A method for identifying battery degradation parameters and estimating battery health status

Abstract

This invention provides a method for identifying battery degradation parameters and estimating battery health status, relating to the field of battery health management technology. This invention constructs an explicit physical correlation between the battery's voltage dynamic response and degradation parameters through an electrochemical model; it significantly improves computational efficiency by replacing the complex calculation process in the electrochemical model with a surrogate model based on a neural network; it achieves identification of battery micro-degradation parameters through self-supervised learning of voltage reconstruction based on physical mechanisms; and it estimates battery state of health (SOH) based on these mechanism-transparent micro-degradation parameters, significantly improving the interpretability and accuracy of battery SOH estimation.

Description

Technical Field

[0001] This invention relates to the field of battery health management technology, and in particular to a method for identifying battery degradation parameters and estimating battery health status. Background Technology

[0002] Lithium-ion batteries have been widely adopted in new energy storage, electric vehicles, portable electronics, and medical equipment due to their high energy density, long cycle life, and lightweight characteristics. However, batteries undergo continuous and irreversible physicochemical changes during use, including loss of active material (LAM) and loss of lithium-ion inventory (LLI). These degradation mechanisms directly lead to changes in key internal parameters of the battery. Microscopically, as the battery ages, key degradation parameters, including the volume fraction of positive and negative electrode active materials and lithium inventory, continuously deteriorate. Macroscopically, as the battery ages, its usable capacity continuously declines, resulting in overall performance degradation. Against this backdrop, accurately identifying microscopic degradation parameters and quantifying the battery's macroscopic state of health (SOH, defined as the difference between the current state and the state of a new battery, usually expressed as a percentage of battery capacity) is crucial for the full lifecycle management of batteries.

[0003] In recent years, breakthroughs in deep learning technology have driven the rapid development of data-driven battery SOH estimation methods. Purely data-driven battery SOH estimation methods estimate SOH by mining the statistical mapping relationship between operating data such as voltage and current and battery capacity decay; however, their black-box nature leads to a lack of model interpretability. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying battery degradation parameters and estimating state of health (SOH), so as to achieve identification of battery micro-degradation parameters and accurate SOH estimation with physical interpretation capabilities.

[0005] In a first aspect, the present invention provides a method for identifying battery degradation parameters and estimating battery health status, comprising:

[0006] Obtain the current voltage-time curve data of the target battery;

[0007] Based on the current voltage-time curve data, the degradation parameters of the target battery are identified using the trained parameter identification model, resulting in the current degradation parameter data of the target battery. Among them, the degradation parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials. The parameter identification model is trained by voltage reconstruction self-supervised learning based on physical mechanism, combining measured voltage-time curve data with a surrogate model based on a neural network. The surrogate model is used to receive the input degradation parameter data and output the corresponding ion concentration data. The surrogate model is trained based on the simulation data output by the electrochemical model of the battery. The electrochemical model is used to construct the explicit physical correlation between the battery voltage dynamic response and the degradation parameters.

[0008] Based on the current attenuation parameter data, the SOH of the target battery is estimated using the trained SOH estimation model to obtain the current SOH value. The SOH estimation model is trained using measured voltage-time curve data and the corresponding true SOH value, combined with the parameter identification model.

[0009] In an optional implementation, the electrochemical model adopts the SPMe model; the battery degradation parameter identification and health status estimation methods also include:

[0010] Based on the battery's fundamental physical parameters, an SPMe model of the battery is established, and simulation data is obtained using the SPMe model.

[0011] Construct a neural network-based agent model and train the agent model using simulation data;

[0012] A parameter identification model with voltage reconstruction error as a self-supervised signal was constructed and trained using measured voltage-time curve data. The parameter identification model includes a parameter estimation network, a trained surrogate model, and an SPMe model physics calculation module. The parameter estimation network receives the input voltage-time curve and outputs estimated decay parameter data. The surrogate model receives the estimated decay parameter data and outputs the corresponding ion concentration data. The SPMe model physics calculation module receives the estimated decay parameter data and the corresponding ion concentration data, and calculates the reconstruction voltage using the electrochemical equations of the SPMe model.

[0013] A SOH estimation model was constructed, and the SOH estimation model was trained by cascading the trained parameter identification model with the SOH estimation model to be trained, using measured voltage-time curve data and corresponding true SOH values.

[0014] In an optional implementation, simulation is performed using the SPMe model to obtain simulation data, including:

[0015] Within the preset range of decay parameters, the SPMe model is used for simulation to generate simulation data covering different aging states, including key decay parameter combinations and corresponding particle surface lithium ion concentration and electrolyte lithium ion concentration.

[0016] In an optional implementation, the proxy model includes two fully connected layers and four output layers. Each fully connected layer is followed by a ReLU activation function. The output layers consist of a fully connected layer, a Sigmoid activation function, and inverse normalization. The four output layers output the lithium ion concentration on the particle surface of the positive and negative electrodes and the lithium ion concentration in the electrolyte of the positive and negative electrodes, respectively.

[0017] During the training of the surrogate model, the error between the output concentration and the actual concentration is used as the loss function, and the weights and biases in the surrogate model are updated through gradient calculation and backpropagation.

[0018] In an optional implementation, the parameter estimation network consists of convolutional layers, pooling layers, fully connected layers, and a ReLU activation function. The convolutional layers are used to extract time-domain features from the voltage-time curve, the pooling layers are used to reduce the dimensionality of the features output by the convolutional layers, and the fully connected layers are used to map the feature vectors output by the pooling layers to the volume fraction and stoichiometric range of the positive and negative electrode active materials.

[0019] During the training of the parameter identification model, the error between the reconstructed voltage and the actual voltage is used as the loss function, and the weights and biases in the parameter identification model are updated through gradient calculation and backpropagation.

[0020] In an optional implementation, the SOH estimation model consists of a fully connected layer, a ReLU activation function, and a Sigmoid activation function; the fully connected layer and the ReLU activation function are used to capture complex nonlinear relationships and map the input decay parameters to SOH values; after the fully connected layer, the Sigmoid activation function restricts the output value of the SOH estimation model to the range of [0,1].

[0021] In an optional implementation, using measured voltage-time curve data and corresponding true SOH values, the SOH estimation model is trained by cascading the trained parameter identification model with the SOH estimation model to be trained, including:

[0022] The trained parameter identification model is cascaded with the SOH estimation model to be trained to obtain a cascaded model.

[0023] The cascaded model is trained using the voltage-time curve from the measured voltage-time curve data as the overall input and the corresponding true SOH value as the expected output. During the training process, the error between the estimated SOH value and the true SOH value is used as the loss function, and the weights and biases of the SOH estimation model are updated through gradient calculation and backpropagation.

[0024] Secondly, the present invention provides a battery degradation parameter identification and health status estimation device, comprising:

[0025] The data acquisition module is used to acquire the current voltage-time curve data of the target battery;

[0026] The decay parameter identification module is used to identify the decay parameters of the target battery based on the current voltage-time curve data and a trained parameter identification model, thereby obtaining the current decay parameter data of the target battery. The decay parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials. The parameter identification model is trained by voltage reconstruction self-supervised learning based on physical mechanisms, using measured voltage-time curve data combined with a surrogate model based on a neural network. The surrogate model receives the input decay parameter data and outputs the corresponding ion concentration data. The surrogate model is trained based on the simulation data output by the battery's electrochemical model. The electrochemical model is used to construct an explicit physical correlation between the battery's voltage dynamic response and the decay parameters.

[0027] The SOH estimation module is used to estimate the SOH of the target battery based on the current decay parameter data and the trained SOH estimation model to obtain the current SOH value. The SOH estimation model is trained using measured voltage-time curve data and the corresponding true SOH value in combination with the parameter identification model.

[0028] Thirdly, the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the battery degradation parameter identification and health status estimation method of any of the foregoing embodiments.

[0029] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs the battery degradation parameter identification and health status estimation method of any of the foregoing embodiments.

[0030] The battery degradation parameter identification and health state estimation method provided by this invention includes: acquiring the current voltage-time curve data of the target battery; identifying the degradation parameters of the target battery using a trained parameter identification model based on the current voltage-time curve data to obtain the current degradation parameter data of the target battery; wherein, the degradation parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials, and the parameter identification model is trained by voltage reconstruction self-supervised learning based on physical mechanism, combining measured voltage-time curve data with a surrogate model based on a neural network, and the surrogate model is used to receive the input degradation parameter data and output the corresponding ion concentration data. The surrogate model is trained based on the simulation data output by the electrochemical model of the battery, and the electrochemical model is used to construct the explicit physical correlation between the battery voltage dynamic response and the degradation parameters; estimating the SOH of the target battery using a trained SOH estimation model based on the current degradation parameter data to obtain the current SOH value; wherein, the SOH estimation model is trained by combining measured voltage-time curve data and the corresponding true SOH value with the parameter identification model. This approach establishes an explicit physical relationship between the battery's voltage dynamic response and degradation parameters through an electrochemical model; it significantly improves computational efficiency by replacing the complex calculation process in the electrochemical model with a surrogate model based on a neural network; it achieves identification of the battery's micro-degradation parameters through self-supervised learning of voltage reconstruction based on physical mechanisms; and it significantly improves the interpretability and accuracy of battery SOH estimation by estimating the battery's SOH based on these mechanism-transparent micro-degradation parameters. Attached Figure Description

[0031] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0032] Figure 1 A flowchart illustrating a battery degradation parameter identification and health status estimation method provided in an embodiment of the present invention;

[0033] Figure 2 This is an overall architecture diagram of a battery degradation parameter identification and health status estimation method provided in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the structure of an SPMe model provided in an embodiment of the present invention;

[0035] Figure 4 This invention provides a comparison result between a reconstructed voltage obtained from a test and the actual voltage, as provided in an embodiment of the invention.

[0036] Figure 5 This invention provides a comparison result between a test-obtained estimated SOH value and an actual SOH value;

[0037] Figure 6 This is a schematic diagram of a battery degradation parameter identification and health status estimation device provided in an embodiment of the present invention;

[0038] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0039] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Capacity decay is essentially the result of the coupling effects of multiple physical mechanisms, such as LLI and LAM. The quantitative parameters of these mechanisms can directly reflect the irreversible evolution of the internal electrochemical processes of the battery. By identifying these parameters online, effective characteristics of the battery's state of decay (SOH) based on physical mechanisms can be obtained, thereby improving the accuracy and interpretability of SOH estimation. Simultaneously, these decay parameters, as microscopic indicators quantifying the internal degradation state of the battery, can overcome the limitations of current battery state assessments that rely on single traditional macroscopic performance parameters (such as capacity and internal resistance). This provides diversified indicator bases for battery state assessment and battery pack consistency assessment in battery health management scenarios, as well as for the screening and reorganization of retired batteries in tiered utilization scenarios. Therefore, integrating data-driven capabilities with the interpretability of physical mechanisms to jointly achieve microscopic decay parameter identification and macroscopic SOH estimation will become a key path to overcome existing technological bottlenecks and achieve precise control of battery state, with significant engineering value for improving the reliability of new energy vehicles and the economics of energy storage power stations.

[0041] Battery models mainly include electrochemical models and equivalent circuit models. Equivalent circuit models use resistance and capacitance to characterize the charge and discharge characteristics of a battery, but their main limitation is the lack of practical physical meaning; changes in the parameters within the model cannot directly reflect the microscopic electrochemical changes inside the battery. Some related literature integrates equivalent circuit models into neural networks in an attempt to optimize the model's lack of interpretability. However, as a simplified equivalent model, the equivalent circuit model cannot actually reflect the essence of battery performance degradation. Therefore, this embodiment of the invention selects an electrochemical model based on the battery reaction mechanism for battery physical modeling.

[0042] Furthermore, battery electrochemical models mainly include pseudo-two-dimensional models, single-particle models, and single-particle electrolyte models. Among them, the pseudo-two-dimensional model can fully describe the lithium-ion diffusion behavior in the solid and liquid phases inside the battery; however, this model involves a large number of partial differential equations, resulting in high computational complexity and generally making it unsuitable for applications requiring rapid calculations. The single-particle model assumes a uniform distribution of electrode potential and electrolyte concentration along the thickness direction, considers only the radial solid-phase diffusion of a single representative particle, and ignores electrolyte dynamics; this simplification reduces the model's computational complexity. However, its voltage simulation results can produce significant errors during high-rate charge and discharge. To ensure modeling accuracy and control computational complexity, this embodiment of the invention uses a single-particle electrolyte model (SPMe) as the physical model foundation for integrating data-driven and physical mechanism methods. The SPMe model, based on the single-particle model, introduces a description of the dynamic changes in electrolyte concentration, solving for the liquid phase potential through analytical methods or simplified approximations. This method effectively controls computational complexity while ensuring modeling accuracy, providing an ideal trade-off between accuracy and efficiency.

[0043] Based on this, embodiments of the present invention provide a method for identifying battery degradation parameters and estimating battery state of health (SOH). This method first utilizes the SPMe model of the battery to construct an explicit physical correlation between the voltage dynamic response and the battery's micro-degradation parameters. Then, a surrogate model based on a neural network is constructed to replace the computational process of solving complex partial differential equations in the SPMe model, significantly improving computational efficiency. Subsequently, self-supervised learning for voltage reconstruction based on physical mechanisms is used to identify the battery's micro-degradation parameters. Finally, accurate estimation of the battery's SOH is achieved based on the battery degradation parameters. This method integrates data-driven approaches with physical mechanisms to achieve identification of battery micro-degradation parameters and estimation of macro-level SOH. It not only overcomes the limitations of traditional battery state assessment relying on a single macro-indicator but also reveals the essential characteristics of the degradation process through deep integration of physical mechanisms, giving the SOH estimation results a robust physical interpretability. This connects the entire chain from micro-mechanisms to system-level health management, providing a new technological path for refined management and control throughout the battery's entire lifecycle.

[0044] To facilitate understanding of this embodiment, a method for identifying battery degradation parameters and estimating health status disclosed in this embodiment of the invention will be described in detail below.

[0045] This invention provides a method for identifying battery degradation parameters and estimating battery health status, which can be executed by an electronic device with data processing capabilities. See also... Figure 1The flowchart shown is a method for identifying battery degradation parameters and estimating health status. The method mainly includes the following steps S110 to S130:

[0046] Step S110: Obtain the current voltage-time curve data of the target battery.

[0047] The target battery described above is one that requires degradation parameter identification and SOH assessment. The target battery type can be, but is not limited to, lithium-ion batteries. In other embodiments, it can also be other types of ion batteries such as sodium-ion, potassium-ion, magnesium-ion, or calcium-ion batteries. This embodiment only uses lithium-ion batteries as an example. Real-time data can be read from the target battery's battery management system or related data acquisition equipment to obtain the current voltage-time curve data of the target battery.

[0048] Step S120: Based on the current voltage-time curve data, the degradation parameters of the target battery are identified using the trained parameter identification model to obtain the current degradation parameter data of the target battery. Among them, the degradation parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials. The parameter identification model is trained by voltage reconstruction self-supervised learning based on physical mechanism, combining the measured voltage-time curve data with a surrogate model based on a neural network. The surrogate model is used to receive the input degradation parameter data and output the corresponding ion concentration data. The surrogate model is trained based on the simulation data output by the electrochemical model of the battery. The electrochemical model is used to construct the explicit physical correlation between the battery voltage dynamic response and the degradation parameters.

[0049] In this embodiment, an electrochemical model corresponding to the battery type of the target battery is pre-constructed. A surrogate model based on a neural network is trained using simulation data obtained from the electrochemical model. Based on the surrogate model and measured voltage-time curve data, a parameter identification model is trained through self-supervised learning of voltage reconstruction based on physical mechanisms. The input to the parameter identification model is the voltage-time curve, and the output is an estimate of the decay parameters. Therefore, the current voltage-time curve data can be input into the trained parameter identification model, and the output of the parameter identification model is the current decay parameter data of the target battery.

[0050] In this embodiment, the volume fraction of the positive and negative electrode active materials is used to quantify the LAM effect, and the stoichiometric ratio range is used to characterize the attenuation caused by LLI. The stoichiometric ratio range can include the positive and negative electrode stoichiometric ratios when the State of Charge (SOC) is 100% (charging cutoff). The stoichiometry of the positive and negative electrodes when SOC is 0% (discharge cutoff) The electrochemical model described above can employ the SPMe model, which can control computational complexity while ensuring modeling accuracy. Taking lithium-ion batteries as an example, the aforementioned ion concentration data can include the lithium-ion concentration on the particle surface and the lithium-ion concentration in the electrolyte. Based on this, the simulation data can include key degradation parameter combinations covering different aging states, along with the corresponding lithium-ion concentrations on the particle surface and in the electrolyte.

[0051] Step S130: Based on the current attenuation parameter data, the SOH of the target battery is estimated using the trained SOH estimation model to obtain the current SOH value; wherein, the SOH estimation model is trained using measured voltage-time curve data and corresponding true SOH values ​​combined with the parameter identification model.

[0052] In this embodiment, an SOH estimation model is constructed with attenuation parameters as input and SOH values ​​as output. After obtaining the trained parameter identification model, the parameter identification model can be cascaded with the SOH estimation model to be trained. The cascaded model is then trained using measured voltage-time curve data and corresponding true SOH values ​​to obtain the trained SOH estimation model. Therefore, the current attenuation parameter data can be input into the trained SOH estimation model, and the output of the SOH estimation model is the current SOH value.

[0053] The battery degradation parameter identification and health state estimation method provided in this invention establishes an explicit physical correlation between the battery voltage dynamic response and degradation parameters through an electrochemical model; it significantly improves computational efficiency by replacing the complex calculation process in the electrochemical model with a surrogate model based on a neural network; it achieves battery micro-degradation parameter identification through voltage reconstruction self-supervised learning based on physical mechanisms; and it estimates battery SOH based on these mechanism-transparent micro-degradation parameters, significantly improving the interpretability and accuracy of battery SOH estimation.

[0054] This invention aims to address the problems of "assessment relying on a single macroscopic indicator" and "lack of interpretability of black-box models" in existing battery management technologies. It proposes a method for identifying battery degradation parameters and estimating state of health (SOH) by integrating data-driven approaches with physical mechanisms. The core of this method is as follows: First, an explicit physical correlation between the battery's voltage dynamic response and key degradation parameters is established using an electrochemical model. Then, a surrogate model based on a neural network is constructed to replace the computational process of solving complex partial differential equations in the electrochemical model, significantly improving computational efficiency. Subsequently, voltage reconstruction error is used as a self-supervised signal to achieve identification of microscopic battery degradation parameters without additional annotation. Finally, SOH estimation is performed based on the identified degradation parameters, improving the accuracy and interpretability of SOH estimation. This method can output diverse degradation quantification indicators that are mechanistically transparent and cover both macroscopic and microscopic levels, providing a reliable decision-making basis for fault warning, strategy optimization, and tiered utilization throughout the battery's entire lifecycle.

[0055] The training process of the above-mentioned related models will be introduced below, taking the SPMe model as an example of an electrochemical model.

[0056] The battery degradation parameter identification and health status estimation method provided in this embodiment of the invention further includes the following steps S01 to S04:

[0057] Step S01: Based on the basic physical parameters of the battery, establish the SPMe model of the battery, and use the SPMe model to perform simulation to obtain simulation data.

[0058] The SPMe model can be used to perform simulations within a preset range of attenuation parameters, generating simulation data covering key attenuation parameter combinations and corresponding particle surface lithium ion concentration and electrolyte lithium ion concentration for different aging states.

[0059] Step S02: Construct a neural network-based agent model and train the agent model using simulation data.

[0060] In some possible embodiments, the above-described proxy model may include two fully connected layers and four output layers. Each fully connected layer is followed by a ReLU activation function. The output layers consist of a fully connected layer, a Sigmoid activation function, and inverse normalization. The four output layers output the lithium ion concentration on the particle surface of the positive and negative electrodes and the lithium ion concentration in the electrolyte of the positive and negative electrodes, respectively.

[0061] During the training of the surrogate model, the error between the output concentration and the actual concentration is used as the loss function, and the weights and biases in the surrogate model are updated through gradient calculation and backpropagation.

[0062] Step S03: Construct a parameter identification model with voltage reconstruction error as a self-supervised signal, and train the parameter identification model using measured voltage-time curve data; wherein, the parameter identification model includes a parameter estimation network, a trained surrogate model, and an SPMe model physics calculation module; the parameter estimation network is used to receive the input voltage-time curve and output the estimated decay parameter data; the surrogate model is used to receive the estimated decay parameter data and output the corresponding ion concentration data; the SPMe model physics calculation module is used to receive the estimated decay parameter data and the corresponding ion concentration data, and calculate the reconstruction voltage using the electrochemical equation of the SPMe model.

[0063] In some possible embodiments, the parameter estimation network described above can consist of convolutional layers, pooling layers, fully connected layers, and ReLU activation functions. The convolutional layers are used to extract time-domain features from the voltage-time curve, the pooling layers are used to reduce the dimensionality of the features output by the convolutional layers, and the fully connected layers are used to map the feature vectors output by the pooling layers to the volume fraction and stoichiometric range of the positive and negative electrode active materials.

[0064] During the training of the parameter identification model, the error between the reconstructed voltage and the actual voltage can be used as the loss function, and the weights and biases in the parameter identification model can be updated through gradient calculation and backpropagation.

[0065] Step S04: Construct a SOH estimation model and train the SOH estimation model by cascading the trained parameter identification model with the SOH estimation model to be trained using measured voltage-time curve data and corresponding true SOH values.

[0066] In some possible embodiments, the above SOH estimation model can be composed of a fully connected layer, a ReLU activation function, and a Sigmoid activation function; the fully connected layer and the ReLU activation function are used to capture complex nonlinear relationships and map the input decay parameters to SOH values; after the fully connected layer, the Sigmoid activation function is used to restrict the output value of the SOH estimation model to the range of [0,1].

[0067] When training the SOH estimation model, the trained parameter identification model can be cascaded with the SOH estimation model to be trained to obtain a cascaded model. The voltage-time curve in the measured voltage-time curve data is used as the overall input, and the corresponding true SOH value is used as the expected output to train the cascaded model. During the training process, the error between the estimated SOH value and the true SOH value is used as the loss function, and the weights and biases of the SOH estimation model are updated through gradient calculation and backpropagation.

[0068] The core of this invention lies in using the battery's voltage-time curve as input, and by fusing data-driven methods and electrochemical physics models, with voltage reconstruction as the core bridge, to achieve the identification of internal microscopic parameters and the estimation of macroscopic state of harmonic equilibrium (SOH) of the battery. Specifically, this includes the following steps:

[0069] Step 1, Electrochemical Physics Modeling and Correlation Construction:

[0070] First, based on the analysis of lithium-ion battery degradation mechanisms, the volume fraction and stoichiometric range of the positive and negative electrode active materials were selected as key characteristic parameters for characterizing battery degradation. Then, the fundamental physical parameters of the battery were obtained, and an SPMe model of the battery was constructed based on fundamental electrochemical equations (including the Butler-Volmer kinetic equation, mass conservation, and charge conservation). Using this electrochemical model, an explicit physical correlation between the voltage dynamic response and the key battery degradation parameters was established.

[0071] Step 2, Building an efficient proxy model:

[0072] First, a surrogate model based on a neural network is constructed. Then, within a preset range of decay parameter variations, the constructed SPMe model is used for extensive simulations to generate key decay parameter combinations covering different aging states, along with corresponding lithium-ion concentrations on the particle surface and in the electrolyte. Finally, the surrogate model is trained based on the generated high-fidelity simulation dataset. During training, the error between the output concentration and the actual concentration is used as the loss function, and the weights and biases in the surrogate model are updated through gradient calculation and backpropagation. The training objective is to enable the surrogate model to directly and quickly output the corresponding concentration using the decay parameters as input, replacing the computational process of solving complex partial differential equations in the SPMe model, thus significantly improving computational efficiency.

[0073] Step 3, Identification of attenuation parameters based on voltage reconstruction self-supervised learning:

[0074] First, a parameter identification model with voltage reconstruction error as a self-supervised signal is constructed. Its core components include: (1) Parameter estimation network: taking the battery voltage-time curve as input, outputting the estimated value of key decay parameters; (2) surrogate model trained in step two: receiving the decay parameters output by the parameter estimation network as input, and outputting the corresponding lithium ion concentration on the particle surface and the lithium ion concentration in the electrolyte; (3) SPMe model physical calculation module: receiving the decay parameters output by the parameter estimation network and the lithium ion concentration output by the surrogate model as input, and calculating the reconstruction voltage based on the explicit physical correlation between the voltage dynamic response and the key battery decay parameters. Then, the collected measured voltage-time curve is input into the parameter identification model for model training. The self-supervised training objective of the parameter identification model is to minimize the error between the reconstruction voltage and the measured voltage. By using only the voltage signal, the weights and biases of the parameter estimation network can be optimized to achieve unlabeled identification of micro-degradation parameters through gradient calculation and backpropagation.

[0075] Step 4, SOH estimation based on battery degradation parameters:

[0076] The output of the parameter estimation network trained in step three, i.e., the identified micro-degradation parameters, is used as an effective feature characterizing the battery's state of equilibrium (SOH). First, an SOH estimation model is constructed, receiving the identified degradation parameters as input. Then, the parameter estimation network trained in step three is cascaded with the SOH estimation model to be trained. Subsequently, the cascaded model is trained using a dataset containing measured voltage-time curves and corresponding SOH values. The training objective is to minimize the error between the estimated SOH value and the true SOH value. Through gradient calculation and backpropagation, an integrated end-to-end model of "parameter identification—SOH estimation" is finally obtained.

[0077] Once the above model construction and training are completed, the resulting integrated end-to-end model of "parameter identification - SOH estimation" can be used to achieve online identification of battery degradation parameters and estimation of health status.

[0078] The overall architecture of this invention embodiment is as follows: Figure 2 As shown below, this method is described in detail from four aspects: electrochemical physics modeling and correlation construction, construction of efficient surrogate models, identification of decay parameters based on voltage reconstruction self-supervised learning, and SOH estimation based on battery decay parameters.

[0079] I. Electrochemical Physics Modeling and Correlation Construction:

[0080] First, the key characteristic parameters characterizing battery degradation are analyzed and determined. During the long-term operation of lithium-ion batteries, performance degradation mainly originates from the LAM and LLI of the positive and negative electrodes. This invention aims to quantify these degradation mechanisms to provide microscopic indicators for battery state assessment, overcoming the limitations of traditional battery state assessment relying on a single macroscopic indicator, and providing a solid physical basis for the macroscopic assessment of battery state of equilibrium (SOH).

[0081] The LAM mechanism is mainly manifested in the reduction of active materials in the positive and negative electrodes within the battery. This is typically caused by factors such as particle breakage, electrode structure collapse, or side reactions with the electrolyte. From an electrochemical modeling perspective, the reduction in active material directly leads to a decrease in the effective volume within the electrode available for lithium ion insertion / extraction, thereby reducing the electrode capacity. Therefore, this embodiment selects the volume fraction of active materials in the positive and negative electrodes. Used as a core feature parameter to quantify the LAM effect.

[0082] The core of the LLI mechanism is the reduction in the total number of lithium ions available for insertion / extraction at the positive and negative electrodes. This mainly stems from the side reactions between lithium ions and the electrolyte to form a solid electrolyte interface film, or from irreversible reactions such as lithium metal deposition, which permanently consume lithium ions. In the electrochemical model, the total amount of lithium ions directly determines the range of change in the stoichiometry (the ratio of the actual lithium ion concentration to the maximum lithium concentration at the electrode) of the positive and negative electrode materials during charge and discharge. When LLI occurs, due to the reduction in the total number of cyclic lithium ions, the dynamic range of the stoichiometry between the positive and negative electrodes will shift and narrow during battery charge and discharge. Therefore, this embodiment uses the stoichiometry of the positive and negative electrodes at a SOC of 100% (charge cutoff). The stoichiometry of the positive and negative electrodes when SOC is 0% (discharge cutoff) It is used as a key parameter to characterize the attenuation caused by LLI.

[0083] In summary, this embodiment selects the volume fraction of positive and negative electrode active materials. and stoichiometric range and As a key characteristic parameter for characterizing battery degradation.

[0084] Then, the basic physical parameters of the battery are obtained to build the SPMe model of the battery. The required parameters are shown in Table 1.

[0085] Table 1

[0086]

[0087] The following section uses the battery SPMe model to construct the voltage dynamic response and key battery degradation parameters. The explicit physical relationship between them.

[0088] like Figure 3 As shown, the SPMe model includes a current collector, and negative and positive electrodes located within the current collector, with a separator between the negative and positive electrodes. The SPMe model assumes that the positive and negative electrodes of the battery are each composed of a single spherical particle. According to Fick's diffusion law, the lithium-ion solid-phase diffusion process can be represented as:

[0089] (1);

[0090] Its boundary conditions are:

[0091] (2);

[0092] (3);

[0093] in, It refers to the lithium-ion concentration at the positive and negative electrodes. It is the electron diffusion coefficient between the positive and negative electrodes. Indicates the radius of the positive and negative pole particles. , This represents the current density at the positive and negative electrodes. It is the reaction area per unit volume of the positive and negative electrodes. F It is Faraday's constant.

[0094] Based on the Butler-Volmer equation, the current density of the positive and negative electrodes It can be represented as:

[0095] (4);

[0096] Then overpotential It can be represented as:

[0097] (5);

[0098] in, The current density exchanged between the positive and negative electrodes can be expressed as:

[0099] (6);

[0100] In the above formula, R and T These are the gas constant and temperature, respectively. and These are the charge transfer coefficients of the anode and cathode, respectively. This represents the rate constant of the positive and negative electrode reactions. This indicates the lithium-ion concentration in the electrolyte at the positive and negative current collectors. This indicates the maximum lithium concentration at both the positive and negative electrodes. This indicates the lithium ion concentration on the surface of the positive and negative electrode particles.

[0101] In the SPMe model, assuming the electrode reaction current density is uniformly distributed along the electrode thickness direction, then:

[0102] (7);

[0103] in, A Indicates the cross-sectional area of ​​the electrode. This indicates the thickness of the positive and negative electrodes.

[0104] Furthermore, the SPMe model considers the ion transport dynamics in the electrolyte based on the single-particle model. The lithium-ion diffusion process in the positive and negative electrodes and the membrane electrolyte can be represented as follows:

[0105] (8);

[0106] (9);

[0107] Its boundary conditions are

[0108] (10);

[0109] (11);

[0110] (12);

[0111] (13);

[0112] (14);

[0113] in, It is the thickness of the diaphragm. It is the lithium-ion transference number. and These are the porosity of the positive and negative electrodes and the membrane, respectively. and These represent the lithium-ion concentrations at the positive and negative electrodes and the electrolyte in the separator, respectively. The effective ion diffusion coefficients of the positive and negative electrodes and the diaphragm electrolyte are given by the Bruggeman relation. Brugg This represents the Bruggeman coefficient. The diffusion of lithium ions in the electrolyte is affected by porosity; therefore, the Bruggeman relation, an empirical equation, is used to adjust the diffusion coefficient.

[0114] In the SPMe model, based on empirical assumptions and the conservation of liquid phase charge, the liquid phase potential difference can be obtained as follows:

[0115] (15);

[0116] in, The liquid phase potential, This is the effective electrolyte conductivity. In the SPMe model, it is assumed that... = It is a constant.

[0117] Therefore, the battery terminal voltage can be expressed as:

[0118] (16);

[0119] in and These represent the open-circuit voltages of the positive and negative terminals, respectively. and These represent the stoichiometry of the positive and negative electrode particle surfaces, respectively. and .

[0120] In summary, based on the battery-based SPMe model, the voltage dynamic response was constructed using equations (1)-(16). With key battery degradation parameters The explicit physical relationship between them.

[0121] II. Construction of an efficient proxy model:

[0122] The computational difficulty of the SPMe model mentioned above mainly lies in solving the partial differential equations (equations (1)-(3) and (8)-(14)) describing lithium-ion transport in the solid and liquid phases, in order to obtain the lithium-ion concentration on the particle surface. and the concentration of lithium ions in the electrolyte The accurate numerical solution of these partial differential equations typically requires significant computational resources, especially in scenarios involving numerous iterations or real-time parameter identification, which can become a significant bottleneck. Traditional order reduction methods (such as simplifying partial differential equations to ordinary differential equations) can improve computational speed, but often sacrifice model accuracy. To significantly reduce computation time while maintaining model accuracy, this embodiment proposes an innovative solution: introducing a data-driven neural network as a surrogate model. The core idea of ​​this surrogate model is to use offline simulation data for training to establish a direct nonlinear mapping from microscopic decay parameters to key concentrations.

[0123] The proxy model in this embodiment, namely Figure 2Model 1 in the model consists of fully connected layers, ReLU activation functions, and Sigmoid activation functions. To eliminate the influence of differences in the units and numerical ranges of the input parameters, all parameters are normalized before being input into the surrogate model. This helps improve the training efficiency and convergence stability of the model, preventing some parameters from dominating gradient calculations due to excessively large values. The normalized parameters are input into one or more fully connected layers to capture the complex interactions between the input parameters. Each fully connected layer is followed by a ReLU activation function to introduce necessary nonlinearity into the model. This allows the surrogate model to learn and fit the nonlinear mapping from micro-decay parameters to complex concentrations. Finally, the features extracted through the fully connected layers and ReLU activation functions are input into four output layers consisting of fully connected layers, Sigmoid activation functions, and inverse normalization, outputting four estimated concentrations. , , and .

[0124] Unlike traditional regression tasks, the lithium-ion concentration variation inside a battery has a defined physical boundary. To ensure that the output of the surrogate model always remains within its reasonable physical range, this embodiment employs range constraint processing at the model's output layer. First, a Sigmoid activation function is applied after the fully connected layer of the output layer to compress the original output values ​​to the range [0,1]. Subsequently, the max-min inverse normalization method shown in Equation (17) is used to further compress the output values ​​within the range [0,1]. Map back to its valid range This method fundamentally constrains the model's output within its physical range, greatly enhancing the physical plausibility and robustness of the prediction results.

[0125] (17);

[0126] in, This represents the final output concentration.

[0127] To establish the mapping relationship between parameters and concentrations in the surrogate model, large-scale simulations were performed using the SPMe model in the offline phase to generate key decay parameter combinations covering different aging states, as well as corresponding data on lithium ion concentrations on particle surfaces and in the electrolyte.

[0128] Subsequently, the surrogate model was trained using the aforementioned simulation data. During training, the error between the output concentration and the actual concentration was used as the loss function. The weights and biases in the surrogate model were updated through gradient calculation and backpropagation until the required training epochs were reached, completing the model training. The training objective was to enable the surrogate model to directly and quickly output the corresponding concentration using the decay parameter as input, replacing the computational process of solving complex partial differential equations in the SPMe model, thus significantly improving computational efficiency.

[0129] III. Identification of attenuation parameters based on voltage reconstruction self-supervised learning:

[0130] One of the core technologies of this invention is a decay parameter identification model based on voltage reconstruction self-supervised learning, used to identify the micro-degradation parameters of the battery from the input voltage-time curve. The design of this model ingeniously combines a data-driven method with an electrochemical physics model, and its core components include: (1) a parameter estimation network: using voltage-time curves As input, extract the feature information related to battery degradation and output the estimated micro-degradation parameters, namely the volume fraction and stoichiometric range of the positive and negative electrode active materials. (2) Pre-trained surrogate model: Receive the estimated micro-degradation parameters as input and output the corresponding lithium ion concentration on the particle surface and the lithium ion concentration in the electrolyte. (3) SPMe model physical calculation module: Receive the estimated micro-degradation parameters and the concentration output by the surrogate model as input, use the electrochemical equation of the SPMe model, calculate the positive and negative electrode open circuit voltage based on the expression of the positive and negative electrode open circuit voltage, calculate the positive and negative electrode overpotential based on equation (5)-(7), calculate the liquid phase potential difference based on equation (15), and finally obtain the reconstructed voltage based on equation (16). .

[0131] The parameter estimation network in this embodiment, namely Figure 2 Model 2, designed specifically for voltage time-series data, consists of convolutional layers, pooling layers, fully connected layers, and a ReLU activation function. The convolutional layers extract time-domain features from the voltage-time curve, effectively capturing its local dynamics. These features include the slope of the voltage curve, inflection points, and plateau length, all closely related to the battery's internal degradation state. Pooling layers reduce the dimensionality of the features output by the convolutional layers, decreasing the model's computational cost and number of parameters. Through downsampling, pooling layers retain the most important feature information while enhancing the model's robustness, thus avoiding overfitting. The fully connected layers, located at the model's output, take the flattened pooling layer output feature vector as input and map these abstract features to the final output, namely key degradation parameters such as the volume fraction of positive and negative electrode active materials and the stoichiometric range. Both the convolutional and fully connected layers are followed by the ReLU activation function, introducing necessary nonlinearity into the model.

[0132] After completing the model construction, the acquired voltage-time curves are used to train the model. During training, the error between the reconstructed voltage and the actual voltage is used as the loss function. Gradient calculation and backpropagation are used to iteratively update the weights and biases in the parameter identification model until the required training epochs are reached, completing the model training. This self-supervised training process guides the parameter identification model to learn and identify accurate micro-degradation parameters from the input voltage curves, thereby minimizing the error between the reconstructed voltage and the actual voltage. This process requires no additional labeled data, achieving efficient and robust identification of battery micro-degradation parameters.

[0133] IV. SOH estimation based on battery degradation parameters:

[0134] The identified volume fractions and stoichiometric ratios of the positive and negative electrode active materials directly quantify the LAM and LLI mechanisms leading to battery performance degradation. Therefore, these microscopic degradation parameters have a strong physical correlation with macroscopic SOH, enabling accurate mapping of the battery's health status. Thus, in this embodiment, the identified microscopic degradation parameters are used as effective characteristics of battery SOH to construct an SOH estimation model based on these parameters, thereby improving the accuracy and interpretability of battery SOH estimation.

[0135] Battery SOH estimation model, i.e. Figure 2 Model 3 in this paper consists of a fully connected layer, a ReLU activation function, and a Sigmoid activation function. The fully connected layer and the ReLU activation function are used to capture complex nonlinear relationships, mapping the input battery decay parameters to the SOH value. This architecture enables the model to learn and fit the complex functional relationship between microscopic parameters and macroscopic SOH. Furthermore, to ensure the reasonableness of the SOH, this embodiment uses a Sigmoid activation function after the fully connected layer in the model's output layer, limiting the model's output value to the range [0,1]. This processing ensures that the estimated SOH value is always within the effective range of 0% to 100%, thus guaranteeing the physical reasonableness of the results.

[0136] After completing the model construction, the model is trained using the collected voltage-time curves and corresponding SOH data. First, the parameter identification model obtained through self-supervised training is cascaded with the SOH estimation model to be trained. Then, the cascaded model is trained using the collected voltage-time curves as the overall input and the corresponding SOH as the expected output. During training, the error between the estimated SOH and the true SOH value is used as the loss function. Gradient calculation and backpropagation are used to iteratively update the weights and biases of the SOH estimation model, aiming to learn the accurate mapping relationship from the voltage curve to microscopic parameters and then to macroscopic SOH, until the required training rounds are reached, completing model training. Through cascaded supervised training, the final model can efficiently and accurately estimate the battery SOH directly from the input voltage curve, and this process has a strong physical basis.

[0137] The embodiments of the present invention have the following beneficial effects:

[0138] This invention proposes a method for identifying battery degradation parameters and estimating State of Health (SOH) by integrating data-driven approaches with physical mechanisms. The core of this method lies in: first, establishing an explicit physical correlation between the battery's voltage dynamic response and key degradation parameters using a battery SPMe model; then, constructing a neural network-based surrogate model to replace the computational process of solving complex partial differential equations in the electrochemical model, significantly improving computational efficiency; subsequently, using voltage reconstruction error as a self-supervised signal, achieving battery degradation parameter identification without additional annotation. These parameters, as microscopic indicators quantifying the internal degradation state of the battery, overcome the limitations of traditional assessments relying on single macroscopic parameters (such as capacity and internal resistance), providing diverse indicators for battery health management. Finally, battery SOH estimation is performed based on these mechanistically transparent microscopic degradation parameters, significantly improving the interpretability and accuracy of battery SOH estimation. The method provided by this invention can output mechanistically transparent, diversified degradation quantification indicators covering both macroscopic and microscopic levels, providing a new technical path for refined management of the entire battery lifecycle and possessing significant engineering application value.

[0139] The above method will be further explained below with reference to specific embodiments.

[0140] Step 1, Electrochemical Physics Modeling and Correlation Construction:

[0141] First, based on the analysis of the lithium-ion battery degradation mechanism, the volume fraction and stoichiometric ratio range of the positive and negative electrode active materials are selected as key characteristic parameters to characterize the battery aging state.

[0142] Next, the basic physical parameters of the battery are obtained. In this example, a LiFePO4 (LFP) battery with a rated capacity of 1.1 Ah and charge / discharge cutoff voltages of 3.6 V and 2 V are analyzed. Based on references, its basic physical parameters are shown in Table 2.

[0143] Table 2

[0144]

[0145] The electrolyte ion diffusion coefficient and electrolyte conductivity are shown below:

[0146] (18);

[0147] (19);

[0148] The expressions for the open-circuit voltages of the positive and negative terminals are as follows:

[0149] (20);

[0150] (twenty one);

[0151] Subsequently, based on the fundamental electrochemical equations (including the Butler-Volmer kinetic equation, mass conservation, and charge conservation), an SPMe model of the battery was established. Using this electrochemical model, an explicit physical correlation between the voltage dynamic response and the key battery degradation parameters, as shown in equations (1)-(16), was constructed.

[0152] Step 2, Building an efficient proxy model:

[0153] First, a surrogate model based on a neural network is constructed. In this embodiment, the surrogate model consists of fully connected layers, ReLU activation functions, and Sigmoid activation functions. To eliminate the influence of differences in the dimensions and numerical ranges of the input parameters, all parameters are normalized before being input into the surrogate model. This helps improve the training efficiency and convergence stability of the model, preventing certain parameters from dominating gradient calculations due to excessively large values. In this embodiment, the normalized parameters are input into two fully connected layers with 64 and 128 neurons respectively to capture the complex interactions between the input parameters. Each fully connected layer is followed by a ReLU activation function to introduce necessary nonlinearity into the model. This allows the surrogate model to learn and fit the nonlinear mapping from micro-decay parameters to complex concentrations. Finally, the features extracted through the fully connected layers and ReLU activation functions are input into four output layers consisting of fully connected layers, Sigmoid activation functions, and inverse normalization, outputting four estimated concentrations. , , and The number of neurons in the fully connected layer of the output layer is the number of time steps, which is set to 200 in this embodiment.

[0154] Unlike traditional regression tasks, the lithium-ion concentration variation inside a battery has a defined physical boundary. To ensure that the output of the surrogate model always remains within its reasonable physical range, this embodiment employs range constraint processing at the model's output layer. First, a sigmoid activation function is applied after the fully connected layer of the output layer to compress the model's original output values ​​to the range [0,1]. Subsequently, the max-min inverse normalization method is used to further compress the output values ​​within the [0,1] range. Mapping back to its physical effective range Internally. In this embodiment, based on the battery's basic physical parameters, the following settings are configured. The physical effective range is , The physical effective range is , The physical effective range is [1100, 1300].

[0155] Table 3

[0156]

[0157] The preset range of micro-degradation parameters is shown in Table 3, which is used to generate key degradation parameter combinations covering different aging states and corresponding data on lithium ion concentration on particle surface and lithium ion concentration in electrolyte.

[0158] Because this embodiment considers the input battery discharge voltage-time curve, the initial stoichiometric ratio of discharge is selected, i.e. The range of variation. When using the SPMe model for battery discharge simulation, input the set discharge initiation stoichiometry. During simulation, the stoichiometric ratio of the positive and negative electrodes when the terminal voltage reaches the discharge cutoff voltage is the corresponding... Therefore, based on the given... The attenuation parameter can be obtained by simulating the variation range using the SPMe model. and the corresponding lithium ion concentration on the particle surface and the concentration of lithium ions in the electrolyte .

[0159] Accordingly, if the input battery charging voltage-time curve is considered, then the initial stoichiometric ratio for charging should be set accordingly, i.e. Then, the SPMe model was used to simulate battery charging and generate a dataset.

[0160] Based on the above range of attenuation parameters, a large number of simulations were conducted using the SPMe model at a 4C discharge current and an ambient temperature of 30°C. This generated key attenuation parameter combinations covering different aging states, as well as corresponding data on lithium ion concentration on particle surfaces and lithium ion concentration in the electrolyte.

[0161] Subsequently, a surrogate model is trained based on the generated high-fidelity simulation dataset. During training, the error between the output concentration and the actual concentration is used as the loss function. The weights and biases in the surrogate model are updated through gradient calculation and backpropagation until the required training epochs are reached, completing the surrogate model training. The training objective is to enable the surrogate model to directly and quickly output the corresponding concentration using the decay parameter as input, replacing the computational process of solving complex partial differential equations in the SPMe model, thus significantly improving computational efficiency. In this embodiment, the loss function value at the end of training is only 0.00136, indicating that the trained surrogate model has high accuracy.

[0162] Step 3, Identification of attenuation parameters based on voltage reconstruction self-supervised learning:

[0163] A parameter identification model with voltage reconstruction error as a self-supervised signal is constructed. Its core components include: (1) Parameter estimation network: using voltage-time curve as input, extracting feature information related to battery degradation from it, and outputting estimated micro-degradation parameters; (2) Pre-trained surrogate model: receiving the estimated micro-degradation parameters as input, and outputting the corresponding particle surface lithium ion concentration and electrolyte lithium ion concentration; (3) SPMe model physical calculation module: receiving the estimated micro-degradation parameters and the lithium concentration output by the surrogate model as input, using the electrochemical equation of the SPMe model, calculating the positive and negative electrode open circuit voltage based on the positive and negative electrode open circuit voltage expression, calculating the positive and negative electrode overpotential based on equations (5)-(7), calculating the liquid phase potential difference based on equation (15), and finally obtaining the reconstruction voltage based on equation (16).

[0164] The parameter estimation network in this embodiment is designed specifically for voltage time series data and consists of convolutional layers, pooling layers, fully connected layers, and a ReLU activation function. In this embodiment, the input battery discharge voltage-time curve is sequentially passed through two convolutional layers with kernel sizes of 5 and 3, and channel numbers of 16 and 32, respectively. Each convolutional layer is followed by a ReLU activation function and a pooling layer with a pooling window size and stride of 2. Applying these convolutional and pooling layers effectively captures the local dynamic features of the voltage curve and reduces the number of parameters and suppresses overfitting through pooling operations. Subsequently, the features extracted by convolution and pooling are flattened and input into a network consisting of three fully connected layers for nonlinear mapping. The number of neurons in these three fully connected layers are 256, 128, and 6, respectively, and the ReLU activation function is also used between layers to introduce necessary nonlinearity, thereby enhancing the model's expressive power. Through the above parameter estimation network, the purpose of estimating the battery's micro-degradation parameters based on the input battery discharge voltage-time curve is achieved.

[0165] After constructing the model as described above, it is trained using battery voltage-time curves. During training, the error between the reconstructed voltage and the actual voltage is used as the loss function. Gradient calculation and backpropagation are used to iteratively update the weights and biases in the parameter identification model until the required training epochs are reached, completing model training. This self-supervised training process guides the parameter identification model to learn accurate micro-degradation parameters from the input voltage curve, thereby minimizing the error between the reconstructed voltage and the measured voltage. This process requires no additional labeled data, achieving efficient and robust identification of battery micro-degradation parameters.

[0166] In this embodiment, a battery discharge voltage-time curve generated by simulation using the SPMe model was used for training and testing. The root mean square error (RMSE) between the reconstructed voltage curve and the actual voltage curve was 0.003680. The comparison results of some reconstructed voltages and actual voltages during the test are shown below. Figure 4 As shown. Figure 4 The data shows that the reconstructed voltage is in high agreement with the actual voltage, indicating that the identified micro-attenuation parameters have high accuracy.

[0167] Step 4, SOH estimation based on battery degradation parameters:

[0168] The output of the parameter estimation network trained in step three, i.e., the identified micro-degradation parameters, is used as an effective feature characterizing the battery's SOH. An SOH estimation model is constructed, which takes the identified degradation parameters as input and outputs the battery's SOH estimate.

[0169] The battery SOH estimation model consists of fully connected layers and ReLU and Sigmoid activation functions. The fully connected layers and ReLU activation function capture complex nonlinear relationships, mapping the input battery degradation parameters to SOH values. This architecture allows the model to learn and fit the complex functional relationship between microscopic parameters and macroscopic SOH, rather than a simple linear correlation. In this embodiment, the battery SOH estimation model contains three fully connected layers with 64, 32, and 1 neurons, respectively. The ReLU activation function is used after the first two fully connected layers to introduce necessary nonlinear characteristics into the model. To ensure the physical meaning and reasonableness of the SOH, a Sigmoid activation function is used after the third fully connected layer to limit the model's output value to the range [0,1]. This ensures that the estimated SOH value is always within the effective range of 0% to 100%, thus guaranteeing the physical reasonableness of the results.

[0170] After constructing the model as described above, the model is trained using the voltage-time curve and the corresponding true SOH value. First, the parameter identification model obtained through self-supervised training is cascaded with the SOH estimation model to be trained. Then, the cascaded model is trained using the voltage-time curve as the overall input and the true SOH value as the expected output. During training, the error between the estimated SOH and the true SOH value is used as the loss function. Gradient calculation and backpropagation are used to iteratively update the weights and biases of the SOH estimation model, aiming to learn the accurate mapping relationship from the voltage curve to microscopic parameters and then to macroscopic SOH, until the required training epochs are reached, completing the model training. Through cascaded supervised training, the final model can efficiently and accurately estimate the battery SOH directly from the input voltage curve, and this process has a strong physical basis.

[0171] In this embodiment, the discharge voltage-time curve and corresponding SOH generated by the SPMe model are used as the dataset for training and testing. The test yielded an RMSE of 0.001025 between the estimated SOH and the actual SOH. The comparison results between the estimated SOH and the actual SOH are as follows: Figure 5 As shown. Figure 5 The results show that the estimated SOH is in high agreement with the actual SOH, therefore, this SOH estimation method has high accuracy.

[0172] Furthermore, using the same dataset, we trained and tested a SOH estimation method based solely on convolutional, pooling, and fully connected layers. This method directly extracts features from the input voltage data for SOH estimation without incorporating a physical model. Test results show that the RMSE between the estimated SOH and the actual SOH is 0.007833. Therefore, this embodiment improves the accuracy and interpretability of battery SOH estimation by incorporating physical mechanisms and extracting effective features characterizing battery SOH, namely microscopic decay parameters.

[0173] Corresponding to the above-described method for identifying battery degradation parameters and estimating battery health status, this embodiment of the invention also provides a device for identifying battery degradation parameters and estimating battery health status. See also... Figure 6 The diagram shows a structural schematic of a battery degradation parameter identification and health status estimation device, which includes:

[0174] The data acquisition module 601 is used to acquire the current voltage-time curve data of the target battery;

[0175] The attenuation parameter identification module 602 is used to identify the attenuation parameters of the target battery based on the current voltage-time curve data and the trained parameter identification model, thereby obtaining the current attenuation parameter data of the target battery. The attenuation parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials. The parameter identification model is trained by voltage reconstruction self-supervised learning based on physical mechanisms, combining measured voltage-time curve data with a surrogate model based on a neural network. The surrogate model is used to receive the input attenuation parameter data and output the corresponding ion concentration data. The surrogate model is trained based on the simulation data output by the electrochemical model of the battery. The electrochemical model is used to construct an explicit physical correlation between the battery voltage dynamic response and the attenuation parameters.

[0176] The SOH estimation module 603 is used to estimate the SOH of the target battery based on the current decay parameter data and the trained SOH estimation model to obtain the current SOH value. The SOH estimation model is trained using measured voltage-time curve data and the corresponding true SOH value in combination with the parameter identification model.

[0177] The battery degradation parameter identification and health state estimation device provided in this invention establishes an explicit physical correlation between the battery voltage dynamic response and degradation parameters through an electrochemical model; it significantly improves computational efficiency by replacing the complex calculation process in the electrochemical model with a surrogate model based on a neural network; it achieves battery micro-degradation parameter identification through voltage reconstruction self-supervised learning based on physical mechanisms; and it estimates battery SOH based on these mechanism-transparent micro-degradation parameters, significantly improving the interpretability and accuracy of battery SOH estimation.

[0178] Furthermore, the aforementioned battery degradation parameter identification and health state estimation device also includes a training module. The training module is used to: establish an SPMe model of the battery based on its fundamental physical parameters, and perform simulation using the SPMe model to obtain simulation data; construct a surrogate model based on a neural network, and train the surrogate model using the simulation data; construct a parameter identification model with voltage reconstruction error as a self-supervised signal, and train the parameter identification model using measured voltage-time curve data; wherein, the parameter identification model includes a parameter estimation network, a trained surrogate model, and an SPMe model physical calculation module; the parameter estimation network receives the input voltage-time curve and outputs estimated degradation parameter data; the surrogate model receives the estimated degradation parameter data and outputs the corresponding ion concentration data; the SPMe model physical calculation module receives the estimated degradation parameter data and the corresponding ion concentration data, and calculates the reconstruction voltage using the electrochemical equation of the SPMe model; construct a SOH estimation model, and train the SOH estimation model by cascading the trained parameter identification model with the SOH estimation model to be trained using measured voltage-time curve data and the corresponding true SOH value.

[0179] Furthermore, the aforementioned training module is specifically used to: perform simulation using the SPMe model within a preset range of attenuation parameter variations, and generate simulation data covering key attenuation parameter combinations for different aging states, as well as the corresponding lithium ion concentrations on the particle surface and in the electrolyte.

[0180] Furthermore, the aforementioned surrogate model includes two fully connected layers and four output layers. Each fully connected layer is followed by a ReLU activation function. The output layers consist of a fully connected layer, a Sigmoid activation function, and inverse normalization. The four output layers output the lithium-ion concentration on the particle surface of the positive and negative electrodes and the lithium-ion concentration in the electrolyte of the positive and negative electrodes, respectively. During the training of the surrogate model, the error between the output concentration and the actual concentration is used as the loss function, and the weights and biases in the surrogate model are updated through gradient calculation and backpropagation.

[0181] Furthermore, the parameter estimation network described above consists of convolutional layers, pooling layers, fully connected layers, and a ReLU activation function. The convolutional layers are used to extract time-domain features from the voltage-time curve, the pooling layers are used to reduce the dimensionality of the features output by the convolutional layers, and the fully connected layers are used to map the feature vectors output by the pooling layers to the volume fraction and stoichiometric range of the positive and negative electrode active materials. During the training process of the parameter identification model, the error between the reconstructed voltage and the actual voltage is used as the loss function, and the weights and biases in the parameter identification model are updated through gradient calculation and backpropagation.

[0182] Furthermore, the SOH estimation model described above consists of a fully connected layer, a ReLU activation function, and a Sigmoid activation function. The fully connected layer and the ReLU activation function are used to capture complex nonlinear relationships and map the input decay parameters to SOH values. After the fully connected layer, the Sigmoid activation function is used to restrict the output value of the SOH estimation model to the range of [0,1].

[0183] Furthermore, the training module is also used to: cascade the trained parameter identification model with the SOH estimation model to be trained to obtain a cascaded model; train the cascaded model using the voltage-time curve in the measured voltage-time curve data as the overall input and the corresponding true SOH value as the expected output; wherein, during the training process, the error between the estimated SOH value and the true SOH value is used as the loss function, and the weights and biases of the SOH estimation model are updated through gradient calculation and backpropagation.

[0184] The battery degradation parameter identification and health status estimation device provided in this embodiment has the same implementation principle and technical effect as the aforementioned battery degradation parameter identification and health status estimation method embodiment. For the sake of brevity, any parts not mentioned in the battery degradation parameter identification and health status estimation device embodiment can be referred to the corresponding content in the aforementioned battery degradation parameter identification and health status estimation method embodiment.

[0185] like Figure 7 As shown, an electronic device 700 provided in this embodiment of the invention includes: a processor 701, a memory 702 and a bus. The memory 702 stores a computer program that can run on the processor 701. When the electronic device 700 is running, the processor 701 and the memory 702 communicate through the bus. The processor 701 executes the computer program to implement the above-mentioned battery degradation parameter identification and health status estimation method.

[0186] Specifically, the memory 702 and processor 701 mentioned above can be general-purpose memory and processor, without any specific limitations here.

[0187] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program performs the battery degradation parameter identification and health status estimation method described in the preceding method embodiments. The computer-readable storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), RAM, magnetic disks, or optical disks.

[0188] In all examples shown and described herein, any specific values ​​should be interpreted as merely exemplary and not as limitations; therefore, other examples of exemplary embodiments may have different values.

[0189] 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 the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying battery degradation parameters and estimating battery health status, characterized in that, include: Obtain the current voltage-time curve data of the target battery; Based on the current voltage-time curve data, the degradation parameters of the target battery are identified using the trained parameter identification model to obtain the current degradation parameter data of the target battery. The degradation parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials. The parameter identification model is trained using a surrogate model based on a neural network, combined with measured voltage-time curve data, through self-supervised learning of voltage reconstruction based on physical mechanisms. The surrogate model receives the input degradation parameter data and outputs the corresponding ion concentration data. The surrogate model is trained using simulation data output from the battery's electrochemical model. The electrochemical model is used to construct an explicit physical correlation between the battery's voltage dynamic response and the degradation parameters. Based on the current attenuation parameter data, the SOH of the target battery is estimated using the trained SOH estimation model to obtain the current SOH value; wherein, the SOH estimation model is trained using measured voltage-time curve data and corresponding true SOH values ​​combined with the parameter identification model; The electrochemical model adopts the SPMe model; the battery degradation parameter identification and health status estimation method further includes: Based on the battery's fundamental physical parameters, an SPMe model of the battery is established, and simulation data is obtained using the SPMe model. Construct a neural network-based agent model and train the agent model using the simulation data; A parameter identification model is constructed using voltage reconstruction error as a self-supervised signal, and trained using measured voltage-time curve data. The parameter identification model includes a parameter estimation network, a trained surrogate model, and an SPMe model physics calculation module. The parameter estimation network receives the input voltage-time curve and outputs estimated decay parameter data. The surrogate model receives the estimated decay parameter data and outputs the corresponding ion concentration data. The SPMe model physics calculation module receives the estimated decay parameter data and the corresponding ion concentration data, and calculates the reconstruction voltage using the electrochemical equations of the SPMe model. A SOH estimation model is constructed, and the SOH estimation model is trained by cascading the trained parameter identification model with the SOH estimation model to be trained using measured voltage-time curve data and corresponding true SOH values.

2. The method for identifying battery degradation parameters and estimating health status according to claim 1, characterized in that, The simulation data obtained by using the SPMe model includes: Within the preset range of attenuation parameter variations, the SPMe model is used for simulation to generate simulation data covering different aging states, including key attenuation parameter combinations and corresponding particle surface lithium ion concentration and electrolyte lithium ion concentration.

3. The method for identifying battery degradation parameters and estimating health status according to claim 1, characterized in that, The proxy model includes two fully connected layers and four output layers. Each fully connected layer is followed by a ReLU activation function. The output layers consist of a fully connected layer, a Sigmoid activation function, and inverse normalization. The four output layers output the lithium ion concentration on the particle surface of the positive and negative electrodes and the lithium ion concentration in the electrolyte of the positive and negative electrodes, respectively. During the training of the surrogate model, the error between the output concentration and the actual concentration is used as the loss function, and the weights and biases in the surrogate model are updated through gradient calculation and backpropagation.

4. The method for identifying battery degradation parameters and estimating health status according to claim 1, characterized in that, The parameter estimation network consists of convolutional layers, pooling layers, fully connected layers, and a ReLU activation function. The convolutional layers are used to extract time-domain features from the voltage-time curve, the pooling layers are used to reduce the dimensionality of the features output by the convolutional layers, and the fully connected layers are used to map the feature vectors output by the pooling layers to the volume fraction and stoichiometric range of the positive and negative electrode active materials. During the training process of the parameter identification model, the error between the reconstructed voltage and the actual voltage is used as the loss function, and the weights and biases in the parameter identification model are updated through gradient calculation and backpropagation.

5. The method for identifying battery degradation parameters and estimating health status according to claim 1, characterized in that, The SOH estimation model consists of a fully connected layer, a ReLU activation function, and a Sigmoid activation function. The fully connected layer and the ReLU activation function are used to capture complex nonlinear relationships and map the input decay parameters to SOH values. After the fully connected layer, the Sigmoid activation function restricts the output value of the SOH estimation model to the range of [0,1].

6. The method for identifying battery degradation parameters and estimating health status according to claim 1, characterized in that, The step of training the SOH estimation model by cascading the trained parameter identification model with the SOH estimation model to be trained, using measured voltage-time curve data and corresponding true SOH values, includes: The trained parameter identification model is cascaded with the SOH estimation model to be trained to obtain a cascaded model. The cascaded model is trained using the voltage-time curve from the measured voltage-time curve data as the overall input and the corresponding true SOH value as the expected output. During the training process, the error between the estimated SOH value and the true SOH value is used as the loss function, and the weights and biases of the SOH estimation model are updated through gradient calculation and backpropagation.

7. A device for identifying battery degradation parameters and estimating battery health status, characterized in that, include: The data acquisition module is used to acquire the current voltage-time curve data of the target battery; The decay parameter identification module is used to identify the decay parameters of the target battery based on the current voltage-time curve data and a trained parameter identification model, thereby obtaining the current decay parameter data of the target battery. The decay parameters include the volume fraction and stoichiometric range of the positive and negative electrode active materials. The parameter identification model is trained using a surrogate model based on a neural network, combined with measured voltage-time curve data, through self-supervised learning of voltage reconstruction based on physical mechanisms. The surrogate model receives the input decay parameter data and outputs the corresponding ion concentration data. The surrogate model is trained using simulation data output from the battery's electrochemical model. The electrochemical model is used to construct an explicit physical correlation between the battery's voltage dynamic response and the decay parameters. The SOH estimation module is used to estimate the SOH of the target battery based on the current decay parameter data and the trained SOH estimation model to obtain the current SOH value; wherein, the SOH estimation model is trained using measured voltage-time curve data and the corresponding true SOH value in combination with the parameter identification model; The electrochemical model adopts the SPMe model; the battery degradation parameter identification and health state estimation device further includes a training module, which is used to: establish an SPMe model of the battery based on the battery's basic physical parameters, and perform simulation using the SPMe model to obtain simulation data; construct a surrogate model based on a neural network, and train the surrogate model using the simulation data; construct a parameter identification model with voltage reconstruction error as a self-supervised signal, and train the parameter identification model using measured voltage-time curve data; wherein, the parameter identification model includes a parameter estimation network, the trained surrogate model, and the SPMe model. The model physics calculation module includes: a parameter estimation network that receives the input voltage-time curve and outputs estimated decay parameter data; a surrogate model that receives the estimated decay parameter data and outputs corresponding ion concentration data; an SPMe model physics calculation module that receives the estimated decay parameter data and corresponding ion concentration data, and calculates the reconstructed voltage using the electrochemical equation of the SPMe model; a SOH estimation model is constructed, and the SOH estimation model is trained by cascading the trained parameter identification model with the SOH estimation model to be trained using measured voltage-time curve data and corresponding true SOH values.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the battery degradation parameter identification and health status estimation method according to any one of claims 1-6.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when executed by the processor, performs the battery degradation parameter identification and health status estimation method according to any one of claims 1-6.

Images