Lithium battery health state estimation method and system based on relaxation voltage

Abstract

The invention relates to a lithium battery health state estimation method and system based on relaxation voltage, and the method comprises the steps: determining relaxation estimation voltage data based on voltage response data, a preset solid-phase diffusion model and a preset electrochemical model; performing Radon transformation on the heat dissipation information, the structure information and the appearance image, and performing multi-domain alignment with the current change information, the first temperature information, the second temperature information, the humidity information and the relaxation estimation voltage data based on the same time axis to generate a state characteristic matrix of the lithium battery; performing convex optimization solution on the state characteristic matrix based on a Nernst equation as a constraint to obtain a health state value attenuation rate value of the lithium battery; generating a relaxation voltage curve based on the state characteristic matrix and the health state value attenuation rate value; and substituting the relaxation voltage curve into a preset deep learning analysis model, and combining historical cycle data and historical environment working condition data of the lithium battery to generate a health state estimation value of the lithium battery.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and in particular to a method and system for estimating the state of health of lithium batteries based on relaxation voltage. Background Technology

[0002] Lithium-ion batteries are rechargeable batteries that achieve charging and discharging by moving lithium ions between the positive and negative electrodes. They are characterized by high energy density, long lifespan, and mature technology, and are primarily used in new energy vehicles, energy storage systems, and consumer electronics. However, during the charge-discharge cycle of lithium-ion batteries, the continuous chemical reactions cause corrosion of internal electrode materials, decomposition and deposition of the solid electrolyte interface film, and precipitation of metal materials in the electrolyte, leading to battery aging and affecting safe use and remaining lifespan. To ensure the battery maintains good operating condition in the usage environment, a rapid and accurate assessment of the battery's State of Health (SOH) is essential. Current technologies typically estimate the battery's state of health based on its measurable capacity or terminal voltage. However, capacity decay or voltage shift is merely a symptom of declining battery health. The complexity of the internal electrochemical reactions and aging mechanisms of lithium batteries reduces the accuracy of SOH estimation. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for estimating the state of health of lithium batteries based on relaxation voltage.

[0004] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0005] This invention is achieved through the following technical solution:

[0006] Firstly, this embodiment provides a method for estimating the state of health of a lithium battery based on relaxation voltage, including the following steps:

[0007] The voltage response data, operating data, and surrounding environment data of the lithium battery after the current input is stopped are obtained. The operating data includes: current change information, first temperature information, heat dissipation information, structural information, and appearance image. The first temperature information is the temperature information of the lithium battery itself. The surrounding environment data includes: second temperature information and humidity information. The second temperature information is the external ambient temperature of the lithium battery.

[0008] Based on the voltage response data, the preset solid-phase diffusion model, and the preset electrochemical model, the relaxation estimation voltage data is determined.

[0009] After performing Radon transform on the heat dissipation information, the structural information, and the appearance image, they are aligned with the current change information, the first temperature information, the second temperature information, the humidity information, and the relaxation estimated voltage data on the same time axis in a multi-domain manner to generate the state feature matrix of the lithium battery.

[0010] The state feature matrix is ​​solved by convex optimization based on the Nernst equation as a constraint to obtain the health state value and decay rate value of the lithium battery.

[0011] Based on the state feature matrix and the decay rate of the health state value, a relaxation voltage curve is generated;

[0012] The relaxation voltage curve is substituted into a preset deep learning analysis model, and combined with the historical cycle data and historical environmental condition data of the lithium battery, an estimated value of the health status of the lithium battery is generated.

[0013] Furthermore, the step of acquiring the voltage response data, operating data, and surrounding environmental data of the lithium battery after the current input is stopped includes:

[0014] When the lithium battery stops receiving current, the start time, terminal voltage change data, current change information, and first temperature information are recorded. The first temperature information includes: outer shell temperature change data and internal temperature change data.

[0015] The heat dissipation information, appearance image and structural information of the lithium battery casing are obtained, as well as the external ambient temperature (i.e., the second temperature) and humidity information of the surrounding environment of the lithium battery are obtained.

[0016] Record the current time when the terminal voltage of the lithium battery is detected to be in a stable state;

[0017] Based on the start time and the current time, the terminal voltage change data is time-calibrated to generate voltage response data;

[0018] Based on the start time and the current time, the current change information, the first temperature information, the heat dissipation information, the appearance image, and the structural information are time-calibrated to generate operating data;

[0019] Based on the start time and the current time, the second temperature information and the humidity information are time-calibrated to generate surrounding environment data.

[0020] Furthermore, the step of determining the relaxation estimation voltage data based on the voltage response data, the preset solid-state diffusion model, and the preset electrochemical model includes:

[0021] The response data is divided based on the current change state to obtain the target data;

[0022] Based on the target data and the corresponding operating state of the lithium battery, the concentration distribution data of the lithium battery in the relaxation stage is determined;

[0023] Based on the preset solid-phase diffusion model, the concentration distribution data is solved in the time domain to generate a diffusion surface state sequence;

[0024] The diffusion surface sequence is mapped to a voltage output sequence based on a preset electrochemical model to obtain a simulated voltage curve;

[0025] The voltage response data is paired with the simulated voltage curve to obtain the relaxation estimated voltage data.

[0026] Furthermore, the step of performing Radon transform on the heat dissipation information, the structural information, and the appearance image, and then aligning them with the current change information, the first temperature information, the second temperature information, the humidity information, and the relaxation estimation voltage data on the same time axis to generate the state feature matrix of the lithium battery includes:

[0027] The heat dissipation information is subjected to Radon transform to generate a heat conduction feature projection matrix;

[0028] The structural information is subjected to a Radon transform to obtain the internal structural feature matrix of the lithium battery, and the transformation process is subject to structural constraints based on the standard model of the lithium battery.

[0029] The appearance image is subjected to Radon transform to obtain the outer surface morphology feature tensor. Based on the feature point fusion strategy, the outer surface morphology feature tensor is edge-consistently aligned to generate an outer surface optimized feature matrix.

[0030] Feature extraction is performed on the heat conduction feature projection matrix, the internal structure feature matrix, and the external surface optimization feature matrix to obtain structural contour features, heat distribution features, and surface morphology features.

[0031] Based on the preset coordinate frame, the structural contour features, the thermal distribution features, and the surface morphology features, a spatial feature set is determined.

[0032] After synchronizing the current change information, the first temperature information, the second temperature information, the humidity information, and the relaxation estimated voltage data in time sequence, they are geometrically aligned with the spatial feature set to form a multidimensional feature set.

[0033] The multidimensional feature set is subjected to hierarchical mapping and normalization to form the state feature matrix of the lithium battery.

[0034] Furthermore, the step of performing a Radon transform on the appearance image to obtain an outer surface morphology feature tensor, and then performing edge consistency alignment on the outer surface morphology feature tensor based on a feature point fusion strategy to generate an optimized outer surface feature matrix includes:

[0035] The appearance image is projected from multiple angles to obtain surface texture data from different perspectives;

[0036] A two-dimensional Radon transform is performed based on the surface texture data to generate an initial surface feature tensor;

[0037] The boundary line information in the initial surface feature tensor is extracted based on the edge detection algorithm, and a set of feature points is established.

[0038] The feature point set is subjected to feature point fusion processing to obtain an edge consistency reference template;

[0039] The edge consistency reference template is matrix registered with the initial surface feature tensor to generate the outer surface optimized feature matrix.

[0040] Furthermore, the step of performing convex optimization on the state feature matrix based on the Nernst equation as a constraint to obtain the degradation rate value of the lithium battery's health state includes:

[0041] After separating the electrochemical and structural features in the state feature matrix, a target constraint equation is formed with the Nernst equation.

[0042] Based on the target constraint equation, the current change information, and the surrounding environment data, a convex optimization objective function is established.

[0043] Using the current change information, the first temperature information, the heat dissipation information, the second temperature information, and the humidity information as a set of boundary constraints, the feasible region of the convex optimization objective function is determined, forming a feature constraint matrix;

[0044] The convex optimization objective function is iteratively optimized and solved based on the preset Lagrange multiplier algorithm and the feature constraint matrix, and the potential difference term in the Nernst equation is gradually embedded into the iterative process as an update factor.

[0045] Once the iterative optimization converges, the optimal solution of the convex optimization objective function is extracted, and the decay rate value of the health state of the lithium battery is determined based on the optimal solution of the convex optimization objective function.

[0046] Furthermore, the step of separating the electrochemical and structural features in the state feature matrix and forming the target constraint equation with the Nernst equation includes:

[0047] Extract the electrochemical feature subset, structural feature subset, and environmental feature subset from the state feature matrix, respectively;

[0048] Electrochemical parameter variables are established based on the electrochemical feature subset, and thermodynamic constraint variables are established by combining them with the structural feature subset and the environmental feature subset.

[0049] Substitute the electrochemical parameter variables and the thermodynamic constraint variables into the Nernst equation to construct the initial constraint equation;

[0050] The initial constraint equations are converted into linear inequality forms to form the target constraint equations.

[0051] Furthermore, the step of generating the relaxation voltage curve based on the state feature matrix and the health state value decay rate value includes:

[0052] The state feature matrix is ​​used to extract electrochemical parameters associated with voltage changes, and these electrochemical parameters are mapped to the decay rate values ​​of the healthy state values ​​to construct a relaxation voltage generation model.

[0053] The relaxation voltage generation model is solved based on the electrochemical and structural parameters in the state feature matrix, and the solution is used as an intermediate feature set for calculating the voltage response.

[0054] A time-domain model is established based on the intermediate feature set, and the decay rate value of the health state value is projected onto the time-domain model to generate a voltage response time series.

[0055] The voltage response time series is projected and a multi-domain mapping matrix is ​​formed based on the heat dissipation information in the state feature matrix.

[0056] The multi-domain mapping matrix is ​​reconstructed in reverse to output the relaxation voltage curve corresponding to the decay rate value of the health state value.

[0057] Furthermore, the step of substituting the relaxation voltage curve into a preset deep learning analysis model and combining it with the historical cycle data and historical environmental condition data of the lithium battery to generate a health status estimate of the lithium battery includes:

[0058] The relaxation voltage curve is time-series encoded to form an input feature vector, and the input feature vector is concatenated with the running parameter vector in the historical loop data and the feature vector in the historical environmental condition data to obtain the concatenated vector.

[0059] After initializing the parameters of the multi-layer network structure of the preset deep learning analysis model based on the spliced ​​vector, the input feature vector is input into the deep learning analysis model for forward propagation calculation to generate an initial estimate of the health status.

[0060] A loss function is established based on the initial health status estimate, the historical cyclic data, and the historical environmental condition data.

[0061] The preset deep learning analysis model is then updated via backpropagation based on the loss function to adjust the model parameters of the preset deep learning analysis model.

[0062] When the model parameters meet the preset convergence conditions, a target deep learning analysis model is generated. Based on the target deep learning analysis model, the relaxation voltage curve is re-estimated to output the health state estimate of the lithium battery.

[0063] Secondly, this embodiment also provides a lithium battery health state estimation system based on relaxation voltage, applied to the lithium battery health state estimation method based on relaxation voltage as described above, including:

[0064] The system includes a data acquisition module, a data processing module, a data transformation module, a data calculation module, a data matching module, and a state estimation module.

[0065] The data acquisition module is used to acquire voltage response data, operating data and surrounding environment data of the lithium battery after the current input is stopped. The operating data includes: current change information, first temperature information, heat dissipation information, structural information and appearance image. The first temperature information is the temperature information of the lithium battery itself. The surrounding environment data includes: second temperature information and humidity information. The second temperature information is the external ambient temperature of the lithium battery.

[0066] The data processing module is used to determine the relaxation estimation voltage data based on the voltage response data, the preset solid-phase diffusion model, and the preset electrochemical model.

[0067] The data conversion module is used to perform Radon transformation on the heat dissipation information, the structural information and the appearance image, and then align them with the current change information, the first temperature information, the second temperature information and humidity information and the relaxation estimated voltage data on the same time axis in a multi-domain manner to generate the state feature matrix of the lithium battery.

[0068] The data calculation module is used to perform convex optimization on the state feature matrix based on the Nernst equation as a constraint to obtain the health state value and decay rate value of the lithium battery.

[0069] The data matching module is used to generate a relaxation voltage curve based on the state feature matrix and the health state value decay rate value.

[0070] The state estimation module is used to substitute the relaxation voltage curve into a preset deep learning analysis model, and combine it with the historical cycle data and historical environmental condition data of the lithium battery to generate a health state estimate of the lithium battery.

[0071] The beneficial effects of this invention compared with the prior art are: based on the cooperation of relaxation voltage, electrochemical model, thermal diffusion model and deep learning model, it realizes bidirectional estimation of the state of health (SOH) of lithium battery from the physical mechanism layer to the data-driven layer. After acquiring multi-dimensional operating data, the estimated value of the state of health of lithium battery is obtained through convex optimization solution of Nernst constraint and deep learning, thus realizing high-precision estimation of the state of health.

[0072] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other objects, features and advantages of the present invention more obvious and understandable, preferred embodiments are described in detail below. Attached Figure Description

[0073] Figure 1 A flowchart illustrating the lithium battery health state estimation method based on relaxation voltage provided in an embodiment of the present invention;

[0074] Figure 2 This is a schematic block diagram of a lithium battery health state estimation system based on relaxation voltage, provided in an embodiment of the present invention. Detailed Implementation

[0075] Please see Figure 1 The specific embodiment shown in this invention discloses a method for estimating the state of health of a lithium battery based on relaxation voltage, comprising the following steps:

[0076] Step S1: Obtain the voltage response data, operating data and surrounding environment data of the lithium battery after the current input is stopped. The operating data includes: current change information, first temperature information, heat dissipation information, structural information and appearance image. The first temperature information is the temperature information of the lithium battery itself. The surrounding environment data includes: second temperature information and humidity information. The second temperature information is the external ambient temperature of the lithium battery.

[0077] Understandably, when the lithium battery under test is at its operating point where the current is about to stop, the voltage sampler records the voltage channel at a preset sampling rate (1kHz), the current meter records the instantaneous current and its minute fluctuations before the current stops, the temperature sensor array continuously records the temperature of the cell surface and nearby points, the environmental sensor records the external temperature and humidity, and the industrial camera takes pictures of the battery casing and key parts and saves the original images. The instant when the lithium battery stops current (t0) is used as the global time reference. All sensors record at multiple time points on and after t0 (e.g., linear and logarithmic scale time series: t0+1s, t0+2s, t0+60s, ...) to generate voltage response data, operating data, and surrounding environmental data.

[0078] Step S2: Based on the voltage response data, the preset solid-phase diffusion model, and the preset electrochemical model, determine the relaxation estimation voltage data;

[0079] Understandably, the presupposed solid-state diffusion model is: Where 0≤r≤R represents the particle radius, and D S Let be the solid-phase diffusion coefficient, and the boundary conditions are: Where j(t) is the surface reactive flux density, F is the Faraday constant, and a s Let be the active specific surface area. During the relaxation phase of the stopped current (i.e., (j(t) = 0)), the boundary is a flux-free boundary: Shortly after the current is stopped, the particle surface concentration c s (t)=c(R,t), if there is an instantaneous disturbance before the current stops causing a concentration deviation Δc s (0), which decays with time satisfying When (t→0) + )hour, This indicates that voltage relaxation typically exhibits a behavior similar to t over a short period of time. -1 / 2 The relevant decay; the preset electrochemical model is: Among them, U 0 This is the standard electrode potential. For the correction value, the solution c of the concentration over time is used. s (t) After setting the electrochemical model, the recovery function U(t) of the open-circuit potential over time after polarization removal can be obtained; the terminal voltage V meas The expression for (t) is: V meas (t)=U pos (t)-U neg (t)+I(t)R ohm -η ct (t), where U pos (t) represents the positive voltage, U neg The negative voltage, R ohm To reduce the voltage of the ohm, ηct I(t) represents the electrode polarization overpotential. When the stopping current I(t) = 0, the ohmic voltage drop disappears, but there is an incompletely recovered interface overpotential and an internal potential difference caused by diffusion. Based on the above model, the relaxation estimated voltage data V can be obtained through parameter fitting and numerical solution. relax (t).

[0080] Step S3: After performing Radon transform on the heat dissipation information, the structural information, and the appearance image, align them with the current change information, the first temperature information, the second temperature information, the humidity information, and the relaxation estimated voltage data on the same time axis to generate the state feature matrix of the lithium battery.

[0081] Understandably, the Radon transform maps the projection set of a two-dimensional image f(x,y) to a parameterized projection domain R. f (ρ,θ), where ρ is the radial coordinate of the projection, and the projection angle and the projection on the polar coordinate displacement are defined as... After performing radiometric and geometric corrections on the thermal image from the appearance image and heat dissipation information, the first image f is generated. i (x,y,t k ), where t k For each sampling time point, a Radon transform is performed to obtain R. i (ρ,θ,t k To reduce dimensionality, a preset angle set {θ} is used. j By sampling and projecting onto ρ and statistically aggregating ρ, the time series feature vector r is obtained. i (t k After mapping the numerical fields of the structural information to a two-dimensional space, the vector feature s(t) is obtained. k Similarly, mapping the numerical fields in the heat dissipation information to a two-dimensional space yields the vector feature h(t). k ), will transmit current change information I(t) k First temperature information T cell (t k ), second temperature information T enu (t k ), humidity H(t) k ) and relaxation estimation voltage V relax (t k Mapping to a time grid based on the same timestamp The state feature matrix X∈R of the lithium battery is then constructed. N×m Each row corresponds to a time point t. k The feature components of the column are spliced ​​in a preset fixed order: X(K,:)=I(t) k ),T cell (t k ),Tenu (t k ),H(t k ),V relax (t k ),r1(t k ),…,h(t k ),s(t k )}.

[0082] Step S4: The state feature matrix is ​​solved by convex optimization based on the Nernst equation as a constraint to obtain the decay rate value of the health state of the lithium battery.

[0083] Understandably, the Nernst equation is: In the case of lithium ions, the function is U(x), where x is the lithium content. To constrain the state characteristic matrix, a function is introduced. in, Let p be the preset basis function and p be the coefficient vector; the state feature matrix X∈R N×m The linear relationship with the equivalent open-circuit voltage is y≈Xw+ε, where w∈R m Let be the parameter to be estimated, and ε be the noise; from the above, the objective function is... Furthermore, this function is constrained by the physical consistency constraint Aw≤b and the boundary monotonicity constraint Cw≥0; if the health status value is {qk} in the time series. k =1 N The corresponding model is: q k =q0-α·η k +η k , where n k Let be the number of lithium battery cycles, and α be the rate of decay of the lithium battery's health state. {qk} is solved using matrix form (Xw), and then the solution is obtained based on least squares fitting.

[0084] Step S5: Generate a relaxation voltage curve based on the state feature matrix and the health state value decay rate.

[0085] Understandably, this refers to the parameterized relaxation curve: Among them, V ∞ Let A be the long-term limiting voltage, and t be the voltage induced within the corresponding short time. -1 / 2 The exponential term corresponds to multi-scale diffusion and interface recovery. To avoid singular constants, the terms obtained in the above steps are... Substituting the values ​​will yield the relaxation voltage curve V. relax est (t).

[0086] Step S6: Substitute the relaxation voltage curve into a preset deep learning analysis model, and combine it with the historical cycle data and historical environmental condition data of the lithium battery to generate an estimated health status value of the lithium battery.

[0087] It is understandable that we let V∈R T The discrete values ​​of the relaxation voltage curve at T time points are represented, and the historical cyclic feature set is C∈R. p The environmental history characteristics are e∈R q Deep model f θ (θ) Outputs the health status estimate The loss function is Where y i The health value labels are derived from historical operational data. Batch gradient descent is used to optimize the aforementioned loss function using labeled historical data to generate the target model. Early stopping and cross-validation are employed during the target model generation process to prevent overfitting, based on the real-time generated data. With the target model Output lithium battery health status estimate

[0088] Through steps S1 to S6, based on the cooperation of relaxation voltage and electrochemical model, thermal diffusion model and deep learning model, bidirectional estimation of the state of health (SOH) of lithium battery is achieved from the physical mechanism layer to the data-driven layer. After acquiring multi-dimensional operating data, the estimated value of the state of health of lithium battery is obtained through convex optimization solution of Nernst constraint and deep learning, thus realizing high-precision estimation of the state of health.

[0089] In one embodiment, the step of acquiring the voltage response data, operating data, and surrounding environment data of the lithium battery after the current input is stopped includes:

[0090] When the lithium battery stops receiving current, the start time, terminal voltage change data, current change information, and first temperature information are recorded. The first temperature information includes: outer shell temperature change data and internal temperature change data.

[0091] Step S11: Obtain heat dissipation information, appearance image and structural information of the lithium battery casing, and obtain the external ambient temperature (i.e., the second temperature) and humidity information of the surrounding environment of the lithium battery.

[0092] Understandably, during initialization, the data acquisition system aligns the time base of each input channel to the same moment to ensure that the time points recorded by all channels are synchronized. Heat dissipation information from the casing is acquired by a surface temperature sensor or a heat flow sensor, and the raw timing signal T is recorded during acquisition. shell (t) and heat flow q(t), based on a preset frame rate, acquire appearance image Ishell (x, y, t) (including visible light and infrared thermal images), and timestamp calibration, read and record structural information S, and simultaneously acquire external ambient temperature T. enu (t) and humidity H(t).

[0093] Step S12: When the terminal voltage of the lithium battery is detected to be in a stable state, record the current time;

[0094] Understandably, based on the sampling sequence With a preset time length of T w Internal judgment: (1) Maximum absolute value derivative: (2) Absolute value variance: If the above two conditions are met at a certain point in time and the subsequent duration is not less than t. hold The time of (holding time) is recorded as the stationary time t. stable .

[0095] Step S13: Based on the start time and the current time, perform time calibration on the terminal voltage change data to generate voltage response data;

[0096] It is understandable that the start time t for reading records... start With stationary time t stable Transform the original voltage sample timestamps into relative times: t'k = t k -t start , 0≤t'k≤t stable Select a uniform sampling grid {τ j}j=0 N The interval is Δτ, and for any τ j Perform linear interpolation if tk≤τ j ≤t k+1 ,but To generate calibrated voltage response data

[0097] Step S14: Based on the start time and the current time, perform time calibration on the current change information, the first temperature information, the heat dissipation information, the appearance image, and the structural information to generate running data;

[0098] It is understandable that the start time t for reading records... start With stationary time t stable The information on the change in current I(t) at the current time. k First temperature information T cell (t k ), second temperature information T enu (t k ), humidity H(t)k ) and heat dissipation information h(t) k The timestamps are transformed into relative times: t'k,i = tk,it start Where i∈{,,,,,}, a uniform time grid {τ} is selected. j Consistent with the voltage interpolation grid. Interpolation is performed on input channel i for each feature: in This indicates the interpolation operator used for channel i; the image data I(x,y,t) is selected in time according to frame timestamps, and spatial correction and pixel-to-physical coordinate mapping functions are applied to each frame. Obtain the eigenvector r(τ) j The information regarding the current I(t) and the first temperature includes: the surface temperature T of the outer casing. shell (t) and internal temperature T int (t), heat dissipation information h(t), and second temperature information T enu (t) performs interpolation to generate time series I(τ) j ),T shell (τ j ),T int (τ j ),h(τ j ),T enu (τ j The structural information is mapped onto the time grid as S(τ). j ), will I(τ j ),T shell (τ j ),T int (τ j ),h(τ j ),T enu (τ j ),S(τ j It is encapsulated to generate runtime data.

[0099] Step S15: Based on the start time and the current time, time-calibrate the second temperature information and the humidity information to generate surrounding environment data.

[0100] It is understandable that reading the raw timestamps and measured values ​​from environmental sensors {(t)} env,k ,T env (t env,k ),H(t env,k ))}, convert the timestamp to a relative time t' env,k Where t'env,k=tenv,kt start In a unified time grid (τ) j The temperature and humidity are interpolated separately to generate T. env (τj ) and H(τ) j ), The generated T env (τ j ) and H(τ) j It is encapsulated to generate surrounding environment data.

[0101] Through steps S11 to S15, multi-dimensional time calibration of the lithium battery's terminal voltage change, temperature change, and environmental parameters is performed at the instant the current stops. This achieves high-precision alignment of voltage response data, operational data, and environmental data in the time domain, ensuring time consistency and measurement synchronization of different data sources. This provides an accurate data foundation for subsequent relaxation voltage modeling and health status estimation, thereby significantly reducing estimation errors caused by data drift.

[0102] In one embodiment, the step of determining the relaxation estimation voltage data based on the voltage response data, a preset solid-state diffusion model, and a preset electrochemical model includes:

[0103] Step S21: Divide the voltage response data based on the current change state to obtain target data;

[0104] It is understandable that when the current I(t) changes from non-zero to zero, the battery enters the relaxation phase from the charging / discharging phase. The time series is segmented based on the rate of change of current. Among them I th By setting a preset threshold and detecting the time interval in which this condition is met, the same time period corresponding to the voltage response data V(t) can be extracted as the target data: D target ={(t,V(t))|t∈[t stop ,t end ]}.

[0105] Step S22: Based on the target data and the operating state of the lithium battery corresponding to the target data, determine the concentration distribution data of the lithium battery in the relaxation stage;

[0106] The solid-phase diffusion model for lithium batteries is understandable: Where c(r,t) is the lithium concentration at radius r, and D s Given the solid-phase diffusion coefficient and the current density j(0) before relaxation, determine the initial concentration gradient boundary conditions: Where R is the particle radius, F is the Faraday constant, and a s The specific surface area is given. From the target data, the operating state at the relaxation initiation moment is obtained: j(0), temperature T, and SOC, from which the temperature diffusion coefficient can be derived. Where D0 is the pre-diffusion factor, e aAs the activation energy, R g As a gas constant, an initial concentration distribution c(r,0) is established for the above solid-phase diffusion model based on the temperature diffusion coefficient and the initial concentration gradient boundary conditions. Based on the initial concentration distribution c(r,0), the initial concentration gradient boundary conditions, the temperature diffusion coefficient, the target data, and the target data corresponding to the operating state of the lithium battery, the solid-phase diffusion model is solved to determine the concentration distribution data {c(r,t)} of the lithium battery in the relaxation stage.

[0107] Step S23: Based on the preset solid-phase diffusion model, the concentration distribution data is solved in the time domain to generate a diffusion surface state sequence;

[0108] Understandably, the radial dispersion step size Δr and the time step size Δt are set based on the initial concentration c(r,0) and the diffusion coefficient D. s Iterative solution of the solid-phase diffusion model: Then, the surface concentration c is output at each time step. s (t), and record them to generate a diffusion surface state sequence {(t,c s (t))}.

[0109] Step S24: Based on a preset electrochemical model, the diffusion surface sequence is mapped to a voltage output sequence to obtain a simulated voltage curve;

[0110] Understandably, the presupposed electrochemical model is as follows: Where U0 is the open-circuit reference voltage, c max Given the maximum lithium concentration and η(t) as the polarization overpotential, based on Butler–Volmer kinetics: When the lithium battery is in the relaxation phase, j(t)≈0, therefore η(t)≈0, and the pre-defined electrochemical model simplifies to: Diffusion surface state sequence {c s (t)} and preset U0 and c max Substitution Output V(t) and record it to generate the analog voltage curve {(t,V(t))}.

[0111] Step S25: Pair the voltage response data with the simulated voltage curve to obtain the relaxation estimated voltage data.

[0112] It is understandable that the simulated voltage curve V sim (t) and voltage response data V meas (t) Matching is performed within the same time interval to construct the objective function J(θ): The parameters of the objective function are iteratively updated based on the gradient descent algorithm. The gradient calculation formula is: When the gradient calculation formula converges to obtain θ * Then, substitute it into the model to calculate V(t; θ) * Therefore, the curve V(t; θ) * () represents the relaxation estimation voltage data.

[0113] Through steps S21 to S25, based on the preset solid-phase diffusion model and electrochemical model, the concentration distribution data is transformed into a voltage output sequence, realizing the theoretical derivation of relaxation voltage estimation and the fusion calculation of experimental data. This achieves multi-scale modeling from concentration gradient to potential difference, effectively capturing the coupling characteristics of internal diffusion behavior and electrochemical reaction during the battery relaxation stage, thereby improving the accuracy of relaxation voltage prediction.

[0114] In one embodiment, the step of generating the state feature matrix of the lithium battery by performing Radon transform on the heat dissipation information, the structural information, and the appearance image, and then aligning them with the current change information, the first temperature information, the second temperature information, the humidity information, and the relaxation estimated voltage data on the same time axis in a multi-domain manner includes:

[0115] Step S31: Perform a Radon transform on the heat dissipation information to generate a heat conduction feature projection matrix; it can be understood that the Radon transform is used to project the two-dimensional thermal field T(x,y) to obtain the parameterized projection domain R. T (ρ,θ), where ρ is the radial position of the projection and θ is the projection angle. The Radon transform is defined as... Based on preset angle By sampling with discrete ρ, the heat conduction characteristic projection matrix P is obtained. T ∈R M×L Where L is the number of sampling points ρ. In this embodiment, firstly, the original thermal field T(x,y,t) is... k Radiometric correction and denoising are performed to obtain a two-dimensional heatmap at each time point. For each time point, the heatmap is then processed according to the angle set {θ}. j Calculate the Radon transform to obtain R. T (ρ,θ j ,t k The projection results at each time point are organized into a time series of thermal conduction characteristic projection matrix sequences {P}. T (t k )}.

[0116] Step S32: Perform a Radon transform on the structural information to obtain the internal structural feature matrix of the lithium battery, and impose structural constraints on the transformation process based on the standard model of the lithium battery;

[0117] It is understandable that if the structural information is two-dimensional, the internal structural projection matrix P can be obtained by performing a Radon transform. S (ρ,θ). To ensure consistency between the transformation result and the actual structure, structural constraints are imposed on the transformation process based on the standard lithium battery model. That is, known geometric features are introduced during the projection mapping and back projection processes to ensure the numerical stability of the projection features in the spatial reference frame. Therefore, its expression is: P S (ρ,θ)=∫∫S(x,y)δ(ρ-xcosθ-ysinθ)dxdy, where the structural constraint is a linear constraint on the projection matrix, i.e., C structvec (PS) = d, where matrix C structvec The vector d originates from the geometric features of the standard model. If the structural information is a three-dimensional volume S(x,y,z), it can be sliced ​​several times and subjected to Lardon transformations to obtain the internal structure projection matrix. In the projection domain, this matrix is ​​applied to P. S After adjusting the constraints to satisfy the structural characteristics, the internal structural feature matrix is ​​output.

[0118] Step S33: Perform Radon transform on the appearance image to obtain the outer surface morphology feature tensor, and perform edge consistency alignment on the outer surface morphology feature tensor based on the feature point fusion strategy to generate an outer surface optimized feature matrix.

[0119] It is understandable that performing a Radon transform on a two-dimensional appearance image I(x,y,t) yields the external surface morphology feature tensor. Subsequently, based on feature point detection and matching, the continuity of edges and contours in the projection domain is maintained. Feature point fusion: Let the set of edge projections obtained from different frames or viewpoints be εk(θ)={ρk,i(θ)}, then edge consistency is achieved by the following equation. in Let wk,i be the edge position after fusion, wk,i be the confidence weight, and γ be the regularization weight. Solving this equation yields... The optimized outer surface feature matrix is ​​then constructed.

[0120] Step S34: Extract features from the heat conduction feature projection matrix, the internal structure feature matrix, and the external surface optimization feature matrix to obtain structural contour features, heat distribution features, and surface morphology features.

[0121] Understandably, regarding Photoshop opt The boundary peak position ρpeak(θ) at each angle is calculated, and the profile width w(θ) forms the structural profile feature. The projected energy spectrum E(θ) = ∑ρR at each angle is calculated for PT. T (ρ,θ) 2 The first moment μ1(θ) forms the heat distribution characteristics, for Texture energy, edge density, and morphological spectral coefficients are extracted to form surface morphological features.

[0122] Step S35: Determine the spatial feature set based on the preset coordinate frame, the structural contour features, the heat distribution features, and the surface morphology features;

[0123] Understandably, a preset coordinate frame is established (in this embodiment, the origin is the geometric center of the battery cell and the shell boundary is the reference axis), and the angle-radial projection features are back-projected or mapped to the parameterized space of this coordinate system. If the projection domain information is represented by angle θ and radial direction ρ, then the mapping to the Cartesian coordinate pair (x, y) is: x = ρcosθ, y = ρsinθ. The spatial feature set consists of the mapped feature fields and the corresponding coordinate set, and can be represented as a set. Where f i This is the multimodal feature vector at this location.

[0124] Step S36: After synchronizing the current change information, the first temperature information, the second temperature information, the humidity information, and the relaxation estimation voltage data in time sequence, geometrically align them with the spatial feature set to form a multidimensional feature set;

[0125] Understandably, using a preset coordinate frame as a reference, the time series scalar (τ) is... j ),T cell (τ j ),T env (τ j ),H(τ j ),V relax (τ j ) through mapping function Assigned to different locations in the spatial feature set. The mapping function can be a space-time coupling kernel K((x,y),τ). Temporal information is weighted and integrated into a spatial point (x, y). Geometric alignment is represented by rigid body transformation, using homogeneous coordinates and a transformation matrix. T is determined by minimizing the Euclidean distance between corresponding points in space, and T is used to combine all time sets {τ} j Substituting these values ​​into the mapping function sequentially yields the set g = {(x...} i ,y i ,τ j ,g i,j )}.

[0126] Step S37: Perform hierarchical mapping and normalization on the multidimensional feature set to form the state feature matrix of the lithium battery.

[0127] It is understandable that hierarchical mapping divides multidimensional spatiotemporal features into several levels according to function or semantics, and applies appropriate mapping-pooling operations to each level to obtain a hierarchical representation; normalization: applying normalization operators to features of different dimensions and scales, in this embodiment using Min-Max normalization: Min-Max: Feature concatenation is the process of concatenating the feature vectors mapped from each layer in a predetermined order to obtain the final state feature matrix X∈R. N×m , where N is the number of time steps or samples, and m is the feature dimension.

[0128] In this embodiment, the multidimensional feature set S = {(x i ,y i ,τ j ,g i,j Then, define the spatial point layer (local), the region layer (partition aggregation), and the global layer (overall statistics); in the spatial point layer, perform local mapping on gi,j to obtain point-level features ui,j; in the region layer, perform statistical summarization of point-level features according to a preset grid to obtain region-level features v. r,j ; Calculate the overall statistical characteristics w at the global level j Min-Max normalization is performed on all hierarchical features, and point-level, region-level, and global-level features are concatenated along the feature dimension to form the feature vector x at each time point. j Finally, {xj}j = 1 N The combination forms a state feature matrix X∈R N×m .

[0129] Through steps S31 to S37, by performing Radon transform on the heat dissipation information, structural information, and appearance image respectively and aligning them with multi-domain time-series signals in multiple dimensions, a unified state feature matrix is ​​generated. This achieves feature fusion of the battery across multiple physical domains, including thermal, structural, appearance, and electrochemical aspects, thereby improving the accuracy of the lithium battery health state estimation.

[0130] In one embodiment, the step of performing a Radon transform on the appearance image to obtain an outer surface morphology feature tensor, and then performing edge consistency alignment on the outer surface morphology feature tensor based on a feature point fusion strategy to generate an optimized outer surface feature matrix includes:

[0131] Step S331: Project the appearance image from multiple angles to obtain surface texture data from different perspectives;

[0132] It is understandable that the planning perspective set {φ k The process includes at least a frontal view and several side views, recording the external pose calibration parameters for each frame, and acquiring appearance images from each viewpoint at a predetermined frame rate or under trigger conditions. kThe image (x, y) is used to associate timestamps with calibration parameters. After performing radiometric and perspective distortion correction on the original image, the corrected multi-view surface texture dataset is output.

[0133] Step S332: Perform a two-dimensional Radon transform based on the surface texture data to generate an initial surface feature tensor;

[0134] Understandably, for each corrected image In the angle set {θ j}j=1 M The discrete Radon becomes RI. k (ρ i ,θ j After performing necessary filtering or normalization on the ρ and θ directions, the projection results from each viewpoint are organized into a three-dimensional tensor according to the viewpoint index k.

[0135] Step S333: Extract boundary line information from the initial surface feature tensor based on the edge detection algorithm and establish a set of feature points;

[0136] It is understandable that the local gradient in the projection domain is calculated as follows: Gradient magnitude:

[0137] In this embodiment, tensors Discrete gradients are calculated in the dimensions ρ and θ, and the gradient magnitude G(i,j,k) is obtained. Non-maximum suppression is applied to the gradient magnitude to refine edge localization. Edge curves are extracted using double thresholding and connected component tracking, generating a set of edge projection lines for each viewpoint k. The edge projection lines are then decomposed into discrete feature points based on local curvature, length, and intensity. Where w k,n As weight.

[0138] Step S334: Perform feature point fusion processing on the feature point set to obtain an edge consistency reference template;

[0139] It is understandable that feature point fusion includes: feature point matching, outlier removal, and weighted fusion. After feature point matching, reference template points are used. For a set of points {pk,m} from several perspectives, their fusion is a weighted average: Where the weight w k,m The gradient strength, view confidence, and inverse function of the matching residual can be used to derive the following: The minimum sample set is randomly sampled from the matching candidate set to fit the local model. The number of points in the inner area is counted and repeated to obtain the maximum set of points, thereby removing erroneous matching points in the projection domain.

[0140] In this embodiment, the feature point sets of each viewpoint Cross-view matching is performed, candidate matching pairs are generated based on local angle-radial neighborhood descriptors and geometric constraints, and a robust estimation algorithm is run on the candidate matching set to eliminate outliers and determine the set of inliers. For each corresponding point family {p k,m The weighted fusion value is calculated, with weights set based on gradient strength, projection view confidence, and matching residuals, to obtain an edge consistency reference template.

[0141] Step S335: Perform matrix registration between the edge consistency reference template and the initial surface feature tensor to generate the outer surface optimized feature matrix.

[0142] It can be understood that, based on the initial surface feature tensor Edge Consistency Reference Template from After extracting the corresponding set of edge points and establishing a preliminary match with the template points, a transformation model is selected and an objective function is constructed. The objective function is solved based on the SVD analytical method, and the point error after registration is calculated. The resulting transformation T is then applied to the tensor. For each slice, a set of registered projection matrices is obtained. Based on the registered projection matrix, the outer surface optimization feature matrix is ​​reconstructed in the ρ and θ dimensions.

[0143] Through steps S331 to S335, by performing multi-angle projection and two-dimensional Radon transform on the appearance image and achieving edge consistency alignment based on the feature point fusion strategy, a high-precision outer surface optimized feature matrix is ​​constructed to reduce the feature deviation caused by imaging angle, lighting conditions and surface defects, realize the unified expression of appearance structure at the morphological layer, and significantly improve the stability of appearance features in battery health analysis.

[0144] In one embodiment, the step of performing convex optimization on the state feature matrix based on the Nernst equation as a constraint to obtain the health state degradation rate value of the lithium battery includes:

[0145] Step S41: After separating the electrochemical features and structural features in the state feature matrix, a target constraint equation is formed with the Nernst equation.

[0146] It is understandable that the state feature matrix X is X∈R N×m Divided into electrochemical characteristic sub-matrices by columns With structural feature submatrix That is, X = [X e X s ],m=me +m s Based on the parameter vector to be estimated For X e ,X s This allows the Nernst equation to relate the electrode potential U to the surface material fraction x and the temperature T: This equation is represented as a linear basis function of the electrochemical characteristics under equality constraints, i.e.: y U ≈Φ(c s ,T)p, where Φ is the basis function mapping and p is the coefficient.

[0147] Step S42: Based on the target constraint equation, the current change information, and the surrounding environment data, establish a convex optimization objective function;

[0148] Understandably, the fitting benchmark is based on a quadratic objective function, and Tikhonov regularization and convex constraints on inequality constraints are applied to achieve the desired fitting of the terminal voltage sequence y∈R. N To fit the target, the objective function is established: Based on Nernst constraints and environment / current-dependent inequality constraints To incorporate the current change information I and the surrounding environment data into the objective function, a weighting matrix W(I,Tenv,H) can be assigned to the fitting residuals to obtain the optimized result.

[0149] Step S43: Using the current change information, the first temperature information, the heat dissipation information, the second temperature information, and the humidity information as a boundary constraint set, determine the feasible region of the convex optimization objective function and form a feature constraint matrix;

[0150] Understandably, current change information, first temperature information, heat dissipation information, second temperature information, and humidity information are used to construct a boundary constraint set to prevent parameters and variables from exceeding the allowable range during the solution process within the feasible region. In this embodiment, the current change information, first temperature information, heat dissipation information, second temperature information, and humidity information are transformed into convex inequality constraints, forming a characteristic constraint matrix (A,b) that makes the feasible region... For a convex set, each row in Aw≤b corresponds to a measurement boundary:

[0151] Step S44: Based on the preset Lagrange multiplier algorithm and the feature constraint matrix, the convex optimization objective function is iteratively optimized and solved, and the potential difference term in the Nernst equation is gradually embedded into the iterative process as an update factor.

[0152] It is understandable that the augmented Lagrange algorithm will constrain... Based on the inequality Aw≤b, construct the augmented Lagrange function: Where λ is the Lagrange multiplier, and ρ>0 is the penalty factor. If For nonlinear problems, linearization is performed in each iteration to maintain convexity, and the potential difference term ΔU of the Nernst equation is used as an update factor to enhance the constraint approximation.

[0153] In this embodiment, the linearization equality constraint (if nonlinear) is applied to the current solution w. (t) Location: in Solving the convex subproblem, i.e., the problem concerning w: Since this subproblem is a convex quadratic programming problem (QP), it is solved using a quadratic programming solver, and then the Lagrange multipliers are updated: The potential difference term ΔU(t) in the Nernst equation is used as an update factor: a potential difference correction factor βΔU is added to the constraint linearization term. (t) (β is the scaling factor), that is, changing the right-hand side of the linearization to... Enhancing the role of electrochemical potential difference in iteration so that variable updates gradually satisfy the Nernst equation, when ||w (t+1) -w (t) ||<ε w and When the time is reached, stop the iteration and output the current solution.

[0154] Step S45: After the iterative optimization converges, extract the optimal solution of the convex optimization objective function, and determine the decay rate value of the health state of the lithium battery based on the optimal solution of the convex optimization objective function.

[0155] It is understandable that the optimal parameters obtained through optimization are Based on the mapping function Convert to health state time series {q k} after {q k The decay rate was fitted using a linear model: Establish the design matrix N = [1, n] and estimate α using the least squares solution:

[0156] Through steps S41 to S45, the Nernst equation is introduced into the convex optimization solution process. The potential difference term is used as a physical constraint factor to optimize the calculation of the state feature matrix, thereby realizing the integration of electrochemical constraints and statistical learning, and thus improving the accuracy of health degradation estimation.

[0157] In one embodiment, the step of separating the electrochemical and structural features in the state feature matrix and forming the target constraint equation with the Nernst equation includes:

[0158] Step S411: Extract the electrochemical feature subset, structural feature subset, and environmental feature subset from the state feature matrix, respectively;

[0159] Understandably, based on preset features, the physical category label of each column of features is determined, and the state feature matrix is ​​split by column: X = [X e X s X env The algorithm records the column index and meaning of each submatrix. After preprocessing each subset (such as missing value imputation, normalization and noise reduction), it outputs the electrochemical feature subset, structural feature subset and environmental feature subset.

[0160] Step S412: Establish electrochemical parameter variables based on the electrochemical feature subset, and combine them with the structural feature subset and the environmental feature subset to establish thermodynamic constraint variables;

[0161] It is understandable that the set of electrochemical parameter variables θ ec Generated from an electrochemical subset via parameterized mapping. Where D s Let be the solid-state diffusion coefficient, j0 be the exchange current density, and c be the solid-state diffusion coefficient. s Given the initial surface concentration, the subset of structural features and the subset of environmental features are mapped to form a set of thermodynamically constrained variables θ. th , Where k th For thermal conductivity, C th For heat capacity, h conv Here, denoted as convective heat transfer coefficient, and H represents humidity.

[0162] Step S413: Substitute the electrochemical parameter variables and the thermodynamic constraint variables into the Nernst equation to construct the initial constraint equation;

[0163] It is understandable that electrochemical parameters and thermodynamic constraints are substituted into the Nernst equation to construct an initial constraint equation that reflects the relationship between electrode potential and concentration and temperature.

[0164] In this embodiment, the electrode potential is: Where c s The surface concentration c in the electrochemical parameter variables s,0 With diffusion kinetics D s The relationship between temperature T and time shows that temperature T is determined by the thermodynamic variable θ. th Determine to form equality constraints: Among them Λ(c) s U is a synonym for the activity ratio function. meas The voltage mapping can be derived from the relaxation estimation. As can be seen from the above, the terminal voltage constraint for a lithium battery is: V meas =U pos (c s,pos ,T)-U neg (c s,neg ,T)+η, let the Nernst function of the two electrodes of the lithium battery be expressed as θ ec ,θ th After being represented and merged, the initial set of constraint equations e is obtained. init (·) = 0.

[0165] Step S414: The initial constraint equation is converted into a linear inequality form to form the target constraint equation.

[0166] It is understandable that for the nonlinear constraint e init First-order Taylor expansion of (θ) = 0: in Replace the equality constraints with equivalent two-sided linear inequalities and allow small relaxations ∈: ∈ ≤ e init (θ (0) )+Je(θ (0) )(θ-θ (0) )≤∈;To ensure convexity and conservatism, the nonlinear logarithmic term ln(c) can be used. s Use convex or concave boundaries to approximate the upper or lower bound. If a transformation of the logarithmic term is needed, then base it on z = ln(c s At the same time, a consistency constraint z = ln(c) is added. s Then, by replacing the domain of z with piecewise linearization, for the function f(c) = ln(c), at point c (0) Perform a first-order expansion: After substituting the Nernst term, the original nonlinear relationship is approximated as an affine function, thereby transforming the initial constraint equation into a linear inequality form to form the target constraint equation.

[0167] Through steps S411 to S414, the state feature matrix is ​​separated into electrochemical and structural features, and combined with environmental feature variables, it is introduced into the Nernst equation to establish the target constraint equation. This achieves parameter linearization and optimizability of the constraint model, and improves the accuracy of health degradation estimation.

[0168] In one embodiment, the step of generating a relaxation voltage curve based on the state feature matrix and the health state value decay rate value includes:

[0169] Step S51: Extract electrochemical parameters associated with voltage changes from the state feature matrix, and map the electrochemical parameters to the decay rate of the healthy state value to construct a relaxation voltage generation model.

[0170] It is understandable that from the state feature matrix X∈R N×m Columns with high correlation to voltage changes were identified, and a subset X of electrochemical parameters was obtained through sparse regression (LASSO). ec The extracted candidate electrochemical parameters include: surface concentration estimation c s Exchange current density j0, equivalent internal resistance Req, and approximate diffusion coefficient D s Define mapping function The extracted electrochemical parameter vector θ ec Mapping the health decay rate γ to the generative model parameters φ:

[0171] Step S52: Solve the relaxation voltage generation model based on the electrochemical parameters and structural parameters in the state feature matrix, and use the solution result as an intermediate feature set for calculating the voltage response;

[0172] It is understandable that the relaxation voltage generation model can be reduced to an ordinary differential equation: Where φ={V ∞ A, {B i ,τ i Structural parameters (e.g., geometric factors, heat capacity, contact resistance) affect the value of φ and the time scale. In this embodiment, solid-phase diffusion is performed based on modal expansion, so the surface concentration c s (t) can be written as: coefficient α n Related to the initial disturbance and influenced by structural parameters, the voltage is mapped through the potential U(c) s The combination of T and the Ohm term produces: V sim (t)=U(c s (t),T(t))-I(t)R eq (t)≈U(c s (t),T0)+small terms, perform a first-order expansion of U(·) and substitute it into c s (t) gives V sim The exponential model has model coefficients that are intermediate features.

[0173] Step S53: Establish a time-domain model based on the intermediate feature set, and project the decay rate value of the health state value onto the time-domain model to generate a voltage response time series;

[0174] Understandably, the time-domain solution based on modal expansion maps the intermediate feature set Z to a voltage time series V(t). Based on the linear time-invariant (LT) form, the time-domain solution is: V(t) = Cx(t) + Du(t). In this embodiment, the strategy for projecting the health decay rate γ into the time domain is to scale the modal amplitude: Among them κ i As the modal sensitivity coefficient, substituting the projection into the first ordinary differential equation of the relaxation voltage generation model yields: In the required time grid {t j Calculate V(t) on} j ;γ) forms a voltage response time series.

[0175] Step S54: Project the voltage response time series and form a multi-domain mapping matrix based on the heat dissipation information in the state feature matrix;

[0176] Understandably, in the time grid {t j The voltage response sequence V(t) on the} j Projecting the vector v(t) yields the discrete vector v(t). j (For example, projecting onto a preset (ρ,θ) sampling point), and reading the heat dissipation projection matrix PT(t) at the same time. j ), and construct a weighting function ωij(t) based on location thermal intensity and geometric distance. j Normalize the weights and form a mapping matrix M(t). j ), calculate the multi-domain mapping result z(t) j )=M(t j )v(t j Form a multi-domain mapping matrix sequence {Z(t)} j )}.

[0177] Step S55: Reconstruct the multi-domain mapping matrix in reverse and output the relaxation voltage curve corresponding to the decay rate value of the health state value.

[0178] It is understandable that for each time point t j Take the multi-domain mapping matrix Z(t) j This is treated as Radon transform projection data. If the projection is complete and the noise is low, then back-projection is performed using the FBP algorithm based on a preset filter. If the projection is incomplete or the noise is significant, establish a regularized inverse problem and solve it iteratively. Where A is the discrete Radon transform matrix, and Z is the vectorized multi-domain mapping data, from the reconstructed spatial field The voltage value is extracted at the electrode endpoint or a designated measurement point to form the final relaxation voltage curve V.relax (t).

[0179] Through steps S51 to S55, the state feature matrix and the health state value decay rate value are mapped and reconstructed in multiple domains to generate a relaxation voltage curve with time continuity and physical consistency. This achieves the collaborative reconstruction of multiple features, making the generated relaxation voltage curve highly accurate and improving the accuracy of health decay estimation.

[0180] In one embodiment, the step of substituting the relaxation voltage curve into a preset deep learning analysis model and combining it with the historical cycle data and historical environmental condition data of the lithium battery to generate a health status estimate of the lithium battery includes:

[0181] Step S61: Perform time-series encoding on the relaxation voltage curve to form an input feature vector, and then concatenate the input feature vector with the running parameter vector in the historical loop data and the feature vector in the historical environmental condition data to obtain the concatenated vector;

[0182] It is understandable that the relaxation voltage curve V relax (t) in a given time grid {t j}j=1 T Add timing coding to obtain voltage feature vectors Right now and position code p j The concatenation of the voltage feature vector v with the running parameter vector from the historical loop data. and historical environmental conditions vector The input vector is formed by concatenating elements along the same feature dimension. Where d = d v +d r +d e .

[0183] Step S62: After initializing the parameters of the multi-layer network structure of the preset deep learning analysis model based on the spliced ​​vector, the input feature vector is input into the deep learning analysis model for forward propagation calculation to generate an initial estimate of the health status.

[0184] Understandably, the parameters of the pre-defined deep learning analysis model are initialized to ensure the gradient scale of the activation function. Let the network function be f. θ (x)(parameter θ), forward propagation produces the initial estimate: If the network contains a time series branch, the voltage sequence v is first processed by the sequence module to generate a time series representation, and then combined with the static features into the fully connected layer to output the initial estimate of the health state.

[0185] Step S63: Establish a loss function based on the initial health status estimate, the historical cyclic data, and the historical environmental condition data;

[0186] Understandably, the loss function consists of several terms: a data fitting term, a physical consistency term, and a regularization term. Let the true health label be y, and the model output be... The loss function is: Where e is often taken data Mean Squared Error (MSE):

[0187] Step S64, and backpropagation update the preset deep learning analysis model based on the loss function to adjust the model parameters of the preset deep learning analysis model;

[0188] Understandably, the parameter θ is updated based on backpropagation combined with gradient basis optimization algorithm. For each parameter update, the loss gradient is calculated and updated according to the optimization rules: The updated formula is (element-by-element): Where η is the learning rate, β1 and β2 are the first and second momentum coefficients, and ∈ is the numerical stability term.

[0189] In this embodiment, forward propagation is performed to calculate the loss within each training iteration. The gradient is calculated using backpropagation. The application optimizer updates the parameter θ according to the above update formula.

[0190] Step S65: When the model parameters meet the preset convergence conditions, a target deep learning analysis model is generated. Based on the target deep learning analysis model, the relaxation voltage curve is re-estimated to output the health state estimate corresponding to the lithium battery.

[0191] Understandably, the preset convergence condition is that the relative decrease rate of the validation set loss lual is less than a threshold δ, i.e. The trained target model By performing forward propagation on the input relaxation voltage curve, the final estimate is obtained.

[0192] Through steps S61 to S65, the relaxation voltage curve is input into the deep learning analysis model and trained by combining historical cycle and environmental operating condition data. The model achieves joint learning of physical features and data features in a multi-layer network. Based on the physical consistency correction of the model through the physical constraint term in the loss function, the accuracy, adaptability and cross-operating condition generalization ability of lithium battery health prediction are improved.

[0193] Please see Figure 2 The present invention also discloses a lithium battery health state estimation system based on relaxation voltage, which is applied to the lithium battery health state estimation method based on relaxation voltage as described above, including: a data acquisition module 10, a data processing module 20, a data conversion module 30, a data calculation module 40, a data matching module 50, and a state estimation module 60.

[0194] The data acquisition module 10 is used to acquire voltage response data, operating data and surrounding environment data of the lithium battery after the current input is stopped. The operating data includes: current change information, first temperature information, heat dissipation information, structural information and appearance image. The first temperature information is the temperature information of the lithium battery itself. The surrounding environment data includes: second temperature information and humidity information. The second temperature information is the external ambient temperature of the lithium battery.

[0195] The data processing module 20 is used to determine the relaxation estimation voltage data based on the voltage response data, the preset solid-phase diffusion model, and the preset electrochemical model.

[0196] The data conversion module 30 is used to perform Radon transformation on the heat dissipation information, the structural information and the appearance image, and then align them with the current change information, the first temperature information, the second temperature information and humidity information and the relaxation estimated voltage data on the same time axis to generate the state feature matrix of the lithium battery.

[0197] The data calculation module 40 is used to perform convex optimization on the state feature matrix based on the Nernst equation as a constraint to obtain the health state value decay rate value of the lithium battery.

[0198] The data matching module 50 is used to generate a relaxation voltage curve based on the state feature matrix and the health state value decay rate value.

[0199] The state estimation module 60 is used to substitute the relaxation voltage curve into a preset deep learning analysis model, and combine it with the historical cycle data and historical environmental condition data of the lithium battery to generate a health state estimate of the lithium battery.

[0200] The above embodiments are preferred implementations of the present invention. In addition, the present invention can be implemented in other ways. Any obvious substitutions without departing from the concept of the present technical solution are within the protection scope of the present invention.

Claims

1. A relaxation voltage based lithium battery state of health estimation method, characterized in that, The method comprises the following steps: Obtain voltage response data, running data and surrounding environment data of the lithium battery after stopping current input, the running data comprising: current change information, first temperature information, heat dissipation information, structure information and appearance image, the first temperature information being the self-temperature information of the lithium battery, the surrounding environment data comprising: second temperature information and humidity information, the second temperature information being the external environment temperature of the lithium battery; Determine relaxation estimation voltage data based on the voltage response data, a preset solid-phase diffusion model and a preset electrochemical model; Perform Radon transformation on the heat dissipation information, the structure information and the appearance image, and perform multi-domain alignment on the current change information, the first temperature information, the second temperature information and humidity information and the relaxation estimation voltage data based on the same time axis to generate a state feature matrix of the lithium battery; Conduct convex optimization solution on the state feature matrix based on the Nernst equation as a constraint to obtain a health state decay rate value of the lithium battery; Generate a relaxation voltage curve based on the state feature matrix and the health state decay rate value; Substitute the relaxation voltage curve into a preset deep learning analysis model, and combine historical cycle data and historical environment working condition data of the lithium battery to generate a health state estimation value of the lithium battery.

2. The relaxation voltage based lithium battery state of health estimation method of claim 1, wherein, The step of obtaining voltage response data, running data and surrounding environment data of the lithium battery after stopping current input comprises: When it is detected that the lithium battery stops current input, record the starting time, the terminal voltage change data, the current change information and the first temperature information, the first temperature information comprising: shell temperature change data and internal temperature change data; Obtain the heat dissipation information, the appearance image and the structure information of the shell of the lithium battery, obtain the external environment temperature, i.e. the second temperature, and the humidity information of the surrounding environment; When it is detected that the terminal voltage of the lithium battery is in a stable state, record the current time; Based on the starting time and the current time, time mark the terminal voltage change data to generate voltage response data; Based on the starting time and the current time, time mark the current change information, the first temperature information, the heat dissipation information, the appearance image and the structure information to generate running data; Based on the starting time and the current time, time mark the second temperature information and the humidity information to generate surrounding environment data.

3. The relaxation voltage based lithium battery state of health estimation method of claim 1, wherein, The step of determining relaxation estimation voltage data based on the voltage response data, a preset solid-phase diffusion model and a preset electrochemical model comprises: Divide the voltage response data based on the current change state to obtain target data; Determine concentration distribution data of the lithium battery in the relaxation phase based on the target data and the running state of the lithium battery corresponding to the target data; Conduct time domain solution on the concentration distribution data based on the preset solid-phase diffusion model to generate a diffusion surface state sequence; Map the diffusion surface sequence to a voltage output sequence based on a preset electrochemical model to obtain a simulated voltage curve; The voltage response data is matched with the analog voltage curve to obtain the relaxation estimated voltage data.

4. The relaxation voltage based lithium battery state of health estimation method of claim 1, wherein, The step of generating the state feature matrix of the lithium battery based on the radiometric transformation of the heat dissipation information, the radiometric transformation of the heat dissipation information, the structural information and the appearance image, and the multi-domain alignment of the current change information, the first temperature information, the second temperature information and humidity information and the relaxation estimated voltage data based on the same time axis comprises: performing radiometric transformation on the heat dissipation information to generate a heat conduction feature projection matrix; performing radiometric transformation on the structural information to obtain an internal structure feature matrix of the lithium battery, and performing structural constraint on the transformation process based on a standard model of the lithium battery; performing radiometric transformation on the appearance image to obtain an external surface morphology feature tensor, and performing edge consistency alignment on the external surface morphology feature tensor based on a feature point fusion strategy to generate an external surface optimized feature matrix; performing feature extraction on the heat conduction feature projection matrix, the internal structure feature matrix and the external surface optimized feature matrix respectively to obtain a structural contour feature, a heat distribution feature and a surface morphology feature; determining a spatial feature set based on the preset coordinate framework, the structural contour feature, the heat distribution feature and the surface morphology feature; performing time sequence synchronization on the current change information, the first temperature information, the second temperature information and humidity information and the relaxation estimated voltage data, and then performing geometric alignment with the spatial feature set to form a multi-dimensional feature set; performing hierarchical mapping and normalization processing on the multi-dimensional feature set to form the state feature matrix of the lithium battery.

5. The relaxation voltage based lithium battery state of health estimation method of claim 4, wherein, The step of performing radiometric transformation on the appearance image to obtain an external surface morphology feature tensor, and performing edge consistency alignment on the external surface morphology feature tensor based on a feature point fusion strategy to generate an external surface optimized feature matrix comprises: performing multi-angle projection on the appearance image to obtain surface texture data at different viewing angles; performing two-dimensional radiometric transformation on the surface texture data to generate an initial surface feature tensor; extracting boundary line information in the initial surface feature tensor based on an edge detection algorithm and establishing a feature point set; performing feature point fusion processing on the feature point set to obtain an edge consistency reference template; performing matrix registration on the edge consistency reference template and the initial surface feature tensor to generate the external surface optimized feature matrix.

6. The relaxation voltage based lithium battery state of health estimation method of claim 1, wherein, The step of performing convex optimization solving on the state feature matrix based on the Nernst equation as a constraint to obtain a health state decay rate value of the lithium battery comprises: separating electrochemical features and structural features in the state feature matrix to form a target constraint equation with the Nernst equation; establishing a convex optimization objective function based on the target constraint equation, the current change information and the surrounding environment data; determining a feasible region of the convex optimization objective function to form a feature constraint matrix by taking the current change information, the first temperature information, the heat dissipation information, the second temperature information and humidity information as a boundary constraint set; and The convex optimization objective function is iteratively optimized and solved based on a preset Lagrange multiplier algorithm and the characteristic constraint matrix, and a potential difference term in the Nernst equation is gradually embedded into the iteration process as an update factor; After the iterative optimization converges, the optimal solution of the convex optimization objective function is extracted, and the decay rate value of the state of health of the lithium battery is determined based on the optimal solution of the convex optimization objective function.

7. The relaxation voltage based lithium battery state of health estimation method of claim 6, wherein, The step of forming the target constraint equation by separating the electrochemical characteristics and the structural characteristics in the state characteristic matrix and the Nernst equation comprises: An electrochemical characteristic subset, a structural characteristic subset and an environmental characteristic subset are extracted from the state characteristic matrix; An electrochemical parameter variable is established based on the electrochemical characteristic subset, and a thermodynamic constraint variable is established in combination with the structural characteristic subset and the environmental characteristic subset; The electrochemical parameter variable and the thermodynamic constraint variable are substituted into the Nernst equation to construct an initial constraint equation; The initial constraint equation is converted into a linear inequality form to form the target constraint equation.

8. The relaxation voltage based lithium battery state of health estimation method of claim 1, wherein, The step of generating a relaxation voltage curve based on the state characteristic matrix and the state of health value decay rate value comprises: An electrochemical parameter associated with voltage change is extracted from the state characteristic matrix, and the electrochemical parameter is mapped with the state of health value decay rate value to construct a relaxation voltage generation model; The relaxation voltage generation model is solved based on the electrochemical parameter and the structural parameter in the state characteristic matrix, and the solving result is taken as an intermediate characteristic set for calculating a voltage response; A time domain model is established based on the intermediate characteristic set, and the state of health value decay rate value is projected into the time domain model to generate a voltage response time sequence; The voltage response time sequence is projected and processed, and a multi-domain mapping matrix is formed based on heat dissipation information in the state characteristic matrix; The multi-domain mapping matrix is reconstructed in reverse to output a relaxation voltage curve corresponding to the state of health value decay rate value.

9. The relaxation voltage based lithium battery state of health estimation method of claim 1, wherein, The step of substituting the relaxation voltage curve into a preset deep learning analysis model and combining historical cycle data and historical environmental working condition data of the lithium battery to generate a state of health estimation value of the lithium battery comprises: The relaxation voltage curve is time-series encoded to form an input feature vector, and the input feature vector is spliced with a running parameter vector in the historical cycle data and a performance feature in the historical environmental working condition data to obtain a splicing vector; After the multi-layer network structure of the preset deep learning analysis model is parameterized based on the splicing vector, the input feature vector is input into the deep learning analysis model for forward propagation calculation to generate an initial state of health estimation value; A loss function is established according to the initial state of health estimation value and the historical cycle data and the historical environmental working condition data; And the preset deep learning analysis model is updated by backward propagation based on the loss function to adjust the model parameters of the preset deep learning analysis model. When the model parameters meet the preset convergence condition, a target deep learning analysis model is generated, and the relaxation voltage curve is re-estimated based on the target deep learning analysis model to output the health state estimation value corresponding to the lithium battery.

10. A relaxation voltage-based lithium battery state of health estimation system applied to the relaxation voltage-based lithium battery state of health estimation method according to any one of claims 1 to 9, characterized in that, Comprise: Data acquisition module, data processing module, data conversion module, data calculation module, data matching module and state estimation module; The data acquisition module is used for acquiring voltage response data, running data and surrounding environment data of the lithium battery after stopping current input, wherein the running data comprises current change information, first temperature information, heat dissipation information, structure information and appearance image, the first temperature information is the self temperature information of the lithium battery, and the surrounding environment data comprises second temperature information and humidity information, and the second temperature information is the external environment temperature of the lithium battery; The data processing module is used for determining relaxation estimation voltage data based on the voltage response data, a preset solid-phase diffusion model and a preset electrochemical model; The data conversion module is used for performing Radon transformation on the heat dissipation information, the structure information and the appearance image, and performing multi-domain alignment on the current change information, the first temperature information, the second temperature information, the humidity information and the relaxation estimation voltage data based on the same time axis to generate a state feature matrix of the lithium battery; The data calculation module is used for performing convex optimization solution on the state feature matrix based on the Nernst equation as a constraint to obtain a health state value decay rate value of the lithium battery; The data matching module is used for generating a relaxation voltage curve based on the state feature matrix and the health state value decay rate value; The state estimation module is used for substituting the relaxation voltage curve into a preset deep learning analysis model, combining historical cycle data and historical environment working condition data of the lithium battery, and generating a health state estimation value of the lithium battery.

Images