State-space-adaptive-filtering-based eVTOL battery life prediction method

Abstract

This invention discloses an eVTOL battery life prediction method based on state-space adaptive filtering. The method involves simultaneously acquiring and preprocessing battery operation data and flight status information, identifying flight phases online, and dynamically extracting phase-specific temporal features. Next, a selective state-space model is constructed, outputting the data-driven prediction mean and its uncertainty variance of the battery health status. In parallel, a physical model of the battery equivalent circuit is constructed and updated online using recursive least squares to generate physical observations of the battery health status and their observation error variance. A state-space model with the battery health status as the state variable is constructed, using the two types of observations and their variances as input and observation noise. An adaptive unscented Kalman filter algorithm is used to recursively fuse the two types of observation sources to obtain the optimal estimate of the battery health status and its estimation error. Based on the optimal estimation history sequence, a forward Monte Carlo simulation is performed using a stochastic degradation model to calculate the probability distribution of the remaining battery life.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology for electric vertical takeoff and landing (eVTOL) aircraft, specifically to a method for predicting eVTOL battery life based on state-space adaptive filtering. Background Technology

[0002] As a core carrier of future urban air mobility, the safety, reliability, and economy of electric vertical takeoff and landing (EVTOL) aircraft heavily rely on the performance of their onboard lithium-ion battery systems. Unlike ground-based electric vehicles, EVTOL aircraft exhibit extreme dynamic characteristics in their flight profile: during vertical takeoff and landing, the battery needs to discharge at a high rate of up to 5C or even higher; while during cruise, the discharge rate decreases significantly. This drastically changing high-rate charging and discharging condition accelerates internal side reactions within the battery, leading to a more complex and rapid aging mechanism.

[0003] Accurate prediction of remaining battery life is crucial for implementing predictive maintenance, ensuring flight safety, and optimizing operating costs. Existing technologies mainly fall into two categories: one is based on physical or empirical degradation models, whose prediction accuracy heavily depends on the accuracy of model parameters and the degree of matching with operating conditions, making it difficult to adapt to the complex dynamic missions of electric vertical takeoff and landing (EVA) aircraft; the other is based on data-driven machine learning methods, which can learn aging patterns from historical data, but their black-box nature leads to poor interpretability of prediction results and a lack of ability to quantify prediction uncertainties, failing to meet the requirements of high-reliability decision-making.

[0004] In the prior art, patent CN119758146A discloses a battery health state prediction method based on a DNN-UKF model. This method constructs an equivalent circuit physical model and uses unscented Kalman filtering for state estimation. However, for novel and complex applications such as electric vertical take-off and landing aircraft, the physical model parameters are difficult to update adaptively online, and the model cannot cover all aging modes, resulting in limited prediction accuracy. Another patent, CN121679344A, proposes a neural network battery health state estimation method based on deep learning model fusion. Although it can learn patterns from data, it does not consider the noise characteristics of the prediction results, nor does it combine with traditional optimal estimation algorithms to improve the stability and accuracy of online estimation, and it lacks the ability to quantify prediction uncertainty.

[0005] The existing technologies mentioned above either rely excessively on accurate mechanistic models that are difficult to obtain, or fail to effectively utilize the uncertainty information in data-driven model predictions. Furthermore, they lack a dedicated fusion architecture for the high dynamic and high safety requirements of eVTOL, making it difficult to meet the urgent needs of eVTOL battery management systems for high accuracy, strong robustness, and online self-adaptation in lifetime prediction. Summary of the Invention

[0006] This invention relates to the field of battery health management technology for electric vertical takeoff and landing (eVTOL) aircraft, specifically to a method for predicting eVTOL battery life based on state-space adaptive filtering, comprising the following steps: Step S1: Collect and preprocess the operating data of the eVTOL battery, synchronously acquire flight status information, identify the flight mission stage online based on the flight status information, output flight stage labels including high-rate stage and cruise stage, and dynamically extract stage-specific temporal features. Step S2: Construct an uncertainty quantification predictor based on a selective state-space model; input the stage-specific temporal features containing flight stage labels into the uncertainty quantification predictor, and output the data-driven prediction mean and prediction uncertainty variance of the battery health status; Step S3: Construct an online updated physical model of the battery equivalent circuit. During the cruise phase, the recursive least squares method is used to dynamically update the model parameters online. Based on the pre-calibrated internal resistance-health state correlation model, physical observations of the battery health state and the observation error variance are generated. Step S4: Construct a state-space model with battery health status as the state variable, take the data-driven prediction mean and physical observation as the observation input, and dynamically construct the observation noise covariance matrix based on the prediction uncertainty variance corresponding to the data-driven prediction mean and the observation error variance corresponding to the physical observation. Step S5: Adaptive unscented Kalman filter algorithm is used to recursively perform time update and measurement update on the state space model, and the data-driven prediction mean and physical observation value are fused to obtain the optimal estimate of the battery health state and the estimation error covariance matrix. Step S6: Based on the historical sequence formed by the optimal estimate of the battery health status and the estimation error covariance matrix, a forward Monte Carlo simulation is performed using a stochastic degradation model to calculate the probability distribution of the remaining battery life.

[0007] Furthermore, in step S1, the online identification of flight mission phases is based on threshold segmentation of the coordinated change patterns of barometric altitude, airspeed, vertical speed, pitch angle, throttle command, and battery current. The phase-specific temporal features include pulse voltage drop, dynamic internal resistance, and temperature rise rate in the high-rate phase, and average current, current variance, average temperature, and energy efficiency features in the cruise phase.

[0008] Furthermore, the discretized state equation of the selective state-space model described in step S2 includes an input selection matrix and an output selection matrix, and the discretized state-space equation is expressed as: ; ; ; ; In the formula, This represents the input feature vector at time k; This represents the hidden state vector at time k; express The hidden state vector at time t; This represents the element-wise multiplication operation; Indicates the block output vector; Represents the learnable diagonal state transition matrix; This represents the input selection matrix at time k, determined by the current input. Through a linear projection layer Dynamically generated; This represents the output selection matrix at time k, determined by the current input. Through another linear projection layer Dynamically generated; the uncertainty quantification predictor is connected to a dual parallel fully connected network at the back end to map the prediction mean and logarithmic variance respectively. During online inference, a Monte Carlo random deactivation mechanism is enabled to perform multiple forward propagations, and the prediction uncertainty variance is obtained by statistical fusion of multiple outputs.

[0009] Furthermore, the equivalent circuit physical model described in step S3 is a first-order RC model. After being discretized into a linear regression form during the current steady-state phase, the parameter vector and the estimation error covariance matrix are identified online using the recursive least squares method with a forgetting factor. Based on the identified internal resistance Physical observations were calculated using a pre-calibrated internal resistance-health state correlation model. : ; In the formula, Indicates in At any given time, the battery internal resistance is identified online using the recursive least squares method; This indicates the reference value of the battery's internal resistance when it is in brand new condition; , Indicates the calibration coefficient; The observation error variance Calculated by the law of error propagation: ; In the formula, This represents the initial internal resistance of the battery. Represents the calibration coefficient; Here is the error covariance matrix; Represents the error covariance matrix Chinese correspondence The estimated variance.

[0010] Furthermore, the system state vector of the state-space model described in step S4 at time k is: ; in, Let k be the system state vector at time k; The battery health status at time k; Let be the adaptive degradation rate coefficient at time k.

[0011] Furthermore, the dynamically constructed observation noise covariance matrix directly participates in the calculation of the Kalman gain in step S5. The Kalman gain is calculated as follows: when the data-driven prediction uncertainty variance increases, the diagonal elements in the observation noise covariance matrix corresponding to the data-driven prediction mean increase, thereby reducing the weight of the data-driven prediction mean in the Kalman gain; when the data-driven prediction uncertainty variance decreases, the diagonal elements in the observation noise covariance matrix corresponding to the data-driven prediction mean decrease, thereby increasing the weight of the data-driven prediction mean in the Kalman gain.

[0012] Furthermore, the adaptive unscented Kalman filter algorithm described in step S5 includes a Sage-Husa adaptive adjustment mechanism, which recursively updates the noise variance of the process based on the statistical characteristics of the innovation sequence and introduces a non-negative definite constraint on the covariance matrix for correction.

[0013] Furthermore, the stochastic degradation model described in step S6 is a Wiener process with drift; when using the maximum likelihood estimation method to identify the drift rate and diffusion coefficient of the Wiener process, the estimation error covariance matrix output in step S5 is used as the prior uncertainty embedded in the likelihood function.

[0014] Furthermore, the parameter update of the recursive least squares method described in step S3 is controlled by the flight phase label: online identification and estimation error covariance matrix update is performed only when the cruise phase is identified; when the high-magnification phase is identified, the model parameters are frozen and the observation error variance of the previous time step is maintained.

[0015] Furthermore, the battery remaining lifespan probability distribution output in step S6 includes a cumulative distribution function and a preset confidence level risk quantile; the risk quantile is used to generate eVTOL task planning decision instructions.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a dynamic feature extraction and physical model parameter identification start-stop mechanism guided by flight phase perception. During high-magnification phases, model parameters are frozen while maintaining the observation error variance from the previous moment. During the cruise phase, recursive least squares identification is performed, ensuring that physical observations have a stable confidence level across different flight phases. Furthermore, based on data-driven feature input from flight phase perception, the fusion system can autonomously adjust weights during takeoff, landing, and cruise transitions in eVTOL missions.

[0017] This invention incorporates the variance of data-driven prediction uncertainty and the variance of physical observation errors into a unified dynamic framework for constructing observation noise covariance. It forward-propagates the uncertainty from the estimation phase to the likelihood function and Monte Carlo simulation of the remaining lifetime probability prediction, causing the diffusion interval of the remaining lifetime probability distribution to adaptively widen or shrink with the real-time estimation confidence level. The output cumulative distribution function and risk quantile can be directly embedded into the eVTOL task planning loop, triggering an early warning when the failure probability within the confidence interval exceeds the safety tolerance.

[0018] The modules of this invention are deeply coupled around robust state estimation under eVTOL alternating conditions: the output of online identification during the flight phase directly determines the feature input of the selective state-space model, the start-stop logic of physical model parameter identification, and the dynamic construction basis of the observation noise covariance matrix; dual-source uncertainty directly drives the dynamic allocation of weights, forming an adaptive fusion closed loop without human intervention; estimation uncertainty is forward-propagated to remaining lifetime prediction, enabling the probability interval to be widened and the estimation confidence to be adaptively coupled. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process of the present invention. Detailed Implementation

[0020] To overcome the shortcomings of existing technologies, this invention provides a method for predicting the battery life of electric vertical takeoff and landing (EVTOL) aircraft based on the fusion of state-space models and adaptive filtering. This invention constructs a complete technical chain from perception, estimation, correction to prediction by introducing online flight mission perception, probabilistic prediction using selective state-space models, online observation generation of physical models, optimal fusion using adaptive unscented Kalman filtering, and state-based closed-loop correction. The final output is a remaining battery life with a probability distribution, thus solving the problems of poor adaptability to complex operating conditions, high prediction uncertainty, low fusion reliability, and difficulty in supporting risk decision-making in existing technologies.

[0021] like Figure 1 As shown, the present invention provides the following technical solution: a method for predicting the battery life of an electric vertical takeoff and landing aircraft based on the fusion of state-space model and adaptive filtering, comprising the following steps: Step S1: Collect and preprocess the operating data of the eVTOL battery, synchronously acquire flight status information, identify the flight mission stage online based on the flight status information, output flight stage labels including high-rate stage and cruise stage, and dynamically extract stage-specific temporal features. Step S2: Construct an uncertainty quantification predictor based on a selective state-space model; input the stage-specific temporal features containing flight stage labels into the uncertainty quantification predictor, and output the data-driven prediction mean and prediction uncertainty variance of the battery health status; Step S3: Construct an online updated physical model of the battery equivalent circuit. During the cruise phase, the recursive least squares method is used to dynamically update the model parameters online. Based on the pre-calibrated internal resistance-health state correlation model, physical observations of the battery health state and the observation error variance are generated. Step S4: Construct a state-space model with battery health status as the state variable, take the data-driven prediction mean and physical observation as the observation input, and dynamically construct the observation noise covariance matrix based on the prediction uncertainty variance corresponding to the data-driven prediction mean and the observation error variance corresponding to the physical observation. Step S5: Adaptive unscented Kalman filter algorithm is used to recursively perform time update and measurement update on the state space model, and the data-driven prediction mean and physical observation value are fused to obtain the optimal estimate of the battery health state and the estimation error covariance matrix. Step S6: Based on the historical sequence formed by the optimal estimate of the battery health status and the estimation error covariance matrix, a forward Monte Carlo simulation is performed using a stochastic degradation model to calculate the probability distribution of the remaining battery life.

[0022] Furthermore, the online identification of flight phases and dynamic feature extraction in step S1 are specifically carried out as follows: Through high-precision sensors and flight control bus, the total battery pack voltage, battery current, highest single cell temperature, lowest single cell temperature, as well as airspeed, barometric altitude, vertical speed, pitch angle and normalized throttle command at time k are collected synchronously. The raw battery operation data is preprocessed, and a dynamic Gaussian distribution discrimination method based on a sliding window is used to remove outliers. The length of the sliding window is fixed to the amount of data sampled per second; that is, the system sampling frequency determines the number of data points included in each window, ensuring that each window exactly covers all data points collected within one second. For a sampled value of a certain feature at the current moment, the algorithm calculates the mean and standard deviation of all data for that feature within the current sliding window. When the absolute value of the difference between the sampled value and the mean within the window is greater than three times the standard deviation, the data point is determined to be an outlier and is removed. The missing values ​​are filled in using the interpolated estimates at time k using the time series linear interpolation method. A first-order low-pass digital filter is used to smooth and reduce noise in the data. The filter coefficient is dynamically adjusted according to the current change rate: when the current change rate is large, the filter coefficient is larger to improve the response speed, and when the current change rate is small, the filter coefficient is smaller to obtain a better smoothing effect.

[0023] Based on the preprocessed flight status data, continuous flight profiles are segmented and classified in real time using preset threshold rules, and flight phase labels are output. The flight phase labels include: ground standby, vertical takeoff, climb, cruise, descent, and landing; among which, the vertical takeoff, climb, and descent phases are high-rate phases; the threshold rules are based on a combination of changes in barometric altitude, airspeed, vertical speed, pitch angle, throttle command, and battery current.

[0024] In this embodiment, the threshold parameters for each flight phase are designed as shown in Table 1: Table 1 Threshold Rule Table

[0025] When multiple phase conditions conflict, a three-tiered hierarchical adjudication process is initiated. Air pressure and altitude are used as the primary criterion to define the candidate phase set; vertical speed is used as a secondary criterion to eliminate inconsistencies in motion direction; and throttle commands are used to verify the power intention. If conflicts still exist after the three-tiered screening, a hard priority adjudication is applied: ground standby > landing > vertical takeoff / descent > climb > cruise. Phase switching is limited to adjacent phases and requires confirmation from N=5 consecutive sampling points; cross-phase jumps are considered abnormal, the original label is maintained, and recursive least squares parameter updates are frozen.

[0026] Based on the identified flight phase, phase-specific feature vectors are dynamically extracted. During the high-rate phase, pulse voltage drop, dynamic internal resistance, and temperature rise rate are extracted; during the cruise phase, average current, current variance, and average temperature are extracted. Specifically, the pulse voltage drop is obtained by comparing the terminal voltage changes before and after the application of high-rate current; the dynamic internal resistance is calculated as the ratio of the voltage change to the absolute value of the corresponding current change; and the temperature rise rate is obtained by dividing the difference in the highest temperature within the battery pack at adjacent times by the sampling interval.

[0027] Furthermore, the uncertainty quantification prediction based on the selective state-space model in step S2 is specifically performed as follows: A selective state-space model probabilistic predictor is constructed, whose core computation is implemented by a selective scanning mechanism. The discretized state-space equation is expressed as follows: ; ; ; ; In the formula, This represents the input feature vector at time k; For time k; The flight phase label at time k; Indicates in The hidden state vector at time t is the model's representation of the system's internal state; express The hidden state vector at time t; This represents the element-wise multiplication operation; This represents the block output vector at time k; Represents the learnable diagonal state transition matrix; Indicates in The input selection matrix at time step 1 is determined by the current input. Through a linear projection layer Dynamically generated, controlling the current input How much information is incorporated into the new hidden state; Indicates in The output selection matrix at time step is determined by the current input. Through another linear projection layer Dynamically generated, controlling the current hidden state. How much information is used to generate the output?

[0028] An uncertainty quantization output layer is connected to the back end of the selective state-space model. This layer contains two parallel fully connected networks to obtain the predicted mean of the battery health state at time k. and variance : ; ; In the formula, and This represents two parallel fully connected neural networks, each used to output from the middle. Mapped to the predicted mean and log-variance; This represents the block output vector at time k.

[0029] The selective state-space model is trained using an offline supervised learning method, and the training phase and the online inference phase share the same network structure.

[0030] During training, a historical flight mission dataset is first constructed, containing phase-specific input features and corresponding true health status values. The model training employs a negative log-likelihood loss function. During training, random inactivation is enabled with a certain probability before the output layer of the model's uncertainty quantification. The inactivation rate is selected based on joint tuning of the negative log-likelihood loss and the predicted interval coverage probability on the validation set. Specifically, the model is trained on multiple candidate inactivation rates. For each model, Monte Carlo random inactivation is enabled on the validation set. The root mean square error of the predicted mean and the coverage probability of a given confidence interval are calculated. The inactivation rate that minimizes the negative log-likelihood loss and whose coverage probability is closest to the nominal confidence level is prioritized.

[0031] During training, a Dropout layer is added before the uncertainty quantization output layer. The random deactivation probability of the Dropout layer is p, and the value of p ranges from 0.1 to 0.3. The negative log-likelihood loss on the validation set and the prediction interval coverage probability are jointly tuned.

[0032] In this embodiment, models are trained on the candidate value set {0.1, 0.15, 0.2, 0.25, 0.3}. Monte Carlo Dropout is enabled for each model on the validation set. The root mean square error of the predicted mean and the coverage probability of the 90% confidence interval are calculated. The Dropout rate that minimizes the negative log-likelihood loss and closests the coverage probability to the nominal confidence level is preferentially selected. The AdamW optimizer is used during training, with an initial learning rate of... It combines cosine annealing learning rate scheduling and gradient pruning, and terminates early when the negative log-likelihood loss on the validation set does not decrease for 20 consecutive epochs.

[0033] During online inference, Monte Carlo random inactivation is enabled to assess the cognitive uncertainty of the model. In this forward propagation embodiment... ,right The outputs are fused, and the arithmetic mean of the predictions from the T forward propagation outputs is taken to obtain the final prediction mean. The total prediction uncertainty variance is obtained by adding the average variance of the T outputs to the sample variance between each prediction. ; Furthermore, the online identification and physical observation generation of the equivalent circuit model in step S3 is specifically as follows: Construct a first-order RC equivalent circuit model, whose terminal voltage equation is: ; In the formula, Indicates in The battery terminal voltage measured at all times; Indicates in The battery open-circuit voltage at any given time. This represents the battery's state of charge at time k, i.e., the battery's current state of charge. function Indicates in Battery current at any given moment; It represents the ohmic internal resistance of the battery, which represents the resistance to ion and electron conduction of the battery's internal materials; Indicates polarization resistance; This represents the pre-obtained polarization capacitance; Indicates the sampling time interval During the phase of stable current change ( Discretize the model and transform it into a linear regression form. : ; ; In the formula, ; The regression vector at time k is calculated using the polarization resistance value identified at the previous time step. This represents the measured battery current at time k; This represents the measured battery current at time k-1; This represents the polarization resistance value identified at time k-1; This represents the polarization resistance parameter to be updated at time k; This represents the pre-obtained polarization capacitance; Indicates the sampling time interval.

[0034] Using a forgetting factor Online parameter identification using recursive least squares method and its estimated error covariance matrix : ; ; ; In the formula, A discrete-time index representing the online identification process; Indicates in The vector of model parameters that needs to be estimated at any given time; express The observed value at time; express The regression vector at time step; express Kalman gain at time step; Indicates the forgetting factor, The smaller the value, the faster the old data is forgotten, and the stronger the algorithm's ability to track time-varying parameters, but the stability may deteriorate.

[0035] Based on the identified ohmic internal resistance Physical observations were calculated using a pre-calibrated internal resistance-health state correlation model. : ; In the formula, Indicates in At any given time, the battery's ohmic internal resistance is identified online using the recursive least squares method; This represents the reference value of the battery's internal resistance in its brand-new condition. , The calibration coefficient describes the mapping relationship between the increase in internal resistance and the decrease in capacity. , The method for obtaining the data is as follows: Take several battery samples of the same model with different aging levels, and measure the ohmic internal resistance of each sample under standard test conditions. The actual health status of each sample was obtained through standard volume testing. Using the internal resistance increment of each sample as the independent variable and the capacity decay as the dependent variable, the least squares method is used to apply the quadratic polynomial. Perform curve fitting and solve for the calibration coefficients. and The optimal estimate.

[0036] Its observation error variance Calculated by the law of error propagation: ; In the formula, This represents the initial internal resistance of the battery. Represents the calibration coefficient. represent Chinese correspondence The estimated variance.

[0037] Furthermore, the specific process of constructing the state-space model and observation noise in step S4 is as follows: Define the system state vector at time k. : ; In the formula, This represents the battery health status at time k; Represents the degradation rate coefficient of the adaptive estimate at time k; Establish the state equations, i.e., the degradation process model: ; ; In the formula, express The cumulative net discharge ampere-hours over the time interval; This represents process noise, indicating random degradation factors not covered by the model. This represents the battery health status at time k; Represents the degradation rate coefficient of the adaptive estimate at time k; Represents the degradation rate coefficient of the adaptive estimate at time k+1; Establish the observation equation: ; In the formula, For observation vectors; = Let k be the system state vector at time k; The observation matrix; Represents observation noise; The observation noise covariance matrix is ​​dynamically constructed from the variances of the two types of observation sources: ; in, To ultimately drive the prediction of total variance, i.e. The corresponding variance of the prediction uncertainty output at time k; Represents physical observation values The variance of the observation error; The observation noise covariance matrix is ​​constructed, which is crucial for adaptive fusion. The larger the value of its diagonal elements, the greater the noise of the corresponding observation, and the lower the weight they are assigned during fusion.

[0038] Furthermore, the adaptive unscented Kalman filter fusion estimation algorithm in step S5 is specifically as follows: Set the initial value of the state estimate The initial values ​​of its error covariance matrix .

[0039] in, This represents the initial health state of the battery. Let be a diagonal matrix representing the initial estimate uncertainty.

[0040] At every moment Firstly based on Posterior estimation at time A set of Sigma points is selected through unscented transformation. The probability distribution approximates the state. Then, each Sigma point is processed through the state equation. Propagate the data to obtain the propagated Sigma point. Based on this set of propagated Sigma points, calculate... Prior estimation of state at time step and its prior estimation error covariance matrix .

[0041] Based on prior estimation Calculate a new set of Sigma points These Sigma points are propagated to the observation space through the observation equation to obtain the predicted observed Sigma points. Based on this, the predicted observation values ​​are calculated. New covariance matrix and the cross-covariance matrix of state and observation Next, calculate the Kalman gain. Finally, the actual observation vector at the current time is used. Information between predicted observations The prior estimate is corrected to obtain the posterior optimal estimate of the state. and its posterior estimation error covariance matrix The updated formula is: ; ; In the formula, This means that at time k, the system state is estimated a priori based solely on observations at time k-1 and earlier. This represents the Kalman gain matrix at time k; This represents the actual observation vector at time k; This represents the observation predicted based on the prior state estimate at time k; This represents the prior estimation error covariance matrix corresponding to the prior state estimate; This represents the information covariance matrix at time k, i.e., the uncertainty in predicting the observed values; Perform adaptive process noise conditioning: According to the new interest sequence Based on the statistical characteristics, the Sage-Husa estimation algorithm is used to dynamically adjust the process noise variance. This enables the filter to adapt to dynamic changes in the system. ; in, Forgetting factor, usually taken as =(1-b) / (1-b k ), b is the forgetting constant, with a value range of (0.95, 0.99), and k represents the current time index; The Kalman gain at time k; For new information; Let be the posterior state estimation error covariance matrix at time k; Let be the posterior state estimation error covariance matrix at time k-1; This is the physical state transition matrix. This represents the cumulative net discharge ampere-hours from time k to time k+1.

[0042] Furthermore, the probabilistic remaining useful life prediction based on the stochastic degradation model in step S6 is specifically as follows: Using the Wiener process with drift as a measure of battery health Random degradation model: ; In the formula, This represents the drift rate, which is the average rate at which the battery's health condition linearly decays over time. Indicates the initial time; express The battery health status at any given time; k represents the index of the current time. This represents the battery health status at time k; It represents the diffusion coefficient, which measures the intensity of random fluctuations in the degradation process; express For the standard Brownian motion within k, the sampling time interval is considered as a unit time for simplification. Based on the optimal estimation history sequence of health status and its estimation error covariance matrix, the maximum likelihood estimation method is used to identify the drift rate and diffusion coefficient. When constructing the likelihood function, the estimation error covariance matrix is ​​incorporated as a prior uncertainty, ensuring that the variance of incremental observations simultaneously includes both random diffusion and estimation error terms. When the estimation uncertainty is large, the weight of the corresponding incremental observation in the likelihood function is automatically reduced, thereby avoiding misjudging the estimation error as the actual degradation fluctuation of the battery.

[0043] A battery failure threshold is defined; in this embodiment, it is set to 80%, meaning that the battery is considered to have failed when its capacity decays to 80% of its initial capacity. Starting from the current moment and the optimal estimated state, multiple forward Monte Carlo simulations are performed. In each simulation, the drift rate and diffusion coefficient are sampled from the parameter estimation distribution to simulate a random Brownian motion path, generating a future health state trajectory. The time when this trajectory first falls below the failure threshold is recorded, thus obtaining the remaining lifespan for that simulation. The remaining lifespan results from all simulations are collected to form a sample set. Based on this set, the probability density function, cumulative distribution function, and risk quantile at a specified confidence level of the remaining lifespan are calculated, completing the transition from single-point estimation to probabilistic risk prediction.

[0044] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting the lifetime of an eVTOL battery based on state-space adaptive filtering, characterized in that, Includes the following steps: Step S1: Collect and preprocess the operating data of the eVTOL battery, synchronously acquire flight status information, identify the flight mission stage online based on the flight status information, output flight stage labels including high-rate stage and cruise stage, and dynamically extract stage-specific temporal features. Step S2: Construct an uncertainty quantification predictor based on a selective state-space model; input the stage-specific temporal features containing flight stage labels into the uncertainty quantification predictor, and output the data-driven prediction mean and prediction uncertainty variance of the battery health status; Step S3: Construct an online updated physical model of the battery equivalent circuit. During the cruise phase, the recursive least squares method is used to dynamically update the model parameters online. Based on the pre-calibrated internal resistance-health state correlation model, physical observations of the battery health state and the observation error variance are generated. Step S4: Construct a state-space model with battery health status as the state variable, take the data-driven prediction mean and physical observation as the observation input, and dynamically construct the observation noise covariance matrix based on the prediction uncertainty variance corresponding to the data-driven prediction mean and the observation error variance corresponding to the physical observation. Step S5: Adaptive unscented Kalman filter algorithm is used to recursively perform time update and measurement update on the state space model, and the data-driven prediction mean and physical observation value are fused to obtain the optimal estimate of the battery health state and the estimation error covariance matrix. Step S6: Based on the historical sequence formed by the optimal estimate of the battery health status and the estimation error covariance matrix, a forward Monte Carlo simulation is performed using a stochastic degradation model to calculate the probability distribution of the remaining battery life.

2. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The online identification of flight mission phases in step S1 is based on threshold segmentation of the coordinated change patterns of barometric altitude, airspeed, vertical speed, pitch angle, throttle command and battery current. The phase-specific temporal features include pulse voltage drop, dynamic internal resistance and temperature rise rate in the high-rate phase, and average current, current variance, average temperature and energy efficiency features in the cruise phase.

3. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The discretized state equation of the selective state-space model described in step S2 includes an input selection matrix and an output selection matrix. The discretized state-space equation is expressed as: ； ； ； ； In the formula, This represents the input feature vector at time k; This represents the hidden state vector at time k; express The hidden state vector at time t; This represents the element-wise multiplication operation; Indicates the block output vector; Represents the learnable diagonal state transition matrix; This represents the input selection matrix at time k, determined by the current input. Through a linear projection layer Dynamically generated; This represents the output selection matrix at time k, determined by the current input. Through another linear projection layer Dynamically generated; the uncertainty quantification predictor is connected to a dual parallel fully connected network at the back end to map the prediction mean and logarithmic variance respectively. During online inference, a Monte Carlo random deactivation mechanism is enabled to perform multiple forward propagations, and the prediction uncertainty variance is obtained by statistical fusion of multiple outputs.

4. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The equivalent circuit physical model described in step S3 is a first-order RC model. After being discretized into a linear regression form during the current steady phase, the parameter vector and the estimation error covariance matrix are identified online using the recursive least squares method with a forgetting factor. Based on the identified internal resistance Physical observations were calculated using a pre-calibrated internal resistance-health state correlation model. : ； In the formula, Indicates in At any given time, the battery internal resistance is identified online using the recursive least squares method; This indicates the reference value of the battery's internal resistance when it is in brand new condition; , Indicates the calibration coefficient; The observation error variance Calculated by the law of error propagation: ； In the formula, This represents the initial internal resistance of the battery. Represents the calibration coefficient; Here is the error covariance matrix; Represents the error covariance matrix Chinese correspondence The estimated variance.

5. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The state-space model described in step S4 has the following system state vector at time k: ； in, Let k be the system state vector at time k; The battery health status at time k; Let be the adaptive degradation rate coefficient at time k.

6. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 5, characterized in that, The dynamically constructed observation noise covariance matrix directly participates in the calculation of the Kalman gain in step S5. The Kalman gain is calculated as follows: when the data-driven prediction uncertainty variance increases, the diagonal elements of the observation noise covariance matrix corresponding to the data-driven prediction mean increase, thereby reducing the weight of the data-driven prediction mean in the Kalman gain; when the data-driven prediction uncertainty variance decreases, the diagonal elements of the observation noise covariance matrix corresponding to the data-driven prediction mean decrease, thereby increasing the weight of the data-driven prediction mean in the Kalman gain.

7. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The adaptive unscented Kalman filter algorithm described in step S5 includes a Sage-Husa adaptive adjustment mechanism. This mechanism recursively updates the noise variance based on the statistical characteristics of the innovation sequence and introduces a non-negative definite constraint on the covariance matrix for correction.

8. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The stochastic degradation model described in step S6 is a Wiener process with drift. When using the maximum likelihood estimation method to identify the drift rate and diffusion coefficient of the Wiener process, the estimation error covariance matrix output in step S5 is used as the prior uncertainty embedded in the likelihood function.

9. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 4, characterized in that, The parameter update of the recursive least squares method described in step S3 is controlled by the flight phase label: online identification and estimation error covariance matrix update is performed only when the cruise phase is identified; when the high-magnification phase is identified, the model parameters are frozen and the observation error variance of the previous time step is maintained.

10. The eVTOL battery lifetime prediction method based on state-space adaptive filtering according to claim 1, characterized in that, The battery remaining lifespan probability distribution output in step S6 includes a cumulative distribution function and a preset confidence level risk quantile; the risk quantile is used to generate eVTOL task planning decision instructions.

Images