evtol battery gate feedback active probing physical constraint safety gate method

Abstract

This invention discloses a physical constraint safety gating method for active detection in eVTOL batteries using a gated feedback approach. The method includes: acquiring online battery observation data, flight profiles, historical cycle data, and measured impedance / impulse response; inputting the above data into a physically constrained Bayesian neural network to obtain the posterior a posteriori of battery health status and the posterior a posteriori of equivalent circuit parameters; performing battery pack-level task replay, calculating the task failure probability, and determining the upper bound of the task failure probability confidence level; outputting the gating result based on the transfer calibration safety budget and the upper bound confidence level, and determining the detection quantity budget and the weights of energy cost, temperature rise cost, and control intrusion cost; calculating the Fisher information matrix of candidate actions and projecting it onto the task failure probability gradient to obtain the risk information value; constructing an objective function with the risk information value as a positive term and selecting actions to execute; and obtaining the measured response to update the posterior a posteriori. This invention improves the reliability of state estimation through physical constraints and optimizes detection efficiency through gating feedback.

Description

Technical Field

[0001] This invention belongs to the field of electric aircraft power battery management technology, and relates to an eVTOL battery gating feedback active detection physical constraint safety gating method. Background Technology

[0002] eVTOL (Electric Vertical Take-Off and Landing) batteries not only need to provide high-rate peak power, but also need to maintain interpretable power margins, energy margins, and thermal margins across different mission phases such as hovering, climb, cruise, return, and emergency landing. If the battery state assessment is too optimistic, the aircraft may be falsely released, resulting in a false clear; if the assessment is too conservative, it may result in a hold or derate, reducing mission availability. Existing battery management technologies combine predetermined flight profiles with battery models to estimate range or functional state. For impedance observations, the Fisher information matrix and Cramer-Rao lower bound can be used to evaluate the contribution of different EIS frequencies to the identification of equivalent circuit parameters. For neural network estimation, physical information neural networks can use voltage equations, thermal equations, or aging laws in the battery model as soft constraints.

[0003] Furthermore, existing applications of physical information neural networks or Bayesian neural networks in battery management typically focus on the estimation error of health status or remaining useful life (RUL), failing to incorporate safety budgets, upper bounds on mission risk confidence, and posterior updates after active detection into a unified closed-loop framework. This deficiency is particularly pronounced under conditions of cross-cell and cross-data domain migration. The lack of a mission risk-driven closed-loop feedback mechanism makes safety gating systems prone to outputting false CLEAR, i.e., incorrectly allowing full-power operation when the actual battery state does not meet mission safety requirements. In public datasets or offline playback experiments, existing methods do not rigorously distinguish between real measured responses, near-real responses, and synthetic responses, often using simulated synthetic data as real observations for model training or posterior updates, resulting in insufficient generalization ability and reliability of the model in real-world deployment scenarios.

[0004] Therefore, there is an urgent need for a management method for the power batteries of electric aircraft, which can form a clear closed-loop coupling between mission profile, physical constraint learning, migration calibration gating, next-moment action selection and posterior update, effectively control the false CLEAR risk under migration uncertainty, and reduce the energy consumption, temperature rise cost and control intrusion cost brought about by active detection. Summary of the Invention

[0005] This invention provides an eVTOL battery-gated feedback-based active detection physical constraint safety gating method to solve the above-mentioned problems, including: Acquire online observation data, flight mission profile data, battery historical cycle data, and measured impedance response data and impulse response data of the eVTOL battery; Online observation data, historical cycle data, measured impedance response data, and impulse response data of the eVTOL battery are input into a physically constrained Bayesian neural network to obtain the battery health state posterior, equivalent circuit parameter posterior, task risk posterior, and physical residual; the training loss of the physically constrained Bayesian neural network includes at least the physical residual loss. Based on the flight mission profile data, the battery health status posterior is sampled to obtain the SOH sample, and the battery physical parameter sample is determined based on the equivalent circuit parameter posterior. The SOH sample and the battery physical parameter sample are then used to perform battery pack-level mission replay to calculate the mission failure probability. Based on historical battery cycling data, gating calibration data is constructed. The upper confidence bound of the task failure probability is determined by the task failure probability and the gating calibration data. The safety gating threshold that meets the preset migration safety budget is determined by the gating calibration data. Based on the comparison results of the task failure probability, the upper confidence bound of the task failure probability and the safety gating threshold, the safety gating result is output. Based on the security gating results, determine the budget for the number of detection actions, the weight of energy cost, the weight of temperature rise cost, and the weight of control intrusion cost for the next moment; For each candidate detection action in the preset candidate frequency set, calculate the energy cost, temperature rise cost, and control intrusion cost corresponding to each candidate detection action; Calculate the Fisher information matrix for each candidate detection action, and then project the Fisher information matrix onto the equivalent circuit parameters using the gradient of the task failure probability to obtain the risk information value. Active detection actions are selected based on an objective function consisting of risk information value, energy cost, temperature rise cost, and intrusion control cost. Based on the online measured response or bench response corresponding to the active detection action, update the posterior of the equivalent circuit parameters, the probability of task failure, and the upper confidence bound of the probability of task failure.

[0006] Furthermore, the online observation data includes at least one of voltage, current, temperature, state of charge, capacity, cycle number, and measured impedance characteristics; the flight mission profile data includes power requirements, duration, energy requirements, and thermal constraints for at least one stage of the mission, including takeoff, hovering, climb, cruise, return, diversion, or de-rated mission.

[0007] Furthermore, the training loss of the physical constraint Bayesian neural network includes at least the physical residual loss; the physical residual loss includes at least the equivalent circuit model voltage residual, SOC dynamics residual, thermal model residual, EIS complex impedance residual, pulse dynamic resistance residual, and task margin physical consistency residual.

[0008] Furthermore, the task margin physical consistency residual is: ; in, θ represents the physical consistency residual of the task margin; θ represents the battery physical parameter vector, which is determined by the posterior mean of the equivalent circuit parameters. Indicates the power required for the flight mission; This indicates the available power that the battery pack can provide at θ. This indicates the energy required for a flight mission; This indicates the available energy that the battery pack can provide at θ. This represents the predicted temperature of the battery pack at the end of the task at θ. Indicates the upper limit of the battery's safe operating temperature; This represents the positive part of the function, meaning only positive numbers are retained, while negative numbers and zero are both represented as 0.

[0009] Furthermore, the probability of task failure is: ; in, This represents the probability of task failure under the current input state x and the task profile π. Represents probability; This represents the sample of battery physical parameters determined based on the current input state x, the SOH sample, and the posterior distribution of the equivalent circuit parameters. π represents the task failure event corresponding to the task profile π. During task replay, the battery physical parameter samples and SOH samples are replayed at the battery pack level. When the task failure event occurs, the value is 1; q(SOH|x) represents the posterior distribution of SOH obtained from the current input state x. The battery health status (SOH) is indicated from... Sampling.

[0010] Furthermore, the security gating result includes three states: release, de-quota, and hold; the security gating threshold includes a release threshold, a de-quota threshold, and a risk threshold; when the upper confidence bound of the task failure probability is lower than the release threshold, the output is release; when the upper confidence bound of the task failure probability is between the release threshold and the de-quota threshold and the task failure probability is lower than the risk threshold, the output is de-quota; otherwise, the output is hold.

[0011] Furthermore, the energy cost is determined by the estimated additional electrical energy consumption generated by performing the candidate detection action, the temperature rise cost is determined by the estimated additional temperature rise generated by performing the candidate detection action, and the control intrusion cost is determined based on the duration of the candidate detection action and the power margin and battery thermal state of the current task.

[0012] Furthermore, the risk information value is determined based on the Fisher information matrix and the gradient of the task failure probability with respect to the equivalent circuit parameter vector, and is used to estimate the expected decrease in the variance of the task failure probability after performing the candidate detection action.

[0013] Furthermore, the objective function is determined according to the following formula: ; in, Indicates the candidate detection actions under the current security gating level g. The objective function value, This indicates the expected reduction in the uncertainty of mission risk caused by the candidate detection actions. , and These represent the energy cost, temperature rise cost, and intrusion control cost corresponding to the candidate detection actions, respectively. , and These represent the preset energy cost weight, temperature rise cost weight, and control intrusion cost weight, respectively. This represents the risk information gain coefficient determined by the current security gating level.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the eVTOL battery gated feedback active detection physical constraint safety gating method.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention projects the Fisher information matrix onto the gradient of parameters using the mission failure probability, thereby changing the selection criterion for active detection actions from maximizing parameter identifiability to maximizing the reduction of the uncertainty of the mission failure probability, while simultaneously constraining energy cost, temperature rise cost, and control intrusion cost.

[0016] In this invention, the safety gating result is not merely the final output, but a feedback control quantity that adjusts the number of detection actions, energy cost weight, temperature rise cost weight, and control intrusion cost weight for the next moment. After obtaining the mission failure probability, this invention actively selects the next detection action and updates the posterior parameters and the upper bound of the mission failure probability confidence through the actual response or bench response. This creates a two-way closed-loop coupling between mission risk assessment and active detection, rather than the unidirectional serial data flow that predicts mission feasibility from battery state in existing technologies.

[0017] In this invention, the training loss of the physically constrained Bayesian neural network includes at least the physical residual loss, which encompasses at least one of the following: equivalent circuit model voltage residual, state-of-charge dynamics residual, thermal model residual, EIS / pulse impedance consistency residual, pulse dynamic resistance residual, and mission margin physical consistency residual. During the training phase, this physical residual loss constrains the network output's battery health posterior, equivalent circuit parameter posterior, and mission risk posterior to conform to equivalent circuit physical consistency. Furthermore, during the inference phase, this physical residual serves as an observable output, unified with the transfer calibration risk upper bound, safety gating, and active detection posterior update within the electric flight mission safety closed loop, rather than merely serving as a regularization term during training. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0019] Figure 2 This is a comparison chart of the false release rate in an embodiment of the present invention.

[0020] Figure 3 This is a comparison chart of the compliance of the detection budget in the embodiments of the present invention.

[0021] Figure 4 This is a comparison chart of the measured accurate matching rates in the embodiments of the present invention. Detailed Implementation

[0022] To facilitate understanding of the present invention, the following description is provided in conjunction with embodiments. It should be understood that the following embodiments are for illustrative purposes only and do not constitute a limitation on the scope of protection.

[0023] like Figure 1 As shown, this invention provides an eVTOL battery-gated feedback-based active detection physical constraint safety gating method, comprising: Acquire online observation data, flight mission profile data, battery historical cycle data, and measured impedance response data and impulse response data of the eVTOL battery; Online observation data, historical cycle data, measured impedance response data, and impulse response data of the eVTOL battery are input into a physically constrained Bayesian neural network to obtain the battery health state posterior, equivalent circuit parameter posterior, task risk posterior, and physical residual; the training loss of the physically constrained Bayesian neural network includes at least the physical residual loss. Based on the flight mission profile data, the battery health status posterior is sampled to obtain the SOH sample, and the battery physical parameter sample is determined based on the equivalent circuit parameter posterior. The SOH sample and the battery physical parameter sample are then used to perform battery pack-level mission replay to calculate the mission failure probability. Based on historical battery cycling data, gating calibration data is constructed. The upper confidence bound of the task failure probability is determined by the task failure probability and the gating calibration data. The safety gating threshold that meets the preset migration safety budget is determined by the gating calibration data. Based on the comparison results of the task failure probability, the upper confidence bound of the task failure probability and the safety gating threshold, the safety gating result is output. Based on the security gating results, determine the budget for the number of detection actions, the weight of energy cost, the weight of temperature rise cost, and the weight of control intrusion cost for the next moment; For each candidate detection action in the preset candidate frequency set, calculate the energy cost, temperature rise cost, and control intrusion cost corresponding to each candidate detection action; Calculate the Fisher information matrix for each candidate detection action, and then project the Fisher information matrix onto the equivalent circuit parameters using the gradient of the task failure probability to obtain the risk information value. Active detection actions are selected based on an objective function consisting of risk information value, energy cost, temperature rise cost, and intrusion control cost. Based on the online measured response or bench response corresponding to the active detection action, update the posterior of the equivalent circuit parameters, the probability of task failure, and the upper confidence bound of the probability of task failure.

[0024] In this embodiment, the definition of the battery cell model on which the physical constraint Bayesian neural network is based and its training process are given.

[0025] The parameters of the second-order Thevenin equivalent circuit are defined as follows: ; Where θ represents the battery physical parameter vector; This indicates the battery's available capacity in its current health condition; This indicates the ohmic internal resistance of the battery; and These represent the polarization resistances of the first polarization RC branch and the second polarization RC branch, respectively. and These represent the polarization capacitors of the first polarization RC branch and the second polarization RC branch, respectively. This represents the equivalent heat transfer coefficient between the battery and the external environment. η represents the equivalent heat capacity of the battery; η represents the coulombic efficiency. This represents the Warburg diffusion impedance coefficient.

[0026] For the second-order Thevenin model, the terminal voltage satisfies: ; in, z(t) represents the predicted terminal voltage at time t; t represents time; z(t) represents the battery state of charge at time t, i.e., SOC; T(t) represents the cell temperature at time t; SOH represents the battery health state; I(t) represents the battery current at time t. In this embodiment, I(t) > 0 indicates that the battery is discharging to the load, and I(t) < 0 indicates that the battery is charging. It represents the open-circuit voltage function determined by the battery's state of charge, cell temperature, and battery health status. and Let represent the polarization voltages of the first and second polarized RC branches at time t, respectively. The dynamic responses of both polarized RC branches follow the dynamic response law of a first-order linear time-invariant system. The time constants of the first and second polarized RC branches are respectively... and .

[0027] The rate of change of SOC is determined by coulombic efficiency, discharge current, and battery available capacity, following the standard ampere-hour integration principle.

[0028] The thermal model uses a lumped heat capacity model: ; in, This indicates the rate of change of cell temperature over time. This represents the Joule heat generated by the ohmic internal resistance of the battery; i represents the number of the polarized RC branch, and i=1, 2; hA represents the heat dissipation generated by the i-th polarized RC branch; hA represents the equivalent heat transfer coefficient between the battery and the external environment; T(t) represents the cell temperature at time t. Indicates the ambient temperature at time t; This represents the heat transferred from the battery to the external environment. If the online system cannot directly observe this... Alternatively, an equivalent resistance can be used to make a conservative approximation of the heat generation term in the polarization branch.

[0029] For the EIS complex impedance, let ,available: ; in, This represents the EIS complex impedance calculated from θ at an angular frequency ω; ω represents the angular frequency; f represents the frequency of the EIS excitation signal; π represents pi; and j represents the imaginary unit, satisfying j²=-1. This represents the Warburg impedance term used to characterize the low-frequency diffusion characteristics of a battery.

[0030] For a short pulse with a duration of τ, the corresponding pulse dynamic resistance can be calculated using a second-order Thevenin model: ; in, This represents the pulse dynamic resistance calculated from θ under the action of a short pulse of duration τ; τ represents the duration of the short pulse; i represents the number of the polarized RC branch. and Let represent the polarization resistance and polarization capacitance of the i-th polarized RC branch, respectively; Let represent the time constant of the i-th polarized RC branch; the exponent term represents the dynamic response of the i-th polarized RC branch during the pulse duration τ. If the applied pulse current amplitude is ΔI and the corresponding battery terminal voltage change is ΔV, then the observed dynamic resistance is: ; in, ΔI represents the observed dynamic resistance measured based on the pulse current and battery terminal voltage response; ΔV represents the pulse current amplitude; and ΔV represents the change in battery terminal voltage caused by the pulse.

[0031] Furthermore, the Physically Constrained Bayesian Neural Network takes online observation data as input and outputs the posterior mean of SOH, the posterior standard deviation of SOH, the predicted value of task risk probability, and the mean and logarithmic variance of the posterior approximation of physical parameters.

[0032] The SOH posterior follows a Gaussian distribution with the mean of the SOH posterior of the network output as the mean and the variance of the SOH posterior as the variance. Its negative log-likelihood loss is composed of the square of the prediction residual divided by the variance plus the logarithm of the variance. The task risk loss is calculated using binary cross-entropy, based on the risk probability output by the network and the actual risk label. Impedance-assisted loss is the mean square error between the network's predicted impedance and the measured impedance, used to help the network learn impedance-related features. Domain adversarial loss is achieved by using a gradient reversal layer to make it difficult for the latent features learned by the network to distinguish between different data domains, and is implemented using cross-entropy loss. SOH monotonicity loss is used to suppress the abnormal increase of predicted SOH with the number of cycles for the same cell, and to penalize the negative slope of SOH increase with the number of cycles; the number of cycles refers to the number of charge-discharge cycles experienced by the same cell. Uncertainty calibration loss is used to ensure that the predicted confidence interval meets the target empirical coverage. The parameter prior KL divergence is the KL divergence between the posterior distribution of the equivalent circuit parameters of the network output and the standard normal prior. Physical residuals are used to constrain the posterior consistency of the equivalent circuit parameters of the network output with the physical consistency of the equivalent circuit, including: The voltage residual of the equivalent circuit model is determined by the deviation between the predicted terminal voltage and the measured terminal voltage. The dynamic residual of the RC branch is determined by the residual of the differential equation of the dynamic characteristics of the RC network during charging and discharging. The charge conservation residual is determined by the SOC rate of change residual described by the ampere-hour integral model; The residual temperature rise rate is determined by the deviation between the temperature change rate predicted by the lumped heat model and the measured temperature change rate. The EIS / pulse impedance consistency residual is determined by the deviation between the complex impedance or pulse response resistance predicted by the equivalent circuit model and the measured value. The task margin residual consists of three components: power margin, energy margin, and temperature rise margin. A positive residual is generated when any of these margins falls below the requirement, and it is used to constrain the physical consistency of the network output near the boundary of the task feasible region. ; in, θ represents the physical consistency residual of the task margin; θ represents the battery physical parameter vector, which is determined by the posterior mean of the equivalent circuit parameters. Indicates the power required for the flight mission; This indicates the available power that the battery pack can provide at θ. This indicates the energy required for a flight mission; This indicates the available energy that the battery pack can provide at θ. This represents the predicted temperature of the battery pack at the end of the task at θ. Indicates the upper limit of the battery's safe operating temperature; Denotes the positive part function, Equivalent to When a is greater than 0, take a; when a is less than or equal to 0, take 0.

[0033] The total training loss is: ; in, The negative log-likelihood loss is the posterior result of the SOH. The weights for the posterior negative log-likelihood loss of SOH; For mission risk and loss, Assign a risk-loss weight to the task. For impedance-assisted loss, Impedance-assisted loss weighting; The cross-entropy domain migration loss between the CMU and NASA domains. The weights are used for cross-entropy domain migration loss. For the monotonicity loss of SOH, The weights for the monotonicity loss of SOH; For the uncertainty calibration loss, Adjust the loss weights for uncertainty; The Kullback-Leibler divergence is the parametric prior under the standard normal prior. The parameter prior is the Kullback-Leibler divergence weight under the standard normal prior; , , , and These are the voltage residual, temperature rise rate residual, charge conservation residual, EIS / pulse impedance consistency residual, and task margin residual, respectively, of the equivalent circuit model. , , , and These are the residual weights for the equivalent circuit model voltage, the residual weight for the rate of temperature rise, the residual weight for charge conservation, the residual weight for EIS / pulse impedance consistency, and the residual weight for task margin.

[0034] In this embodiment, =1.0, =0.35, =0.10; =0.005, =0.05, =0.12, =0.01; =0.006, =0.003, =0.001, =0.015, =0.020.

[0035] In this embodiment, the battery pack is composed of multiple series and parallel cells. The equivalent internal resistance and open-circuit voltage of the battery pack are determined based on the equivalent internal resistance and open-circuit voltage characteristics of each cell, as well as the state of charge, temperature and health status.

[0036] Based on the minimum bus voltage allowed by the maximum power transfer theorem, the conservative available power of the battery pack is calculated. This conservative available power is the smaller of the theoretical maximum transfer power and the power constrained by the minimum voltage. Simultaneously, based on the battery pack's total nominal capacity, current state of health, average cell operating voltage during the mission phase, and reserved safe state of charge threshold, the available energy of the battery pack is estimated.

[0037] For a mission profile that includes power and energy requirements, temperature limits, and duration for each stage, define various safety margins: The power margin is the available power of the battery pack under the battery physical parameter vector minus the power required for the task. The energy margin is the available energy of the battery pack under the battery physical parameter vector minus the energy required for the task. The temperature margin is the upper limit of the battery's safe operating temperature minus the predicted temperature at the end of the task under the battery's physical parameter vector; The health margin is the SOH sample value obtained from the SOH posterior distribution minus the minimum SOH required for the task; when using the posterior mean for nominal state assessment, the health margin is the SOH posterior mean minus the minimum SOH required for the task. If any of the above margins is less than 0, it is considered a task failure event.

[0038] Under the SOH posterior and current battery parameter state estimation, the task failure probability is calculated through Monte Carlo task replay. For each task replay, an SOH sample is obtained from the SOH posterior distribution, and the battery parameter state used for this task replay is determined based on this SOH sample, the current online state, and the battery model. Battery pack-level task replay is then performed, and it is determined whether a task failure event has occurred. The task failure probability is expressed as: ; in, This represents the probability of task failure under the current input state x and the task profile π. Represents probability; This represents the sample of battery physical parameters determined based on the current input state x, the SOH sample, and the posterior distribution of the equivalent circuit parameters. π represents the task failure event corresponding to the task profile π. During task replay, the battery physical parameter samples and SOH samples are replayed at the battery pack level. When the task failure event occurs, the value is 1; q(SOH|x) represents the posterior distribution of SOH obtained from the current input state x. The battery health status (SOH) is indicated from... Sampling.

[0039] Furthermore, based on the preset migration calibration safety budget and the confidence upper bound of the mission failure probability, the safety gating result is output; After the network training is completed, gating calibration data is constructed from the replay events of the reserved battery cells that have never participated in the training and verification of the physical constraint Bayesian neural network. In each task scenario, a one-cell-reserved approach is adopted, and the safety gating threshold is determined by the replay events of the other reserved battery cells, and an independent evaluation is performed on the current reserved battery cell.

[0040] In this embodiment, the 95th percentile of the SOH standardized residuals in the gated calibration data is taken as the calibration quantile. The SOH calibration confidence interval is the posterior mean of the SOH plus or minus the product of the calibration quantile and the posterior standard deviation of the SOH. Based on the posterior mean of the task failure probability, the posterior standard deviation of the task failure probability, and the calibration quantile corresponding to the task failure probability, an upper confidence bound for the task failure probability is constructed: ; in, This represents the posterior mean of the probability of task failure. Let be the posterior standard deviation of the probability of task failure; The posterior calibration quantile is used to calculate the probability of task failure.

[0041] The migration calibration safety budget includes a mis-release budget and an erroneous de-rating release budget. The mis-release budget is the upper limit of the proportion of infeasible task samples in the gated calibration data that are judged as released, and the erroneous de-rating release budget is the upper limit of the proportion of infeasible task samples that are judged as de-rating. In this embodiment, the mis-release budget is 0, and the erroneous de-rating release budget is 0.01.

[0042] The release threshold, derating threshold, and risk threshold are determined by grid search on the gating calibration data. The grid search traverses all candidate threshold combinations within a preset threshold range and selects the threshold combination that ensures that the false release rate and false derating rate of each cell calibration result do not exceed the above-mentioned safety budget as the gating judgment condition. The release threshold is less than the derating threshold.

[0043] When the upper confidence bound of the task failure probability is lower than the release threshold and the calibration is effective, the output is "release"; when the upper confidence bound of the task failure probability is between the release threshold and the reduction threshold, and the task failure probability is lower than the risk threshold, the output is "reduction"; otherwise, the output remains unchanged. In this embodiment, the release threshold is 0.05, the reduction threshold is 0.35, and the risk threshold is 0.20.

[0044] Furthermore, the budget for the number of detection actions and the risk information gain coefficient for the next moment are determined based on the security gating result. Specifically, when the security gating result is to allow, a lower budget for the number of detection actions and a lower risk information gain are set. In the derated state, a medium budget for the number of detection actions is set and the risk information gain is increased. In the held state, a higher budget for the number of detection actions is set and the risk information gain is further increased to strengthen active detection in the high-risk state.

[0045] For any candidate detection action, a Fisher information matrix is ​​constructed based on the sensitivity of the observation model to the equivalent circuit parameters and the measurement noise covariance. The Fisher information matrix is ​​then projected onto the equivalent circuit parameters using the gradient of the task failure probability to obtain the risk information value of the candidate detection action in relation to the uncertainty of task failure. The system compares the risk information value of each candidate action with the energy consumption, battery temperature rise, and control intrusion cost associated with executing the action, and selects the candidate detection action with the largest comprehensive objective function value for measurement.

[0046] Using the value of the aforementioned risk information as a positive term, and energy consumption, battery temperature rise, and control intrusion costs as penalty terms, a comprehensive objective function is constructed: ; in, This represents the objective function value of candidate detection action a under the current security gating level g; This indicates the expected reduction in uncertainty regarding the direction of mission risk caused by candidate detection actions; This represents the estimated additional energy consumption corresponding to the candidate detection action; This indicates the estimated additional temperature rise caused by the candidate detection action; This represents the control intrusion cost corresponding to the candidate detection action, which is determined based on the duration of the detection action, the current task power margin, and the battery thermal state. , and These represent the preset energy cost weight, temperature rise cost weight, and control intrusion cost weight, respectively. This represents the risk information gain coefficient corresponding to the current security gating level. In the additive objective function of this embodiment, the energy cost weight, temperature rise cost weight, and control intrusion cost weight are preset values, respectively. =0.04、 =0.7 and =0.2.

[0047] In the shadow playback embodiment of measured action alignment, to reduce the impact of different cost item scales on the ranking of candidate frequency points, the candidate frequency points can also be ranked according to the normalized ratio of risk information benefit to comprehensive detection cost: ; in, This represents the normalized ranking score of candidate detection action a under the current security gating level g; This represents the task risk and benefit score corresponding to the candidate detection action; This represents the risk information gain coefficient corresponding to the current security gating level; , and These represent the energy cost, temperature rise cost, and intrusion control cost corresponding to the candidate detection actions, respectively. , and These represent the weights of the three types of costs in the normalized ranking; A positive stabilizing term is included to prevent the denominator from reaching zero. In this embodiment, the positive stabilizing term is taken as... =0.05、 =0.8, =0.2. Within the budgeted number of detection actions given the current security gating level, according to Candidate detection actions are ranked from largest to smallest, with priority given to those that have higher mission risk and information benefits and lower overall detection costs.

[0048] In the open state, set a low budget for the number of detection actions, select 3 candidate EIS frequencies, and take the risk information gain coefficient. =1.0.

[0049] Under reduced load conditions, the system moderately increases active detection to further confirm security boundaries, sets a moderate budget for the number of detection actions, selects 5 candidate EIS frequencies, and takes the risk information gain coefficient. =2.2.

[0050] In the hold-up state, the battery task risk or risk uncertainty is high. The system increases the intensity of active detection, sets a higher budget for the number of detection actions, selects 7 candidate EIS frequency points, and takes the risk information gain coefficient. =4.0.

[0051] Within the action quantity budget, the candidate detection action that maximizes the objective function value is selected as the active detection action, and the weights corresponding to energy consumption, battery temperature rise and control intrusion cost are dynamically adjusted by the current gating level.

[0052] In the offline measured shadow playback of this embodiment, after selecting the active detection action, the corresponding measurement response is matched from a unified measured observation library. For EIS frequency detection, the measured value of complex impedance under the same or similar cell state and detection frequency conditions is matched; for pulse detection, the voltage response characteristics under the corresponding pulse conditions and the available temperature response or estimated additional temperature rise characteristics are matched.

[0053] Fisher information increments are constructed based on the observation model sensitivity and measurement noise level corresponding to the selected active detection action. The prior accuracy matrix corresponding to the current equivalent circuit parameter covariance is added to the Fisher information increments to obtain the posterior covariance matrix of the equivalent circuit parameters after active detection. ; in, This represents the observation sensitivity matrix corresponding to the candidate detection actions. for The transpose of the matrix, This represents the inverse matrix of the measurement noise covariance matrix. This is the inverse of the posterior covariance matrix of the current equivalent circuit parameters. This is the posterior covariance matrix of the equivalent circuit parameters after active detection. Furthermore, the reduction in risk uncertainty resulting from this action is estimated based on the change in the variance of the mission risk direction before and after active detection.

[0054] Based on the decrease ratio of the variance of the mission risk direction before and after active detection, and combined with the matching quality of the measured response, the uncertainty interval of the mission failure probability is updated; while keeping the current central estimate of the mission failure probability unchanged, the updated upper bound of the mission failure probability confidence is used to re-execute the safety gating, thereby forming a closed loop of gating, active detection, uncertainty update and re-gating.

[0055] In this embodiment, the system unifies publicly available data or bench data into a measured_observation_bank.csv file. This file includes at least the following fields: dataset_name, cell_id, cycle_id, timestamp, soc, soh, capacity, voltage, current, temperature, operation_type, eIS_frequency, z_real, z_imag, pulse_width, pulse_current, pulse_voltage_response, pulse_temperature_response, source_file, and source_row.

[0056] Once an action is selected, the system searches for a response in the measured observation database. If a real response exists for the same cell, the same cycle, or the same aging state, the same frequency, or the same pulse width, it is marked as `exact_measured`; if preset similarity conditions are met, it is marked as `nearby_measured`; if no real response exists, it is marked as `unavailable`; if only a model-generated response exists, it is marked as `synthetic_only`. When the system statistically analyzes the technical effect, `synthetic_only` responses are not included in the actual measured responses. Through this structure, each gating decision and each detection response can be traced back to `source_file` and `source_row`.

[0057] The experiments used publicly available datasets from NASA PCoE, CMU eVTOL, CALCE CS2_33, PulseBat, and High Power NMC / MolicellP42A. After impulse response extraction, a measured observation database was formed, containing a total of 118,385 rows of data (NASA 100,435 rows, CMU 8,690 rows, High Power NMC 4,824 rows, PulseBat 2,700 rows, and CALCE 1,736 rows). The training environment consisted of Python 3.12, PyTorch 2.11.0+cu128, CUDA 12.8, and an NVIDIA GeForce RTX 5080.

[0058] The complete dataset contains 9653 cyclic samples and 60 battery cells (22 from the CMU domain and 38 from the NASA domain). The main experiment uses a cell migration test, fixing VAH28 and VAH30 from the CMU domain and B0018 from the NASA domain as test cells (632 cyclic samples in total). The remaining 57 cells are divided into training and validation sets (approximately 15% for the validation set), and all cycles of the same cell do not appear across sets. Under random seeds 20260708, 20260709, and 20260710, the number of training / validation / test samples are 7998 / 1023 / 632, 7461 / 1560 / 632, and 7713 / 1308 / 632, respectively. All model variants use the same partitioning, and feature statistics are calculated only based on the training set. Cross-domain experiments are partitioned separately and not mixed with the main experiment statistics.

[0059] In this embodiment, the physically constrained Bayesian neural network employs a multi-task residual neural network with physical constraints and uncertainty outputs. The network contains four residual modules, a hidden layer width of 256, and a latent feature dimension of 128. A random deactivation mechanism is used to estimate uncertainty. Inputs include voltage, current, temperature, cycle number, capacity, energy, impedance, and missing data markers; outputs are the posterior mean and standard deviation of the SOH (State of Health), task risk probability, impedance-assisted prediction, and the mean and logarithmic variance of the posterior approximations of the physical parameters. The training objective is composed of SOH loss, task risk loss, impedance-assisted loss, domain adversarial loss, SOH monotonicity loss, uncertainty calibration loss, parameter prior loss, and physical residual loss. In this embodiment, real-data training uses cyclic-level surrogate physical residuals, including equivalent circuit voltage consistency, temperature rise rate consistency, SOC and charge integral consistency, EIS / pulse impedance consistency, and task margin consistency.

[0060] Transfer learning employs a domain adversarial strategy based on a gradient inversion layer. The AdamW optimizer is used for 90 epochs of training with a batch size of 512 and an initial learning rate of 0.002. Cosine annealing is used as the learning rate, and the model with the lowest root mean square error (RMSE) of SOH on the validation set is selected as the final model. During the testing phase, 24 Monte Carlo random inactivation predictions are performed, and the 95th percentile of the standardized prediction residuals from the validation set is used to scale the SOH prediction uncertainty. Furthermore, empirical coverage of the 95% prediction interval is statistically analyzed on the test set.

[0061] Under three randomized seed experiments, the root mean square error of SOH for the complete model was 0.02715±0.00120, and after removing the physical residuals, it was 0.02946±0.00407. The physical residual of EIS / pulse for the complete model was 0.23908±0.04341, which was lower than that of the model with physical residuals removed (0.37588±0.00102). This indicates that physical residual constraints can improve the SOH prediction performance and enhance the physical consistency of impedance / pulse correlation.

[0062] In this embodiment, a total of 522 precise anchor point events were selected to construct a precise anchor point measurement grid. The matching conditions are as follows: for EIS actions, the same cell and the same cycle (cycle difference not exceeding 0.5) are required, and the logarithmic relative error between the requested frequency and the measured frequency does not exceed 0.035; for pulse actions, the same cell and the same cycle (cycle difference not exceeding 0.5) are required, the pulse width difference does not exceed 0.5 seconds, and the SOC difference does not exceed 0.05.

[0063] The original risk network, migration calibration only, security gating only, random probe, fixed probe, parameter Fisher, risk coupling (without gating feedback), risk coupling closed loop, and risk coupling closed loop (upgraded) were compared. The experimental configurations were as follows: Original Risk Network: Makes decisions using only the risk-related predictions directly output by the physically constrained Bayesian neural network, without using the task failure probability obtained from battery pack-level task replay, and without performing transfer calibration, security gating, or active detection.

[0064] Migration calibration only: The mission failure probability and its confidence upper bound obtained from battery pack-level mission replay are used to make decisions, and the safety gating threshold obtained from migration calibration is used. No active probing is performed.

[0065] Safety-only gating: It utilizes the task failure probability and its confidence upper bound obtained from battery pack-level task replay, and adopts a preset fixed gating threshold to output the decision of allowing, reducing or maintaining the status quo. It does not perform migration safety budget calibration, nor does it perform active probing.

[0066] Random Probe: The probe action is selected completely randomly from the set of candidate actions, without relying on any optimization criteria, and the gating feedback mechanism is preserved.

[0067] Fixed detection: Select EIS frequency points and pulse actions according to preset fixed rules, without optimizing and sorting candidate actions based on Fisher information or mission risk information, and retain gating feedback to adjust the budget for the number of detection actions.

[0068] Fisher parameters: Fisher information and parameter identifiability provided by candidate detection actions are used as the basis for action selection. Task risk direction information is not introduced, and the gating feedback mechanism is retained.

[0069] Risk Coupling (Ungated Feedback): This method selects actions based on the risk information value criterion, but does not dynamically adjust the budget for the number of detection actions and the risk information gain coefficient for the next moment based on the current task risk status.

[0070] Risk-coupled closed loop: This method selects actions based on the risk information value criterion, and dynamically adjusts the budget for the number of detection actions and the risk information gain coefficient for the next moment based on the current task risk status.

[0071] Risk-coupled closed loop (upgrade): Based on the risk-coupled closed loop, it further allows for the safe upgrade of some reduction decisions to release decisions, provided that the preset security budget is met.

[0072] The experiment used the same model output and the same set of events after freezing, without retraining the model; the unified measured observation library was only used for probing response matching and posterior update.

[0073] The evaluation metrics are: accurate detection matching rate, 95% confidence upper bound, detection budget compliance rate, insufficient detection rate, and false release rate. The precise matching rate represents the proportion of active detection actions that can find the same cell, the same cycle, and whose frequency or pulse conditions meet the precise matching threshold in a unified observation database. The 95% confidence upper bound is calculated using the Wilson interval of the binomial distribution; The detection budget compliance rate represents the proportion of events in which the number of EIS frequency points actually selected meets the minimum requirements of the current gating level; the detection insufficiency rate represents the proportion of events in which the number of actual detection actions did not meet the minimum number of detections required by the current security gating level, out of the total number of events in which active detection was performed. The false release rate is defined as the proportion of events in which the final decision is to release, but the task replay results determine that the event is actually not feasible, out of the total number of evaluated events. In this embodiment, the release, de-quota, and hold status require at least 3, 5, and 7 EIS frequency points to be selected, respectively.

[0074] A total of 18,288 response records were generated across all active detection strategies, of which 12,108 were exact measured matches and 6,180 were neighbor-to-neighbor measured matches. Purely synthetic or unavailable responses were not used, resulting in an overall exact measured match rate of 66.21%. The exact measured match rate for the risk-coupled closed-loop strategy was 78.51%, and for the upgraded closed-loop strategy, it was 78.70%. In 522 events, the false release rate of the original risk network was 15.3%; after incorporating migration calibration, security gating, or closed-loop active detection, the observed false release rate was 0%, with a 95% confidence upper bound of approximately 0.73%. The detection budget compliance rate of the risk-coupled closed-loop strategy was 100%, and the average detection energy was [missing information]. .

[0075] like Figure 2 As shown, the false release rate of the original risk network is significantly higher than that of other strategies. In 522 events, the false release rate is 15.3%, and its 95% confidence upper bound is also the highest. In contrast, the false release rate of the risk-coupled closed-loop strategy is zero, and its confidence upper bound is the lowest.

[0076] like Figure 3 As shown, the original risk network, migration calibration only, and security gating only strategies did not perform active probing, and therefore did not participate in the gating budget compliance experiment for active probing strategies. The budget compliance rates for random probing, fixed probing, parameter Fisher, risk-coupled closed loop, and risk-coupled closed loop (upgrade) were all 100%. After removing gating feedback, the budget compliance rate dropped to 57.09%, and the underprobing rate increased to 42.91%. This result indicates that the feedback of the gating level to the probing budget at the next time step can directly improve the compliance of probing actions with gating requirements.

[0077] like Figure 4As shown, in the measured grid of precise anchor points, the measured accurate matching rates for parameters Fisher, risk-coupled closed loop, and risk-coupled closed loop (upgraded) are 65.4%, 78.5%, and 78.7%, respectively. The results indicate that aligning the candidate action set around the publicly measured frequency, SOC, and impulse conditions can improve the accurate matching between active detection actions and actual measured responses.

[0078] Furthermore, this embodiment sets up a migration safety gating experiment. The posterior results of the test set generated using the reserved cell model are used to generate 5688 decision events through battery pack task replay, covering three scenarios: nominal task, return task with 70% SOH, and return task with 55% SOH. This involves three reserved test cells: VAH28, VAH30, and B0018. Within each scenario, a one-cell-reserved calibration is used; that is, the safety gating threshold is determined using the task replay events of two cells, and a rotation evaluation is performed on the third cell. The safety gating outputs a release, de-rated, or hold result based on the task failure probability, the upper bound of the task failure probability confidence, and the gating threshold determined by the safety budget constraint. The results comparing the original risk network risk, ordinary probability calibration, fixed task replay gating, safety budget calibration gating, and ideal decision criterion show that the average false release rate of the original risk network direct gating is 10.38%, while the average false release rates of the fixed task replay gating and safety budget calibration gating are both about 0.01%. This indicates that the risk output of the original risk network can be used for risk ranking, but it is not suitable for safe release alone without the constraints of task replay, migration calibration, and risk confidence upper bound.

[0079] In three scenarios—nominal mission, SOH 70% return mission, and SOH 55% return mission—the gating budget compliance rate for the risk-coupled closed loop was 100%; the compliance rates for risk coupling (without gating feedback) were approximately 0.89%, 3.11%, and 68.0%, respectively. The results indicate that gating feedback enables the detection actions to adjust according to the mission risk status and maintains budget constraints under different mission severity levels.

[0080] To further analyze the reliability of the risk output of the original risk network and the risk upper bound of the task replay, 522 precise anchor point events were binned according to a fixed risk interval of 0.1 within the range of 0-1. The number of samples, average risk prediction value and actual task infeasibility rate in each bin were counted. The results are shown in Table 1.

[0081] Table 1. Bin reliability statistics based on the original risk network task failure probability and the upper confidence bound of the task failure probability.

[0082] Table 1 shows that the average predicted risk of the original risk network exhibits significant reliability bias across different risk ranges. Particularly in the low predicted risk range of 0-0.1, the average predicted risk is only 2.52%, while the actual task infeasibility rate reaches 32.35%, indicating a significant underestimation of risk in the low-risk region. Therefore, the single-point risk probability of the original risk network should not be directly used as a basis for safe release. In contrast, the upper confidence bound of the task failure probability obtained from battery pack task replay shows a more pronounced risk separation characteristic: no task infeasibility events occurred in the 311 events with an upper confidence bound below 0.3, while the actual task infeasibility rate was 100% for the 193 events with an upper confidence bound in the 0.9-1.0 range.

[0083] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements an eVTOL battery gated feedback active detection physical constraint safety gating method.

[0084] The above description merely illustrates preferred embodiments of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make modifications or alterations to the above-disclosed content to create equivalent embodiments. However, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. An eVTOL battery gate feedback active probing physical constraint safety gate method, characterized in that, The methods include: Acquire online observation data, flight mission profile data, battery historical cycle data, and measured impedance response data and impulse response data of the eVTOL battery; Online observation data, historical cycle data, measured impedance response data, and impulse response data of the eVTOL battery are input into a physically constrained Bayesian neural network to obtain the battery health state posterior, equivalent circuit parameter posterior, task risk posterior, and physical residual; the training loss of the physically constrained Bayesian neural network includes at least the physical residual loss. Based on the flight mission profile data, the battery health status posterior is sampled to obtain the SOH sample, and the battery physical parameter sample is determined based on the equivalent circuit parameter posterior. The SOH sample and the battery physical parameter sample are then used to perform battery pack-level mission replay to calculate the mission failure probability. Based on historical battery cycling data, gating calibration data is constructed. The upper confidence bound of the task failure probability is determined by the task failure probability and the gating calibration data. The safety gating threshold that meets the preset migration safety budget is determined by the gating calibration data. Based on the comparison results of the task failure probability, the upper confidence bound of the task failure probability and the safety gating threshold, the safety gating result is output. Based on the security gating results, determine the budget for the number of detection actions, the weight of energy cost, the weight of temperature rise cost, and the weight of control intrusion cost for the next moment; For each candidate detection action in the preset candidate frequency set, calculate the energy cost, temperature rise cost, and control intrusion cost corresponding to each candidate detection action; Calculate the Fisher information matrix for each candidate detection action, and then project the Fisher information matrix onto the equivalent circuit parameters using the gradient of the task failure probability to obtain the risk information value. Active detection actions are selected based on an objective function consisting of risk information value, energy cost, temperature rise cost, and intrusion control cost. Based on the online measured response or bench response corresponding to the active detection action, update the posterior of the equivalent circuit parameters, the probability of task failure, and the upper confidence bound of the probability of task failure.

2. The eVTOL battery-gated feedback active probing physical constraint safety-gating method of claim 1, wherein, The online observation data includes at least one of voltage, current, temperature, state of charge, capacity, cycle number, and measured impedance characteristics; the flight mission profile data includes power requirements, duration, energy requirements, and thermal constraints for at least one of the following stages: takeoff, hovering, climb, cruise, return, diversion, or de-rated mission.

3. The eVTOL battery-gated feedback active probing physical constraint safety-gating method of claim 1, wherein, The training loss of the physical constraint Bayesian neural network includes at least the physical residual loss; the physical residual loss includes at least the equivalent circuit model voltage residual, SOC dynamics residual, thermal model residual, EIS complex impedance residual, pulse dynamic resistance residual, and task margin physical consistency residual.

4. The eVTOL battery-gated feedback active probing physical constraint safety-gating method of claim 3, wherein, The task margin physical consistency residual is: ； in, θ represents the physical consistency residual of the task margin; θ represents the battery physical parameter vector, which is determined by the posterior mean of the equivalent circuit parameters. Indicates the power required for the flight mission; This indicates the available power that the battery pack can provide at θ. This indicates the energy required for a flight mission; This indicates the available energy that the battery pack can provide at θ. This represents the predicted temperature of the battery pack at the end of the task at θ. Indicates the upper limit of the battery's safe operating temperature; This represents the positive part of the function, meaning only positive numbers are retained, while negative numbers and zero are both represented as 0.

5. The eVTOL battery gated feedback active detection physical constraint safety gating method according to claim 1, characterized in that, The probability of task failure is: ； in, This represents the probability of task failure under the current input state x and the task profile π. Represents probability; This represents the sample of battery physical parameters determined based on the current input state x, the SOH sample, and the posterior distribution of the equivalent circuit parameters. π represents the task failure event corresponding to the task profile π. During task replay, the battery physical parameter samples and SOH samples are replayed at the battery pack level. When the task failure event occurs, the value is 1; q(SOH|x) represents the posterior distribution of SOH obtained from the current input state x. The battery health status (SOH) is indicated from... Sampling.

6. The eVTOL battery gated feedback active detection physical constraint safety gating method according to claim 1, characterized in that, The security gating result includes three states: release, de-rated, and hold. The security gating threshold includes a release threshold, a de-rated threshold, and a risk threshold. When the upper confidence bound of the task failure probability is lower than the release threshold, the output is release. When the upper confidence bound of the task failure probability is between the release threshold and the de-rated threshold and the task failure probability is lower than the risk threshold, the output is de-rated. Otherwise, the output is hold.

7. The eVTOL battery gated feedback active detection physical constraint safety gating method according to claim 1, characterized in that, The energy cost is determined by the estimated additional electrical energy consumption generated by performing the candidate detection action, the temperature rise cost is determined by the estimated additional temperature rise generated by performing the candidate detection action, and the control intrusion cost is determined based on the duration of the candidate detection action and the power margin and battery thermal state of the current task.

8. The eVTOL battery-gated feedback-based active detection physical constraint safety gating method according to claim 1, characterized in that, The risk information value is determined based on the Fisher information matrix and the gradient of the mission failure probability with respect to the equivalent circuit parameter vector, and is used to estimate the expected decrease in the variance of the mission failure probability after performing the candidate detection action.

9. The eVTOL battery-gated feedback-based active detection physical constraint safety gating method according to claim 1, characterized in that, The objective function is determined according to the following formula: ； in, Indicates the candidate detection actions under the current security gating level g. The objective function value, This indicates the expected reduction in the uncertainty of mission risk caused by the candidate detection actions. , and These represent the energy cost, temperature rise cost, and intrusion control cost corresponding to the candidate detection actions, respectively. , and These represent the preset energy cost weight, temperature rise cost weight, and control intrusion cost weight, respectively. This represents the risk information gain coefficient determined by the current security gating level.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the eVTOL battery gating feedback active detection physical constraint safety gating method according to any one of claims 1 to 9.

Images