Lithium ion battery diagnostic method fusing physical information parameters and electrochemical impedance spectroscopy

Abstract

The application provides a lithium ion battery diagnosis method fusing physical information parameters and electrochemical impedance spectroscopy, which comprises the following steps: constructing a lithium ion battery equivalent circuit model by means of electrochemical impedance spectroscopy, and extracting physical information parameters representing the internal state of the lithium ion battery; fusing physical knowledge and electrochemical impedance spectroscopy based on deep learning, and constructing a lithium ion battery physical knowledge deep learning model; fusing physical knowledge, measurement data and domain knowledge, and extracting physically interpretable hidden features; and adopting a deep integration strategy to evaluate the prediction uncertainty of lithium ion battery capacity estimation and the prediction uncertainty of lithium ion battery life characteristic classification. The application constructs a lithium ion battery equivalent circuit model by means of electrochemical impedance spectroscopy, gives specific neurons in the deep learning model physical significance by adopting a physical regularization method, effectively utilizes three types of information sources including physical knowledge, measurement data and domain knowledge, and makes the lithium ion battery diagnosis result have interpretability and high accuracy.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery diagnostic technology, and in particular, it is a lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy. Background Technology

[0002] Lithium-ion batteries are widely used in industry and modern life, such as in energy storage systems, electric vehicles, and medical devices. However, during charging and discharging, irreversible physical and electrochemical processes occur at the electrodes, electrolytes, and separators of lithium-ion batteries, leading to capacity degradation. To ensure the efficiency, safety, and reliability of lithium-ion batteries, lithium-ion battery diagnostics are crucial, involving analysis of their internal states—impedance, interface phenomena, and material properties—and estimation of their capacity. Existing lithium-ion battery diagnostics typically use charge-discharge curves or features derived from them, usually containing only limited information about the battery's internal state. To further analyze the internal physical and electrochemical processes of the battery, electrochemical impedance spectroscopy (EIS) can be used. By recording the voltage response to current perturbations, an EIS spectrum over a wide frequency range is obtained, containing rich information about the battery's internal state.

[0003] Recent research methods for lithium-ion battery diagnostics can be broadly categorized into two types: model-based methods and deep learning. Model-based methods utilize electrochemical models, (semi-)empirical models, and equivalent circuit models to construct mathematical or physical equations describing the general degradation patterns of lithium-ion batteries by analyzing potential degradation mechanisms or current, voltage, and EIS measurement data. While model-based methods have achieved considerable success in lithium-ion battery diagnostics, establishing accurate mathematical or physical models remains challenging. This is because it requires a systematic and comprehensive understanding of the battery's internal state and degradation mechanisms, which are often too complex to be fully described by models. Furthermore, individual batteries exhibit variability due to factors such as material properties, manufacturing quality, and different operating environments, further complicating accurate modeling. Deep learning, on the other hand, analyzes collected lithium-ion battery sensor data to directly map the relationship between sensor data and internal state or capacity. Existing deep learning methods primarily employ charge-discharge curves or features derived from them, providing only limited information about the battery's internal state. Consequently, deep learning methods are insufficient in characterizing lithium-ion battery degradation mechanisms and promoting accurate capacity estimation. Furthermore, a major obstacle to deep learning methods is their limited interpretability, as their structure is often a black box and unconvincing, a phenomenon exacerbated by the randomness of the training process. Therefore, to balance interpretability and accuracy, it is urgent and necessary to seek a lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy, thereby ensuring both interpretability and high accuracy in the diagnostic results. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing a lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy. The method includes: constructing an equivalent circuit model of the lithium-ion battery using electrochemical impedance spectroscopy to extract physical information parameters characterizing the internal state of the lithium-ion battery; constructing a deep learning model of lithium-ion battery physics knowledge by fusing physical knowledge and electrochemical impedance spectroscopy based on deep learning; extracting physically interpretable latent features by integrating physical knowledge, measurement data, and domain knowledge; and employing a deep integration strategy to evaluate the predictive uncertainty of lithium-ion battery capacity estimation and the predictive uncertainty of lithium-ion battery lifetime characteristic classification. This invention constructs an equivalent circuit model of the lithium-ion battery using electrochemical impedance spectroscopy and assigns physical meaning to specific neurons in the deep learning model using physical regularization methods. It effectively utilizes three types of information sources, including physical knowledge, measurement data, and domain knowledge, making the lithium-ion battery diagnostic results interpretable and highly accurate.

[0005] This invention provides a method for diagnosing lithium-ion batteries that integrates physical information parameters and electrochemical impedance spectroscopy, comprising the following steps:

[0006] S1. Extract physical information parameters characterizing the internal state of lithium-ion batteries: Construct an equivalent circuit model of lithium-ion batteries using electrochemical impedance spectroscopy to extract physical information parameters characterizing the internal state of lithium-ion batteries.

[0007] S11. Perform characteristic tests on the lithium-ion battery to obtain the electrochemical impedance spectrum of the lithium-ion battery after n cycles.

[0008] S12. Perform image analysis on the electrochemical impedance spectroscopy of the lithium-ion battery to construct an equivalent circuit model of the lithium-ion battery; the components of the equivalent circuit model of the lithium-ion battery include the ohmic internal resistance R. ohm Solid electrolyte membrane internal resistance R SEI Charge transfer internal resistance R CT Warburg element Z War The first constant-phase element CPE1 and the second constant-phase element CPE2, wherein the ohmic internal resistance R ohm Caused by the movement of electrons in the electrode and ions in the electrolyte, and serving as the first module; the internal resistance R of the solid electrolyte membrane SEI The first impedance Z is formed in parallel with the first constant-phase element CPE1. arc1 And as a second module, it characterizes the diffusion resistance encountered by ions when passing through a solid electrolyte membrane in the high-frequency region; the charge transfer resistance R CT and Warburg component Z War The second impedance Z is formed by connecting it in series with the second constant-phase element CPE2 and then in parallel. arc2And as the third module, it characterizes the internal resistance of charge transfer in the mid-frequency region and the diffusion internal resistance of material transfer in the low-frequency region; the first module, the second module and the third module are connected in series to form the equivalent circuit model of the lithium-ion battery.

[0009] S13. Nonlinear fitting is performed on the electrochemical impedance spectrum of the lithium-ion battery at each cycle number to extract the physical information parameters characterizing the internal state of the lithium-ion battery at each cycle number.

[0010] S2. Constructing a deep learning model of lithium-ion battery physics knowledge: Based on deep learning, a deep learning model of lithium-ion battery physics knowledge is constructed by integrating physical knowledge and electrochemical impedance spectroscopy.

[0011] S3. Integrate physical knowledge, measurement data and domain knowledge to extract physically interpretable latent features, which include a first latent feature, a second latent feature and a third latent feature;

[0012] S31. Extracting physically interpretable first hidden features using a physical regularization method: By assigning physical meaning to specific neurons in the lithium-ion battery physical knowledge deep learning model, the values ​​of the neurons are mapped to the physical information parameters. The neurons are brought closer together so that their values ​​can characterize the internal state of the lithium-ion battery.

[0013] S32. Construct a multi-task learning model based on domain knowledge, jointly train T similar tasks, and extract the second latent feature by sharing neurons among the T similar tasks; extract the physical information parameters of each iteration. Electrochemical impedance spectroscopy is used as the input to the multi-task learning model;

[0014] S33. Based on a multi-task learning model, physical regularization is used as an auxiliary task. P shared neurons are selected from T similar task shared neurons to extract physically interpretable third latent features and assign them physical meaning. The values ​​of the P shared neurons are then mapped to the physical information parameters. By bringing them closer together, the numerical representation of the shared neurons can be used to characterize the internal state of the lithium-ion battery.

[0015] S4. Employ a deep integration strategy to evaluate the predictive uncertainty of lithium-ion battery capacity estimation and the predictive uncertainty of lithium-ion battery life characteristic classification.

[0016] Furthermore, step S13 specifically includes the following steps:

[0017] S131. Based on the equivalent circuit model of a lithium-ion battery, derive the impedance Z(w) of the lithium-ion battery in each cycle number as a function of frequency w:

[0018] Z(w)=R ohm+Z arc1 (w)+Z arc2 (w) (1)

[0019] Among them, Z arc1 (w) represents the first impedance related to frequency, and C1 represents Z arc1 The double-layer capacitance, where j represents the imaginary unit and n1 represents the first glide factor; Z arc2 (w) represents the second impedance related to frequency, and C2 and C W Z arc2 and Z War The double-layer capacitance, where n2 represents the second glide factor;

[0020] S132, The initial physical information parameter characterizing the internal state of a lithium-ion battery is θ. phy =[R ohm ,R SEI ,R CT ,C1,C2,C W The initial physical information parameter θ is obtained by using the least squares method. phy Make an estimate:

[0021]

[0022] Where m represents the number of sampling points for an electrochemical impedance spectroscopy curve; Re(·) and Im(·) represent the sampling points obtained at the current frequency w, respectively. i The impedance Z(w) below i Operations on the real and imaginary parts of (). w represents the current frequency of model fitting. i The impedance below, and Represent θ phy The lower and upper limits are given by domain knowledge;

[0023] S133. Nonlinear fitting is performed on the electrochemical impedance spectroscopy of the lithium-ion battery at each cycle number to analyze the aging phenomenon of each initial physical information parameter as the lithium-ion battery operates. Sensitivity analysis is then used to determine the aging phenomenon of the initial physical information parameter θ. phy P key parameters characterizing the internal state of a lithium-ion battery were selected and used to form the physical information parameters. Used to assess the internal state of lithium-ion batteries and estimate their capacity.

[0024] Preferably, step S2 specifically includes the following steps:

[0025] S21. Based on deep learning, construct a deep learning model of lithium-ion battery physics knowledge, and incorporate the physical information parameters of each cycle. The data is input to the input layer of the deep learning model of lithium-ion battery physics knowledge along with the electrochemical impedance spectroscopy. The output layer of the deep learning model of lithium-ion battery physics knowledge outputs a value to characterize the capacity of the lithium-ion battery at the current cycle number.

[0026] S22. For lithium-ion battery capacity estimation, the mean square error is used as the first loss function L. MSE (θ DL ):

[0027]

[0028] Among them, y k A measured value representing the capacity of a lithium-ion battery; This represents the lithium-ion battery capacity estimate output by the deep learning model of lithium-ion battery physics; N represents the number of samples in the training set; k is a constant; θ DL The parameters of the deep learning model representing the physical knowledge of lithium-ion batteries, including the weight matrix and bias vector, are obtained by iterative optimization using backpropagation, Adam optimization algorithm and its corresponding variants, by minimizing the loss function.

[0029] S23. For the classification of lithium-ion battery life characteristics, cross-entropy is used as the second loss function L. CE (θ DL ):

[0030]

[0031] in, This indicates the actual type of lithium-ion battery entered. This indicates that the deep learning model, which represents the physical knowledge of lithium-ion batteries, classifies the input as a category. The probability; log represents the base-10 log function.

[0032] Preferably, in step S31, a physical regularization method is used, and a loss function L based on the physical regularization method is set. PR (θ DL ), in the first loss function L MSE (θ DL Add a physical regularization penalty term L to ) P (θ DL This is used to penalize the neuron's numerical value in relation to the physical information parameters. The degree of deviation, that is:

[0033] L PR (θ DL ) = L MSE (θ DL )+λ p LP (θ DL (5)

[0034] Where, λ p Hyperparameters representing the degree to which control physics knowledge influences the deep learning model of lithium-ion battery physics;

[0035] In step S32, the third loss function L of the multi-task learning model is set. MTL (θ DL ):

[0036] L MTL (θ DL )=λ1L1(θ DL )+…+λ T L T (θ DL (7)

[0037] Where, λ i Let λ represent the hyperparameters of the i-th similar task and λ represent the hyperparameters of the i-th similar task. i ≥0, i=1,…,T;L i (θ DL () represents the loss function for the i-th corresponding task;

[0038] In step S33, the fourth loss function L of the multi-task learning model, which uses physical regularization as an auxiliary task, is set. MTL,P (θ DL The fourth loss function L MTL,P (θ DL Let L be the loss function and physical regularization penalty term for T similar tasks. P (θ DL The weighted sum of ), that is:

[0039] L MTL,P (θ DL )=λ1L1(θ DL )+…+λ T L T (θ DL )+λ p L P (θ DL (8).

[0040] Preferably, step S4 specifically includes the following steps:

[0041] S41. Construct M identical deep learning models of lithium-ion battery physics knowledge as base models;

[0042] S42. Sequentially input the physical information parameters for each cycle number. Electrochemical impedance spectroscopy is used as input to M basis models, and the input method depends on the structure of the basis models;

[0043] S43. Train and run each base model individually to obtain the outputs of the M base models. …、

[0044] S44. For lithium-ion battery capacity estimation, the mean of the outputs of the M basic models is used as the final output result.

[0045]

[0046] S45. Arrange the outputs of the M base models in ascending order to obtain... …、 Calculate the index index a1 = M × α / 2 for the lower limit of the confidence interval at the α% confidence level, and the index index a2 = M × (1 - α / 2) for the upper limit of the confidence interval at the α% confidence level. Then, the measure of the predictive uncertainty of the capacity estimate is the length L of the confidence interval at the α% confidence level. CI :

[0047]

[0048] in, Indicates the lower limit of the confidence interval; Indicates the upper limit of the confidence interval;

[0049] S46. For the classification of lithium-ion battery life characteristics, a voting method is adopted. The classification result of each base model is recorded as one vote. The classification results of each base model are then counted, and the category with the most votes is taken as the final classification result. The maximum number of votes is H. The measurement of the prediction uncertainty of lithium-ion battery life characteristic classification is accuracy β = H / M.

[0050] Preferably, the physical regularization penalty term L in step S31 P (θ DL It is composed of the sum of P regularization terms:

[0051]

[0052] in, This represents the value of the r-th specific neuron; Representing physical information parameters The r-th parameter in the equation.

[0053] Preferably, considering that electrochemical impedance spectroscopy is high-dimensional, and the physical information parameters... The dimensionality is relatively low, and the physical information parameters The rich amount of information contained is also diluted layer by layer by the neural layers. In the deep learning model of lithium-ion battery physics knowledge in step S2, the physical information parameters are... The input is fed into a neural layer closer to the output layer. In the multi-task learning model of step S32, the physical information parameters are... The input is sent to the neural layer containing the T similar task-shared neurons.

[0054] Compared with the prior art, the technical effects of the present invention are as follows:

[0055] 1. This invention designs a lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy. In order to study the internal state of lithium-ion batteries, an equivalent circuit model of lithium-ion batteries is constructed with the help of electrochemical impedance spectroscopy, and then physical information parameters characterizing the internal state of lithium-ion batteries are extracted. At the same time, a deep learning model of lithium-ion battery physics knowledge is constructed, and the physical information parameters and electrochemical impedance spectroscopy are used as model inputs to estimate the capacity of lithium-ion batteries.

[0056] 2. This invention presents a lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy. The designed lithium-ion battery physical knowledge deep learning model employs a physical regularization method to assign physical meaning to specific neurons in the deep learning model and extract physically interpretable latent features. Simultaneously, based on domain knowledge, multiple similar tasks are combined to achieve multi-task learning, thereby extracting better latent features. Furthermore, physical regularization is used as an auxiliary task, effectively utilizing three types of information sources, including physical knowledge, measurement data, and domain knowledge, thus enabling the diagnostic results of lithium-ion batteries to have interpretability and high accuracy. Attached Figure Description

[0057] Figure 1 This is a flowchart of the lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy according to the present invention.

[0058] Figure 2 This is the lithium-ion battery equivalent circuit model of the present invention;

[0059] Figure 3 This is a sensitivity analysis diagram of the initial physical information parameters of the present invention;

[0060] Figure 4 This is the mean square error of the six models of the present invention when estimating the capacity of the test cell 1 in each cycle;

[0061] Figure 5 This is the mean square error of the six models of the present invention when estimating the capacity of the test cell 2 in each cycle;

[0062] Figure 6This refers to the confidence interval length of the six models of the present invention when estimating the capacity of the test set battery 1 at each cycle number;

[0063] Figure 7 It is the length of the confidence interval for each cycle number when the six models of this invention estimate the capacity of the test set battery 2. Detailed Implementation

[0064] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0065] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0066] Figure 1 This invention illustrates a lithium-ion battery diagnostic method that integrates physical information parameters and electrochemical impedance spectroscopy. The method includes the following steps:

[0067] S1. Extracting physical information parameters characterizing the internal state of lithium-ion batteries: Constructing an equivalent circuit model of the lithium-ion battery using electrochemical impedance spectroscopy to extract physical information parameters characterizing the internal state of the lithium-ion battery: Ohmic internal resistance R ohm Internal resistance R of solid electrolyte membrane SEI Charge transfer internal resistance R CT .

[0068] S11. Perform characteristic tests on the lithium-ion battery to obtain the electrochemical impedance spectrum of the lithium-ion battery after n cycles.

[0069] S12. Perform image analysis on the electrochemical impedance spectroscopy of the lithium-ion battery and construct an equivalent circuit model of the lithium-ion battery, such as... Figure 2 As shown; the components of the equivalent circuit model of a lithium-ion battery include the ohmic internal resistance R. ohm Solid electrolyte membrane internal resistance R SEI Charge transfer internal resistance R CT Warburg element Z War First constant-phase element CPE1 and second constant-phase element CPE2, internal resistance R in ohms ohm Caused by the movement of electrons in the electrode and ions in the electrolyte, and serving as the first module; the internal resistance R of the solid electrolyte membrane SEI The first impedance Z is formed in parallel with the first constant-phase element CPE1. arc1 And as the second module, it characterizes the diffusion resistance encountered by ions when passing through the solid electrolyte membrane in the high-frequency region; the charge transfer resistance RCT and Warburg component Z War The second impedance Z is formed by connecting it in series with the second constant-phase element CPE2 and then in parallel. arc2 The first, second, and third modules are connected in series to form an equivalent circuit model of a lithium-ion battery. The third module is used to characterize the internal resistance of charge transfer in the mid-frequency region and the diffusion resistance of material transfer in the low-frequency region.

[0070] S13. Nonlinear fitting is performed on the electrochemical impedance spectrum of the lithium-ion battery at each cycle number to extract the physical information parameters characterizing the internal state of the lithium-ion battery at each cycle number.

[0071] S131. Based on the equivalent circuit model of a lithium-ion battery, derive the impedance Z(w) of the lithium-ion battery in each cycle number as a function of frequency w:

[0072] Z(w)=R ohm +Z arc1 (w)+Z arc2 (w) (1)

[0073] Among them, Z arc1 (w) represents the first impedance related to frequency, and C1 represents Z arc1 The double-layer capacitance, where j represents the imaginary unit and n1 represents the first glide factor; Z arc2 (w) represents the second impedance related to frequency, and C2 and C W Z arc2 and Z War The double-layer capacitance, n2 represents the second glide factor.

[0074] S132, The initial physical information parameter characterizing the internal state of a lithium-ion battery is θ. phy =[R ohm ,R SEI ,R CT ,C1,C2,C W The least squares method is used to analyze the initial physical information parameter θ. phy Make an estimate:

[0075]

[0076] Where m represents the number of sampling points for an electrochemical impedance spectroscopy curve; Re(·) and Im(·) represent the sampling points obtained at the current frequency w, respectively. i The impedance Z(w) below i Operations on the real and imaginary parts of (). w represents the current frequency of model fitting. i The impedance below, and Represent θ phy The lower and upper limits are given by domain knowledge.

[0077] S133. Nonlinear fitting is performed on the electrochemical impedance spectroscopy of the lithium-ion battery at each cycle number to analyze the aging phenomenon of each initial physical information parameter as the lithium-ion battery operates, and sensitivity analysis is used to determine the aging phenomenon of the initial physical information parameter θ. phy P key parameters characterizing the internal state of a lithium-ion battery were selected and used to form the physical information parameters. Used to assess the internal state of lithium-ion batteries and estimate their capacity.

[0078] S2. Constructing a deep learning model of lithium-ion battery physics knowledge: Based on deep learning, a deep learning model of lithium-ion battery physics knowledge is constructed by integrating physical knowledge and electrochemical impedance spectroscopy.

[0079] S21. Based on deep learning, construct a deep learning model of lithium-ion battery physics knowledge, and incorporate the physical information parameters of each cycle. Along with the electrochemical impedance spectroscopy, the data is input into the input layer of the deep learning model for lithium-ion battery physics. The output layer of the deep learning model for lithium-ion battery physics outputs a numerical value to characterize the capacity of the lithium-ion battery at the current cycle number.

[0080] S22. For lithium-ion battery capacity estimation, the mean square error is used as the first loss function L. MSE (θ DL ):

[0081]

[0082] Among them, y k A measured value representing the capacity of a lithium-ion battery; This represents the lithium-ion battery capacity estimate output by the deep learning model of lithium-ion battery physics; N represents the number of samples in the training set; k is a constant; θ DL The parameters of the deep learning model representing the physical knowledge of lithium-ion batteries include the weight matrix and bias vector. They are obtained by iterative optimization using backpropagation, Adam optimization algorithm and its variants, by minimizing the loss function.

[0083] S23. For the classification of lithium-ion battery life characteristics, cross-entropy is used as the second loss function L. CE (θ DL ):

[0084]

[0085] in, This indicates the actual type of lithium-ion battery entered. This indicates that the deep learning model, which represents the physical knowledge of lithium-ion batteries, classifies the input as a category. The probability; log represents the base-10 log function.

[0086] S3. Integrate physical knowledge, measurement data, and domain knowledge to extract physically interpretable latent features, including the first latent feature, the second latent feature, and the third latent feature.

[0087] S31. Extracting physically interpretable first hidden features using physical regularization: By assigning physical meaning to specific neurons in the deep learning model of lithium-ion battery physics knowledge, the numerical values ​​of neurons are transformed into physical information parameters. By bringing neurons closer together, the numerical values ​​can characterize the internal state of the lithium-ion battery, ensuring the interpretability of the model's capacity estimates. A physical regularization method is employed, and a loss function L based on this method is defined. PR (θ DL ), in the first loss function L MSE (θ DL Add a physical regularization penalty term L to ) P (θ DL This is used to punish neurons based on their numerical and physical information parameters. The degree of deviation, that is:

[0088] L PR (θ DL ) = L MSE (θ DL )+λ p L P (θ DL (5)

[0089] Where, λ p Hyperparameters representing the degree to which physical knowledge influences the deep learning model of lithium-ion battery physics; physical regularization penalty term L. P (θ DL It is composed of the sum of P regularization terms:

[0090]

[0091] in, This represents the value of the r-th specific neuron; Representing physical information parameters The r-th parameter in the equation.

[0092] S32. Construct a multi-task learning model based on domain knowledge, jointly train T similar tasks, and extract the second latent feature by sharing neurons among the T similar tasks; extract the physical information parameters of each iteration. Electrochemical impedance spectroscopy is used as input to the multi-task learning model. A third loss function L is defined for the multi-task learning model. MTL (θ DL ):

[0093] L MTL (θ DL )=λ1L1(θ DL )+...+λ T L T (θ DL (7)

[0094] Where, λ i Let λ represent the hyperparameters of the i-th similar task and λ represent the hyperparameters of the i-th similar task. i ≥0, i=1,...,T;L i (θ DL Let ) represent the loss function for the i-th corresponding task.

[0095] S33. Based on a multi-task learning model, physical regularization is used as an auxiliary task. P shared neurons are selected from T similar task shared neurons to extract physically interpretable third latent features and assign them physical meaning. The values ​​of the P shared neurons are then mapped to physical information parameters. By bringing the shared neurons closer together, the internal state of the lithium-ion battery can be numerically represented. A fourth loss function L is defined for the multi-task learning model that uses physical regularization as an auxiliary task. MTL,P (θ DL ), the fourth loss function L MTL,P (θ DL Let L be the loss function and physical regularization penalty term for T similar tasks. P (θ DL The weighted sum of ), that is:

[0096] L MTL,P (θ DL )=λ1L1(θ DL )+…+λ T L T (θ DL )+λ p L P (θ DL (8).

[0097] Step S3 first employs physical regularization to assign physical meaning to specific neurons in the deep learning model; simultaneously, based on domain knowledge, it combines multiple similar tasks to achieve multi-task learning; then, physical regularization is used as an auxiliary task, thereby effectively utilizing three types of information sources: physical knowledge, measurement data, and domain knowledge, which is one of the key inventive points of this invention.

[0098] S4. Employ a deep integration strategy to evaluate the predictive uncertainty of lithium-ion battery capacity estimation and the predictive uncertainty of lithium-ion battery life characteristic classification.

[0099] S41. Construct M identical deep learning models of lithium-ion battery physics knowledge as base models.

[0100] S42. Sequentially input the physical information parameters for each cycle number. Electrochemical impedance spectroscopy is used as input to M basis models, and the input method depends on the structure of the basis models.

[0101] S43. Train and run each base model individually to obtain the outputs of the M base models.

[0102] S44. For lithium-ion battery capacity estimation, the mean of the outputs of the M basic models is used as the final output result.

[0103]

[0104] S45. Arrange the outputs of the M base models in ascending order to obtain... …、 Calculate the index index a1 = M × α / 2 for the lower limit of the confidence interval at the α% confidence level, and the index index a2 = M × (1 - α / 2) for the upper limit of the confidence interval at the α% confidence level. Then, the measure of the predictive uncertainty of the capacity estimate is the length L of the confidence interval at the α% confidence level. CI :

[0105]

[0106] in, Indicates the lower limit of the confidence interval; This indicates the upper limit of the confidence interval.

[0107] S46. For the classification of lithium-ion battery life characteristics, a voting method is adopted. The classification result of each base model is recorded as one vote. The classification results of each base model are then counted, and the category with the most votes is taken as the final classification result. The maximum number of votes is H. The measurement of the prediction uncertainty of lithium-ion battery life characteristic classification is accuracy β = H / M.

[0108] Considering that electrochemical impedance spectroscopy is high-dimensional, and the physical information parameters... The dimensionality is relatively low, and the physical information parameters The rich amount of information contained is also diluted layer by layer in the neural layers. In step S2, the physical information parameters in the deep learning model of lithium-ion battery physics knowledge are... The input is fed into a neural layer closer to the output layer. In the multi-task learning model of step S32, the physical information parameters are... The input is sent to the neural layer containing the T similar task-shared neurons.

[0109] In one specific embodiment, a detailed comparison is made between the deep learning model of lithium-ion battery physics proposed in this invention and a deep learning model that only uses electrochemical impedance spectroscopy to verify the effectiveness of the method proposed in this invention.

[0110] Experimental case studies were analyzed using the lithium-ion battery electrochemical impedance spectroscopy dataset published in the international journal *Nature Communications*. This dataset contains three prematurely aging lithium-ion batteries and five long-lived lithium-ion batteries. Specifically, the electrochemical impedance spectra and capacity values ​​for each lithium-ion battery at each cycle number were collected. Figure 2 The equivalent circuit model shown is used for nonlinear fitting of the electrochemical impedance spectrum, with initial physical information parameter θ. phy =[R ohm ,R SEI ,R CT ,C1,C2,C W The initial values ​​and upper and lower limits of ] are shown in Table 1:

[0111] Table 1

[0112]

[0113] Sensitivity analysis was conducted by scaling the parameters to be analyzed to 90%, 95%, 105%, and 110% of their original values, respectively. While keeping the other five parameters constant, the changes in the four simulated electrochemical impedance spectra relative to the measured electrochemical impedance spectra were observed. The parameters with the greatest impact on the electrochemical impedance spectra were then selected as the primary parameters, forming the physical information parameters. Figure 3 This is a sensitivity analysis plot of the initial physical information parameters, due to R ohm R SEI and R CT Since these three parameters have a significant impact on the electrochemical impedance spectroscopy, they are used as physical information parameters. The physical information parameters of each lithium-ion battery at each cycle number are obtained. Subsequently, for ease of comparison, a baseline model was constructed, with electrochemical impedance spectroscopy as input and battery capacity as output. A deep learning model-1 of lithium-ion battery physics was then constructed, with physical information parameters as input. Electrochemical impedance spectroscopy and other parameters are input into the input layer of the model; a deep learning model-2 of lithium-ion battery physics is constructed, whose input is physical information parameters. And electrochemical impedance spectroscopy, but physical information parameters are needed. The input is fed into a neural layer near the output layer; a deep learning model-3 of lithium-ion battery physics knowledge is constructed, whose input is physical information parameters. While using electrochemical impedance spectroscopy, a physical regularization method is employed, extracting three neurons from the neural layer closest to the output layer and aligning their values ​​with physical information parameters. A deep learning model-4 for lithium-ion battery physics is constructed, integrating the lifespan characteristic classification task—determining whether a lithium-ion battery is prematurely aged or long-lived—into capacity estimation, with physical information parameters as its input. And electrochemical impedance spectroscopy, but physical information parameters need to be included. The input is fed into the neural layer containing the shared neurons, and the output is the classification results of battery capacity and lifespan characteristics. A deep learning model-5 for lithium-ion battery physics is constructed, integrating the lifespan characteristic classification task into capacity estimation; its input is physical information parameters. While using electrochemical impedance spectroscopy, a physical regularization method is employed to extract three neurons from the neural layer containing the shared neurons and align their values ​​with physical information parameters. The output is a classification result of battery capacity and lifespan characteristics.

[0114] In the deep learning model-3 of lithium-ion battery physics knowledge, the hyperparameter λ controls the degree of influence of physics knowledge on the model. p In the deep learning model of lithium-ion battery physics knowledge-4, the hyperparameter λ2 controlling the influence of the lifetime characteristic classification task on the model is set to 0.0090. Similarly, in the deep learning model of lithium-ion battery physics knowledge-5, the hyperparameter λ2 controlling the influence of the lifetime characteristic classification task on the model is set to 0.0080, and the hyperparameter λ controlling the influence of physical knowledge on the model is set to... p Set to 0.0060.

[0115] The electrochemical impedance spectroscopy (EIS) datasets of six lithium-ion batteries were divided into training and validation sets, while the EIS datasets of two other lithium-ion batteries were used as the test set. The hyperparameter values ​​for the six models are shown in Table 2.

[0116] Table 2

[0117]

[0118] A deep integration strategy is adopted to construct M=500 identical base models. For battery capacity estimation, the mean of the M=500 base models is used as the final estimation result, and the length of the confidence interval at the 90% confidence level is used as a measure of prediction uncertainty. For battery life characteristic classification, the categories determined by the M=500 base models according to a voting method are used as the classification results, and the accuracy is used as a measure of uncertainty.

[0119] The mean squared errors of the above benchmark model and the five deep learning models based on lithium-ion battery physics for each cycle of battery capacity estimation on the test set are shown below. Figure 4 and Figure 5 , Figure 4 and Figure 5 The “Deep Learning Model for Lithium-ion Battery Physics Knowledge-1” will be abbreviated as “Model-1”, and subsequent models will be abbreviated in the same way.

[0120] The mean squared errors of the above-mentioned benchmark model and the five deep learning models of lithium-ion battery physics knowledge in the battery capacity estimation over the entire life cycle of the test set are summarized in Table 3.

[0121] The confidence interval lengths of the aforementioned benchmark model and the five deep learning models based on lithium-ion battery physics for each cycle of battery capacity estimation in the test set are respectively displayed. Figure 6 and Figure 7 .

[0122] The mean values ​​of the confidence interval lengths for the entire lifespan of the battery capacity estimation in the test set for the above-mentioned benchmark model and the five deep learning models of lithium-ion battery physics knowledge are summarized in Table 4.

[0123] Table 3

[0124]

[0125] Table 4

[0126]

[0127] From Table 3, Figure 4 and Figure 5 As can be seen, the proposed model has low mean square error in the test cells and can accurately estimate the cell capacity. (See Table 3.) Figure 4 , Figure 5 Table 4 Figure 6 and Figure 7 In this study, by comparing the benchmark model with five deep learning models based on lithium-ion battery physics knowledge, it was verified that incorporating physical or domain knowledge can improve the accuracy of capacity estimation and reduce prediction uncertainty. A comparison between lithium-ion battery physics knowledge deep learning model-1 and model-2 demonstrates that a reasonable input method for physical information parameters is beneficial to improving the model's performance in capacity estimation. Furthermore, a comparison between lithium-ion battery physics knowledge deep learning model-1, model-4, and model-5 verifies that fusing three types of information sources can significantly improve capacity estimation accuracy and reduce prediction uncertainty. Physical regularization can be considered as a priority strategy to integrate physical knowledge into the deep model.

[0128] Both the lithium-ion battery physics knowledge deep learning model-4 and the lithium-ion battery physics knowledge deep learning model-5 can accurately classify the life characteristics of the test set batteries. The accuracy rates of these two models in classifying the life characteristics of the test set batteries are summarized in Table 5.

[0129] Table 5

[0130]

[0131] As shown in Table 5, incorporating physics knowledge and domain knowledge can ensure a high classification accuracy.

[0132] This invention presents a lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy. To study the internal state of lithium-ion batteries, an equivalent circuit model is constructed using electrochemical impedance spectroscopy, thereby extracting physical information parameters characterizing the internal state of the battery. Simultaneously, a deep learning model of lithium-ion battery physics knowledge is constructed, using the physical information parameters and electrochemical impedance spectroscopy as model inputs to estimate the capacity of the lithium-ion battery. The designed deep learning model employs physical regularization to assign physical meaning to specific neurons in the deep learning model, extracting physically interpretable latent features. Furthermore, based on domain knowledge, multiple similar tasks are combined to achieve multi-task learning, resulting in better latent feature extraction. Physical regularization is then used as an auxiliary task, effectively utilizing three types of information sources: physical knowledge, measurement data, and domain knowledge. This ensures that the diagnostic results for lithium-ion batteries are interpretable and highly accurate.

[0133] Finally, it should be noted that the above embodiments are for illustration only and not for limiting the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention without departing from the spirit and scope of the present invention. Any modifications or partial substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A diagnostic method for lithium-ion batteries that integrates physical information parameters and electrochemical impedance spectroscopy, characterized in that, It includes the following steps: S1. Extract physical information parameters characterizing the internal state of lithium-ion batteries: Construct an equivalent circuit model of lithium-ion batteries using electrochemical impedance spectroscopy to extract physical information parameters characterizing the internal state of lithium-ion batteries. S11. Perform characteristic tests on the lithium-ion battery to obtain the electrochemical impedance spectrum of the lithium-ion battery after n cycles. S12. Perform image analysis on the electrochemical impedance spectroscopy of the lithium-ion battery to construct an equivalent circuit model of the lithium-ion battery; the components of the equivalent circuit model of the lithium-ion battery include the ohmic internal resistance R. ohm Solid electrolyte membrane internal resistance R SEI Charge transfer internal resistance R CT Warburg element Z War The first constant-phase element CPE1 and the second constant-phase element CPE2, wherein the ohmic internal resistance R ohm Caused by the movement of electrons in the electrode and ions in the electrolyte, and serving as the first module; the internal resistance R of the solid electrolyte membrane SEI The first impedance Z is formed in parallel with the first constant-phase element CPE1. arc1 And as a second module, it characterizes the diffusion resistance encountered by ions when passing through a solid electrolyte membrane in the high-frequency region; the charge transfer resistance R CT and Warburg component Z War The second impedance Z is formed by connecting it in series with the second constant-phase element CPE2 and then in parallel. arc2 And as the third module, it characterizes the internal resistance of charge transfer in the mid-frequency region and the diffusion internal resistance of material transfer in the low-frequency region; the first module, the second module and the third module are connected in series to form the equivalent circuit model of the lithium-ion battery. S13. Nonlinear fitting is performed on the electrochemical impedance spectrum of the lithium-ion battery at each cycle number to extract the physical information parameters characterizing the internal state of the lithium-ion battery at each cycle number. S2. Constructing a deep learning model of lithium-ion battery physics knowledge: Based on deep learning, a deep learning model of lithium-ion battery physics knowledge is constructed by integrating physical knowledge and electrochemical impedance spectroscopy. S3. Integrate physical knowledge, measurement data and domain knowledge to extract physically interpretable latent features, which include a first latent feature, a second latent feature and a third latent feature; S31. Extracting physically interpretable first hidden features using physical regularization: By assigning physical meaning to specific neurons in the deep learning model of lithium-ion battery physics knowledge, the numerical values ​​of neurons are transformed into physical information parameters. The neurons are brought closer together so that their values ​​can characterize the internal state of the lithium-ion battery. S32. Construct a multi-task learning model based on domain knowledge, jointly train T similar tasks, and extract the second hidden features by sharing neurons among the T similar tasks. Physical information parameters for each loop iteration Electrochemical impedance spectroscopy is used as the input to the multi-task learning model; S33. Based on a multi-task learning model, physical regularization is used as an auxiliary task. P shared neurons are selected from T similar task shared neurons to extract physically interpretable third latent features and assign them physical meaning. The values ​​of the P shared neurons are then mapped to the physical information parameters. By bringing them closer together, the numerical representation of the shared neurons can be used to characterize the internal state of the lithium-ion battery. S4. Use a deep integration strategy to evaluate the prediction uncertainty of lithium-ion battery capacity estimation and the prediction uncertainty of lithium-ion battery life characteristic classification. An equivalent circuit model of a lithium-ion battery is constructed using electrochemical impedance spectroscopy, thereby extracting physical information parameters that characterize the internal state of the lithium-ion battery. At the same time, a deep learning model of lithium-ion battery physics is constructed, using the physical information parameters and electrochemical impedance spectroscopy as model inputs to estimate the capacity of the lithium-ion battery.

2. The lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy according to claim 1, characterized in that, Step S13 specifically includes the following steps: S131. Based on the equivalent circuit model of a lithium-ion battery, derive the impedance Z(w) of the lithium-ion battery in each cycle number as a function of frequency w: Z(in)=R ohm +Z arc1 (in)+Z arc2 (in) (1) Among them, Z arc1 (w) represents the first impedance related to frequency, and C1 represents Z arc1 The double-layer capacitance, where j represents the imaginary unit and n1 represents the first glide factor; Z arc2 (w) represents the second impedance related to frequency, and C2 and C W Z arc2 and Z War The double-layer capacitance, where n2 represents the second glide factor; S132, The initial physical information parameter characterizing the internal state of a lithium-ion battery is θ. phy =[R ohm ,R SEI ,R CT ,C1,C2,C W The initial physical information parameter θ is obtained by using the least squares method. phy Make an estimate: Where m represents the number of sampling points for an electrochemical impedance spectroscopy curve; Re(·) and Im(·) represent the sampling points obtained at the current frequency w, respectively. i The impedance Z(w) below i Operations on the real and imaginary parts of Z(w); i ) represents the current frequency w of the model fit. i The impedance below, and Represent θ phy The lower and upper limits are given by domain knowledge; S133. Nonlinear fitting is performed on the electrochemical impedance spectroscopy of the lithium-ion battery at each cycle number to analyze the aging phenomenon of each initial physical information parameter as the lithium-ion battery operates. Sensitivity analysis is then used to determine the aging phenomenon of the initial physical information parameter θ. phy P parameters characterizing the internal state of a lithium-ion battery are selected from the data to form physical information parameters. Used to assess the internal state of lithium-ion batteries and estimate their capacity.

3. The lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Based on deep learning, construct a deep learning model of lithium-ion battery physics knowledge, and incorporate the physical information parameters of each cycle. The data is input to the input layer of the deep learning model of lithium-ion battery physics knowledge along with the electrochemical impedance spectroscopy. The output layer of the deep learning model of lithium-ion battery physics knowledge outputs a value to characterize the capacity of the lithium-ion battery at the current cycle number. S22. For lithium-ion battery capacity estimation, the mean square error is used as the first loss function L. MSE (θ DL ): Among them, y k The measured value representing the capacity of a lithium-ion battery; y k This represents the lithium-ion battery capacity estimate output by the deep learning model of lithium-ion battery physics; N represents the number of samples in the training set; k is a constant; θ DL The parameters of the deep learning model representing the physical knowledge of lithium-ion batteries, including the weight matrix and bias vector, are obtained by iterative optimization using backpropagation, Adam optimization algorithm and its corresponding variants, by minimizing the loss function. S23. For the classification of lithium-ion battery life characteristics, cross-entropy is used as the second loss function L. CE (θ DL ): in, This indicates the actual type of lithium-ion battery entered. This indicates that the deep learning model, which represents the physical knowledge of lithium-ion batteries, classifies the input as a category. The probability; log represents the base-10 log function.

4. The lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy according to claim 1, characterized in that, In step S31, a physical regularization method is used, and a loss function L based on the physical regularization method is set. PR (θ DL ), in the first loss function L MSE (θ DL Add a physical regularization penalty term L to ) P (θ DL This is used to penalize the neuron's numerical value in relation to the physical information parameters. The degree of deviation, that is: L PR (i DL )=L MSE (i DL )+λ p L P (i DL ) (5) Where, λ p Hyperparameters representing the degree to which control physics knowledge influences the deep learning model of lithium-ion battery physics; In step S32, the third loss function L of the multi-task learning model is set. MTL (θ DL ): L MTL (i DL )=λ1L1(θ DL )+...+λ T L T (i DL ) (7) Where, λ i Let λ represent the hyperparameters of the i-th similar task and λ represent the hyperparameters of the i-th similar i ≥0, i=1,...,T;L i (θ DL () represents the loss function for the i-th corresponding task; In step S33, the fourth loss function L of the multi-task learning model, which uses physical regularization as an auxiliary task, is set. MTL,P (θ DL The fourth loss function L MTL,P (θ DL Let L be the loss function and physical regularization penalty term for T similar tasks. P (θ DL The weighted sum of ), that is: L MTL,P (i DL )=λ1L1(θ DL )+...+λ T L T (i DL )+λ p L P (i DL ) (8).

5. The lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Construct M identical deep learning models of lithium-ion battery physics knowledge as base models; S42. Sequentially input the physical information parameters for each cycle number. Electrochemical impedance spectroscopy is used as input to M basis models, and the input method depends on the structure of the basis models; S43. Train and run each base model individually to obtain the output y of the M base models. k,1 ... y k,M ; S44. For lithium-ion battery capacity estimation, the mean of the outputs of the M basic models is used as the final output result. S45. Arrange the outputs of the M base models in ascending order to obtain... Calculate the index index a1 = M × α / 2 for the lower limit of the confidence interval at the α% confidence level, and the index index a2 = M × (1 - α / 2) for the upper limit of the confidence interval at the α% confidence level. Then, the measure of the predictive uncertainty of the capacity estimate is the length L of the confidence interval at the α% confidence level. CI : in, Indicates the lower limit of the confidence interval; Indicates the upper limit of the confidence interval; S46. For the classification of lithium-ion battery life characteristics, a voting method is adopted. The classification result of each base model is recorded as one vote. The classification results of each base model are then counted, and the category with the most votes is taken as the final classification result. The maximum number of votes is H. The measurement of the prediction uncertainty of lithium-ion battery life characteristic classification is accuracy β = H / M.

6. The lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy according to claim 4, characterized in that, The physical regularization penalty term L in step S31 P (θ DL It is composed of the sum of P regularization terms: in, This represents the value of the r-th specific neuron; Representing physical information parameters The r-th parameter in the equation.

7. The lithium-ion battery diagnostic method integrating physical information parameters and electrochemical impedance spectroscopy according to claim 1, characterized in that, Considering that electrochemical impedance spectroscopy is high-dimensional, and the physical information parameters... The dimensionality is relatively low, and the physical information parameters The rich amount of information contained is also diluted layer by layer by the neural layers. In the deep learning model of lithium-ion battery physics knowledge in step S2, the physical information parameters are... The input is fed into a neural layer closer to the output layer. In the multi-task learning model of step S32, the physical information parameters are... The input is sent to the neural layer containing the T similar task-shared neurons.

Images