A novel health feature-based method for estimating the health status of lithium-ion batteries

Abstract

This invention presents a novel health status estimation method for lithium-ion batteries based on novel health features. First, it obtains the charging voltage curve under constant current and constant voltage charging conditions, using the voltage difference at equal starting voltages and time intervals as the novel health feature. Then, since the starting voltage and time interval are two parameters affecting the novel health feature, they are optimized, resulting in an optimized starting voltage and multiple optimized time intervals from the same charging voltage curve. Combining the optimized starting voltage with each optimized time interval yields multiple novel health features extracted from the same charging voltage curve. Finally, a lithium battery health status estimation model is constructed based on a multiple linear regression model. The model is trained using the extracted novel health features and corresponding health status values, and the trained model is then used for lithium battery health status estimation. This method has low computational complexity and significantly improves estimation speed while maintaining accuracy, making it applicable even when current battery management systems have insufficient computing power.

Description

Technical Field

[0001] This invention belongs to the field of battery management system health status estimation technology, specifically relating to a method for estimating the health status of lithium-ion batteries based on novel health features. Background Technology

[0002] The development of lithium-ion battery technology has led to the rapid growth of the electric vehicle industry. However, lithium-ion batteries experience performance and capacity degradation during long-term cycling, and improper operation can result in fires or explosions. To ensure the reliable and safe operation of lithium batteries, the Battery Management System (BMS) needs to monitor the battery's State of Health (SOH). SOH reflects the battery's current ability to store and supply energy / power relative to the beginning of its lifespan, and is typically characterized by the battery's available capacity. When the SOH declines below a certain level, the battery is considered to have reached its maximum lifespan and needs to be replaced. Therefore, assessing the battery's state of health is crucial for the safe operation of the system.

[0003] On the one hand, in existing research on SOH estimation for lithium-ion batteries, the development and application of intelligent algorithms have greatly improved estimation accuracy. However, these algorithms are highly complex, especially those based on deep learning, which require significant computational resources. Since current battery management systems are primarily microcontroller-based, the memory space available for computation is limited, making it difficult to deploy these intelligent algorithms within battery management systems. On the other hand, most algorithms directly use complete charging data (SOH data from 0% to 100%) as input or for feature extraction. However, in actual battery use, the initial and final states of charge during charging are difficult to fix due to operating conditions, resulting in fragmented charging data and compromised estimation accuracy. Based on these two reasons, simplifying the model and reducing computational load are crucial factors for deploying estimation algorithms within battery management systems without sacrificing accuracy. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the technical problem to be solved by the present invention is to propose a method for estimating the health status of lithium-ion batteries based on novel health characteristics.

[0005] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows: The present invention provides a method for estimating the health status of lithium-ion batteries based on novel health characteristics, specifically including the following steps:

[0006] Step 1: Obtain the charging voltage curve under constant current and constant voltage charging conditions. Select the voltage difference at equal starting voltage and equal time intervals as a novel health characteristic, expressed as:

[0007] ΔV=V(t+N)-V0(t) (1)

[0008] Where V0(t) represents the initial voltage, V(t+N) represents the charging voltage at time t+N, and N represents the time interval;

[0009] Step 2: Analysis shows a linear relationship between the new health characteristics and the health status of lithium batteries, which is verified using the correlation coefficient method.

[0010] Step 3: Since different lithium batteries have different charge and discharge platforms, the intervals on the charging voltage curve where the new health features and the health status of the lithium battery are linearly related are different. In order to determine the intervals where the linear relationship is found, the starting voltage and time interval are optimized to obtain an optimized starting voltage and multiple optimized time intervals from the same charging voltage curve. The optimized starting voltage and each optimized time interval are combined to obtain multiple new health features extracted from the same charging voltage curve.

[0011] Step 3-1, Optimization of starting voltage: Set the selection range of starting voltage, traverse the voltage values ​​within the range at certain intervals, and roughly select the time interval. Calculate the Pearson correlation coefficient between the new health characteristics obtained by different starting voltages and time intervals and the health status of the lithium battery. Take the voltage corresponding to the maximum absolute value of the Pearson correlation coefficient as the optimized starting voltage.

[0012] Step 3-2, Time Interval Optimization: Set a time interval range containing multiple time intervals. Combine each time interval with the optimized starting voltage to obtain a new set of health features. Use the new set of health features as input and the lithium battery health status as output for OLS regression analysis to obtain the significance of the linear relationship between each new health feature and the lithium battery health status. Remove the new health feature with the lowest significance, and then perform OLS regression analysis on the remaining features with the health status again. Repeat this process to gradually remove new health features with low significance until the test values ​​of each remaining new health feature are all less than the set threshold. The time intervals corresponding to the remaining new health features are the optimized time intervals.

[0013] Step 4: Select a multiple linear regression model to construct a lithium battery health status estimation model, the expression of which is:

[0014] y≈w T x+b (2)

[0015] Where y = {y1, y2, ..., y i ,…,y m Let} be the lithium battery health state matrix corresponding to m charging voltage curves, x = {(x 11,x 12 ,…,x 1n ),(x 21 ,x 22 ,…,x 2n ),…,(x m1 ,x m2 ,…,x mn )} is the novel health feature matrix corresponding to y, (x i1 ,x i2 ,…,x in ) represents the charging voltage curve of the i-th line within the optimal time interval N1, N2, ..., N. n The new health feature matrix at time, i∈m, b is a constant coefficient matrix, T is the matrix transpose, w is the feature coefficient matrix, is determined by the least squares method according to equation (3);

[0016]

[0017] Where ||·||2 represents the 2-norm;

[0018] The novel health features and corresponding health state values ​​extracted from the charging voltage curve in step three are input into the lithium battery health state estimation model to train the model, and the trained model is used for lithium battery health state estimation.

[0019] Furthermore, for lithium batteries of the same model, the health status can be estimated using equation (2) by obtaining the charging voltage data under the same charging conditions; for lithium batteries of different models or charging voltage data obtained by changing the charging conditions, it is necessary to re-optimize the starting voltage and time interval, and then retrain the lithium battery health status estimation model.

[0020] Compared with the prior art, the beneficial effects of the present invention are:

[0021] 1. By using the voltage difference at equal initial voltage and equal time intervals as a health feature, and based on the linear relationship between the feature and the health state, it is only necessary to find the linearly related interval from the charging voltage curve, rather than using the voltage data of the entire curve. This reduces the computational load. The estimation model uses a multiple linear regression model, which does not require the complex calculation and training of neural networks. Experimental results show that, while ensuring estimation accuracy, the method of this invention improves the estimation speed by more than 2 times compared with the KRR method. It can be deployed in battery management systems and can be applied when the computing power of current battery management systems is insufficient.

[0022] 2. To extract features that exhibit a significant linear relationship with health status, the two influencing parameters—initial voltage and time interval—are optimized. Parameter optimization occurs before model training. Therefore, for lithium batteries of the same model and under the same charging conditions, the optimized parameters can be used directly without further optimization, further improving estimation speed. For different battery models or changes in charging conditions, the parameters need to be re-optimized, and the model retrained before it can be used for health status estimation. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of the present invention;

[0024] Figure 2(a) is a graph showing the change of charging voltage over time for lithium batteries in the Oxford University aging dataset under different state of harmonics (SOH).

[0025] Figure 2(b) is a graph showing the change of charging voltage over time for lithium batteries in NASA's random usage dataset under different state of harmonics (SOH).

[0026] Figure 3(a) shows the actual capacity decay curves of each lithium battery in the Oxford University aging dataset at different cycle numbers;

[0027] Figure 3(b) shows the actual capacity decay curves of each lithium battery in the NASA random use dataset at different cycle numbers;

[0028] Figure 4 These are fitting curves of novel health characteristics and health status for various lithium batteries in the Oxford University Aging Data Set.

[0029] Figure 5(a) shows the Pearson correlation coefficient of each lithium battery in the Oxford University battery aging dataset as a function of the initial voltage.

[0030] Figure 5(b) shows the Pearson correlation coefficient of each lithium battery in the NASA random use dataset as a function of the initial voltage.

[0031] Figure 6(a) shows the health status estimation results of battery No. 4 in the Oxford University aging dataset;

[0032] Figure 6(b) shows the health status estimation results of battery No. 8 in the Oxford University aging dataset. Detailed Implementation

[0033] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of protection of this application.

[0034] The present invention provides a method for estimating the health status of lithium-ion batteries based on novel health characteristics (hereinafter referred to as the method), the specific steps of which are as follows:

[0035] Step 1: Processing and analyzing lithium battery aging data: Obtain the charging voltage curve under constant current and constant voltage charging conditions, and select the voltage difference at equal intervals of the same initial voltage (VDEISV) as a new health characteristic; constant current and constant voltage charging conditions refer to constant current charging to the cutoff voltage, and then constant voltage charging until the current drops to 0.01A;

[0036] This embodiment uses the Oxford University Aging Dataset and the NASA Random Use Dataset as examples for analysis. These two datasets record the charging voltage data during lithium battery use, and after each cycle, the battery capacity is measured using the ampere-hour integration method during the discharge phase to obtain the actual State of Charge (SOH), which can be used for subsequent analysis and verification. Figures 2(a) and (b) are the charging voltage variation curves of lithium batteries under different SOH values ​​in the two datasets, respectively. The SOH is represented by a gradual color change from light to dark as the battery's lifespan decreases. It can be seen from the figures that the change in the lithium battery charging voltage curve is relatively stable due to the fixed charging method, but increases with the aging of the lithium battery. There is a regular pattern of change, specifically: when the lithium battery has been used for a short time, i.e., when the battery's state of harm (SOH) is high (light-colored curve), the process of the charging voltage rising to the cutoff voltage (4.2V) is relatively slow, and the constant current charging time is longer. This is consistent with the fact that the newer the battery, the longer the internal chemical reaction time and the larger the capacity. However, when the lithium battery has been used for a period of time, the SOH decreases (dark-colored curve), the process of the charging voltage rising to the cutoff voltage is accelerated, and the constant current charging time is significantly shortened. At this time, the lithium battery capacity also decreases. Through the analysis of the charging voltage curve, it can be seen that the method of directly estimating the SOH by analyzing and processing the charging voltage curve is feasible.

[0037] The Oxford University Aging Dataset collected data from eight Kokam lithium cobalt oxide pouch batteries with a rated capacity of 740 mA·h. All data were obtained under a constant ambient temperature of 40°C by repeatedly discharging and recharging the lithium-ion batteries at a constant current of 2C (1.48A) using ARTEMIS urban driving conditions. The minimum capacity decay was 0.43 A·h. During charging, the voltage sampling frequency was 1Hz, resulting in 519 complete charging voltage curves. Figure 3(a) shows the actual capacity decay curves of each lithium battery in this dataset at different cycle numbers. From the figure, the latter part of the capacity abrupt changes in Cell2 and Cell5 was removed, resulting in 490 charging voltage curves that can be used for SOH estimation. The NASA Random Use Dataset used LG Chem 18650 cylindrical batteries with a rated capacity of 2.1 A·h and a normal operating voltage range of 3.2–4.2V. Although each experiment in this dataset used different random conditions during discharge, the charging conditions were the same constant current and constant voltage conditions. This embodiment selects eight batteries tested at room temperature for analysis: RW13, RW14, RW15, and RW16 with low-current random discharge, and RW17, RW18, RW19, and RW20 with high-current random discharge. A standard charge-discharge cycle was performed after every 50 cycles. The discharge conditions were 1A constant current to 3.2V for calculating the actual capacity, and the charging conditions were 2A constant current to 4.2V, followed by constant voltage charging until the current dropped to 0.01A. A total of 203 charging voltage curves under constant charging conditions were obtained. NASA's random dataset considered that a standard charge-discharge test was performed after every 50 random discharge cycles. After the 450th cycle, the first sampled voltage value reached 4.0V, and subsequent cycle curves were discarded. This resulted in 80 usable curve data points for the eight batteries. Figure 3(b) shows the actual capacity decay curves of different lithium batteries in this dataset at different cycle counts.

[0038] The voltage difference of the charging voltage curve increasing over time is used as a metric feature of the curve itself. Combined with the time interval of equal charging voltage difference, a new health feature is constructed. That is, the voltage difference ΔV of equal starting voltage and equal time interval is selected from the charging voltage curve as the new health feature, and its expression is:

[0039] ΔV=V(t+N)-V0(t) (1)

[0040] Where V0(t) represents the initial voltage, i.e. the charging voltage at time t; V(t+N) represents the charging voltage at time t+N, and N represents the time interval;

[0041] Step 2: Analyze the relationship between the novel health characteristics and the health status of lithium batteries through curve fitting, and verify it using the correlation coefficient method;

[0042] Based on the inherent characteristics of the feature, it is known that extracting the health status of a lithium battery requires determining two parameters: the initial voltage and the time interval. Therefore, we roughly selected an initial voltage of 3.8V and a time interval of 600Hz, and performed curve fitting between the new health feature and the corresponding lithium battery health status. Figure 4 The figure shows the fitting curves of novel health characteristics and health status of various lithium batteries in the Oxford University Aging Data Set. A significant linear relationship between the two can be clearly seen from the figure.

[0043] The relationship between the novel health characteristics and the health status was verified by using the Pearson correlation coefficient. The absolute values ​​of the Pearson correlation coefficients were all above 0.99. Therefore, the verification showed that there is a linear relationship between the novel health characteristics and the health status of lithium batteries.

[0044] Step 3: Since different lithium batteries have different charge and discharge platforms, the intervals on the charging voltage curve where the novel health features show a significant linear relationship with the lithium battery health status are different. Therefore, in order to accurately determine the intervals where the linear relationship exists, it is necessary to optimize the starting voltage and time interval to obtain optimized starting voltage and time interval. The optimized starting voltage of the same charging voltage curve is combined with different optimized time intervals to obtain multiple novel health features extracted from the same charging voltage curve, which are used to train the lithium battery health status estimation model.

[0045] Step 3-1, Optimization of starting voltage V0: Set the selection range of starting voltage V0, and iterate through the voltage values ​​within the range at certain intervals. At this time, the time interval N is roughly selected. Calculate the Pearson correlation coefficient between the new health characteristics obtained by different starting voltages V0 and corresponding time intervals N and the health status of lithium batteries. The optimization result is the one with the largest Pearson correlation coefficient (showing a negative correlation, i.e., the smallest negative value). The voltage corresponding to the maximum value of the Pearson correlation coefficient is taken as the starting voltage V0 that has the most significant impact on the linear relationship.

[0046] Step 3-1-1: The voltage selection range of the Oxford University aging dataset is 3.75 to 3.81V. The voltage is taken as the starting voltage every 0.01 units. The time interval N is 600Hz. The curve of the Pearson correlation coefficient with the starting voltage is shown in Figure 5(a). It can be seen that the optimized starting voltage is 3.8V.

[0047] Step 3-1-2: NASA randomly selected the voltage range of the dataset from 3.87 to 4.02V, with a time interval N of 4 and a charging voltage sampling interval of 60s (actually 240s). The curve of the Pearson correlation coefficient with the initial voltage is shown in Figure 5(b). It can be seen that the optimized initial voltage is 3.89V.

[0048] Step 3-2, Optimization of Time Interval N: Set a time interval range, which contains multiple time intervals. Each time interval, combined with the optimization starting voltage V0, yields a new health feature, thus creating a set of new health features. Using this set of new health features as input and the lithium battery health status as output, perform OLS (Ordinary Least Squares) regression analysis to obtain the significance of the linear relationship between each feature in the new health feature set and the lithium battery health status. Remove the new health feature with the lowest significance, and then perform OLS regression analysis on the remaining features with the health status again. Repeat this process, gradually removing new health features with lower significance, until the test values ​​of all remaining new health features are less than the set threshold. The time intervals corresponding to the remaining new health features are the optimized time intervals.

[0049] Step 3-2-1: The time interval range of the Oxford University aging dataset is 100 to 1000, with an interval of 100, so it includes a total of 10 time intervals. The results of the first OLS regression analysis are shown in Table 1. The threshold is set to 0.05. According to the principle of stepwise elimination, the time intervals of 300 and 500 are eliminated, and the optimized time intervals are finally obtained as 100, 200, 400, 600, 700, 800, 900, and 1000.

[0050] Table 1. OLS analysis results of the Oxford University aging dataset.

[0051] P>|t| 0.000 0.006 0.143 0.001 0.007 N 600 700 800 900 1000 P>|t| 0.034 0.002 0.019 0.008 0.000

[0052] Step 3-2-2: NASA randomly used the dataset, considering that a standard charge-discharge test was performed after every 50 random discharge cycles. After the 450th cycle, the first sampled voltage value reached 4.0V, and the subsequent cycle curves were discarded. OLS analysis was performed using 80 curves from 8 batteries, with a voltage sampling interval of 60 seconds, N being one sampling interval, and the time interval ranging from 1 to 10. The results of the first OLS regression analysis are shown in Table 2. Considering that a large data sampling interval would lead to an overestimation of the OLS analysis results, a threshold of 0.2 was selected, resulting in optimized time intervals of 1, 2, 3, 5, 6, 7, and 9.

[0053] Table 2. OLS analysis results of NASA's randomly used dataset.

[0054] P>|t| 0.093 0.077 0.223 0.887 0.301 N 6 7 8 9 10 P>|t| 0.062 0.284 0.349 0.310 0.366

[0055] Step 4: Considering the current limitations of computing resources in battery management systems, and based on the relationship between the novel health characteristics obtained in Step 2 and the lithium battery health status, a multiple linear regression (MLR) model is selected as the lithium battery health status estimation model. The expression is as follows:

[0056] y≈w T x+b (2)

[0057] Where y = {y1, y2, ..., y i ,…,y m Let} be the lithium battery health state matrix corresponding to m charging voltage curves, x = {(x 11 ,x 12 ,…,x 1n ),(x 21 ,x 22 ,…,x 2n ),…,(x m1 ,x m2 ,…,x mm )} is the novel health feature matrix corresponding to y, (x i1 ,x i2 ,…,x in ) represents the charging voltage curve of the i-th line in the optimization time interval N1, N2, ..., N n The new health feature matrix at time, i∈m, b is a constant coefficient matrix, T is the matrix transpose, w is the feature coefficient matrix, is determined by the least squares method according to equation (3);

[0058]

[0059] Where ||·||2 represents the second norm;

[0060] The novel health features and corresponding health state values ​​extracted from the charging voltage curve in step three are input into the lithium battery health state estimation model to train the model, and the trained model is used for lithium battery health state estimation.

[0061] For lithium batteries of the same model, the health status can be estimated using equation (2) by obtaining the charging voltage data under the same charging conditions; for lithium batteries of different models or charging voltage data obtained by changing the charging conditions, it is necessary to re-determine the optimized starting voltage and time interval according to step three, and retrain the lithium battery health status estimation model according to step four.

[0062] To verify the effectiveness of the method of this invention, Mean Absolute Error (MAE) and Maximum Error (MAX) were selected as evaluation metrics to quantitatively assess the method, and its estimation accuracy was compared with that of the existing Kernel Ridge Regression (KRR) method. Simultaneously, the computational power was compared through model training and estimation time. The expressions for the two evaluation metrics are as follows:

[0063]

[0064] MAX = max|y j -y′ j | (5)

[0065] In the formula: m is the number of samples to be tested, y j y′ j These represent the estimated and actual values ​​of SOH, respectively.

[0066] To compare the results with those obtained by the Core Ridge Regression (KRR) method, data from batteries 4 and 8 in the Oxford University Aging Dataset were used as test samples, and the remaining six batteries as training samples. NASA randomly used data from its dataset, using RW16 (low-rate discharge) and RW20 (high-rate discharge) batteries as test samples, and the remaining six batteries as training samples. Because NASA's randomly used dataset uses different battery data, only the global quantity evaluation results were compared. The estimation results of the two datasets are compared in Tables 3 and 4, where MLR represents the method of this invention.

[0067] Table 3. SOH estimation results from the Oxford University aging dataset.

[0068]

[0069] Table 4 Evaluation of NASA Random Use Data SOH Estimation Results

[0070]

[0071] Figures 6(a) and (b) show the health status estimation results of batteries 4 and 8 in the Oxford University aging dataset. As can be seen from Figure 6, the estimation results of the two methods are basically consistent, verifying the reliability of the method proposed in this invention. Table 3 shows that compared with the KRR method, the method of this invention not only reduces computational complexity and increases computational speed by more than two times, but also does not suffer much loss in accuracy; the MAX error differs by only 0.01%, and the MAE of global quantities is even reduced, making it applicable even when the computing power of current battery management systems is insufficient. Table 4 shows that for the NASA Random Use dataset, even with a large battery charging voltage sampling interval and less precise data calculation, the method proposed in this invention still achieves good estimation results. The mean absolute error (MAE) of global quantities is 1.65%, and the maximum error (MAX) is 3.80%, showing that both speed and accuracy are improved. This demonstrates the strong generalization ability of this method.

[0072] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for estimating the health status of lithium-ion batteries based on novel health characteristics, characterized in that, The method includes the following steps: Step 1: Obtain the charging voltage curve under constant current and constant voltage charging conditions. Select the voltage difference at equal starting voltage and equal time intervals as a novel health characteristic, expressed as: ΔV=V(t+N)-V0(t) (1) Where V0(t) represents the initial voltage, V(t+N) represents the charging voltage at time t+N, and N represents the time interval; Step 2: Analysis shows a linear relationship between the new health characteristics and the health status of lithium batteries, which is verified using the correlation coefficient method. Step 3: Since different lithium batteries have different charge and discharge platforms, the intervals on the charging voltage curve where the new health features and the health status of the lithium battery are linearly related are different. In order to determine the intervals where the linear relationship is found, the starting voltage and time interval are optimized to obtain an optimized starting voltage and multiple optimized time intervals from the same charging voltage curve. The optimized starting voltage and each optimized time interval are combined to obtain multiple new health features extracted from the same charging voltage curve. Step 3-1, Optimization of starting voltage: Set the selection range of starting voltage, traverse the voltage values ​​within the range at certain intervals, and roughly select the time interval. Calculate the Pearson correlation coefficient between the new health characteristics obtained by different starting voltages and time intervals and the health status of the lithium battery. Take the voltage corresponding to the maximum absolute value of the Pearson correlation coefficient as the optimized starting voltage. Step 3-2, Time Interval Optimization: Set a time interval range containing multiple time intervals. Combine each time interval with the optimized starting voltage to obtain a new set of health features. Use the new set of health features as input and the lithium battery health status as output for OLS regression analysis to obtain the significance of the linear relationship between each new health feature and the lithium battery health status. Remove the new health feature with the lowest significance, and then perform OLS regression analysis on the remaining features with the health status again. Repeat this process to gradually remove new health features with low significance until the test values ​​of each remaining new health feature are all less than the set threshold. The time intervals corresponding to the remaining new health features are the optimized time intervals. Step 4: Select a multiple linear regression model to construct a lithium battery health status estimation model, the expression of which is: y≈w T x+b (2) Where y = {y1, y2, ..., y} i , ..., y m Let} be the lithium battery health state matrix corresponding to m charging voltage curves, x = {(x 11 x 12 , ..., x 1n ), (x 21 x 22 , ..., x 2n ), ..., (x m1 x m2 , ..., x mn )} is the novel health feature matrix corresponding to y, (x i1 x i2 , ..., x in ) represents the charging voltage curve of the i-th line within the optimal time interval N1, N2, ..., N. n The new health feature matrix at time, i∈m, b is a constant coefficient matrix, T is the matrix transpose, w is the feature coefficient matrix, is determined by the least squares method according to equation (3); Where ||·||2 represents the 2-norm; The novel health features and corresponding health state values ​​extracted from the charging voltage curve in step three are input into the lithium battery health state estimation model to train the model, and the trained model is used for lithium battery health state estimation.

2. The lithium-ion battery health status estimation method based on novel health characteristics according to claim 1, characterized in that, For lithium batteries of the same model, the health status can be estimated using equation (2) by obtaining the charging voltage data under the same charging conditions; for lithium batteries of different models or charging voltage data obtained by changing the charging conditions, it is necessary to re-optimize the starting voltage and time interval, and then retrain the lithium battery health status estimation model.

Images