A lithium battery state of health localized evaluation method based on feature extraction and light gated neural network

Abstract

The application discloses a lithium battery health state localized evaluation method based on feature extraction and a light gated neural network, and belongs to the technical field of lithium battery health state detection. The application adopts an improved Lorenz fitting function constructed by a capacity increment domain Lorenz function through voltage integration, directly completes fitting in a capacity-voltage domain, introduces a charge-discharge cutoff capacity boundary condition to eliminate an integral constant and normalize fitting parameters, adopts a light gated neural network with a removal reset gate and a timing cycle mechanism, takes single-time static health features as input, completes adaptive weighted mapping through an update gating mechanism, and outputs a health state evaluation result. The application can avoid differential noise interference, greatly reduces model computing power requirements, and is suitable for localized evaluation deployment of a battery management system.

Description

Technical Field

[0001] This invention belongs to the field of lithium battery health status detection technology, specifically relating to a localized assessment method for lithium battery health status based on feature extraction and lightweight gated neural networks. Background Technology

[0002] Lithium-ion batteries, as energy storage units with high energy density and low self-discharge rate, have been widely used in energy storage power stations, electric vehicles, electric ships, and other fields. During long-term use, batteries undergo cyclic aging and calendar aging, characterized by increased internal resistance, decreased usable capacity, and a gradual decline in overall performance. When batteries age to a certain extent, they are prone to safety accidents such as internal short circuits and thermal runaway under extreme conditions such as low temperatures, overcharging and over-discharging, and high-rate charging and discharging. Therefore, accurate monitoring of the State of Health (SOH) of lithium-ion batteries and timely retirement of aging batteries are crucial for ensuring the safe operation of battery energy storage systems.

[0003] In the industry, the State of Harmony (SOH) is typically characterized by the ratio of the battery's current maximum usable capacity to its nominal capacity. A direct method for measuring maximum capacity is the constant voltage, constant current charge-discharge method, which calculates capacity by integrating the current. However, this method relies on a complete charge-discharge process, is time-consuming, requires specialized equipment, and makes it difficult to obtain full-process charge-discharge data in practical applications, thus hindering convenient and localized SOH detection.

[0004] Currently, lithium battery SOH assessment methods are mainly divided into two categories: physical model-based methods and data-driven methods.

[0005] Physical model-based methods model the internal electrochemical reactions of a battery, establish a correspondence between the state of health (SOH) and model parameters, and then obtain the SOH through curve fitting or parameter identification. Typical methods include electrochemical models, equivalent circuit models, and capacity increment models. Among them, electrochemical models involve numerous parameters and have high computational complexity, and the evaluation accuracy is highly dependent on the modeling accuracy. Equivalent circuit models establish state equations by simplifying reaction links, which are highly interpretable but have limited accuracy. Capacity increment models are based on the Gibbs free energy change law of lithium-ion intercalation and deintercalation processes, and use peak function fitting of incremental capacity analysis (ICA) curves to extract health characteristics. They have physical meaning and good generalization, but the ICA curve is obtained by differentiating the capacity-voltage (QV) curve. The original measurement noise will be amplified after differentiation, requiring additional filtering. The selection of filtering parameters directly affects the feature extraction results. Inappropriate parameters can easily lead to the loss of effective information and insufficient stability of feature extraction. To avoid the influence of filtering parameters, existing technologies also include methods that directly fit the peak function of the QV curve. These methods often use Lorentz, Gaussian, and other function prototypes for fitting without simplifying the parameters to be determined by combining the physical constraints of charging and discharging. As a result, such methods have a large number of parameters to be fitted, the fitting results are greatly affected by the initial values, and the computational efficiency and stability are poor.

[0006] Data-driven approaches utilize artificial intelligence models to train on historical charge-discharge data, directly establishing a mapping between input data and State of Harmony (SOH). Common models include SVR, GPR, RNN, CNN, LSTM, and GRU. While these methods do not require precise modeling of the battery's internal reactions and can achieve high evaluation accuracy in specific scenarios, they suffer from poor interpretability, high dependence on data volume and quality, large model parameter count, and high computational requirements. The trained models are difficult to deploy directly on ordinary battery management systems with limited computing and memory resources, and their adaptability and generalization ability under real-world operating conditions need further verification.

[0007] To balance interpretability and evaluation accuracy, existing technologies have also seen fusion schemes combining physical models and neural networks. This involves first extracting health features using physical methods, and then using neural networks to perform regression predictions. However, most existing fusion schemes still rely on ICA curves for feature extraction, failing to fundamentally eliminate the influence of differential noise and filtering parameters. Furthermore, the neural networks used are mostly standard temporal network structures, resulting in a large number of parameters and computational overhead, making local deployment on embedded systems difficult. In addition, fusion schemes directly based on QV curve fitting generally suffer from numerous fitting parameters and high computational complexity, making it difficult to balance extraction efficiency and feature stability.

[0008] In summary, there is currently no lithium battery SOH evaluation scheme in this field that can fundamentally avoid the influence of differential noise, achieve stable and efficient feature extraction, while having low model computational cost, local deployment capability, and evaluation accuracy that meets actual needs. There is still room for further improvement. Summary of the Invention

[0009] To address the shortcomings and deficiencies of existing technologies, this invention provides a localized assessment method for the health status of lithium batteries based on feature extraction and a lightweight gated neural network. This method first acquires voltage and current data during the constant-current charge-discharge phase of the target lithium battery, constructing a capacity-voltage curve by integrating the current data. Then, an improved Lorentz fitting function is used to directly fit this curve in the capacity-voltage domain, thus fundamentally avoiding the high-frequency noise amplification problem introduced by the traditional method of first differentiating the capacity-voltage curve into a capacity increment curve. The improved Lorentz fitting function is constructed from the Lorentz function in the capacity increment domain through voltage integration, and the boundary condition of the lithium battery's charge-discharge cutoff capacity is introduced to eliminate the integration constant, converting the amplitude coefficient into a normalized capacity contribution coefficient. This reduces the number of parameters required for fitting, improves the stability and efficiency of feature extraction, and directly extracts voltage platform health feature parameters related to the health status, including the maximum usable capacity. Furthermore, considering that the extracted health feature parameters are static degradation features at a single time step and lack temporal correlation between time steps, this invention performs a lightweight reconstruction of the standard gated recurrent unit network, blocking the state loop transmission of historical time steps and retaining only the update gating mechanism to adaptively weight and modulate the spatial dimension of the input features. This significantly reduces the number of parameters and computational complexity required for model training while maintaining nonlinear representation capabilities. Finally, using the aforementioned health feature parameters as input, the pre-trained lightweight gated neural network outputs the lithium battery health status assessment result. This invention deeply couples improvements in physical feature extraction with lightweight improvements in neural network structure, balancing interpretability, feature extraction stability, and low computational cost. It is suitable for deployment in embedded chips of battery management systems to achieve localized calculation and assessment of lithium battery health status.

[0010] The technical solutions adopted by this invention to solve its technical problems include:

[0011] This invention provides a localized assessment method for the health status of lithium batteries based on feature extraction and lightweight gated neural networks. It adopts a technical approach that combines physical-driven feature extraction with lightweight neural network regression, and simultaneously optimizes the two stages of feature extraction and network modeling. This reduces computing power and storage requirements while ensuring assessment accuracy, making it suitable for localized deployment scenarios of battery management systems.

[0012] This method specifically includes the following steps:

[0013] The first step is to acquire voltage and current data of the target lithium battery during the constant current charge and discharge phases. The current data is then integrated to construct a capacity-voltage curve. This curve visually reflects the change in cumulative capacity with terminal voltage during charge and discharge, serving as the data foundation for subsequent voltage platform feature extraction.

[0014] The second step involves directly fitting the capacity-voltage curve in the capacity-voltage domain using an improved Lorentz fitting function to extract health characteristic parameters of the voltage plateau, including the maximum usable capacity. The equilibrium potential of a lithium-ion battery electrode is determined by the Gibbs free energy change during the lithium-ion insertion / extraction process. A voltage plateau is formed during the constant free energy phase of charging and discharging, serving as an inherent electrochemical fingerprint characterizing battery degradation. Traditional methods often convert the capacity-voltage curve into a capacity increment curve using differential operations before peak fitting. However, differential operations amplify high-frequency noise in the original measurement, requiring additional filtering. The selection of filtering parameters directly affects the characteristic results. The improved Lorentz fitting function used in this invention is constructed by integrating the Lorentz function in the capacity increment domain with voltage. This eliminates the need to perform differential operations on the capacity-voltage curve, thus avoiding the problem of differential amplification noise in principle. At the same time, it introduces the boundary condition of the lithium battery charge and discharge cutoff capacity, namely, the cumulative capacity at the upper cutoff voltage is equal to the current maximum usable capacity, and the cumulative capacity at the lower cutoff voltage is zero. This physical constraint eliminates the constant term introduced by the integration and converts the original amplitude coefficient into a normalized capacity contribution coefficient, simplifying the scale of the parameters to be fitted and improving the convergence stability of the fitting process.

[0015] The third step involves inputting the health feature parameters obtained at a single time step into a pre-trained lightweight gated neural network, which then outputs the lithium battery health status assessment result. The aforementioned fitted health feature parameters are static degradation features extracted from a single charge-discharge curve, independent of multi-cycle time-series data. Therefore, this invention restructures the standard gated recurrent unit network to address this attribute, blocking the state loop transmission of historical time steps, removing the reset gate structure used for time-series state modulation, and retaining only the update gate mechanism. This allows for adaptive weighted modulation of the spatial dimension of the input single-time step static degradation features. This design retains the nonlinear representation advantages of the gated structure while eliminating redundant parameters and computational overhead caused by time-series loops, forming a complementary overall design with the front-end static feature extraction.

[0016] Furthermore, a nonlinear least squares algorithm is used to fit the capacity-voltage curve. To constrain the physical rationality of the fitting results and improve convergence efficiency, corresponding boundary values ​​are set for each fitting parameter: the upper limit of the maximum usable capacity is the battery's nominal maximum capacity at the factory, and the lower limit is the charge-discharge capacity during the constant current stage; the normalized capacity contribution coefficient ranges from [0,1]; the center voltage is located between the upper and lower limits of the battery's operating voltage; and the peak width coefficient ranges from [0, 0.5], a constraint interval set based on empirical values. These boundaries are derived from the battery's electrochemical characteristics and the statistical regularities of the existing system, which can narrow the parameter optimization range and reduce the probability of the fitting falling into a local optimum.

[0017] Furthermore, the number of voltage plateaus is determined based on the properties of the battery's positive and negative electrode materials, taking the number of significant peaks in the initial charge-discharge capacity increment curve. Too many voltage plateaus would cause the parameters to be fitted to increase exponentially, introducing additional fitting errors. Moreover, some plateaus show weak characteristic changes with aging, contributing little to the characterization of the battery's health. Therefore, only plateaus corresponding to significant peaks are selected for extraction. The significant peak refers to a peak corresponding to the battery's inherent voltage plateau that stands out in the capacity increment curve; its determination can be achieved by using a threshold representing the proportion of the peak height to the global maximum peak height. Analysis of the capacity increment characteristics of a large amount of battery data shows that most batteries have only two effective peaks; therefore, the number of voltage plateaus is set to 2.

[0018] Furthermore, the extracted health characteristic parameters include maximum usable capacity, normalized capacity contribution coefficient, center voltage, and peak width coefficient. These parameters reflect the changes in electrochemical characteristics during battery aging from four dimensions: global capacity decay, capacity contribution ratio of each platform, characteristic potential shift, and platform steepness, respectively, and can comprehensively characterize the battery's degradation state.

[0019] Furthermore, the lightweight gated neural network is composed of an input layer, a gated network layer, a Dropout layer (a random deactivation layer used to randomly disconnect some neurons during model training according to a preset probability to suppress overfitting and improve the network's generalization performance), a fully connected layer, and an output layer; the calculation of the gated network layer satisfies the following relationship:

[0020]

[0021]

[0022] Where z is the update gate output vector, h is the hidden layer output state vector, and x is the input feature vector. To update the gate weight matrix, The hidden layer weight matrix is... To update the gate bias vector, Let σ be the hidden layer bias vector, σ be the Sigmoid activation function, tanh be the hyperbolic tangent activation function, and × be element-wise multiplication. The elements of the update gate output vector are mapped to the [0,1] interval by the Sigmoid function, which can dynamically weight the contribution of each dimension of input features and achieve adaptive purification of feature dimensions. Compared with the computation logic of the standard gated recurrent unit, this process no longer introduces the hidden state of the previous time step into the operation, which is a concrete manifestation of the network removing temporal loops and adapting to static feature inputs.

[0023] Furthermore, the lightweight gated neural network employs a single hidden layer structure with a hidden layer dimension of 64. This configuration, while ensuring sufficient nonlinear mapping capability, minimizes the overall network size. Under the same input and hidden layer dimensions, compared to conventional gated recurrent unit networks, this network structure reduces the number of parameters requiring training by approximately 94.78%, and the trained model file size is approximately 8 KB, effectively adapting to the storage and computing power limitations of embedded chips.

[0024] Furthermore, the training hyperparameters of the lightweight gated neural network include: a mean squared error loss function, an Adam optimizer, a learning rate of 0.001, batching by individual battery cells, 2000 training epochs, a dropout rate of 0.2, and uniform weight initialization. These hyperparameter configurations ensure stable convergence during model training. The batching by individual battery cells allows gradient updates to better reflect the complete degradation trajectory of a single battery, and combined with the Dropout mechanism, it enhances the model's generalization ability.

[0025] Furthermore, after extracting the health feature parameters and before inputting them into the lightweight gated neural network, a feature selection step is included: calculating the Pearson correlation coefficient between each health feature parameter and the health status of the lithium battery, selecting features with absolute correlation coefficient values ​​higher than a preset threshold to form an input feature set, and then inputting the input feature set into the lightweight gated neural network. Correlation selection can eliminate redundant features with weak correlation to the health status, simplifying the input dimensions and reducing computational overhead while minimizing the interference of irrelevant features on the evaluation results.

[0026] Furthermore, the method is deployed in the embedded chip of the battery management system to achieve localized calculation and evaluation of the lithium battery's health status. Thanks to the simplified parameters at the feature extraction end and the lightweight design of the network end, the overall computational and storage overhead of the solution is low, and the entire evaluation process can be completed locally without relying on a remote computing platform.

[0027] Based on the same inventive concept, the present invention also provides a lithium battery health status assessment system, including a data acquisition module, a feature extraction module, and a status assessment module.

[0028] The data acquisition module is used to acquire voltage and current data during the constant current charging and discharging phase of the target lithium battery, and to perform integral processing on the current data to construct a capacity-voltage curve.

[0029] The feature extraction module is used to directly fit the capacity-voltage curve in the capacity-voltage domain using an improved Lorentz fitting function to extract voltage platform health feature parameters including the maximum available capacity. The improved Lorentz fitting function is constructed by integrating the Lorentz function in the capacity increment domain with voltage, introducing boundary conditions for the lithium battery charge and discharge cutoff capacity to eliminate the integration constant, and converting the amplitude coefficient into a normalized capacity contribution coefficient.

[0030] The state evaluation module is used to input the health feature parameters at a single moment into a pre-trained lightweight gated neural network and output the lithium battery health state evaluation result. The lightweight gated neural network blocks the state loop transmission of historical time steps and only performs spatial dimension adaptive weighted modulation on the input single-moment static degradation features through the retained update gating mechanism.

[0031] Compared with the prior art, the present invention and its preferred embodiments have at least the following beneficial effects:

[0032] 1. Enhanced stability of feature extraction. Fitting calculations are performed directly in the original capacity-voltage domain, avoiding the amplification effect of differential operations on measurement noise. No additional filtering is required, thus eliminating the problem of loss of effective information due to improper selection of filtering parameters. The reliability and consistency of feature extraction results are superior.

[0033] 2. Higher fitting calculation efficiency. The simplified fitting parameter system based on the physical boundary constraints of charging and discharging reduces the number of parameters to be solved, lowers the dependence of the fitting results on the initial values, and improves the convergence speed and stability of curve fitting.

[0034] 3. Lower deployment and operation threshold. The network structure is customized for static feature input, removing redundant computing units related to time-series dependencies. The number of network parameters and computing power requirements are significantly reduced, the model size is small, and it can be directly deployed in resource-constrained embedded chips. Health status assessment can be completed locally in the battery management system without relying on remote computing resources.

[0035] 4. More balanced overall performance. The overall solution combines the interpretability of the physical model with the evaluation accuracy of the data-driven method, demonstrating stable evaluation results on multiple types of public datasets, and is adaptable to the health monitoring needs of lithium batteries in various scenarios such as energy storage, automotive, and marine applications. Attached Figure Description

[0036] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0037] Figure 1 This is a flowchart illustrating the overall process of an embodiment of the present invention.

[0038] Figure 2 This is a comparison chart of SOH prediction results based on a lightweight gated neural network on the NASA dataset in this embodiment of the invention. Figure 2 (a) is a comparison chart of the predicted and reference values ​​of SOH for B5 battery. Figure 2 (b) is a comparison chart of the predicted and reference values ​​of SOH for B6 batteries.

[0039] Figure 3 This is a comparison chart of SOH calculation results based on pure Lorentz function fitting on the NASA dataset in this embodiment of the invention. Figure 3 (a) is a comparison chart of the calculated and reference values ​​of SOH for B5 battery. Figure 3 (b) is a comparison chart of the calculated and reference values ​​of SOH for B6 battery.

[0040] Figure 4 This is a comparison chart of SOH prediction results based on a lightweight gated neural network on the Oxford University dataset, as shown in this embodiment of the invention. Figure 4 (a) is a comparison chart of the predicted and reference values ​​of SOH for Cell4 batteries. Figure 4 (b) is a comparison chart of the predicted and reference values ​​of SOH for Cell5 battery.

[0041] Figure 5 This is a comparison chart of SOH calculation results based on pure Lorentz function fitting using the Oxford University dataset in this embodiment of the invention. Figure 5 (a) is a comparison chart of the calculated and reference values ​​of SOH for Cell4 batteries. Figure 5 (b) is a comparison chart of the calculated SOH value and the reference value of Cell5 battery. Detailed Implementation

[0042] To make the features and advantages of the present invention more apparent and understandable, specific embodiments are described below in detail:

[0043] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0045] This invention provides a localized assessment method for the state of health (SOH) of lithium batteries based on feature extraction and a lightweight gated neural network. Based on the physical characteristics of the voltage plateau during the charging and discharging process of lithium batteries, and combined with a lightweight gated neural network, a lithium battery SOH assessment model deployable on embedded chips is proposed. This invention employs a fusion of a physical model and an artificial neural network. Based on the voltage plateau reaction process of the lithium battery, health factors are extracted from the QV curve, and then an artificial neural network is used to establish a health status assessment model. The main improvements of this invention include: 1. Direct fitting of the QV curve, avoiding high-frequency noise problems in the ICA curve. 2. Simplification of the parameters in the Lorentz function based on the capacity change trend during the charging and discharging process of the lithium battery, reducing the number of parameters that need to be fitted. 3. The use of a lightweight neural network to establish the assessment model. Compared with the linear model, this model has higher accuracy for the same type of lithium battery, and the total number of parameters is smaller than that of the currently widely used complex models. First, the improved Lorentz function is used to fit the voltage-capacity curve of the lithium battery, and voltage plateau features are obtained through screening. Then, the GRU network is modified to be lightweight, removing the reset gate while retaining the update gate mechanism, reducing the number of parameters that need to be trained. Finally, the results were validated on the NASA and Oxford University datasets. The results show that the calculated SOH is close to the actual reference value, with the maximum SOH difference within 1.2%. The MAE and RMSE obtained from the NASA dataset are all within 0.0050 and 0.0056, respectively. The MAE and RMSE obtained from the Oxford University dataset are all within 0.0019 and 0.0050, respectively. The model proposed in this invention reduces the number of parameters required for training by 94.78% and consumes fewer computational resources, making it promising for localized deployment of lithium battery SOH in ordinary battery management systems.

[0046] The implementation of the present invention will be further described below with reference to the accompanying drawings:

[0047] 1. QV model based on voltage platform

[0048] The equilibrium potential of a lithium-ion battery electrode is determined by the change in Gibbs free energy during the lithium-ion insertion / extraction process. During charging and discharging, if the free energy remains constant, the electrode potential does not change with the amount of lithium inserted, resulting in a corresponding voltage plateau in the voltage curve. Generally, during the charging and discharging process of a lithium-ion battery, the lithium-ion insertion / extraction reaction occurs in multiple stages, each with a constant free energy, resulting in multiple independent voltage plateaus in the voltage characteristic curve. Therefore, the voltage plateau is essentially an electrochemical fingerprint of the lithium insertion phase transition behavior of the electrode, an inherent property directly determined by the material's crystal structure and phase transition mode. During cycle aging or calendar aging, the SEI film continuously thickens, and the active lithium and electrode active materials undergo irreversible loss. The overall potential of the voltage plateau will shift slightly, the effective length of the plateau will shorten, and the slope will increase slightly. However, since the lithium-ion insertion and extraction process generally occurs within a narrow voltage range, the voltage plateau differences under different aging states are not significant, and noise is generally present during measurement. Therefore, it is difficult to directly obtain health characteristics from the voltage curve.

[0049] To better extract voltage plateau characteristics, voltage and capacity characteristics are typically converted into capacity increment curves. The originally smooth voltage plateau becomes a distinct curve peak. By fitting the capacity increment curve with a Lorentz function, the voltage plateau characteristics can be represented by corresponding parameters such as peak size, peak width, and peak position. The capacity increment curve can be composed of the superposition of several Lorentz functions, as follows:

[0050] (1)

[0051] in, is the capacity increment value, i.e., the rate of change of battery capacity with terminal voltage; is the vertical axis of the capacity increment (ICA) curve; n is the total number of lithium battery voltage plateaus, i.e., the number of Lorentz peaks; i is the sequence number of the Lorentz component corresponding to the i-th voltage plateau. The amplitude coefficient of the i-th Lorentz component is positively correlated with the capacity contribution of the corresponding voltage plateau. is the peak width coefficient of the i-th Lorentz component, which corresponds to the steepness of the voltage plateau and is directly related to the peak half-width at half maximum (FWHM); V is the terminal voltage of the lithium battery. Let be the center voltage of the i-th Lorentz component, and be the characteristic potential of the corresponding voltage plateau.

[0052] To avoid fitting errors caused by high-frequency measurement noise due to differentiation, integrating equation (1) yields the improved Lorentz function:

[0053] (2)

[0054] Where Q is the cumulative charge and discharge capacity of the battery when the terminal voltage is V; C is the integration constant.

[0055] When the voltage V approaches the lower cutoff voltage, Q should be close to 0 Ah, while when V approaches the upper cutoff voltage, Q should be close to the maximum capacity value Q. max Therefore, the parameters in equation (2) should have the following relationship:

[0056] (3)

[0057] in, The maximum usable capacity of the battery is currently, which is the total capacity under full charge. The above formula corresponds to the boundary condition that the cumulative capacity at the upper cutoff voltage is equal to the maximum usable capacity, and the below formula corresponds to the boundary condition that the cumulative capacity at the lower cutoff voltage is 0.

[0058] It can be organized into the following:

[0059] (4)

[0060] Substituting equation (4) into equation (2), we obtain the capacity-voltage model based on the improved Lorentz function:

[0061] (5)

[0062] In the formula, =A i / Q max , is the normalized capacity contribution coefficient for the i-th voltage platform, with a value range of [0,1]. By analyzing the capacity relationship during the charging and discharging process, the number of parameters that need to be fitted in the improved Lorentz function is reduced, which helps to reduce the dimension of health features that need to be screened. In addition, the fitted Q max It has a high correlation with battery health status.

[0063] 2. Extraction of health features of voltage platform

[0064] To obtain the key feature Q contained in the voltage plateau max , V 0i and ω iThe QV curve of the lithium battery can be fitted according to equation (5). Before fitting, it is necessary to determine the number of voltage plateaus of the lithium battery, i.e., the number of peaks n of the QV curve. Generally, n is determined by the material properties of the positive and negative electrodes in the battery. For example, the LiFePO4 (LFP) positive electrode has a flat voltage plateau, while the graphite negative electrode has four voltage plateaus. However, if the value of n is too large, the parameters that need to be fitted will increase exponentially, which will cause greater errors during fitting. Moreover, some voltage plateaus do not have obvious health characteristics, and their parameters do not change much during aging. In addition, during the battery aging process, some peaks may disappear or new peaks may appear. Therefore, in order to estimate SOH, n can be taken as the number of relatively obvious peaks in the ICA data of the first charge and discharge (for example, the ICA peak height exceeds 20% of the global maximum peak height as the judgment criterion), and only the relatively obvious voltage plateaus are used for feature extraction. This invention analyzes the ICA data characteristics of a large dataset, and most batteries only have two effective peaks, so n=2 is selected.

[0065] A nonlinear least squares algorithm is used to fit the QV curve to extract the health feature Q. max , V 0i and ω i The selected charging or discharging curve data should be the capacity Q and voltage V of the battery when discharging from the upper cutoff voltage to the lower cutoff voltage, or charging from the lower cutoff voltage to the upper cutoff voltage. The upper and lower cutoff voltages of the battery can be obtained from the factory nominal parameters of different batteries. The sampling rate of the data used for fitting needs to be greater than 0.1 Hz. As a further preferred embodiment, in this embodiment, the voltage range of 2.7-4.2 V in the constant current charging section is selected, the number of sampling points is not less than 2000, and the fitting determination coefficient R is... 2 A value >0.98 is required to proceed to SOH inference.

[0066] Since the fitting results are easily affected by the upper and lower bounds of the fitting parameters, it is necessary to analyze each parameter individually. Based on the analysis in Section 1, Q can be obtained. max The maximum value should be the factory-set maximum capacity, and the minimum value should be the charging / discharging capacity value during the constant current charging phase. =A i / Q max According to equation (5), it should satisfy ∈[0,1]. And V 0i It should be between the upper and lower limits of the battery's operating voltage. ω iRelated to the full width at half maximum (FWHM) of the peak, a fixed empirical upper and lower limit [0, 0.5] can be taken based on the characteristics of the ICA curve. During the aging process, the characteristic changes of some voltage plateaus may be less significant. Therefore, it is necessary to analyze the correlation between the fitted health characteristics and the health status, and to select key health characteristics based on the magnitude of the correlation. This invention uses the Pearson correlation coefficient to characterize the degree of correlation, which can be calculated by the following formula:

[0067] (6)

[0068] In the formula, F represents health characteristics, S represents health status, and R is the Pearson correlation coefficient, ranging from -1 to 1. The closer the absolute value of R is to 1, the stronger the linear correlation between the corresponding health characteristic and SOH; the closer the absolute value of R is to 1, the greater the correlation between the health characteristic and health status. The closer the absolute value of R is to 0, the less obvious the linear relationship between the health characteristic and health status. Let j be the health feature value of the j-th sample; This is the arithmetic mean of the health characteristics of all samples; This represents the true health status (SOH) value of the j-th sample. is the arithmetic mean of the SOH values ​​of all samples; m is the total number of samples involved in the correlation calculation.

[0069] 3. Lightweight gated neural network

[0070] Gated Recurrent Unit (GRU) neural networks are an improved structure of recurrent neural networks and long short-term memory networks. By introducing update and reset gates, GRU effectively solves the problems of vanishing gradients and insufficient capture of long-term dependencies in traditional recurrent neural networks. Compared with LSTM, GRU has only two gate structures, which can maintain modeling accuracy while having advantages such as high computational efficiency and small storage space.

[0071] Traditional GRUs employ a multi-timestep sequence input model, relying on temporal correlations between consecutive time steps to complete feature modeling. The health features extracted in this invention have no temporal correlations, thus eliminating the temporal dependency modeling approach of the reset gate. Instead, the network input is solely based on single-time-step static degradation features, simplifying the GRU network and establishing a lightweight gated neural network. This network eliminates the need for multi-time-step state loop iterations, significantly reducing the number of network parameters and computational complexity. Simultaneously, it retains the update gating mechanism of GRU, enabling adaptive filtering and weighted purification of input features. Compared to conventional fully connected multilayer perceptron networks, the proposed lightweight gated neural network possesses superior nonlinear representation capabilities and adaptive feature modeling performance, better fitting the complex intrinsic relationship between battery degradation features and health status.

[0072] The lightweight gated neural network established in this invention uses the health feature at a single time step as input x, calculates and updates the gate output z using equation (7), and then updates the hidden layer state h using equation (8).

[0073] (7)

[0074] (8)

[0075] In the formula, W zx and W hx These are the weights for updating the gate and the hidden layer, respectively, b z and b h These are the bias vectors for the update gate and the hidden layer, respectively. z is the output vector of the update gate, with elements ranging from [0,1], used to control the weighting of the corresponding features; is the Sigmoid activation function, which maps the input to the [0,1] interval; x is the health feature vector of the network input; h is the output state vector of the hidden layer; tanh is the hyperbolic tangent activation function, which maps the input to the [-1,1] interval; This is element-wise multiplication (Hadamard product), which means that the update gate weights are multiplied element-wise with the candidate features.

[0076] The SOH estimation model established in this invention consists of an input layer, a gated network layer, a Dropout layer, a fully connected layer, and an output layer. It uses the previously proposed health features as input. The gated network layer is responsible for establishing the nonlinear relationship between the battery SOH and the input data. The addition of the Dropout layer avoids overfitting, and the predicted SOH value is obtained through the fully connected layer. Currently, the number of parameters in widely used time-series prediction models for SOH assessment is related to the time step, generally exceeding 5000. The input feature dimension of this invention's model is 4, with one hidden layer of dimension 64. If a conventional GRU model is used, a total of 13505 parameters need to be trained. However, by simplifying the gating mechanism, this invention's model ultimately only requires training 705 parameters, reducing the number of parameters required by 94.78%. The trained model size is approximately 8 KB, reducing the memory footprint of the battery management system for SOH assessment and facilitating localized SOH assessment.

[0077] The steps and framework of the method used in this invention are as follows: Figure 1As shown, the lithium battery charging and discharging data are first preprocessed to obtain current and voltage data during the constant current charging or discharging phase. Next, the current data is integrated to obtain the QV curve, and the least squares method is used to fit the QV curve to the Lorentz function, obtaining the parameters of the Lorentz function. Then, correlation analysis is performed on the fitted parameters to obtain the health factor. Finally, a lightweight gated neural network is established, using the health factor as feature input and health status as data label, to train a health status assessment model.

[0078] Preferably, preprocessing includes, but is not limited to, removing missing values ​​and outlier points, and smoothing the original measurement data using moving average or low-pass filtering to eliminate interference from high-frequency random noise.

[0079] To verify the effectiveness of the proposed method, the NASA battery dataset and the Oxford University battery dataset were selected for validation. B5, B6, B7, and B18 from the NASA dataset were chosen as the total dataset, with three battery datasets selected as the training set and one battery dataset as the test set. The Oxford University dataset contains eight battery datasets; Cell1, Cell2, and Cell3 were selected as the training set, and the remaining individual battery datasets were used as the test set. The hyperparameters during model training are shown in Table 1.

[0080] Table 1 Model Hyperparameters

[0081]

[0082] In cross-validation on the NASA dataset, the SOH calculated using the lightweight gated neural network of this invention showed a high degree of agreement with the actual SOH. Some validation results are as follows: Figure 2 As shown, where Figure 2 (a) shows the comparison between the predicted and reference values ​​of SOH for the B5 battery. Figure 2 (b) Comparison of the predicted and reference values ​​of SOH for battery B6. The calculated SOH can accurately track the capacity decay trend of lithium batteries and effectively capture and predict the temporary capacity recovery phenomenon that occurs during aging. The mean absolute error (MAE) obtained from different cross-validation groups is within 0.0050, and the root mean square error (RMSE) is within 0.0056; among them, the maximum difference between the calculated and actual SOH values ​​is 0.0080 for battery B5, 0.0093 for battery B6, 0.011 for battery B7, and 0.0085 for battery B18.

[0083] Figure 3 The comparison results between the SOH obtained by the pure Lorentz function fitting method and the actual value are shown below. Figure 3 (a) shows the comparison results for the B5 battery. Figure 3(b) shows the comparison results for B6 batteries; the calculated values ​​of this fitting method deviate significantly from the actual values. Through... Figure 2 and Figure 3 The comparison shows that the method using the lightweight gated neural network of this invention has a significant improvement in the accuracy of SOH calculation.

[0084] The calculation results from the Oxford University dataset also show a high degree of agreement with the actual values, and some validation results are as follows: Figure 4 As shown, where Figure 4 (a) shows the comparison between the predicted and reference values ​​of SOH for Cell4 batteries. Figure 4 (b) shows the comparison between the predicted and reference values ​​of SOH for Cell5 battery. The maximum mean absolute error (MAE) of this verification experiment was 0.0019, the maximum root mean square error (RMSE) was 0.005, and the maximum SOH calculation error was 0.0074.

[0085] Figure 5 The comparison results between the SOH obtained by the pure Lorentz function fitting method and the actual value are shown below. Figure 5 (a) shows the comparison results of Cell4 batteries. Figure 5 (b) shows the comparison results for Cell5 batteries. The calculated values ​​of this fitting method deviate significantly from the actual values ​​and cannot be directly used for evaluating the SOH of lithium batteries. Figure 4 and Figure 5 The comparison shows that the lightweight gated neural network method of this invention also achieves a significant improvement in accuracy on the Oxford University dataset.

[0086] In summary, the effectiveness of the proposed method was verified on both the NASA dataset and the Oxford University dataset, demonstrating higher computational accuracy compared to traditional Lorentz function fitting methods. Finally, the parameter count of the proposed model was analyzed to verify its lightweight nature. Currently, the parameter count of widely used time-series prediction models for SOH evaluation is related to the time step, typically exceeding 5000 parameters. The model in this invention has an input feature dimension of 4 and a hidden layer of dimension 64. Using a conventional GRU model, a total of 13505 parameters need to be trained. However, by simplifying the gating mechanism, the model in this invention requires only 705 parameters to be trained, reducing the number of parameters by 94.78%. The trained model is approximately 8 KB in size, reducing the memory footprint of battery management system SOH evaluation and facilitating localized SOH computation evaluation.

[0087] In summary, this invention proposes a lithium battery SOH evaluation model with low computational resource requirements, based on the physical characteristics of the voltage plateau during the charging and discharging process of lithium batteries and combined with a lightweight gated neural network. This model is expected to be deployed in lithium battery management systems for localized computational evaluation. First, by adjusting the upper and lower bounds of the parameters in the Lorentz function through the charging and discharging constraints of the lithium battery, the number of features requiring fitting is reduced. Then, the voltage plateau characteristics in the QV curve of the lithium battery are obtained through fitting. In the NASA dataset and the Oxford University dataset, Q... max , V 0i and ω i The isovoltage plateau features all showed good correlation with SOH. Finally, by lightweighting the GRU network, the number of parameters required for training was reduced by 94.78%. The calculated SOH was close to the actual reference value. The MAE obtained from the NASA dataset was within 0.0050, and the RMSE was within 0.0056. The MAE obtained from the Oxford University dataset was within 0.0019, and the RMSE was within 0.0050. The model proposed in this invention consumes fewer computational resources and is expected to be directly deployed in ordinary automotive BMS systems, realizing localized calculation and evaluation of lithium battery SOH.

[0088] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0089] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A localized assessment method for the health status of lithium batteries based on feature extraction and lightweight gated neural networks, characterized in that, Includes the following steps: S1. Obtain voltage and current data of the target lithium battery during constant current charging and discharging, perform integration processing on the current data, and construct a capacity-voltage curve; S2. The capacity-voltage curve is directly fitted in the capacity-voltage domain using an improved Lorentz fitting function to extract voltage platform health characteristic parameters including the maximum available capacity. The improved Lorentz fitting function is constructed by integrating the Lorentz function in the capacity increment domain with voltage. The boundary condition of the lithium battery charge and discharge cutoff capacity is introduced to eliminate the integration constant, and the amplitude coefficient is converted into a normalized capacity contribution coefficient. S3. Input the health feature parameters at a single time step into the pre-trained lightweight gated neural network and output the lithium battery health status assessment result; the lightweight gated neural network blocks the state loop transmission of historical time steps, removes the reset gate, and only uses the retained update gating mechanism to perform spatial dimension adaptive weighted modulation on the input single-time static degradation features. The lightweight gated neural network is composed of an input layer, a gated network layer, a Dropout layer, a fully connected layer, and an output layer connected sequentially; the calculation of the gated network layer satisfies the following relationship: Where z is the update gate output vector, h is the hidden layer output state vector, and x is the input feature vector. To update the gate weight matrix, The hidden layer weight matrix is... To update the gate bias vector, σ is the hidden layer bias vector, σ is the Sigmoid activation function, tanh is the hyperbolic tangent activation function, and × is element-wise multiplication.

2. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: In step S2, a nonlinear least squares algorithm is used to fit the capacity-voltage curve; the boundary values ​​of the fitting parameters satisfy the following: the upper limit of the maximum usable capacity is the battery's nominal maximum capacity at the factory, and the lower limit is the charge and discharge capacity during the constant current stage. The normalized capacity contribution coefficient ranges from [0,1]; the center voltage is located between the upper and lower limits of the battery operating voltage; and the peak width coefficient ranges from [0, 0.5], a constraint range set based on empirical values.

3. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: In step S2, the number of voltage plateaus is determined based on the properties of the positive and negative electrode materials of the battery, taking the number of significant peaks in the first charge-discharge capacity increment curve; the significant peaks refer to the peaks that correspond to the inherent voltage plateaus of the battery and have prominent characteristics in the capacity increment curve, and their determination is achieved by the threshold of the proportion of the peak height to the global maximum peak height.

4. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: The health characteristic parameters extracted in step S2 include maximum available capacity, normalized capacity contribution coefficient, center voltage, and peak width coefficient.

5. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: The lightweight gated neural network adopts a single hidden layer structure with a hidden layer dimension of 64.

6. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: The training hyperparameters of the lightweight gated neural network include: the loss function is the mean squared error loss function, the optimizer is the Adam optimizer, the learning rate is 0.001, the batch is divided by battery cell, the training epochs are 2000, the dropout rate is 0.2, and the weights are initialized using a uniform distribution.

7. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: After extracting the health feature parameters in step S2 and before inputting them into the lightweight gated neural network in step S3, a feature selection step is also included: calculating the Pearson correlation coefficient between each health feature parameter and the health status of the lithium battery, selecting features whose absolute values ​​of the correlation coefficients are higher than a preset threshold to form an input feature set, and then inputting the input feature set into the lightweight gated neural network.

8. The method for localized assessment of lithium battery health status based on feature extraction and lightweight gated neural networks according to claim 1, characterized in that: The method is deployed in the embedded chip of the battery management system to realize localized calculation and evaluation of the health status of lithium batteries.

9. A lithium battery health status assessment system, used to implement the method as described in claim 1, characterized in that, It includes a data acquisition module, a feature extraction module, and a status assessment module; The data acquisition module is used to acquire voltage and current data during the constant current charging and discharging phase of the target lithium battery, and to perform integral processing on the current data to construct a capacity-voltage curve. The feature extraction module is used to directly fit the capacity-voltage curve in the capacity-voltage domain using an improved Lorentz fitting function to extract voltage platform health feature parameters including the maximum available capacity. The improved Lorentz fitting function is constructed by integrating the Lorentz function in the capacity increment domain with voltage, introducing boundary conditions for the lithium battery charge and discharge cutoff capacity to eliminate the integration constant, and converting the amplitude coefficient into a normalized capacity contribution coefficient. The state evaluation module is used to input the health feature parameters at a single moment into a pre-trained lightweight gated neural network and output the lithium battery health state evaluation result. The lightweight gated neural network blocks the state loop transmission of historical time steps and only performs spatial dimension adaptive weighted modulation on the input single-moment static degradation features through the retained update gating mechanism.

Images