A Method for Constructing a Battery Degradation Model Based on a Bayesian Physical Information Neural Network

Abstract

This invention discloses a method for constructing a battery degradation model based on a Bayesian physical information neural network, comprising the following steps: constructing a pseudo-two-dimensional battery data generation module based on the battery aging mechanism; designing a feature extraction network to extract IC feature parameters; constructing an aging parameter mapping network to characterize the relationship between IC feature parameters and battery aging parameters; and constructing a Bayesian battery health state inference network through parameter randomization. This invention employs the above-mentioned method for constructing a battery degradation model based on a Bayesian physical information neural network. While achieving explicit mapping between IC features and aging parameters, it utilizes the internal electrochemical equations of the battery to construct physical residual constraints, enabling the battery degradation model to possess high physical interpretability and avoiding the risks of black-box models. Through a Bayesian framework based on random parameter distribution modeling, the uncertainty in the model modeling process is quantified, thereby achieving quantified confidence of model predictions and supporting risk-sensitive decision-making.

Description

Technical Field

[0001] This invention relates to the field of battery degradation modeling technology, and in particular to a method for constructing battery degradation models based on Bayesian physical information neural networks. Background Art

[0002] With the transformation of the global energy structure and the rapid development of electric vehicles, renewable energy storage, and other fields, battery technology, as one of the core supporting technologies, has received widespread attention. Battery performance directly affects the driving range of electric vehicles, the driving time of electronic devices, and the energy storage efficiency of the power grid. However, battery performance degradation is inevitable during use, which not only affects battery lifespan but may also lead to safety hazards. Therefore, accurately predicting the battery degradation process and optimizing battery use and maintenance strategies have become critical issues that the battery industry urgently needs to address.

[0003] Currently, research on battery degradation models mainly focuses on two categories: physical models and data-driven models. Physical models describe the battery degradation process by establishing electrochemical reaction equations within the battery. While they offer high interpretability, their accuracy and generalization ability are often limited due to the complex chemical reactions and material properties within the battery. Data-driven models, on the other hand, utilize extensive experimental data and deep learning techniques to construct multi-layer neural networks, automatically extracting complex features in the battery degradation process and improving prediction accuracy to some extent. However, most existing deep learning models rely on deterministic predictions based on precise data and lack the ability to quantify uncertainties. This is particularly important in battery degradation prediction, as the battery degradation process is influenced by various uncertain factors, such as temperature, charge / discharge rates, and their extent.

[0004] To address the shortcomings of existing research, this invention proposes a battery degradation model construction method based on a Bayesian physical information neural network. This method combines the powerful feature extraction capabilities of deep learning with the uncertainty quantification advantages of Bayesian inference, enabling more accurate prediction of battery degradation trends and providing uncertainty estimates for degradation predictions. Through the implementation of this invention, battery manufacturers, electric vehicle manufacturers, and energy storage system operators can be provided with more accurate battery degradation prediction tools, thereby optimizing battery usage and maintenance strategies, extending battery lifespan, reducing operating costs, and improving system safety and reliability. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a battery degradation model based on a Bayesian physical information neural network. While achieving explicit mapping between IC features and aging parameters, it utilizes the electrochemical equations inside the battery to construct physical residual constraints, enabling the battery degradation model to have high physical interpretability and avoiding the risks of black box models. Through a Bayesian framework based on random parameter distribution modeling, the uncertainty in the model modeling process is quantified, thereby realizing the quantification of the confidence of model predictions and supporting risk-sensitive decision-making.

[0006] To achieve the above objectives, this invention provides a method for constructing a battery degradation model based on a Bayesian physical information neural network, comprising the following steps:

[0007] Step S1: Based on the battery aging mechanism, construct a pseudo-two-dimensional battery data generation module to generate training data for the battery degradation model;

[0008] Step S2: Design a feature extraction network to extract IC feature parameters from the voltage-capacity curve;

[0009] Step S3: Construct an aging parameter mapping network to characterize the relationship between IC feature parameters and battery aging parameters;

[0010] Step S4: Construct a Bayesian battery health state inference network by randomizing the parameters. Its input is the battery aging parameters and its output is the probability distribution of the battery SOH.

[0011] Step S5: During model training, the loss function is designed using the Butler-Volmer equation and the diffusion equation, physical constraints are added to the aging parameter mapping network, and the stochastic gradient variational Bayesian method is used to estimate the posterior distribution of the Bayesian neural network parameters, ultimately obtaining the battery degradation model.

[0012] Preferably, in step S1, a pseudo-two-dimensional battery data generation module is constructed based on the battery aging mechanism to generate multi-dimensional training data for the battery degradation model. The specific process is as follows:

[0013] Step S11: Determine battery and operating parameters;

[0014] Step S111, battery parameters, specifically including: positive / negative electrode materials, active material volume fraction, and initial lithium concentration;

[0015] Step S112, operating parameters, specifically including: charge / discharge rate, temperature, and number of cycles;

[0016] Step S12: Using the above parameters and the PyBamm toolkit, introduce the SEI growth and active material loss and aging mechanism to improve the pseudo-two-dimensional battery model;

[0017] Step S13: By combining different operating condition parameters, multiple sets of aging state data are generated. Each set includes: time series voltage data. Capacity decay curve Voltage-capacity curve;

[0018] Step S14: Perform numerical differentiation on the voltage-capacity curve and calculate the IC feature parameter labels for subsequent feature extraction network training. Specifically, this includes: peak voltage. Peak area Half-peak width .

[0019] Preferably, in step S2, a feature extraction network is designed to extract IC feature parameters from the voltage-capacity curve. The specific process is as follows:

[0020] Step S21: Perform data cleaning and interpolation on the voltage-capacity curve to obtain a uniformly spaced voltage V and capacity Q sequence;

[0021] Step S22: Construct a feature extraction network, which specifically includes an input layer, a 1D convolutional layer, a max pooling layer, a physically constrained layer, and a fully connected layer;

[0022] Step S23: Input the voltage-capacity curve processed in step S21 into the input layer of the feature extraction network;

[0023] Step S24: Extract local voltage-capacity correlation features from the 1D convolutional layer;

[0024] Step S25: Dimensionality reduction of the extracted local voltage-capacity correlation features is performed using a max pooling layer while retaining salient features;

[0025] Step S26: In the physical constraint layer, the mathematical characteristics of the network output IC curve are constrained by designing a loss function;

[0026] Step S27: Directly regress IC feature parameters through a fully connected layer. , , .

[0027] Preferably, in step S26, the loss function is designed, specifically including:

[0028] Step S261, Basic Loss The difference between the predicted IC parameters and the tag data is calculated using the mean square error (MSE).

[0029] Step S262, Physical Constraints By using the Savitsky-Golay differential operator, a differential relation constraint is added to the loss function to ensure that the mathematical properties of the intermediate features of the network are consistent with those of the IC curve, as shown below:

[0030] ;

[0031] in, The capacity sequence predicted by the network; This is the actual capacity sequence, i.e., the label data; For Savitzky-Golay differentiation operations; This represents the total number of data points.

[0032] Step S263, Total Loss Function As shown below:

[0033] ;

[0034] in, This is a hyperparameter that balances the weights of the mean squared error loss and the physical constraint term.

[0035] Preferably, in step S3, an aging parameter mapping network is constructed to characterize the relationship between IC feature parameters and battery aging parameters. The specific process is as follows:

[0036] Step S31: Determine the relationship between IC characteristic parameters and battery aging parameters;

[0037] Step S311: Battery aging reduces the exchange current density drop, overpotential The rise eventually led to Offset;

[0038] Step S312, SEI growth or particle breakage reduces the concentration of active material The decline directly led to reduce;

[0039] Step S313: Aging leads to pore blockage, which reduces the diffusion coefficient. The decline led to rise;

[0040] Step S32: Use a multi-layer MLP to construct an aging parameter mapping network, whose structure includes an input layer, a hidden layer, and an output layer;

[0041] Step S33: Input IC feature parameters, and based on the relationship between IC feature parameters and battery aging parameters, add constraints to the mapping network to achieve the mapping output of battery aging parameters.

[0042] Preferably, in step S33, residual constraints are constructed using the Butler-Volmer equation. As shown below:

[0043] ;

[0044] in, For exchange current density; and It is a symmetry factor; It is Faraday's constant; This is an overpotential; It is the ideal gas constant; Absolute temperature; Tag current density; It is an L2 norm;

[0045] Total loss function As shown below:

[0046] ;

[0047] in, For residual constraints The weighting coefficients.

[0048] Preferably, in step S4, a Bayesian battery health state inference network is constructed by randomizing the parameters. The input is the battery aging parameters, and the output is the probability distribution of the battery's state of health (SOH). The specific process is as follows:

[0049] Step S41: Define Bayesian network node parameters using Gaussian distribution. The prior distribution is as follows:

[0050] ;

[0051] ;

[0052] in, These are the weights of the network node parameters; It is the bias of network node parameters; Indicates variance; Represents weight parameters The prior probability density function; Represents weight parameters The prior probability density function; Indicates a Gaussian distribution;

[0053] Step S42: Assume the noise in the data follows a normal distribution. Choose a Gaussian distribution and define the likelihood function. The likelihood function represents the probability of observing correct data given the network parameters, as shown below:

[0054] ;

[0055] in, This represents a dataset of aging parameters; Represents the likelihood function; This represents the true value of SOH; This represents the model's predicted value SOH;

[0056] Step S43: Input battery aging parameters and output the probability distribution of battery health status. , ), used to quantify the uncertainty of forecasts; among which, This represents the mean. Represents standard deviation.

[0057] Preferably, in step S5, the constructed battery degradation model is trained using the stochastic gradient variational Bayesian method to estimate the posterior distribution of the Bayesian neural network parameters, ultimately obtaining the battery degradation model. The specific process includes:

[0058] Step S51: Data preprocessing;

[0059] Step S52: Optimize the algorithm;

[0060] Step S53: Set training parameters;

[0061] In step S52, the specific process of performing the optimization algorithm is as follows:

[0062] Step S521: Based on Bayesian inference theorem, use variational inference to construct a set of parameters. Parameterized variational distribution ( ), used to fit the posterior distribution to be estimated;

[0063] The posterior distribution fitted based on the Bayesian inference theorem is as follows:

[0064] ;

[0065] in, As the evidence normalization factor, the product of the likelihood function and the prior distribution is... Integral in space; It is the prior distribution; The posterior distribution to be estimated;

[0066] Variational distribution used ( The mean field is a Gaussian distribution, as shown below:

[0067] ;

[0068] in, Network node parameters The decomposition results in multiple independent sub-variables; represents the mean parameter of the variational distribution; is the variance parameter of the variational distribution;

[0069] Step S522: The optimization objective for fitting the posterior distribution is to maximize the lower bound of evidence ELBO, i.e., minimize the negative ELBO, as shown below:

[0070] ;

[0071] ;

[0072] in, Represents the variational distribution Below, the log-expected value of the data likelihood; KL represents the variational distribution. With prior distribution KL divergence; Represents the joint probability distribution;

[0073] ;

[0074] in, The standard deviation represents the prior distribution; The standard deviation represents the variational distribution; This represents the mean of the prior distribution; Represents the mean of the variational distribution;

[0075] Step S523: Using the stochastic gradient variational Bayesian method, the gradient is calculated through the reparameterization technique, specifically including:

[0076] First, Represented as a deterministic function: ;

[0077] in, For element-wise multiplication, Standard Gaussian noise; A random parameter vector generated through reparameterization; is the mean of the variational distribution; The standard deviation of the variational distribution;

[0078] Then, estimate gradient As shown below:

[0079] ;

[0080] in, Indicates the total number of samples; , indicating the first The parameters of the next sample; For variational parameters; Indicates the given parameters Time data The probability of; Indicates the variational distribution at The probability density; Indicates parameters The prior probability density;

[0081] Step S524: Using the minimization of negative ELBO as the loss calculation method, the gradient is updated using the backpropagation algorithm, and finally the variational distribution at the convergence of the loss is used as an approximation of the posterior distribution.

[0082] A battery degradation model construction system based on Bayesian physical information neural network includes a pseudo-two-dimensional battery data generation module, a feature extraction network module, a battery aging parameter mapping module, and a Bayesian battery health status inference module.

[0083] Among them, the pseudo-two-dimensional battery data generation module is used to generate training data for the battery degradation model;

[0084] The feature extraction network module is used to extract IC feature parameters from the voltage-capacity curve;

[0085] The battery aging parameter mapping module is used to explicitly map aging parameters. At the same time, it uses the physical partial differential equations in the battery degradation process to construct physical constraint loss and incorporate it into the model training process, so that the battery degradation model conforms to physical laws and has physical interpretability.

[0086] The Bayesian battery health status inference module is used for randomized parameter modeling of the model. It utilizes battery aging parameters and prior knowledge to construct and realize battery health inference and model uncertainty quantification.

[0087] A computer device includes a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of a method for constructing a battery degradation model based on a Bayesian physical information neural network.

[0088] Therefore, this invention adopts the above-mentioned battery degradation model construction method based on Bayesian physical information neural network. While realizing the explicit mapping between IC features and aging parameters, it uses the electrochemical equations inside the battery to construct physical residual constraints, so that the battery degradation model has high physical interpretability and avoids the risks of black box models. Through the Bayesian framework based on random parameter distribution modeling, the uncertainty in the model modeling process is quantified, thereby realizing the confidence quantification of model prediction and supporting risk-sensitive decision-making.

[0089] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0090] Figure 1 This is a flowchart of the battery degradation model construction method based on Bayesian physical information neural network of the present invention;

[0091] Figure 2 This is a flowchart illustrating the implementation of the feature extraction network designed in this invention to extract IC feature parameters;

[0092] Figure 3 This is a flowchart illustrating the implementation of the present invention in constructing an aging parameter mapping network to characterize the relationship between IC feature parameters and battery aging parameters. Detailed Implementation

[0093] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0094] like Figure 1 As shown, the present invention provides a method for constructing a battery degradation model based on a Bayesian physical information neural network, comprising the following steps:

[0095] Step S1: Construct a pseudo-two-dimensional (P2D) battery data generation module based on the battery aging mechanism to generate training data for the battery degradation model;

[0096] Step S2: Design a feature extraction network to extract IC feature parameters from the voltage-capacity (VQ) curve;

[0097] Step S3: Construct an aging parameter mapping network to characterize the relationship between IC feature parameters and battery aging parameters;

[0098] Step S4: Construct a Bayesian SOH (Battery Health Status) inference network by randomizing the parameters. Its input is the battery aging parameters, and its output is the probability distribution of the battery SOH.

[0099] Step S5: During model training, the Butler-Volmer equation and diffusion equation are used to design a loss function to add physical constraints to the aging parameter mapping network. The stochastic gradient variational Bayesian (SGVB) method is used to estimate the posterior distribution of the Bayesian neural network parameters, and finally the battery degradation model is obtained.

[0100] Example 1

[0101] Step S1: Based on the battery aging mechanism, construct a pseudo-two-dimensional (P2D) battery data generation module to generate multi-dimensional training data for the battery degradation model.

[0102] Step S11: Determine battery and operating parameters.

[0103] Step S111, battery parameters, specifically including:

[0104] (1) The positive electrode material is The active material volume fraction is 0.65, and the initial lithium concentration is 2000 mol / m³. 3 .

[0105] (2) The negative electrode material is graphite, the volume fraction of the active material is 0.70, and the initial lithium concentration is 0 mol / m³. 3 .

[0106] Step S112, operating parameters, specifically including:

[0107] (1) Charge / discharge rates: 1C, 2C, 3C.

[0108] (2) Temperature: 25℃, 35℃, 45℃.

[0109] (3) Number of cycles: 50, 100, 150, 200.

[0110] Step S12: Using the above parameters and the PyBamm toolkit, introduce aging mechanisms such as SEI growth and active material loss to improve the P2D battery model.

[0111] SEI film growth is described by the Butler-Volmer equation, which includes parameters such as the reaction rate constant and activation energy. Active material loss simulates the concentration decay of active materials due to particle breakage, exhibiting an exponential relationship with the number of cycles.

[0112] Step S13: By combining different operating parameters, 3×3×4=36 sets of aging state data are generated, each set containing:

[0113] (1) Time series voltage data The sampling frequency is 10Hz.

[0114] (2) Capacity decay curve Record once every 5 cycles.

[0115] (3) Voltage-capacity curve (VQ curve), which is the curve of voltage change with capacity during charging and discharging.

[0116] Step S14: Perform numerical differentiation on the VQ curve, set the window size of the Savitzky-Golay filter to 7, and calculate the IC characteristic parameter labels, specifically including:

[0117] (1) Peak voltage This represents the voltage corresponding to the peak value of the differential curve.

[0118] (2) Peak area The integral area of ​​the peak region of the differential curve;

[0119] (3) Half-peak width The width is the width at half the height of the peak.

[0120] The calculated IC feature parameter labels are used for training the subsequent feature extraction network, so that the model is not limited to a specific dataset, and the neural network mapping method has high computational efficiency.

[0121] Step S2: Design a feature extraction network to extract IC feature parameters from the voltage-capacity (VQ) curve, such as... Figure 2 As shown.

[0122] Step S21: Perform data cleaning and interpolation on the VQ curve to obtain uniformly spaced V and Q sequences.

[0123] Step S22: Construct a feature extraction network, which specifically includes an input layer, a 1D convolutional layer, a max pooling layer, a physical constraint layer, and a fully connected layer, as shown in Table 1.

[0124] Table 1 Feature Extraction Network Architecture

[0125] ;

[0126] Step S23: Input the preprocessed VQ curve into the input layer of the feature extraction network.

[0127] Step S24: Extract local voltage-capacity correlation features from the 1D convolutional layer.

[0128] Step S25: Dimensionality reduction of the extracted local voltage-capacity correlation features is performed using a max pooling layer while retaining significant features.

[0129] Step S26: In the physical constraint layer, a loss function is designed to constrain the network output to conform to the mathematical characteristics of the IC feature curve; wherein, the loss function consists of the true value and the predicted value.

[0130] Step S261, Basic Loss The difference between the predicted IC parameters and the tag data is calculated using the mean square error (MSE).

[0131] Step S262, Physical Constraints By using the Savitsky-Golay differential operator, a differential relation constraint is added to the loss function to ensure that the mathematical properties of the intermediate features of the network are consistent with those of the IC curve, as shown below:

[0132] ;

[0133] in, The capacity sequence predicted by the network; This is the actual capacity sequence, i.e., the label data; For Savitzky-Golay differentiation operations; This represents the total number of data points.

[0134] Step S263, Total Loss Function As shown below:

[0135] ;

[0136] in, For hyperparameters, the weights of the mean squared error loss and the physical constraint term are balanced, and the following values ​​are taken: .

[0137] Step S27: Directly regress IC feature parameters through a fully connected layer. , , .

[0138] Step S3: Construct an aging parameter mapping network to characterize the relationship between IC feature parameters and battery aging parameters, such as... Figure 3 As shown.

[0139] Step S31: Determine the relationship between IC characteristic parameters and battery aging parameters.

[0140] Step S311: Battery aging reduces the exchange current density drop, overpotential The rise eventually led to Offset;

[0141] Step S312, SEI growth or particle breakage reduces the concentration of active material The decline directly led to reduce;

[0142] Step S313, aging leading to pore blockage and other phenomena reduces the diffusion coefficient The decline led to rise.

[0143] Step S32: Construct an aging parameter mapping network using a multi-layer MLP, the structure of which is as follows:

[0144] (1) Input layer: 3 neurons, each corresponding to the IC feature parameters , , ;

[0145] (2) Hidden layers: 2 layers, 32 neurons per layer, using the ReLU activation function;

[0146] (3) Output layer: 3 neurons, each corresponding to the aging parameter. , , .

[0147] Step S33: Input IC feature parameters, and add constraints to the mapping network according to the relationship between IC feature parameters and battery aging parameters to realize the mapping output of battery aging parameters.

[0148] Using the Butler-Volmer equation, construct residual constraints. As shown below:

[0149] ;

[0150] in, For exchange current density; and Let be the symmetry factor, and take... ; It is Faraday's constant; This is an overpotential, calculated from voltage data; It is the ideal gas constant; Absolute temperature; Tag current density; It is an L2 norm.

[0151] Total loss function As shown below:

[0152] ;

[0153] in, For residual constraints The weighting coefficients are taken as follows: .

[0154] Step S4: Construct a Bayesian SOH (Battery Health Status) inference network by randomizing the parameters. Its input is the battery aging parameters and its output is the probability distribution of the battery SOH.

[0155] Step S41: Define Bayesian network node parameters using Gaussian distribution. The prior distribution is as follows:

[0156] ;

[0157] ;

[0158] in, These are the weights of the network node parameters; It is the bias of network node parameters; Indicates variance; Represents weight parameters The prior probability density function; Represents weight parameters The prior probability density function; This represents a Gaussian distribution.

[0159] Step S42: Assuming the noise in the data follows a normal distribution, we choose a Gaussian distribution to define the likelihood function. The likelihood function represents the probability of observing correct data given the network parameters, as shown below:

[0160] ;

[0161] in, This represents a dataset of aging parameters; Represents the likelihood function; This represents the true value of SOH; This represents the model's predicted value, SOH.

[0162] Step S43: Input battery aging parameters and output the probability distribution of SOH (Solution to Overheating). , ), used to quantify the uncertainty of forecasts; among which, This represents the mean. Represents standard deviation.

[0163] Step S5: Based on the above process, the constructed battery degradation model is trained, and the stochastic gradient variational Bayesian (SGVB) method is used to estimate the posterior distribution of the Bayesian neural network parameters, and finally the battery degradation model is obtained.

[0164] Step S51: Data preprocessing.

[0165] Step S511: Interpolate the VQ curve to a uniform interval of 500 points;

[0166] Step S512: Standardize IC characteristic parameters and aging parameters to the [0,1] interval.

[0167] Step S52: Optimize the algorithm.

[0168] Step S521: Based on Bayesian inference theorem, use variational inference to construct a set of parameters. Parameterized variational distribution ( ), used to fit the posterior distribution to be estimated.

[0169] The posterior distribution to be fitted based on the Bayesian inference theorem is:

[0170] ;

[0171] in, As the evidence normalization factor, the product of the likelihood function and the prior distribution is... Integral in space; It is the prior distribution; The posterior distribution to be estimated is denoted as .

[0172] Variational distribution used ( The mean field is a Gaussian distribution, as shown below:

[0173] .

[0174] in, Network node parameters The decomposition results in multiple independent sub-variables; represents the mean parameter of the variational distribution; is the variance parameter of the variational distribution.

[0175] Step S522: Use maximizing the lower bound of evidence (ELBO) as the optimization objective for fitting the posterior distribution, i.e., minimizing the negative ELBO, as shown below:

[0176] ;

[0177] ;

[0178] in, Represents the variational distribution Below, the log-expected value of the data likelihood; KL represents the variational distribution. With prior distribution KL divergence; This represents the joint probability distribution.

[0179] ;

[0180] in, The standard deviation of the prior distribution is represented (usually taken as 1). The standard deviation represents the variational distribution; This represents the mean of the prior distribution; Represents the mean of the variational distribution;

[0181] Step S523: Calculate the gradient using the stochastic gradient variational Bayes (SGVB) method through reparameterization techniques.

[0182] Will Represented as a deterministic function: ;

[0183] in, For element-wise multiplication, Standard Gaussian noise; This is a random parameter vector generated through reparameterization (as a sampled output of the variational distribution, which is the actual parameter input to the model network nodes). is the mean of the variational distribution; denoted as the standard deviation of the variational distribution.

[0184] estimate gradient As shown below:

[0185] ;

[0186] in, Indicates the total number of samples; , indicating the first The parameters of the next sample; For variational parameters; Indicates the given parameters Time data The probability of; Indicates the variational distribution at The probability density; Indicates parameters The prior probability density.

[0187] Step S524: Using the minimization of the negative ELBO as the loss calculation method, the gradient is updated using the backpropagation algorithm, and finally the variational distribution at the convergence of the loss is used as an approximation of the posterior distribution.

[0188] Step S53: Set training parameters:

[0189] (1) Batch size = 32, epoch = 200;

[0190] (2) Learning rate = 1e-3, using Adam optimizer;

[0191] (3) Verify the physical residuals and prediction accuracy every 10 epochs.

[0192] Example 2

[0193] This embodiment constructs a test scenario to experimentally verify the physical interpretability of the model construction method proposed in this invention.

[0194] 1. Experimental environment.

[0195] (1) Hardware configuration, as shown in Table 2.

[0196] Table 2 Hardware Configuration

[0197] ;

[0198] (2) Software implementation, as shown in Table 3.

[0199] Table 3 Software Configuration

[0200] ;

[0201] 2. Set training parameters.

[0202] A set of operating conditions not used in the training was selected: temperature = 30℃, charge / discharge rate = 2C, number of cycles = 180. The key performance indicators compared with those of the existing model are shown in Table 4.

[0203] Table 4 Key Indicators

[0204] ;

[0205] (1) VQ curve comparison: The IC features (peak voltage shift, peak area reduction) extracted by the model of this invention are highly consistent with the actual aging data.

[0206] (2) SOH uncertainty distribution: The output includes the mean (e.g., 85.3%) and standard deviation (e.g., 1.2%), providing a basis for risk decision-making for the battery management system.

[0207] This embodiment achieves high-precision prediction and confidence assessment of battery degradation models by integrating physical constraints and Bayesian uncertainty quantification. It has high physical interpretability and is suitable for battery health management in scenarios such as electric vehicles and energy storage systems.

[0208] Based on the battery degradation model construction method based on Bayesian physical information neural network proposed in this invention, a battery degradation model construction system based on Bayesian physical information neural network is also proposed, including a pseudo two-dimensional (P2D) battery data generation module, a feature extraction network module, a battery aging parameter mapping module, and a Bayesian battery health status inference module.

[0209] Among them, the pseudo-two-dimensional (P2D) battery data generation module is used to generate training data for the battery degradation model.

[0210] The feature extraction network module is used to extract IC feature parameters from the voltage-capacity (VQ) curve.

[0211] The battery aging parameter mapping module is used to explicitly map aging parameters. At the same time, it uses the physical partial differential equations in the battery degradation process to construct physical constraint loss and incorporates it into the model training process, so that the battery degradation model conforms to physical laws and has physical interpretability.

[0212] The Bayesian battery health status inference module is used for randomized parameter modeling of the model. It utilizes battery aging parameters and prior knowledge to construct and realize battery health inference and model uncertainty quantification.

[0213] Therefore, this invention adopts the above-mentioned battery degradation model construction method based on Bayesian physical information neural network. While realizing the explicit mapping between IC features and aging parameters, it uses the electrochemical equations inside the battery to construct physical residual constraints, so that the battery degradation model has high physical interpretability and avoids the risks of black box models. Through the Bayesian framework based on random parameter distribution modeling, the uncertainty in the model modeling process is quantified, thereby realizing the confidence quantification of model prediction and supporting risk-sensitive decision-making.

[0214] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for constructing a battery degradation model based on a Bayesian physical information neural network, characterized in that, Includes the following steps: Step S1: Based on the battery aging mechanism, construct a pseudo-two-dimensional battery data generation module to generate training data for the battery degradation model; Step S2: Design a feature extraction network to extract IC feature parameters from the voltage-capacity curve; Step S3: Construct an aging parameter mapping network to characterize the relationship between IC feature parameters and battery aging parameters; Step S4: Construct a Bayesian battery health state inference network by randomizing the parameters. Its input is the battery aging parameters and its output is the probability distribution of the battery SOH. Step S5: Train the constructed battery degradation model, design a loss function using the Butler-Volmer equation and diffusion equation, add physical constraints to the aging parameter mapping network, and use the stochastic gradient variational Bayesian method to estimate the posterior distribution of the Bayesian neural network parameters, finally obtaining the battery degradation model. The specific process includes: Step S51: Data preprocessing; Step S52: Optimize the algorithm; Step S53: Set training parameters; In step S52, the specific process of performing the optimization algorithm is as follows: Step S521: Based on Bayes' inference theorem, use variational inference to construct a set of parameters Parameterized variational distribution ( ), used to fit the posterior distribution to be estimated; The posterior distribution fitted based on the Bayesian inference theorem is as follows: ； in, As the evidence normalization factor, the product of the likelihood function and the prior distribution is... Integral in space; It is the prior distribution; The posterior distribution to be estimated; Variational distribution used ( The mean field is a Gaussian distribution, as shown below: ； in, Network node parameters The decomposition results in multiple independent sub-variables; is the mean parameter of the variational distribution; is the variance parameter of the variational distribution; Step S522: The optimization objective for fitting the posterior distribution is to maximize the lower bound of evidence ELBO, i.e., minimize the negative ELBO, as shown below: ； ； in, Represents the variational distribution Below, the log-expected value of the data likelihood; KL represents the variational distribution. With prior distribution KL divergence; Represents the joint probability distribution; ； in, The standard deviation represents the prior distribution; The standard deviation represents the variational distribution; This represents the mean of the prior distribution; Represents the mean of the variational distribution; Step S523: Using the stochastic gradient variational Bayesian method, the gradient is calculated through the reparameterization technique, specifically including: First, Represented as a deterministic function: ; in, For element-wise multiplication, Standard Gaussian noise; A random parameter vector generated through reparameterization; is the mean of the variational distribution; The standard deviation of the variational distribution; Then, estimate gradient As shown below: ； in, Indicates the total number of samples; , indicating the The parameters of the next sample; ={ } represents the variational parameters; Indicates the given parameters Time data The probability of; Indicates the variational distribution at The probability density; Indicates parameters The prior probability density; Step S524: Using the minimization of negative ELBO as the loss calculation method, the gradient is updated using the backpropagation algorithm, and finally the variational distribution at the convergence of the loss is used as an approximation of the posterior distribution.

2. The method for constructing a battery degradation model based on a Bayesian physical information neural network according to claim 1, characterized in that, In step S1, based on the battery aging mechanism, a pseudo-two-dimensional battery data generation module is constructed to generate multi-dimensional training data for the battery degradation model. The specific process is as follows: Step S11: Determine battery and operating parameters; Step S111, battery parameters, specifically including: positive / negative electrode materials, active material volume fraction, and initial lithium concentration; Step S112, operating parameters, specifically including: charge / discharge rate, temperature, and number of cycles; Step S12: Using the above parameters and the PyBamm toolkit, introduce the SEI growth and active material loss and aging mechanism to improve the pseudo-two-dimensional battery model; Step S13: By combining different operating condition parameters, multiple sets of aging state data are generated. Each set includes: time series voltage data. Capacity decay curve Voltage-capacity curve; Step S14: Perform numerical differentiation on the voltage-capacity curve and calculate the IC feature parameter labels for subsequent feature extraction network training. Specifically, this includes: peak voltage. Peak area Half-peak width .

3. The method for constructing a battery degradation model based on a Bayesian physical information neural network according to claim 1, characterized in that, In step S2, a feature extraction network is designed to extract IC feature parameters from the voltage-capacity curve. The specific process is as follows: Step S21: Perform data cleaning and interpolation on the voltage-capacity curve to obtain a uniformly spaced voltage V and capacity Q sequence; Step S22: Construct a feature extraction network, which specifically includes an input layer, a 1D convolutional layer, a max pooling layer, a physically constrained layer, and a fully connected layer; Step S23: Input the voltage-capacity curve processed in step S21 into the input layer of the feature extraction network; Step S24: Extract local voltage-capacity correlation features from the 1D convolutional layer; Step S25: Dimensionality reduction of the extracted local voltage-capacity correlation features is performed using a max pooling layer while retaining salient features; Step S26: In the physical constraint layer, the mathematical characteristics of the network output IC curve are constrained by designing a loss function; Step S27: Directly regress IC feature parameters through a fully connected layer. , , .

4. The method for constructing a battery degradation model based on a Bayesian physical information neural network according to claim 3, characterized in that, In step S26, the loss function is designed, specifically including: Step S261, Basic Loss The difference between the predicted IC parameters and the tag data is calculated using the mean square error (MSE). Step S262, Physical Constraints By using the Savitsky-Golay differential operator, a differential relation constraint is added to the loss function to ensure that the mathematical properties of the intermediate features of the network are consistent with those of the IC curve, as shown below: ； in, The capacity sequence predicted by the network; This is the actual capacity sequence, i.e., the label data; For Savitzky-Golay differentiation operations; This represents the total number of data points. Step S263, Total Loss Function As shown below: ； in, This is a hyperparameter that balances the weights of the mean squared error loss and the physical constraint term.

5. The method for constructing a battery degradation model based on a Bayesian physical information neural network according to claim 1, characterized in that, In step S3, an aging parameter mapping network is constructed to characterize the relationship between IC feature parameters and battery aging parameters. The specific process is as follows: Step S31: Determine the relationship between IC characteristic parameters and battery aging parameters; Step S311: Battery aging reduces the exchange current density drop, overpotential The rise eventually led to Offset; Step S312, SEI growth or particle breakage reduces the concentration of active material The decline directly led to reduce; Step S313: Aging leads to pore blockage, which reduces the diffusion coefficient. The decline led to rise; Step S32: Use a multi-layer MLP to construct an aging parameter mapping network, whose structure includes an input layer, a hidden layer, and an output layer; Step S33: Input IC feature parameters, and based on the relationship between IC feature parameters and battery aging parameters, add constraints to the mapping network to achieve the mapping output of battery aging parameters.

6. The method for constructing a battery degradation model based on a Bayesian physical information neural network according to claim 5, characterized in that, In step S33, residual constraints are constructed using the Butler-Volmer equation. As shown below: ； in, For exchange current density; and It is a symmetry factor; It is Faraday's constant; This is an overpotential; It is the ideal gas constant; Absolute temperature; Tag current density; It is an L2 norm; Total loss function As shown below: ； in, For residual constraints The weighting coefficients.

7. The method for constructing a battery degradation model based on a Bayesian physical information neural network according to claim 1, characterized in that, In step S4, a Bayesian battery health state inference network is constructed by randomizing the parameters. Its input is the battery aging parameters, and its output is the probability distribution of the battery's state of health (SOH). The specific process is as follows: Step S41: Define Bayesian network node parameters using Gaussian distribution. The prior distribution is as follows: ； ； in, These are the weights of the network node parameters; It is the bias of network node parameters; Indicates variance; Represents weight parameters The prior probability density function; Represents weight parameters The prior probability density function; Indicates a Gaussian distribution; Step S42: Assume the noise in the data follows a normal distribution. Choose a Gaussian distribution and define the likelihood function. The likelihood function represents the probability of observing correct data given the network parameters, as shown below: ； in, This represents a dataset of aging parameters; Represents the likelihood function; This represents the true value of SOH; This represents the model's predicted value SOH; Step S43: Input battery aging parameters and output the probability distribution of battery health status. , ), used to quantify the uncertainty of forecasts; among which, This represents the mean. It represents the standard deviation.

8. The system for constructing a battery degradation model based on a Bayesian physical information neural network according to any one of claims 1-7, characterized in that, It includes a pseudo-two-dimensional battery data generation module, a feature extraction network module, a battery aging parameter mapping module, and a Bayesian battery health status inference module; Among them, the pseudo-two-dimensional battery data generation module is used to generate training data for the battery degradation model; The feature extraction network module is used to extract IC feature parameters from the voltage-capacity curve; The battery aging parameter mapping module is used to explicitly map aging parameters. At the same time, it uses the physical partial differential equations in the battery degradation process to construct physical constraint loss and incorporate it into the model training process, so that the battery degradation model conforms to physical laws and has physical interpretability. The Bayesian battery health status inference module is used for randomized parameter modeling of the model. It utilizes battery aging parameters and prior knowledge to construct and realize battery health inference and model uncertainty quantification.

9. A computer device, characterized in that: It includes a memory and a processor; the memory stores a computer program, and the processor executes the computer program to implement the steps of the method as described in any one of claims 1-7.

Images