A method for estimating the health status of lithium-ion batteries based on fragment multi-charge feature fusion

Abstract

This invention relates to a method for estimating the state of health (SQH) of lithium-ion batteries based on the fusion of multi-segment charging features. The method includes: acquiring charging voltage data, corresponding time data, and maximum discharge capacity data during lithium-ion battery charge-discharge cycles; extracting equal voltage difference time data, equal time difference voltage data, and battery SQH data for each cycle's voltage and time data; processing the extracted equal voltage difference time data and equal time difference voltage data using canonical correlation analysis to extract fused feature data, and combining this with the battery SQH data to form a lithium-ion battery SQH dataset, which is then divided into a training set and a test set; establishing a long short-term memory (LSTM) recurrent neural network model; training the model using the training set and adjusting the model parameters; and testing the model's estimation accuracy using the test set. This invention improves the accuracy of lithium-ion battery SQH estimation, reduces the impact of data acquisition errors and model complexity, and accelerates the estimation speed.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery technology, and in particular to a method for estimating the state of health of lithium-ion batteries based on the fusion of multi-segment charging features. Background Technology

[0002] Lithium-ion batteries are widely used in various fields, such as electric vehicles, electrochemical energy storage, and aerospace. However, in practical applications, the performance of lithium-ion batteries continuously degrades, and their usable capacity decreases. Therefore, it is necessary to continuously estimate the battery's state of health during operation.

[0003] Currently, methods for estimating the state of health (SQH) of lithium-ion batteries mainly include model-based methods and data-driven methods. Model-based methods require first constructing an equivalent circuit model or electrochemical model of the battery, then identifying model parameters using methods such as least squares, and finally estimating the battery SQH using methods such as Kalman filtering and particle filtering. The estimation accuracy of these methods depends on the accuracy of the model and its parameter identification. Data-driven methods do not require the construction of an equivalent circuit model of the battery. Existing data-driven SQH estimation methods based on segmented charging data generally extract equal voltage difference time features or equal time difference voltage features as input to a deep learning model to estimate the SQH. However, extracting time or voltage features alone results in relatively singular features, and the features have high similarity, leading to model overfitting and affecting the accuracy of SQH estimation. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for estimating the health status of lithium-ion batteries based on the fusion of multi-segment charging features.

[0005] To achieve the above objectives, the technical solution of the present invention is: a method for estimating the state of health of a lithium-ion battery based on the fusion of multiple charging features of different segments, comprising the following steps:

[0006] Step 1: Perform multiple charge-discharge cycles on the lithium-ion battery, collect the charging voltage data and corresponding time data for each charge-discharge cycle, as well as the maximum discharge capacity data for that cycle.

[0007] Step 2: For the charging voltage and time data of each charge-discharge cycle, extract equal voltage difference time data, equal time difference voltage data, and battery health status data respectively. Specifically: To extract equal voltage difference time data, first determine the starting voltage, ending voltage, and sampling voltage interval. Starting from the starting voltage, extract time data once every sampling voltage until the ending voltage to obtain equal voltage difference time data; To extract equal time difference voltage data, first determine the starting voltage, ending voltage, and sampling time interval. Starting from the starting voltage, extract voltage data once every sampling time until the ending voltage to obtain equal time difference voltage data; To extract battery health status data, divide the maximum discharge capacity of each cycle by the rated capacity to obtain the battery health status data for the current cycle.

[0008] Step 3: Process the equal voltage difference time data and equal time difference voltage data extracted in Step 2 using typical correlation analysis to extract fusion feature data, and combine it with the battery health status data of the corresponding cycle to form a lithium-ion battery health status dataset. Divide the dataset into a training set and a test set.

[0009] Step 4: Establish a long short-term memory recurrent neural network model. The input of this model is set to fused feature data, the output is set to the health status of the lithium-ion battery, and the initial parameters of the long short-term memory recurrent neural network model are set.

[0010] Step 5: Train the long short-term memory recurrent neural network model using the training set, and adjust the model parameters according to the estimation error to reduce the estimation error; use the test set to test the estimation accuracy of the model and evaluate the estimation effect of the model.

[0011] In one embodiment of the present invention, step 1 includes the following sub-steps:

[0012] Step 101: Use a battery charge and discharge tester to perform charge and discharge cycles on the newly manufactured lithium-ion battery. The charging method is constant current and constant voltage charging, and the discharging method is constant current discharging. The charge and discharge cycle experiment ends when the maximum discharge capacity of the lithium-ion battery drops to 70% of the rated capacity. The total number of cycles is defined as D.

[0013] Step 102: Record the charging voltage data and corresponding time data during the constant current charging process of each cycle, and record the maximum discharge capacity data of this cycle.

[0014] In one embodiment of the present invention, step 2 includes the following sub-steps:

[0015] Step 201: When extracting the equal voltage difference time data, for each loop, determine the starting voltage V0 and ending voltage V of the segment charging voltage. nExtract the starting voltage V0 and the ending voltage V n The corresponding time T0 and And determine the sampling voltage interval Δv as

[0016]

[0017] Where n is the number of features in the equal voltage difference time data;

[0018] Starting from the initial voltage V0, for each additional sampling voltage interval Δv, the voltage value and corresponding time are recorded, i.e., for Obtain voltage data sequence and the corresponding time series The time difference corresponding to the equal voltage interval during the charging process is calculated based on the time series, i.e. This allows for the extraction of isovoltage difference time data features for a single charging cycle.

[0019] Step 202: When extracting the equal time difference voltage data, for each cycle, determine the starting voltage V0 of the segment charging voltage and its corresponding time T0, and determine the sampling time interval Δt. Based on the number of features n in step 201, determine the termination time.

[0020] Starting from the initial voltage V0, for each additional sampling time interval Δt, the time and the corresponding voltage value are recorded, i.e., for Obtain the corresponding time series and voltage data sequence The voltage difference corresponding to equal time intervals during the charging process is calculated based on the voltage data sequence, i.e. This allows for the extraction of isochronous voltage data features from a single charging cycle.

[0021] Step 203: When extracting battery health status data, for each cycle, based on the maximum discharge capacity C of each cycle... max Given the battery's rated capacity C0, calculate the lithium-ion battery's state of health for each cycle as follows:

[0022]

[0023] In one embodiment of the present invention, step 3 includes the following sub-steps:

[0024] Step 301: For the 2n features extracted in Step 2, namely the isoelectric difference time data features Features of Equal Time Difference Voltage Data The feature data were dimensionality reduced by using canonical correlation analysis, and a multivariate random variable Z = (X, Y) was constructed, where X is the feature matrix with a dimension of (2n, D), where 2n is the number of features and D is the number of cycles, and Y is the health status data of lithium-ion batteries with a dimension of (1, D).

[0025] Step 302: Construct the linear transformation P = a T X and Q = b T Y performs standardization on the original data to obtain the standardized matrix Z of the multivariate random variable Z = (X, Y). * And calculate the normalized matrix Z. * The covariance matrix is ​​obtained as follows:

[0026]

[0027] Among them, S XX S XY S YX S YY This is the result of covariance calculation;

[0028] To maximize the correlation between P and Q, we need to choose vectors a and b that maximize the objective function, i.e.:

[0029]

[0030] Define the Rayleigh entropy matrix as:

[0031]

[0032] Define u j and λ j Let R be the j-th eigenvector and the square root of its eigenvalues, and thus solve the maximization problem. Vectors a and b can be represented as:

[0033]

[0034] Based on vectors a and b, calculate P = a T X and Q = b T Y, yielding the final fused feature vector FF, with dimension D, i.e.:

[0035] FF = XP + YQ

[0036] Step 303: Based on the fused features and lithium-ion battery health status data, construct a lithium-ion battery health status dataset:

[0037]

[0038] Use N data samples from the dataset Data as the training set Data1, and the remaining data samples as the test set Data2.

[0039] In one embodiment of the present invention, the long short-term memory recurrent neural network model in step 4 has 1 input node, 1 output node, and the optimization algorithm is Adam.

[0040] In one embodiment of the present invention, step 5 includes the following sub-steps:

[0041] Step 501: Normalize the fused feature data and health status data in the training set Data1. Use the normalized fused feature data as the input of the long short-term memory recurrent neural network model and the normalized health status data as the output of the long short-term memory recurrent neural network model to train the model.

[0042] Step 502: After normalizing the fused feature data in the test set Data2, input it into the trained long short-term memory recurrent neural network model, and perform inverse normalization on the health status estimate output by the model. Compare it with the health status data in the test set Data2, calculate the root mean square error RMSE and the mean absolute error MAE, and evaluate the accuracy of the long short-term memory recurrent neural network model.

[0043] Step 503: If the RMSE and MAE calculated in step 502 do not meet the expected targets, return to step 4 to readjust the model parameters and train until the model's test error meets the expected targets.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] This invention extracts equal voltage difference time feature data and equal time difference voltage feature data from fragmented charging data of lithium-ion batteries. Compared with single time features or voltage features, it can extract more battery aging information and is less affected by battery charging data acquisition errors. Canonical correlation analysis is used to reduce the dimensionality of the feature data to obtain fused features, which can reduce feature data redundancy and improve the accuracy of lithium-ion battery health state estimation. Since the fused features are one-dimensional data, the input nodes of the long short-term memory recurrent neural network model only need to be set to 1, which can effectively reduce model complexity and speed up model estimation. Attached Figure Description

[0046] Figure 1 This is a flowchart of a lithium-ion battery health state estimation method based on the fusion of multi-segment charging features proposed in this invention. Detailed Implementation

[0047] To describe the advantages and features of the present invention in more detail, the following description will be provided in conjunction with the accompanying drawings and specific embodiments. The described embodiment is one example of the present invention; the present invention can also be implemented in other ways, all of which fall within the scope of protection of the present invention.

[0048] In one specific embodiment, such as Figure 1 As shown, a method for estimating the state of health of a lithium-ion battery based on the fusion of multiple charging features of different segments includes the following steps:

[0049] 1. Use a battery charge-discharge tester to perform charge-discharge cycles on newly manufactured lithium-ion batteries. The charging method is constant current and constant voltage charging, and the discharging method is constant current discharging. The charge-discharge cycle experiment ends when the maximum discharge capacity of the lithium-ion battery drops to 70% of the rated capacity. Define the total number of cycles as D. Record the charging voltage data and corresponding time data during the constant current charging process of each cycle, and record the maximum discharge capacity data of this cycle.

[0050] 2. For the voltage and time data of each cycle, extract the equal voltage difference time data, equal time difference voltage data, and battery health status data respectively.

[0051] When extracting equal voltage difference time data, for each loop, determine the starting voltage V0 and ending voltage V of the segment charging voltage. n Extract the starting voltage V0 and the ending voltage V n The corresponding time T0 and And determine the sampling voltage interval Δv as

[0052]

[0053] Where n is the number of features in the equal voltage difference time data;

[0054] Starting from the initial voltage V0, for each additional sampling voltage interval Δv, the voltage value and corresponding time are recorded, i.e., for Voltage data sequence can be obtained and the corresponding time series The time difference corresponding to the equal voltage interval during the charging process is calculated based on the time series, i.e. This allows for the extraction of isovoltage difference time data features for a single charging cycle.

[0055] When extracting equal-time-difference voltage data, for each cycle, determine the starting voltage V0 of the segment charging voltage and its corresponding time T0, and determine the sampling time interval Δt. Based on the number of features n in step 201, determine the termination time.

[0056] Starting from the initial voltage V0, for each additional sampling time interval Δt, the time and the corresponding voltage value are recorded, i.e., for The corresponding time series can be obtained. and voltage data sequence The voltage difference corresponding to equal time intervals during the charging process is calculated based on the voltage data sequence, i.e. This allows for the extraction of isochronous voltage data features from a single charging cycle.

[0057] When extracting battery health status data, for each cycle, the maximum discharge capacity C of each cycle is used. max Given the battery's rated capacity C0, calculate the lithium-ion battery's state of health for each cycle as follows:

[0058]

[0059] 3. Regarding the 2n features extracted in step 2, namely the isoelectric difference time data features... Features of Equal Time Difference Voltage Data The feature data were dimensionality reduced by using canonical correlation analysis to construct a multivariate random variable Z = (X, Y), where X is the feature matrix with dimensions (2n, D), where 2n is the number of features, D is the number of cycles, and Y is the health status data of lithium-ion batteries.

[0060] Construct a linear transformation P = a T X and Q = b T Y performs standardization on the original data to obtain the standardized matrix Z of the multivariate random variable Z = (X, Y). * And calculate the normalized matrix Z. * The covariance matrix is ​​obtained as follows:

[0061]

[0062] Among them, S XX S XY S YX S YY This is the result of covariance calculation.

[0063] To maximize the correlation between P and Q, we need to choose appropriate vectors a and b that maximize the objective function, i.e.:

[0064]

[0065] Define the Rayleigh entropy matrix as:

[0066]

[0067] Define u jand λ j Let R be the j-th eigenvector and the square root of its eigenvalues, and thus solve the maximization problem. Vectors a and b can be represented as:

[0068]

[0069] Based on vectors a and b, P = a can be calculated. T X and Q = b T Y, thus obtaining the final fused feature vector FF, whose dimension is D, i.e.:

[0070] FF = XP + YQ

[0071] Based on the fused features and lithium-ion battery health status data, a lithium-ion battery health status dataset is constructed:

[0072]

[0073] Use N data samples from the dataset Data as the training set Data1, and the remaining data samples as the test set Data2.

[0074] 4. Establish a Long Short-Term Memory Recurrent Neural Network (LSTM) model. The input of this model is set to fused feature data, with 1 input node. The output is set to the health status of the lithium-ion battery, with 1 output node. Set the initial parameters of the LSM model: single hidden layer, 250 initial nodes, Adam optimization algorithm, 120 initial iterations, and an initial learning rate of 0.005. After 90 iterations, the learning rate becomes 0.2.

[0075] 5. Train the Long Short-Term Memory Recurrent Neural Network (LSTM) model using the training set, and adjust the model parameters based on the estimation error to reduce the estimation error. Normalize the fused feature data and health status data in the training set Data1. Use the normalized fused feature data as the input to the LTM model and the normalized health status data as the output to train the model.

[0076] The above data normalization method is as follows:

[0077]

[0078] Where x k Let x be the k-th value of the sequence x, which represents the feature data and health status data in the text, where 0 < k ≤ D. min and x max These are the minimum and maximum values ​​in the x sequence, respectively.

[0079] The estimation accuracy of the model is tested using a test set to evaluate its estimation performance. After normalizing the fused feature data in the test set Data2, it is input into the trained Long Short-Term Memory Recurrent Neural Network (LSN) model. The health status estimates output by the model are then denormalized and compared with the health status data in the test set Data2. The root mean square error (RMSE) and mean absolute error (MAE) are calculated using the following formulas to evaluate the accuracy of the LSN model.

[0080]

[0081]

[0082] in S is the estimated health status value, and S is the actual health status value.

[0083] If the calculated RMSE and MAE do not meet the expected targets, return to 4 to readjust the model parameters and train until the model's test error meets the expected targets.

[0084] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for estimating the state of health of a lithium-ion battery based on the fusion of multiple charging features from different segments, characterized in that, Includes the following steps: Step 1: Perform multiple charge-discharge cycles on the lithium-ion battery, collect the charging voltage data and corresponding time data for each charge-discharge cycle, as well as the maximum discharge capacity data for that cycle. Step 2: For the charging voltage and time data of each charge-discharge cycle, extract equal voltage difference time data, equal time difference voltage data, and battery health status data respectively. Specifically: To extract equal voltage difference time data, first determine the starting voltage, ending voltage, and sampling voltage interval. Starting from the starting voltage, extract time data once every sampling voltage interval until the ending voltage to obtain equal voltage difference time data; To extract equal time difference voltage data, first determine the starting voltage, ending voltage, and sampling time interval. Starting from the starting voltage, extract voltage data once every sampling time interval until the ending voltage to obtain equal time difference voltage data; To extract battery health status data, divide the maximum discharge capacity of each cycle by the rated capacity to obtain the battery health status data for the current cycle. Step 3: Process the equal voltage difference time data and equal time difference voltage data extracted in Step 2 using typical correlation analysis to extract fusion feature data, and combine it with the battery health status data of the corresponding cycle to form a lithium-ion battery health status dataset. Divide the dataset into a training set and a test set. Step 4: Establish a long short-term memory recurrent neural network model. The input of this model is set to fused feature data, the output is set to the health status of the lithium-ion battery, and the initial parameters of the long short-term memory recurrent neural network model are set. Step 5: Train the long short-term memory recurrent neural network model using the training set, and adjust the model parameters according to the estimation error to reduce the model estimation error; The estimation accuracy of the model is tested using a test set to evaluate the model's estimation performance.

2. The lithium-ion battery health state estimation method based on fragment multi-charge feature fusion according to claim 1, characterized in that, Step 1 It includes the following steps: Step 101: Use a battery charge and discharge tester to perform charge and discharge cycles on the newly manufactured lithium-ion battery. The charging method is constant current and constant voltage charging, and the discharging method is constant current discharging. The charge and discharge cycle experiment ends when the maximum discharge capacity of the lithium-ion battery drops to 70% of the rated capacity. The total number of cycles is defined as D. Step 102: Record the charging voltage data and corresponding time data during the constant current charging process of each cycle, and record the maximum discharge capacity data of this cycle.

3. The lithium-ion battery health state estimation method based on fragment multi-charge feature fusion according to claim 1, characterized in that, Step 2 includes the following sub-steps: Step 201: When extracting the equal voltage difference time data, for each loop, determine the starting voltage V0 and ending voltage V of the segment charging voltage. n Extract the starting voltage V0 and the ending voltage V n The corresponding time T0 and And determine the sampling voltage interval Δv as Where n is the number of features in the equal voltage difference time data; Starting from the initial voltage V0, for each additional sampling voltage interval Δv, the voltage value and corresponding time are recorded, i.e., for V1 (1) =V0+Δv, Obtain voltage data sequence and the corresponding time series The time difference corresponding to the equal voltage interval during the charging process, ΔT1, is calculated based on the time series. (1) =T1 (1) -T0, This allows for the extraction of isovoltage difference time data features for a single charging cycle. Step 202: When extracting the equal time difference voltage data, for each cycle, determine the starting voltage V0 of the segment charging voltage and its corresponding time T0, and determine the sampling time interval Δt. Based on the number of features n in step 201, determine the termination time. Starting from the initial voltage V0, for each additional sampling time interval Δt, the time and the corresponding voltage value are recorded, i.e., for T1 (2) =T0+Δt, Obtain the corresponding time series and voltage data sequence Calculate the voltage difference ΔV1 corresponding to equal time intervals during the charging process based on the voltage data sequence. (2) =V1 (2) -V0, This allows for the extraction of isochronous voltage data features from a single charging cycle. Step 203: When extracting battery health status data, for each cycle, based on the maximum discharge capacity C of each cycle... max Given the battery's rated capacity C0, calculate the lithium-ion battery's state of health for each cycle as follows:

4. The lithium-ion battery health state estimation method based on fragment multi-charge feature fusion according to claim 3, characterized in that, Step 3 includes the following sub-steps: Step 301: For the 2n features extracted in Step 2, namely the isoelectric difference time data features Features of Equal Time Difference Voltage Data The feature data were dimensionality reduced by using canonical correlation analysis, and a multivariate random variable Z = (X, Y) was constructed, where X is the feature matrix with a dimension of (2n, D), where 2n is the number of features and D is the number of cycles, and Y is the health status data of lithium-ion batteries with a dimension of (1, D). Step 302: Construct the linear transformation P = a T X and Q = b T Y performs standardization on the original data to obtain the standardized matrix Z of the multivariate random variable Z = (X, Y). * And calculate the normalized matrix Z. * The covariance matrix is ​​obtained as follows: Among them, S XX S XY S YX S YY This is the result of covariance calculation; To maximize the correlation between P and Q, we need to choose vectors a and b that maximize the objective function, i.e.: Define the Rayleigh entropy matrix as: Define u j and λ j Let R be the j-th eigenvector and the square root of its eigenvalues, and thus solve the maximization problem. Vectors a and b can be represented as: Based on vectors a and b, calculate P = a T X and Q = b T Y, yielding the final fused feature vector FF, with dimension D, i.e.: FF = XP + YQ Step 303: Based on the fused features and lithium-ion battery health status data, construct a lithium-ion battery health status dataset: Use N data samples from the dataset Data as the training set Data1, and the remaining data samples as the test set Data2.

5. The lithium-ion battery health state estimation method based on fragment multi-charge feature fusion according to claim 1, characterized in that, In step 4, the long short-term memory recurrent neural network model has 1 input node and 1 output node, and the optimization algorithm is Adam.

6. The lithium-ion battery health state estimation method based on fragment multi-charge feature fusion according to claim 4, characterized in that, Step 5 includes the following sub-steps: Step 501: Normalize the fused feature data and health status data in the training set Data1. Use the normalized fused feature data as the input of the long short-term memory recurrent neural network model and the normalized health status data as the output of the long short-term memory recurrent neural network model to train the model. Step 502: After normalizing the fused feature data in the test set Data2, input it into the trained long short-term memory recurrent neural network model, and perform inverse normalization on the health status estimate output by the model. Compare it with the health status data in the test set Data2, calculate the root mean square error RMSE and the mean absolute error MAE, and evaluate the accuracy of the long short-term memory recurrent neural network model. Step 503: If the RMSE and MAE calculated in step 502 do not meet the expected targets, return to step 4 to readjust the model parameters and train until the model's test error meets the expected targets.

Images