing an electro-thermal-mechanical multiphysics coupling model to invert the internal stress distribution state of the battery in real time; when a local hot spot is detected, directional air cooling system is activated to implement precise thermal management; and the parameters of the electro-thermal-mechanical multiphysics coupling model are optimized through reinforcement learning algorithms so that the prediction accuracy continues to improve with the number of charging cycles.

7. The UAV charging control method according to claim 1, characterized in that, The method also includes an abnormal situation handling strategy, the execution steps of which include: deploying an acoustic emission detection device to capture the ultrasonic signal characteristics of lithium plating phenomenon inside the battery; when a specific frequency ultrasonic harmonic is detected, immediately switching to pulse charging mode and reducing the average current; simultaneously starting an X-ray diffraction analyzer to detect changes in the lattice structure of the electrode material; if the lattice distortion rate exceeds the safety threshold, triggering an emergency circuit breaker mechanism and saving the failure analysis data.

8. The UAV charging control method according to claim 1, characterized in that, The method also includes precise determination of charging termination. The control steps include: measuring the AC internal resistance of the battery using the four-terminal method and plotting the internal resistance-charge change curve; starting the relaxation measurement program when the second derivative of the internal resistance-charge change curve shows an inflection point and recording the open-circuit voltage decay characteristics; predicting the duration of the voltage plateau period using the Gaussian process regression algorithm and calculating the optimal termination charging point; maintaining micro-current maintenance after disconnecting the main charging circuit until the battery voltage stabilizes at the rated floating value.

9. A drone charging control system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the drone charging control method as described in any one of claims 1 to 8.

Images