A lithium ion battery state of health (SOH) estimation method based on multi-rate discharge data

Abstract

The application discloses a lithium ion battery state of health (SOH) estimation method based on multi-rate discharge data, and relates to the technical field of battery state evaluation. The method comprises the following steps: acquiring multi-rate cycle discharge data of a lithium ion battery; establishing relevant real state of health labels under different rates; performing unified voltage grid resampling and Savitzky-Golay smoothing processing on the charge-discharge data of the lithium ion battery to obtain a capacity increment curve, extracting interval area, main peak height, main peak half-peak width, main peak area, centroid voltage and other features from the capacity increment curve, and screening out key features; and establishing a physical guided residual network (PGRN) on the basis of sample alignment to perform SOH estimation. The application can realize high-precision estimation of SOH under multi-rate and cross-battery conditions, and can improve the generalization ability and physical rationality of SOH estimation.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery state estimation and health management technology, and in particular to a method for estimating the state of health (SOH) of lithium-ion batteries based on multi-rate discharge data. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, portable electronic devices, and energy storage systems due to their high energy density, long cycle life, and low self-discharge rate. However, with increasing cycle count, batteries experience degradation phenomena such as capacity decay, increased internal resistance, and decreased power performance. Therefore, accurately estimating the State of Health (SOH) is a crucial prerequisite for battery management systems to achieve safe control, lifespan prediction, and maintenance decisions.

[0003] Existing SOH estimation methods mainly include capacity testing-based methods, equivalent circuit and parameter identification-based methods, and data-driven methods. While direct capacity testing has strong physical significance, it often requires a complete charge and discharge process, making it difficult to meet the needs of online applications. Model-based methods are prone to model mismatch under complex operating conditions and multi-rate conditions. Although traditional data-driven methods have a certain nonlinear fitting capability, they are prone to misidentifying rate differences as degradation differences when rate changes are significant, resulting in insufficient cross-rate extrapolation capability.

[0004] Especially in multi-rate discharge scenarios, batteries with the same degradation level exhibit different terminal voltage responses, polarization behaviors, and capacity availability at different rates, resulting in different apparent state of equilibrium (SOH). If samples from different rates are directly mixed without explicitly incorporating rate conditions, the model is prone to systematic bias. Furthermore, high-frequency noise under multi-rate conditions, after amplification through differentiation, interferes with the shape of the capacity increment curve, further reducing the stability of feature extraction and modeling. Therefore, it is necessary to propose a lithium-ion battery SOH estimation method that is suitable for multi-rate discharge conditions, considering both physical interpretability and cross-cell generalization ability, to address the problems of traditional methods in terms of operating condition migration, noise suppression, and long-term trend preservation. Summary of the Invention

[0005] To address the problems of sample mismatch, unclear operating condition coupling, insufficient generalization ability across batteries, and lack of physical consistency in prediction results that exist in existing State of Health (SOH) estimation methods under multi-rate discharge conditions, this invention aims to provide a lithium-ion battery SOH estimation method based on multi-rate discharge data, comprising the following steps:

[0006] Step 1: Under the set ambient temperature, perform a multi-rate cycle discharge test on the target lithium-ion battery to obtain voltage data, current data and discharge capacity data for multiple cycle stages.

[0007] Step 2: Based on the discharge capacity and corresponding initial capacity of each cycle under the corresponding rate conditions, construct a rate-related real health status (SOH) label.

[0008] Step 3: Perform unified voltage grid resampling and smoothing preprocessing on the capacity-voltage data of each cycle to obtain the capacity increment curve.

[0009] Step 4: Extract relevant features characterizing battery degradation based on the capacity increment curve, and screen the key feature set for SOH estimation through correlation analysis.

[0010] Step 5: Align the key feature set, number of iterations, multiplier condition, and temperature condition samples to construct supervised learning samples and apply consistency constraints.

[0011] Step 6: Establish a physically guided residual network, outputting a trend baseline with physical degradation priors, outputting residual compensation with neural networks, and combining physical consistency constraints with conditional residual learning.

[0012] Step 7: Input the sample to be estimated into the trained model and output the SOH estimate under the corresponding scalar conditions.

[0013] Beneficial results:

[0014] 1. This invention defines the SOH at different rates as the true SOH related to the rate, avoiding additional errors in the rate conversion process and preserving the true differences in capacity availability under different rate conditions; it is more in line with the capacity utilization rules in actual engineering applications.

[0015] 2. This invention improves the comparability of capacity increment curves among multiple cycle samples by using unified voltage grid resampling, differential differentiation, and Savitzky-Golay filtering for smoothing. It can more sensitively characterize battery plateau changes, peak shape evolution, and polarization enhancement processes, while also reducing the impact of noise amplification on feature quality.

[0016] 3. By introducing the discharge rate as a conditional variable, this invention solves the problem of the drift of the mapping relationship between features and SOH under multiple discharge rates as the operating conditions change. It can enhance the model's ability to learn uniformly across discharge rates and significantly improve the prediction ability under multiple discharge rate conditions.

[0017] 4. This invention employs a combination of physically meaningful features such as interval area, peak shape, and centroid to simultaneously characterize degradation mechanisms such as capacity decay, polarization enhancement, and reaction redistribution.

[0018] 5. This invention combines physical priors about battery degradation with data-driven compensation through a physically guided residual network, thereby improving the physical rationality, stability, and generalization ability of the SOH estimation results.

[0019] 6. This invention is applicable to SOH estimation across rate, battery type, and even temperature conditions, and can provide effective support for health assessment, life prediction, and maintenance decisions in power battery management systems. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the process of the present invention;

[0021] Figure 2 This is a comparison chart of the battery discharge capacity increment (IC) curves before and after the Savitzky-Golay smoothing process of this invention;

[0022] Figure 3 This is a comparison chart of the actual SOH value, model prediction value, and estimation error under the conditions of 25℃ and 1C according to the present invention;

[0023] Figure 4 This is a comparison chart of the actual SOH value, model prediction value, and estimation error under the conditions of 25℃ and 2C according to the present invention;

[0024] Figure 5 This is a comparison chart of the actual SOH value, model prediction value, and estimation error under the conditions of 25℃ and 4C according to the present invention;

[0025] Figure 6 This is a comparison chart of the actual SOH value, model prediction value, and estimation error under the conditions of 55℃ and 0.5C according to the present invention;

[0026] Figure 7 This is a comparison chart of the actual SOH value, model prediction value, and estimation error under the conditions of 55℃ and 1C according to the present invention;

[0027] Figure 8 This is a comparison chart of the actual SOH value, model prediction value, and estimation error under the conditions of 55℃ and 2C according to the present invention. Detailed Implementation

[0028] To better understand the technical solution of this application, the following description, in conjunction with the accompanying drawings, further illustrates and explains a method for estimating the state of health (SOH) of a lithium-ion battery based on multi-rate discharge data.

[0029] like Figure 1 As shown, a method for estimating the state of health (SOH) of a lithium-ion battery based on multi-rate discharge data includes the following steps:

[0030] Step 1: Under the set ambient temperature, perform a multi-rate cycle discharge test on the target lithium-ion battery to obtain voltage data, current data and discharge capacity data for multiple cycle stages;

[0031] This invention utilizes a cylindrical ternary lithium-ion battery as an example, constructing an experimental platform consisting of a battery charge-discharge tester, a temperature control chamber, and a host computer. The test equipment records the battery's terminal voltage, current, and capacity data during the charge-discharge process, with a sampling period of 1 second. A constant ambient temperature is provided by the temperature control chamber to ensure consistency of test conditions under different operating conditions.

[0032] In the multi-rate discharge experiment, three accelerated discharge rates were set for different temperature groups. Taking the 25℃ temperature group as an example, the accelerated aging rates could be set to 1C, 2C, and 4C; taking the 55℃ temperature group as an example, the accelerated aging rates could be set to 0.5C, 1C, and 2C. All rate groups adopted a unified constant current-constant voltage charging process, with an upper limit charging voltage of 4.2V. The constant voltage phase ended after the current decayed to the cutoff threshold.

[0033] The cyclic discharge phase adopted a pattern of "9 acceleration cycles + 1 low-rate reference cycle". The first 9 cycles used constant current discharge at the corresponding acceleration rate, and the 10th cycle used 1 / 6C low-rate constant current discharge to the same cutoff voltage to reduce the impact of polarization on capacity measurement and periodically obtain reference capacity. The initial capacity of each rate was independently calibrated at the beginning of the experiment. Different calibration procedures were used for the 1 / 6C low-rate and the 1C, 2C, and 4C acceleration rates to ensure the accuracy of the capacity reference. Table 1 shows an example of a discharge experiment at 25℃ and 1C.

[0034] Table 1 Battery testing procedures under 25℃ and 1C conditions

[0035] Step 2: Based on the discharge capacity and corresponding initial capacity of each cycle under the corresponding rate conditions, construct a rate-related real health status (SOH) label;

[0036] To ensure consistency in SOH calculations, at least three complete charge-discharge cycles were performed on batteries under the same rate conditions at the beginning of the experiment, and the average discharge capacity was used as the initial capacity at that rate. The initial capacity at the 1 / 6C low rate is the average of the discharge capacity after three full cycles of constant current and constant voltage charging and three 1 / 6C discharging. The initial capacity at other acceleration rates is the average of the capacity after three full discharge cycles at that rate at the beginning of the experiment.

[0037] Suppose that the battery has a discharge capacity of c times and r times. .

[0038] The SOH at this ratio is defined as:

[0039]

[0040] in, The discharge capacity of the cycle under the corresponding rate condition is obtained by integrating the discharge current or by directly using the discharge capacity reading recorded by the test equipment.

[0041] To avoid introducing additional conversion errors, this implementation does not convert SOH at different ratios to a uniform ratio, but treats it as the ratio-dependent true SOH.

[0042] To reduce outlier interference, samples with capacity or SOH exceeding a reasonable range can be removed, but the original SOH sequence is not monotonically corrected, thus preserving the true fluctuation information under high-magnification conditions.

[0043] Step 3: Perform unified voltage grid resampling and smoothing preprocessing on the capacity-voltage data for each cycle to obtain the capacity increment curve;

[0044] ICA is typically characterized by the first derivative of capacitance with respect to voltage, and its core curve is as follows:

[0045]

[0046] Where Q is the cumulative capacity during the cycle, and V is the terminal voltage.

[0047] Before calculating the IC curve, the Q-V data were subjected to uniform resampling and smoothing preprocessing.

[0048] For each cycle, a segment of capacity change in the same direction is selected, and a discrete point sequence is obtained from the test records. Considering the inconsistent distribution of voltage sampling points in different cycles, the voltage axis is uniformly resampled into a fixed-interval grid, and the voltage step size is set to [value missing]. A voltage grid is constructed within the effective voltage range of this cycle. And by interpolation, the capacity is mapped onto a uniform voltage grid to obtain This aligns the IC curves from different cycles within the same voltage coordinate system, thus avoiding morphological deviations caused by differences in voltage sampling density.

[0049] After obtaining the capacity sequence with equal voltage intervals, the discrete IC values ​​are calculated using the finite difference approximation. To improve numerical stability and reduce the influence of noise, the interior points are calculated using a central difference form:

[0050]

[0051] Boundary points are filled using forward / backward differential methods. In this way, the original discrete Q-V record can be converted into a sequence of IC curves on a unified voltage grid.

[0052] In order to suppress noise while preserving the morphological information such as the peak position, peak height and peak width, this invention uses Savitzky-Golay filtering to smooth the IC curve.

[0053] like Figure 2 As shown, the IC curve after Savitzky-Golay filtering retains key information such as the peak position, peak height, and peak width while suppressing high-frequency noise, providing a stable data foundation for subsequent feature extraction.

[0054] Step 4: Extract relevant features characterizing battery degradation based on the capacity increment curve, and screen the key feature set for SOH estimation through correlation analysis;

[0055] Based on the IC curve of each cycle, a set of statistical and morphological features are constructed, mainly including voltage range area features, main peak correlation features, and centroid voltage. All features are calculated under the same voltage grid and consistent preprocessing procedure to ensure comparability between samples from different cycles and different rate scaling factors.

[0056] Define the area characteristic of the interval [Va, Vb] as follows:

[0057]

[0058] The integration upper and lower limits are given in the direction of voltage from high to low to keep in line with the voltage evolution direction of the discharge / charge data.

[0059] This invention extracts multiple interval area features:

[0060] For example, A[4.2,3.9], A[3.9,3.6], A[3.6,3.3], A[3.3,3.0], A[3.0,2.7], A[2.7,2.5], etc.

[0061] This type of feature can reflect the changing trend of capacity contribution with cycle at different voltage platforms and is suitable for cross-cycle modeling.

[0062] To further characterize the central location of the incremental distribution of the content within the interval, a voltage-weighted mean can be defined within the interval:

[0063]

[0064] When the IC value is small or there are missing points within the interval, a validity check needs to be set for the denominator to avoid numerical instability.

[0065] Define the height of the main peak :

[0066]

[0067] in, This represents the effective voltage range for this cycle. Peak height is used to characterize the change in peak intensity as the cycle decays.

[0068] Define the position of the main peak Used to characterize peak position shift:

[0069]

[0070] Let the half-peak height threshold be... Find the conditions on both sides of the peak position. The left and right intersection points .

[0071] Define the half-peak width (FWHM) of the main peak. :

[0072]

[0073] FWHM describes peak broadening or contraction, which is related to factors such as enhanced polarization and changes in reaction distribution.

[0074] In the vicinity of the main peak Integral with respect to IC, defining the main peak area :

[0075]

[0076] To describe the overall "concentration" of the IC curve on the voltage axis and its migration with cycling, this implementation defines the centroid voltage across the entire voltage range as:

[0077]

[0078] in, This represents the effective voltage range of the cyclic IC curve. The centroid voltage can comprehensively reflect the relative intensity and positional changes of multiple peak structures, and can mitigate the impact of instability in identifying a single peak to some extent.

[0079] All the above features are calculated under a unified voltage grid and the same preprocessing parameters to avoid inconsistencies in feature caliber due to sampling differences between different cycles. For interval features, if a cycle has no effective voltage coverage or missing points within a specified interval, a missing point marker is used or calculation is only performed within the effective intersection range, and the same processing rules are maintained in subsequent modeling stages. For peak features, the main peak identification process is based on the smoothed curve, prioritizing the stability of peak position and peak width. When multiple peaks exist or the peak shape is weak, the "maximum peak height" is used as the main peak determination rule to maintain the determinism and reproducibility of the feature definition.

[0080] After constructing the candidate features of the IC curve, it is still necessary to further select a subset of features that are more closely related to the SOH change and have lower redundancy, in order to reduce the input dimension and improve the stability and interpretability of model training.

[0081] This invention uses Pearson correlation analysis as the initial screening criterion and combines it with score ranking to determine the final feature set.

[0082] The calculation process for the Pearson correlation coefficient is defined as follows:

[0083]

[0084] in, Indicates the HI value. Indicates the battery's SOH value. This represents the average HI value. This represents the average SOH value. This is the Pearson correlation coefficient.

[0085] Pearson correlation analysis was performed on the ICA curves of all rates, and the following features with the strongest correlation were found: A[2.7, 2.5], A[3.6, 3.3], A[3.0, 2.7], peak height of the main peak, half width at half maximum of the main peak, and centroid voltage.

[0086] Step 5: Align the key feature set, number of iterations, multiplier condition, and temperature condition samples to construct supervised learning samples and apply consistency constraints.

[0087] By introducing the scalar ratio as an explicit conditional variable into the feature set and constructing a strict alignment between the rate and the SOH label, the model can simultaneously consider the influence of the scalar ratio when learning the feature SOH mapping, thus achieving unified modeling across scalar ratio samples.

[0088] The rate information is encoded as a numerical value called "rate" and used as input to the model. The value of the rate variable is constructed based on the experimental conditions and cycle numbering rules. For test sequences with periodically alternating rates, the rate corresponding to the cycle is determined according to the cycle numbering rule, and the base rate of the battery is used for other cycles.

[0089] Using a cycle as the basic sample unit, each sample corresponds to an IC feature vector and its SOH label within a cycle. For the c-th cycle, construct the sample group: ,in The feature set extracted from this cyclic IC curve, The temperature condition variable corresponding to this cycle. This is the condition variable for the multiplier corresponding to this loop. This is the SOH label calculated at the corresponding magnification. It ensures a one-to-one correspondence between the SOH label and the feature vector, cycle number, temperature condition, and magnification condition.

[0090] To avoid misalignment of "features, scaling factor, and label" that could cause bias in the supervision signal, the alignment process follows these principles: for any feature vector It must be calculated from the IC curve of the c-th cycle; for the same sample At The data must come from the operating condition and capacity record corresponding to the c-th cycle; cross-cycle splicing or replacing the current cycle label with an adjacent cycle is strictly prohibited; when there is a rate switching in the test sequence... It is uniquely determined by the cycle numbering rule and is consistent with the multiplier on which the cycle capacity calculation is based.

[0091] The calculation of IC features depends on the voltage range and the resampling grid; for features such as the area of ​​the interval, it must be ensured that there is valid data in the corresponding voltage range.

[0092] If insufficient interval coverage renders features uncalcifiable, then a uniform rule should be applied to avoid inconsistent calculation methods for different samples; if a cycle is removed due to abnormal SOH, abnormal volume recording, or unsatisfactory IC curve quality, then the corresponding cycle... , , The entire sample set must be deleted to ensure that the sample index does not shift; the aligned sample set should be stored and tracked using cycle as the unique index for easy reproduction and auditing.

[0093] After alignment, the model input consists of IC features and a rate-conditional variable. The rate-conditional variable, as an explicit condition, helps the model distinguish the sources of IC feature changes at different rates, partly from degradation evolution and partly from differences in operating conditions.

[0094] Step 6: Establish a physical-guided residual network, outputting a trend baseline with physical degradation priors, outputting residual compensation with neural networks, and combining physical consistency constraints with conditional residual learning.

[0095] A physical-guided residual network PGRN is constructed to estimate SOH. This network includes a physical prior branch and a residual learning branch.

[0096] To introduce interpretable degradation priors in SOH prediction tasks across temperature and rate, a mechanism-inspired SOH baseline model was constructed. It is used to characterize the overall degradation trend of SOH as the cycle progresses and the temperature changes.

[0097] This baseline model does not aim for a perfect fit to all details, but rather provides stable trend constraints, enabling subsequent residual networks to learn deviations caused by rate effects, polarization differences, and individual battery differences, thereby improving generalization ability and training stability.

[0098] Let the number of iterations be... The maximum number of loops is Then the normalized loop variable is:

[0099]

[0100] This invention introduces a temperature factor in the form of Arrhenius. The temperature is converted from Celsius to Kelvin:

[0101]

[0102]

[0103] in, The learnable reference temperature parameter in the model is initially set to 40℃.

[0104] And set the apparent activation energy Reasonable range constraints:

[0105]

[0106] Define the temperature acceleration factor:

[0107]

[0108] It is used to describe the scale change of degradation rate caused by temperature changes, so that the decay intensity under the same cycle scale is adaptively adjusted with temperature.

[0109] To maintain model simplicity while covering typical degradation patterns, this invention employs a weighted mixture of exponential and linear decay terms as the degradation baseline. The effective decay rate is defined as follows:

[0110]

[0111] in, Let be the decay rate coefficient to be learned. The exponential and linear components are written as follows:

[0112]

[0113]

[0114] in, The coefficients are linear decay coefficients, and both are weighted according to their respective properties. The mixture yields the original degradation trend:

[0115]

[0116] in, These are constant weights used to control the proportion of the exponential and linear components.

[0117] To address the difference in degradation rates across different temperature groups and avoid the discontinuities and insufficient generalization inherent in piecewise models, a temperature weighting function is employed to achieve continuous parameter switching. Definition And a convex combination is used for the attenuation rate and linearity coefficient:

[0118]

[0119]

[0120]

[0121] Temperature weights enable the model to transition smoothly across different temperature ranges, improving its ability to uniformly describe data across temperature ranges.

[0122] Considering the physical range of SOH and the stability of the training process, for Truncate and linearly map to the interval [0.8, 1.0]:

[0123]

[0124] As a physical prior baseline, it mainly characterizes the macroscopic degradation trend of SOH with changes in cycling and temperature.

[0125] However, since the polarization / plateau evolution information reflected by the rate-of-magnitude variation and IC characteristics was not explicitly modeled in the prior, this invention introduces a residual learning mechanism in subsequent models. Based on this, compensation is made for individual differences and the impact of working conditions.

[0126] By adopting the concept of conditional modeling, operating condition information is used as an explicit conditional input, enabling the model to learn the SOH mapping relationship determined by the input features and operating conditions, thereby achieving cross-operating condition prediction under a unified framework.

[0127] To ensure the consistency of numerical scale for variables with different dimensions during network training, this invention normalizes the condition variables and constructs necessary interaction terms to enhance the expressive power of nonlinear coupling relationships. Specifically, it is defined as follows:

[0128]

[0129]

[0130]

[0131] And construct the condition vector:

[0132]

[0133] in, and Used to characterize the degradation process and temperature coupling effect, as well as the nonlinear trend of degradation as it progresses with cycles. As a scaling factor, it is used to distinguish the systematic differences in characteristic SOH mapping under different loads.

[0134] In terms of model structure, the IC feature x and the condition vector are combined. The data is fed into different feature extraction branches. The data branch is used to extract SOH-related features from x, and the conditional branch is used to extract features from x. The system encodes operating condition information and forms a modulated representation of the degradation trend.

[0135] Subsequently, the two types of representations are concatenated and nonlinearly mapped at the fusion layer to output the residual term. and with physical prior baseline The final prediction is obtained by superimposing the results:

[0136]

[0137] The training objective consists of data items, physical consistency items, boundary penalties, and... Constraints together form the basis. Let the true SOH of the i-th sample be denoted as . The model predicts that The physical prior degradation baseline output is The total loss can then be written as:

[0138]

[0139] The data item is used to measure the fitting error between the prediction and the label. This invention uses the mean absolute error (MAE) form:

[0140]

[0141] To prevent the residual branches from drifting freely in the early stages of training, which could lead to non-physical shapes in the prediction curve, this invention introduces a physical consistency loss to constrain the prediction results to the vicinity of the physical prior trend.

[0142]

[0143] in, Operators that retain the same value but do not return gradients:

[0144]

[0145]

[0146] therefore, During backpropagation, only for The branch in question exerts a constraint effect, but does not drive the physical prior baseline. Reverse updates ensure that the physical prior always serves as a stable trend reference during training.

[0147] To ensure the reasonableness of the physical values ​​of the predicted results, this invention introduces a boundary penalty term to limit the output range of SOH. Considering that the reasonable range of SOH is [0.8, 1.0], and allowing a small numerical margin, the following definition is provided:

[0148]

[0149] in, The range of values ​​for the apparent activation parameter can be constrained as follows:

[0150]

[0151] Physical consistency weight Employing a warmup and linear decay strategy, before training Take the larger epoch. To strengthen prior constraints, which are then gradually reduced as training progresses. This allows the model to rely more on the details of the data fit in later stages. Its expression is:

[0152]

[0153]

[0154] The model was trained using the AdamW optimizer, with an initial learning rate of 0.0003 and a weight decay coefficient of [value missing]. The mini-batch training size was 16, and the maximum number of training epochs was 100. To avoid overfitting, dropout and early stopping strategies were used during training, with dropout set to 0.5 and early stopping patience value set to 30. The physical consistency weight adopted a warmup and linear decay strategy, with a value of 0.8 for the first 20 epochs, and then gradually decayed to 0.1. The other main hyperparameter settings are shown in Table 2.

[0155] Table 2 PGRN Hyperparameter Settings Batch size 16 Small batch training Epochs 100 Maximum number of training rounds Learning rate 0.0003 AdamW Initial Learning Rate Weight decay 1e-3 L2 regularization Early stopping patience 30 Number of early stops Warmup epochs 20 Physical weight preheating rounds λphy 0.8→0.1 linear decay λbound 0.2 Boundary penalty weight λEa 0.001 Parameter constraint weights Dropout 0.5 Preventing overfitting Hidden dim 48 Initial Hidden Dimensions of Data Branches

[0156] Step 7: Input the sample to be estimated into the trained model and output the SOH estimate under the corresponding scalar conditions.

[0157] This invention employs a "leave-one-battery-within-group testing and training set time-segmentation verification" strategy in its data partitioning and training process. It also combines rate-conditionalized input with physically constrained weighted annealing to comprehensively evaluate generalization capabilities across different rate scenarios. The test set construction utilizes a leave-one-battery-within-the-group (LOBO) strategy. Under the same temperature conditions, one battery is selected as the test set at each step, while the remaining batteries serve as the training source.

[0158] In the partitioning of the validation set within the training set, a time-sharing method based on cycle order is adopted. The specific partitioning rule is as follows: under the same temperature conditions, three batteries within that temperature group are selected as experimental subjects. Two of the batteries are sorted in ascending order of cycle count and then partitioned in time. The first 80% of the cycle samples of each battery are used as the training set, the last 20% of the cycle samples are used as the validation set, and the full cycle samples of the remaining battery are used as a separate test set, thus achieving independent partitioning of samples for training, validation, and testing.

[0159] The rate conditional variable is explicitly constructed and used as part of the model input. Each sample carries the true rate information corresponding to that cycle in addition to its feature vector. The model establishes a mapping between SOH and (x, cycle, T, rate) through conditional learning.

[0160] This invention selects data at three rates (1C, 2C, and 4C) at 25℃ and at three rates (0.5C, 1C, and 2C) at 55℃ to conduct SOH estimation verification. Both the 25℃ and 55℃ temperature groups follow the aforementioned dataset partitioning rule of "80 / 20 split of two batteries (training / validation) and full testing of one battery" to ensure the consistency and reproducibility of the experimental design across different temperature groups. To evaluate the model's cross-rate generalization ability under the same temperature conditions, a three-fold cross-validation strategy is used within each temperature group. Each time, the battery data corresponding to one rate is used as the test set, while the data from the other two rates are used together as the training and validation sets.

[0161] Taking 25℃ as an example, the three-fold division scheme is as follows:

[0162] The 1C and 2C data were used as the training and validation sets, respectively, and the 4C data was used as the test set.

[0163] The 1C and 4C data were used as the training and validation sets, respectively, and the 2C data was used as the test set.

[0164] 2C and 4C data were used as the training and validation sets, respectively, and 1C data was used as the test set.

[0165] The division method under 55℃ operating conditions is the same as above, that is, one of the three magnifications of 0.5C, 1C and 2C is reserved for testing, and the other two magnifications are used for training and verification.

[0166] To verify the effectiveness of the method, this invention constructs comparative models such as Lightweight Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGB), Extreme Random Tree (ET), and Ordinary Neural Network (NN).

[0167] like Figures 3 to 8 The figures shown are comparisons of the model's prediction results at different temperatures and magnifications. Figures 3 to 8 As can be seen, the implementation results after adopting the method of the present invention show that the model proposed in this invention can more stably track the long-term degradation trend of SOH under multi-rate and cross-cell conditions, and effectively suppress the estimation error caused by rate difference, late drift and systematic bias. It is especially suitable for online health status assessment and life management under complex working conditions.

Claims

1. A method for estimating the state of health (SOH) of a lithium-ion battery based on multi-rate discharge data, characterized in that, Includes the following steps: Step 1: Under the set ambient temperature, perform a multi-rate cycle discharge test on the target lithium-ion battery to obtain voltage data, current data and discharge capacity data for multiple cycle stages; Step 2: Based on the discharge capacity and corresponding initial capacity of each cycle under the corresponding rate conditions, construct a rate-related real health status (SOH) label; Step 3: Perform unified voltage grid resampling and smoothing preprocessing on the capacity-voltage data for each cycle to obtain the capacity increment curve; Step 4: Extract relevant features characterizing battery degradation based on the capacity increment curve, and screen the key feature set for SOH estimation through correlation analysis; Step 5: Align the key feature set, number of iterations, multiplier condition, and temperature condition samples to construct supervised learning samples and apply consistency constraints. Step 6: Establish a physical-guided residual network, outputting a trend baseline with physical degradation priors, outputting residual compensation with neural networks, and combining physical consistency constraints with conditional residual learning. Step 7: Input the sample to be estimated into the trained model and output the SOH estimate under the corresponding scalar conditions.

2. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that... Step 1 specifically includes a periodic reference cycle strategy in the multi-rate cyclic discharge experiment, where a low-rate reference discharge cycle is inserted after several accelerated rate discharge cycles. The periodic reference cycle strategy involves inserting one low-rate discharge cycle after nine accelerated rate discharge cycles, where the low-rate discharge cycle uses a 1 / 6C discharge rate. The initial capacity under different rate conditions is determined by the rate: the initial capacity at the 1 / 6C low rate is the average of the discharge capacity of three complete 1 / 6C discharge cycles; the initial capacity at other accelerated rates is the average of the capacity of three complete discharge cycles at that rate at the beginning of the experiment. The capacity of the current complete discharge data at each different rate is taken as the true value for that rate.

3. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that... Step 2 specifically includes: defining the SOH of the c-th cycle under the rate r condition as the discharge capacity of that cycle. With this ratio, the rated initial capacity The ratio; The discharge capacity is calculated by integrating the discharge current under the corresponding rate conditions in this cycle, or by directly using the discharge capacity reading recorded by the test equipment. The State of Emergency (SOH) at different rates is not forcibly converted to a unified reference rate, but is instead used as the rate-dependent real SOH in the modeling.

4. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that... Step 3 specifically includes: setting the voltage step size of the unified voltage grid to 5mV, and constructing the voltage grid within the effective voltage range of each cycle. And by interpolation, the capacity is mapped onto a uniform voltage grid to obtain The discrete IC values ​​are approximated using finite difference, and the capacity increment at internal grid points is calculated using central difference. Boundary points are filled using forward or backward difference. The original discrete Q-V records are converted into a sequence of IC curves on a unified voltage grid. Smoothing preprocessing employs Savitzky-Golay filtering, maintaining the same set of smoothing parameters across all cyclic samples to suppress high-frequency noise amplified by derivatives and preserve morphological information such as peak position, peak height, and peak width.

5. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that... Step 4 specifically includes: constructing a set of statistical and morphological features based on the IC curve of each cycle, mainly including voltage range area features, main peak correlation features, and centroid voltage and other indicators. All features are calculated under the same voltage grid and consistent preprocessing procedure. Then, the extracted health feature parameters are screened using Pearson correlation analysis and scoring ranking methods to obtain a key feature set with higher correlation to SOH changes and lower redundancy.

6. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that... Step 5 specifically includes: using a cycle as the basic sample unit, meaning each sample corresponds to an IC feature vector and its SOH label within a cycle. For the c-th cycle, construct a sample group: ,in The feature set extracted from this cyclic IC curve, The temperature condition variable corresponding to this cycle. This is the condition variable for the multiplier corresponding to this loop. The SOH label is calculated for the corresponding magnification. It is ensured that the SOH label corresponds one-to-one with the feature vector, cycle number, temperature condition, and magnification condition. Simultaneously, abnormal samples are removed entirely. These abnormal samples include at least SOH abnormal samples, capacity record abnormal samples, and samples whose incremental capacity curve quality does not meet requirements. Furthermore, the corresponding features, magnification, cycle number, and label are deleted simultaneously when an abnormal sample is removed.

7. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that... The physical-guided residual network in step 6 includes a physical prior branch and a residual learning branch; The physical prior branch outputs the SOH trend baseline using a function that includes the Arrhenius temperature factor and an exponential-linear hybrid degeneracy form. The residual learning branch outputs the compensation amount based on the key feature set, number of iterations, multiplier conditions, and temperature conditions. The final output of the model is the sum of the trend baseline and the compensation amount.

8. The method for estimating the state of health (SOH) of a lithium-ion battery based on multi-rate discharge data as described in claim 1, characterized in that... The physical consistency constraints in step 6 include at least one of the following: the constraint that SOH decreases overall with the number of iterations, the SOH output boundary constraint, and the residual compensation magnitude constraint, in order to reduce the risk of unreasonable drift and overfitting during model training.

9. The lithium-ion battery state of health (SOH) estimation method based on multi-rate discharge data as described in claim 1, characterized in that: In steps 6 and 7, the model is trained and validated using a leave-one-cell cross-validation method within the same temperature group, and the estimated value of SOH is finally obtained to determine the generalization performance of the physical guided residual network in SOH estimation across rate and across cells.

Images