A battery state of health estimation method fusing mechanism and data-driven model

Abstract

A battery health pipeline technology field fusion mechanism and data driven model battery health state estimation method, including the following steps: a strong battery response behavior is described, a reduced order electrochemical model is constructed, and a rich training data set is generated; a data driven model for estimating the SOH of the battery is constructed; based on the large number of charge and discharge data segments generated by the reduced order electrochemical model, the source data set is constructed, the adaptive momentum estimation algorithm is used to optimize the free variables in the data driven model, and the data driven model pre-training is completed; after obtaining a small amount of real current, voltage data and SOH label, part of the source data set is mixed with the real data, and the pre-trained data driven model is used for transfer learning to quickly realize the SOH estimation of the target domain. The present application only needs a small part of the real data, and can complete the transfer learning task of the SOH estimation of the target domain, solving the problem of less label data in the actual use scene.

Description

Technical Field

[0001] This invention relates to an assessment method in the field of battery health management technology, and in particular to a battery health state estimation method that uses a fusion mechanism and a data-driven model to complete the target domain SOH estimation with only a very small amount of real data. Background Technology

[0002] As an important electrochemical energy storage device, lithium-ion batteries are widely used in energy storage systems and electric vehicles due to their high energy density, low self-discharge rate, lack of memory effect, and long cycle life. However, during long-term use, a series of aging side reactions inevitably occur inside lithium-ion batteries, leading to irreversible degradation of their energy storage and power characteristics. With the widespread use of lithium-ion batteries in various practical application environments, the aging characteristics of battery systems have become exceptionally complex. Different ambient temperatures, charge / discharge rates, and inconsistencies between batteries all have varying impacts on the battery's state of health (SOH). To ensure the long-term stability and reliability of the system, accurate assessment and long-term tracking of lithium-ion battery SOH is a core issue in current battery management systems. As a typical unobservable, time-varying, and nonlinear system, the internal SOH characteristics of lithium-ion batteries cannot be directly monitored using sensors. Instead, external assessment relies on characteristic data such as current and voltage, combined with specific estimation methods. Existing lithium-ion battery SOH estimation schemes can be mainly divided into two categories: model-based state observation and parameter identification algorithms, and data-driven algorithms. The former primarily simulates a real battery system by establishing a physical model of the battery and describing its voltage response under different current inputs. It then applies state observation algorithms or parameter identification algorithms to obtain optimal estimates of the parameters in the model and establishes a mapping relationship between these estimates and the battery's state of equilibrium (SOH). The latter, however, is based on external characteristic data such as current, voltage, and temperature of the battery. It utilizes sufficient data to autonomously learn the nonlinear relationship between external characteristics and the battery's internal SOH index. Model-based SOH estimation algorithms have strong generalization capabilities and can be applied to various complex charging and discharging operating conditions. However, this method is highly dependent on the accuracy of the battery's physical model. As a complex system, lithium-ion batteries have multiple physical field distributions and various aging mechanisms that are difficult to fully and accurately describe with a single model, thus limiting the overall accuracy of this method. In contrast, data-driven methods avoid establishing precise battery models and directly use machine learning algorithms to extract hidden information from the raw data and map it directly to the battery's SOH index. Although data-driven methods can achieve extremely high estimation performance when there is sufficient training data, the contradiction between the diversity of battery aging mechanisms and charge / discharge operating conditions and the scarcity of existing datasets means that machine learning models trained on partial data can only be adapted to some scenarios or some battery models and operating conditions. It is difficult to guarantee the generalization of the learned feature combinations across multiple battery differences, thus limiting the use of this method in real-world scenarios.

[0003] In the prior art, patent document CN201710226895.4 identifies four degradation characteristic parameters—SEI film resistance, initial lithium-ion concentration of the negative electrode, and diffusion coefficients of the positive and negative electrodes—based on an electrochemical model and intelligent optimization algorithm. It also completes the SOH assessment by combining the degradation characteristic parameter-usable capacity change roadmap. However, the evolution law of the selected characteristic parameters with the decay of usable capacity varies greatly under different aging modes. The roadmap, which relies on the data extracted from laboratory aging experiments, is difficult to characterize the correspondence between battery internal parameters and usable capacity under the coupled influence of multiple aging mechanisms in real-world scenarios. Patent document CN202210393541.X describes a method for extracting characteristic parameters such as maximum lithium-ion concentration, cyclic lithium-ion loss, and active material loss of the battery's positive and negative electrodes using an electrochemical model and IC / DV curves, and mapping the manually extracted features to the SOH index using a backpropagation neural network. However, the confidence level of the maximum lithium-ion concentration parameters extracted by the electrochemical model is directly affected by the model's accuracy, making it difficult to accurately represent the battery's true state. Furthermore, the required IC / DV curves are sensitive to sensor measurement noise and charge / discharge current rates, making them difficult to extract in practical applications. Patent document CN201911038721.0 utilizes a fuzzy neural network to learn the voltage response sequence of a lithium-ion battery under different load currents, temperatures, and SOCs. By simulating a standard capacity test process, it obtains the battery's virtual voltage response curve, thereby acquiring the battery's current capacity. However, the actual vehicle SOC data used is mostly estimated internally by the BMS, making its accuracy difficult to assess and introducing uncontrollable interference to the establishment of the fuzzy neural network. Patent document CN202011604636.9 introduces a method for estimating the state of charge (SOH) by using a multidimensional scaling method to reduce the dimensionality of five feature parameter sequences extracted under charging conditions and establishing a support vector machine regression model. However, while extracting specific feature parameter sequences reflects the SOH state of the battery, the single-rate constant current and constant voltage charging conditions used for feature extraction are not representative of actual usage scenarios. Furthermore, obtaining features such as the constant current charging current cutoff time requires the battery to start charging from a completely discharged state, which places high demands on the operating conditions and is difficult to meet in actual usage scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a battery state of health estimation method that integrates mechanism and data-driven model. This invention comprehensively considers the generalizability of the lithium-ion battery physical model under operating conditions and the estimation accuracy of the data-driven algorithm. By integrating the battery electrochemical mechanism model and the data-driven model, it effectively improves the accuracy and reliability of battery SOH estimation, while reducing the dependence of the data-driven algorithm on the sample size of the dataset.

[0005] This invention is achieved through the following technical solution, comprising the following steps: Step 1, constructing a reduced-order electrochemical model with strong characterization of battery response behavior. Based on the ordered random combination of different feature parameters, it can accurately simulate the current and voltage response data of the battery throughout its entire life cycle under different aging modes, and construct a rich training dataset; Step 2, constructing a data-driven model that can estimate the battery's state of charge (SOH) using fragmented charge-discharge data under random initial charge; Step 3, constructing a source dataset based on a large number of charge-discharge data fragments generated by the reduced-order electrochemical model, and using an adaptive momentum estimation algorithm to optimize the free variables within the data-driven model, thus completing the pre-training of the data-driven model; Step 4, after obtaining a small amount of real current, voltage data and SOH labels, mixing part of the source dataset with real data, and performing transfer learning on the pre-trained data-driven model to quickly achieve SOH estimation for the target domain.

[0006] Furthermore, in this invention, the construction of the reduced-order electrochemical model includes the following steps: Step 1.1, establishing the governing equations and boundary conditions for the lithium-ion concentration distribution and potential distribution in the positive and negative electrode plates, separator, and electrolyte inside the battery, and solving them using methods such as Laplace transform and Pad approximation; Step 1.2, designing the experiment by combining multi-factor orthogonal aging cycle experiments and referencing performance tests; Step 1.3, obtaining the battery parameter change paths under different aging paths based on an intelligent parameter optimization algorithm, adding random perturbations, and constructing a simulation parameter set; Step 1.4, injecting the required operating conditions, solving the constrained nonlinear programming problem, and obtaining simulated charge and discharge data.

[0007] Furthermore, in this invention, the intelligent parameter optimization algorithm is based on the particle swarm optimization algorithm.

[0008] Furthermore, in this invention, the constrained nonlinear programming solution uses the monotonically decreasing current decay rate as a constraint condition, the root mean square error between the simulated voltage during the constant voltage stage and the preset target constant voltage as a cost function, and employs optimization algorithms including the stochastic gradient method to solve the nonlinear normative problem.

[0009] Furthermore, in this invention, the construction of the data-driven model includes the following steps: Step 2.1, using ampere-hour integration and nonlinear interpolation methods to convert the time-series signals of current and voltage of all simulated data into discrete data based on charge-discharge ampere-hours; Step 2.2, randomly extracting charge-discharge segments and processing the data through data normalization and resampling; Step 2.3, constructing the data-driven model; Step 4, optimizing the data-driven model through structural optimization, hyperparameter optimization, and network pruning.

[0010] Furthermore, in this invention, the data-driven model is a SOH data-driven regression model with residual convolutional neural networks as its basic structure.

[0011] Furthermore, in this invention, structural optimization and hyperparameter optimization are performed using a Bayesian optimization algorithm.

[0012] Furthermore, in this invention, the implementation of the particle swarm optimization algorithm includes the following steps: Step 1, initializing all particles in a preset parameter search space; Step 2, calculating the fitness of each particle, i.e., the root mean square error between the simulated predicted voltage and the actual voltage; Step 3, determining the optimal values ​​for individuals and the population; Step 4, updating the velocity and position of each particle; Step 5, repeating steps 2 to 4 until the allowable error threshold or the maximum number of iterations is reached.

[0013] Furthermore, in this invention, a portion of the source dataset is mixed with the real data at a ratio of 25:1.

[0014] The purpose of this invention is to accurately assess the current state of equilibrium (SOH) of a battery using only segmented charge-discharge data with random initial charge levels, through a data-driven model. Furthermore, it utilizes a reduced-order electrochemical model to generate a pre-trained dataset covering different aging modes and charge-discharge conditions, thereby improving the applicability and accuracy of the data-driven model-based SOH estimation method in real-world scenarios, while significantly reducing the dependence of the data-driven method on the amount of real-world data samples.

[0015] To achieve the above objectives, the present invention includes the following components: the establishment of a reduced-order electrochemical model, parameter identification algorithms and simulation condition data generation algorithms, the construction and pre-training of data-driven algorithms such as deep learning, and a transfer learning method that integrates some real data.

[0016] The reduced-order electrochemical model is a mechanistic battery model based on the internal kinetics and thermodynamic processes of the battery. This model uses a series of partial differential equations and algebraic equations to describe the diffusion and migration behavior of lithium ions in the electrolyte within the positive and negative electrode and separator regions, the lithium ion diffusion process within the solid-phase active particles, and the electrochemical reactions of lithium ion insertion / extraction on the particle surface—all the physical and chemical processes within the battery. By calculating the establishment process of the battery's internal potential distribution, the battery's terminal voltage response under different current inputs can be simulated. Combined with multi-factor orthogonal aging cycle experiments at different operating temperatures and charge / discharge rates, parameter identification algorithms based on particle swarm optimization can extract the variation patterns of aging-related parameters within the battery under different aging paths. Random perturbations are added to the acquired parameter variation curves to improve the coverage of the parameter set for different aging states of the battery. To obtain the current and voltage data required for SOH estimation, the external current input conditions can be set according to the battery's charge / discharge strategy in real-world scenarios, and the battery voltage response sequences under different parameter sets can be obtained. To accurately reflect the variation of external current under constant-voltage charging conditions, this invention uses a monotonically decreasing current decay rate as a constraint and the root mean square error between the simulated voltage and the preset target constant voltage during the constant-voltage phase as a cost function. Optimization algorithms, including the stochastic gradient method, are employed to solve this nonlinear normative problem, thereby ensuring that the simulated current and voltage curves closely resemble the characteristics of real-world control scenarios. Subsequently, ampere-hour integration and nonlinear interpolation methods are used to convert the time-series current and voltage signals of all simulated data into discrete data based on charge / discharge ampere-hours to avoid interference from uneven sampling frequencies in actual applications. Based on the converted discrete sequence, a fixed charge / discharge ampere-hour increment is used... To segment the data, charging and discharging data are randomly divided to extract data fragments with different initial capacities, and a resampling method is used for sample equalization. To extract implicit capacity-related information from these random data fragments and achieve rapid SOH estimation, a data-driven regression model for SOH, including but not limited to one based on a residual convolutional neural network, is constructed. For the neural network structure and hyperparameters, hyperparameter optimization methods, including Bayesian optimization algorithms, and model sparsification methods, including network pruning algorithms, are employed to optimize the model's nonlinear representation capability, while removing redundant structures and reducing the model's dependence on computational resources. Based on a large number of charging and discharging data fragments generated by the electrochemical model, a source dataset is constructed. An adaptive momentum estimation algorithm (Adam) is used to optimize the free variables such as convolutional kernel weights and biases within the data-driven model, completing model pre-training. Subsequently, after obtaining a small amount of real current, voltage data, and SOH labels, a portion of the source dataset is mixed with the real data for transfer learning, quickly achieving the SOH estimation task for the target domain. The aforementioned battery SOH estimation framework can be well applied to BMS and cloud computing platforms. The real current and voltage data segments on which SOH estimation transfer learning depends can be directly obtained from BMS. The pre-trained model obtained offline can be deployed on the cloud platform. Through the real data uploaded by BMS, the current residual convolutional neural network can be iteratively optimized using transfer learning to ensure accurate evaluation and long-term tracking of battery SOH.

[0017] This invention leverages the generalization ability of battery models to describe the external voltage response of batteries under different aging states or charge / discharge conditions, thereby generating rich simulated datasets. Furthermore, by utilizing the powerful nonlinear representation capabilities of data-driven models, it learns the accurate mapping law from charge / discharge data segments with arbitrary initial charge levels to the battery's state of equilibrium (SOH). Based on a pre-trained model constructed from a large simulated dataset, in real-world applications, only a minimal amount of real data and labels are required to complete the SOH estimation task in the target domain through transfer learning. This framework effectively combines the generalization ability of battery models with the accuracy of data-driven models in SOH estimation, significantly reducing the dependence of data-driven models on the amount of real data samples. The reduced-order electrochemical model and residual convolutional neural network used in the framework are merely examples of how to implement related functions. This framework can be combined with various battery models and data-driven models, including but not limited to P2D models, single-particle models, multilayer perceptron models, and support vector machine regression models.

[0018] In this invention, a reduced-order electrochemical model establishes a complete mathematical description of the lithium-ion insertion / extraction, diffusion, migration processes, and potential distribution within the battery based on a series of partial differential equations and algebraic equations. The model is solved by reducing its order through methods including, but not limited to, polynomial approximation, Padre approximation, and Laplace transform. Starting from the underlying electrochemical mechanism, this model can accurately characterize the voltage response characteristics of the battery under different aging states and charge / discharge conditions, while significantly reducing the computational load. This allows for the efficient generation of simulation data from a large set of parameters.

[0019] In this invention, the parameter identification and simulation parameter set construction method can effectively suppress local optima problems through parameter identification algorithms including particle swarm optimization, accurately extract the variation path of aging-related parameters in different aging experiments in the reduced-order electrochemical model, and ensure the accuracy and representativeness of the constructed simulation parameter set. At the same time, random perturbation is added to the original parameter set to improve the richness of the parameter set.

[0020] In this invention, the method for generating simulated current and voltage data during the constant-voltage charging stage, compared to traditional voltage-based feedback control algorithms, additionally considers the problem of limited expressive power of the open-loop model in the closed-loop control stage, leading to distortion of the current curve. It uses the monotonicity of the current decay rate as a constraint term, transforming the closed-loop control problem into a constrained nonlinear programming optimization problem. This problem uses the root mean square error between the simulated voltage during the constant-voltage stage and the preset target constant voltage as the cost function, and iteratively solves the problem using methods including but not limited to stochastic gradient descent. The resulting simulated current and voltage curves closely resemble the current and voltage characteristics under real-world control scenarios.

[0021] In this invention, the application method of the SOH estimation framework on BMS and cloud computing platforms requires real charge and discharge current and voltage data segments that can be directly obtained from the BMS, avoiding additional data extraction processes and computational costs. Simultaneously, it can effectively utilize the powerful computing capabilities of cloud computing platforms, combining real data to iteratively optimize offline pre-trained models, thereby achieving accurate SOH assessment of batteries in the current scenario.

[0022] Compared with existing technologies, this invention has the following advantages: First, the simulated data generated by the electrochemical model can cover real battery usage scenarios with different aging paths and different charge-discharge conditions, and the data-driven model trained based on this has very wide applicability. Second, the data-driven model used can quickly estimate SOH using partial charge-discharge data under random initial charge levels. The data fragments used are easy to obtain and can guarantee high SOH estimation accuracy throughout the entire battery life cycle. Third, the constructed source dataset and pre-trained model only require a very small amount of real data to complete the transfer learning task of SOH estimation in the target domain, solving the problem of limited labeled data in actual usage scenarios. Attached Figure Description

[0023] Figure 1 This is a flowchart of the present invention; Figure 2 This is a comparison chart of simulated current and voltage and experimental data under different SOH conditions in this embodiment of the invention; Figure 3 This is a schematic diagram of charge / discharge data segmentation and preprocessing in an embodiment of the present invention; Figure 4 The image shows the SOH estimation results in the target domain before and after transfer learning of the pre-trained model in this embodiment of the invention. Detailed Implementation

[0024] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0025] Example like Figures 1 to 4 As shown, a reduced-order electrochemical model is first constructed. Constructing this model requires solving the governing equations within the battery in a reduced-order manner, thereby establishing the transfer function relationship between the input current and the output voltage. The terminal voltage of the battery system mainly consists of components such as the open-circuit potentials of the positive and negative electrodes, the overpotentials of the positive and negative electrodes, the liquid phase potential difference, and the DC impedance voltage. The open-circuit potentials of the positive and negative electrodes are functions of the lithium-ion chemical equivalent on the surface of the active particles of the positive and negative electrodes, and their relationship is often obtained through half-cell open-circuit voltage (OCV) experiments. Here, the lithium-ion chemical equivalent is defined as the ratio between the lithium-ion concentration on the material surface and the maximum lithium-ion concentration that the electrode material can accommodate. To obtain the lithium-ion chemical equivalent on the surface of the active particles, Fick's second law is introduced to describe the diffusion process of lithium-ion concentration inside the electrode:

[0026] in, For lithium ions in different time dimensions Spatial dimension and the inner radius dimension of spherical active particles Concentration distribution on (unit: mol / m³) 3 ), The diffusion coefficient of lithium ions within the active material (unit: m) 2 / s). The boundary conditions corresponding to this equation are:

[0027] in, Electrochemical reaction rate (unit: A / m) 2 ), It is Faraday's constant. The particle size of the active particles is in meters (m).

[0028] The governing equation can be solved in the s-domain using the Laplace transform, and its analytical solution is:

[0029] Substituting the boundary conditions, we get:

[0030] Here, the Pader approximation can be used to transform the above transcendental function into a rational fractional transfer function:

[0031] Due to the electrochemical reactions of lithium-ion insertion / extraction on the surface of the positive and negative electrode active material particles, the potential between the solid-phase active material and the liquid-phase electrolyte will deviate from the equilibrium potential. This deviation can be described by the Butler-Volmer equation:

[0032]

[0033] In the formula, This is the offset overpotential (unit: V). The gas constant is Temperature (in K). The reaction rate kinetic constant is Exchange current density (unit: A / m) 2 ), Lithium ion concentration on the surface of active material particles (unit: mol / m³) 3 ), Lithium ion concentration in liquid electrolyte (unit: mol / m³) 3 ), and Let be the anode and cathode transfer coefficients, respectively, and their values ​​can be set to 0.5. Solving for , we get:

[0034] The lithium-ion concentration and potential distributions in the electrolyte phase can be described by Fick's second law and Ohm's law, respectively. The equation for Fick's second law and the boundary conditions are as follows:

[0035]

[0036]

[0037] in, This represents the integral of the electrolyzed liquid. Specific surface area of ​​active material (unit: m²) -1 ), Equivalent liquid phase lithium-ion diffusion coefficient (unit: m) 2 / s), The transference number of ions. The thickness of the negative electrode plate (in meters). The thickness of the diaphragm is expressed in meters (m).

[0038] The governing equation for the potential distribution in the electrolyte along the electrode thickness direction is as follows:

[0039]

[0040] In the formula, The electric potential distribution function (unit: V) Effective ionic conductivity (unit: S / m) The effective conductivity is given by (S / m). Its boundary conditions are:

[0041]

[0042] Integrating the original governing equations twice in the x-direction and substituting the boundary conditions, we can obtain the following solution:

[0043] If we only consider the electrolyte potential difference in the terminal voltage, we can substitute... have to:

[0044] Finally, by combining all the above voltage components, the terminal voltage can be calculated as follows:

[0045] Where R0 is the lumped DC impedance.

[0046] Based on the above electrochemical model, and combined with multi-factor orthogonal aging experimental data, a particle swarm optimization algorithm can be used to extract the parameter variation patterns under different aging paths. The implementation steps of the particle swarm algorithm are as follows: Step 1: Initialize all particles in the preset parameter search space (each particle corresponds to a parameter set). Step 2: Calculate the fitness of each particle, i.e., the root mean square error between the simulated predicted voltage and the actual voltage. Step 3: Determine the optimal values ​​for individuals and the group; Step 4: Update the velocity and position of each particle; Step 5: Repeat steps 2-5 until the allowable error threshold or the maximum number of iterations is reached.

[0047] The aging-related parameter change paths extracted from different aging experiments can represent the degradation law of the battery's internal state under different aging modes. Adding random perturbations to the existing parameter set can cover more aging scenarios and improve the richness of subsequent simulation data. By using a reduced-order electrochemical model and a large set of simulation parameters, and injecting preset charge and discharge current conditions that conform to actual scenarios, the battery's terminal voltage response can be simulated. In the constant-voltage charging stage, to match the current and voltage curves in actual scenarios and avoid distortion of the current curve shape in the pure voltage closed-loop feedback control, an iterative optimization algorithm is used to find the optimal current and voltage sequence. The specific implementation details are as follows: using 10s as the dividing node, the current value in each 10s segment is taken as the same value and recorded as . , and The difference between them is denoted as To ensure that the current decay rate exhibits a monotonically decreasing trend, the following constraint matrix is ​​constructed:

[0048] Simultaneously, using the root mean square error between the simulated voltage during the constant voltage stage and the preset constant target voltage as the cost function, the constrained optimization problem is solved iteratively using the stochastic gradient method. Based on the above process, rapid batch generation of simulated data can be achieved, and the simulated data has a high degree of accuracy in matching with real data, such as... Figure 2 As shown Subsequently, as Figure 3 As shown, by using ampere-hour integration and nonlinear interpolation, the current and voltage time-series signals of all analog data can be converted into discrete data currents based on charge-discharge ampere-hours, with fixed charge-discharge ampere-hour increments. To segment the data, it is randomly divided into data segments with different initial charge levels. These segments are then resampled and reconstructed to form the source dataset. To extract the implicit capacity information from these random data segments and complete the SOH estimation, this invention chooses to build an SOH data-driven regression model based on a convolutional neural network. To suppress network degradation and gradient vanishing problems as the network deepens, residual connections are introduced between adjacent convolutional blocks to reduce information loss during convolution operations and optimize the gradient backpropagation path. Furthermore, a Bayesian optimization algorithm is introduced to select the neural network structure and hyperparameters to improve the model's SOH estimation accuracy. The optimization process is as follows: Step 1: Initialize the prior estimate of the surrogate function using a Gaussian distribution; Step 2: Based on the principle of maximizing the surrogate function and acquisition function in Step 1, select the next evaluation point (the set of parameters to be optimized in the neural network). Step 3: Complete the neural network performance verification at the evaluation point and update the posterior estimate of the surrogate function; Step 4: Repeat steps 2-3 until the maximum number of iterations is reached.

[0049] To reduce the computational resource dependence of the data-driven model, the optimized neural network is pruned to remove redundant structures. Specifically, the importance of each convolutional kernel or neuron to subsequent structures can be evaluated based on the scaling factor in the batch normalization layer, and redundant channels with low importance are removed. Subsequently, the pruned model is pre-trained on a large amount of simulated data using the adaptive momentum estimation algorithm (Adam). Then... Figure 4 As shown, in different application scenarios, a small number of real current and voltage data fragments and SOH labels can be mixed with a portion of the source dataset at a ratio of 1:25, and transfer learning can be performed on the basis of the pre-trained model to improve the SOH estimation accuracy in the target domain.

[0050] The above embodiments are merely illustrative of the design principles and uses of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A battery health state estimation method integrating mechanism and data-driven model, characterized in that, Includes the following steps: Step 1: Construct a reduced-order electrochemical model with strong characterization of battery response behavior. Based on the ordered random combination of different feature parameters, it can accurately simulate the current and voltage response data of the battery throughout its entire life cycle under different aging modes and construct a rich training dataset. Step 2: Construct a data-driven model that can use segmented charge-discharge data under random initial charge to estimate the battery's state of charge (SOH). Step 3: Based on the large number of charge and discharge data fragments generated by the reduced-order electrochemical model, construct the source dataset, use the adaptive momentum estimation algorithm to optimize the free variables inside the data-driven model, and complete the pre-training of the data-driven model. Step 4: After obtaining a small amount of real current and voltage data and SOH labels, mix part of the source dataset with the real data, and perform transfer learning on the basis of the pre-trained data-driven model to quickly achieve SOH estimation of the target domain.

2. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 1, characterized in that... The construction of the reduced-order electrochemical model described in step 1 includes the following steps: Step 1.1: Establish the governing equations and boundary conditions for the lithium ion concentration distribution and potential distribution in the positive and negative electrode plates, separator and electrolyte inside the battery, and solve them by order reduction using the Laplace transform and Pad approximation methods. Step 1.2: Design the experiment by combining the multi-factor orthogonal aging cycle experiment with reference to the performance test; Step 1.3: Based on the intelligent parameter optimization algorithm, obtain the battery parameter change path under different aging paths, add random perturbation, and construct the simulation parameter set; Step 1.4: Inject the required operating conditions, solve the constrained nonlinear programming problem, and obtain simulated charging and discharging data.

3. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 2, characterized in that... The intelligent parameter optimization algorithm described in step 1.3 is a particle swarm optimization algorithm.

4. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 2, characterized in that... The constrained nonlinear programming solution described in step 1.4 uses the monotonically decreasing current decay rate as a constraint condition, the root mean square error between the simulated voltage during the constant voltage stage and the preset target constant voltage as a cost function, and employs optimization algorithms including the stochastic gradient method to solve the nonlinear programming problem.

5. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 1, characterized in that... The construction of the data-driven model in step 2 includes the following steps: Step 2.1: Use ampere-hour integration and nonlinear interpolation methods to convert the time-series signals of current and voltage from the original data into discrete data based on charge-discharge ampere-hours; Step 2.2: Randomly extract charge / discharge segments and process the data through data normalization and resampling; Step 2.3, Build a data-driven model; Step 2.4: Optimize the data-driven model through structural optimization, hyperparameter optimization, and network pruning.

6. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 5, characterized in that... The data-driven model described in step 2.3 is a SOH data-driven regression model with residual convolutional neural networks as its basic structure.

7. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 5, characterized in that... The structural optimization and hyperparameter optimization described in step 2.4 are performed using a Bayesian optimization algorithm.

8. The battery health state estimation method based on the fusion mechanism and data-driven model according to claim 1, characterized in that... In step 4, the source dataset is mixed with the real data at a ratio of 25:1.

Images