Lithium battery health state estimation method fusing local and global dependency information

Abstract

The invention relates to the technical field of lithium ion batteries, and discloses a lithium battery health state estimation method fusing local and global dependency information. The method comprises the following steps: firstly, collecting capacity increment curve data generated in the charging process of a battery, performing index sliding filtering on an extracted capacity increment curve, and extracting tangent angle characteristics of a voltage curve in combination with information of the capacity increment curve and a charging voltage curve, so as to solve the problems that an IC curve peak value disappears and data is distorted after filtering. And then, fusing a long short-term memory (LSTM) network and an Informer model, constructing a parallel model structure, respectively extracting local dependency and global information in a battery recession process, and realizing comprehensive capture of the local and global dependency information of the battery. And then, a self-attention mechanism layer is added into the model to fuse local and global time sequence information, so that an LSTM-Informer-SA battery health state estimation model is established, and lithium ion battery SOH estimation is realized based on the model.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion batteries, and in particular to a lithium battery health state estimation method that integrates local and global dependency information. Background Art

[0002] Lithium-ion batteries are widely used in the field of pure electric vehicles due to their high energy density, long cycle life and low self-discharge rate. Battery state of health (SOH) is one of the key indicators to ensure the reliability of battery performance. Accurate estimation of SOH can effectively predict the best time to replace the battery and identify performance anomalies in a timely manner. This not only helps to extend battery life, but also reduces potential safety risks. However, due to the complex reaction mechanism of the battery, SOH estimation still faces many challenges. Currently, data-driven methods are widely used in SOH estimation, but such methods require high-quality aging features and suitable algorithm models. In terms of feature extraction, existing methods usually extract aging features such as peak value, peak position and peak area through incremental capacity (IC) curves.

[0003] However, as the battery ages, the peak of the IC curve gradually disappears, and these features will also weaken. In terms of model construction, existing methods include time series prediction models based on recurrent neural networks (RNNs) and attention mechanisms. Traditional RNNs are susceptible to the problem of gradient vanishing or exploding, and can usually only capture short-term dependency information; while models based on attention mechanisms perform well in capturing long-distance dependency information, but have relatively low accuracy in local and short-term information processing. Summary of the invention

[0004] In order to solve the above-mentioned problem of gradient disappearance or explosion, the present invention proposes a lithium battery health state estimation method that integrates local and global dependency information.

[0005] To implement the above technical solution, the details are as follows:

[0006] A lithium battery health state estimation method integrating local and global dependency information, comprising:

[0007] S1. Collect battery power data during the battery charge and discharge cycle; the details are as follows:

[0008] The collected batteries are lithium-ion batteries of the same model;

[0009] The experimental process is as follows: each single cell is subjected to an aging cycle test until the overall capacity of each battery is less than 80% of the nominal capacity;

[0010] The collected battery power data include: real-time recording of battery current, voltage, capacity and data collection time;

[0011] The present invention records eight single-cell battery data in total;

[0012] S2. Extract the IC curve and voltage curve during the battery charging cycle and perform smoothing. The steps are as follows:

[0013] S2.1, dividing charging segment data for each battery according to charging current;

[0014] Since the direction of the current is different during charging and discharging, that is, when the battery is charging, the current is positive, and when it is discharging, the current is negative; therefore, when the current is positive, it is marked as charging data; when the current is negative, it is marked as discharging data;

[0015] S2.2, extracting the IC curve and voltage curve during the battery charging cycle according to the charging segment data;

[0016] Here are the steps:

[0017] S2.2.1. Determine the battery capacity increment IC by the battery capacity and voltage value; the expression is as follows:

[0018]

[0019] In the formula, Q c represents the charging capacity, V represents the voltage, I represents the current, T represents the charging time interval, dQ c represents the differential of charge capacity, and dV represents the differential of voltage;

[0020] S2.2.2.2, introduce fitting error, and change dQ in S2.1.1 c and dV are replaced by ΔQ c and ΔV;

[0021] Introduce a fitting error into the expression of capacity increment, that is, dQ c and dV are replaced by ΔQ c and ΔV, replacing dV, ΔQ with a fixed voltage interval ΔV c To correspond to the capacity change in the ΔV range, when ΔV approaches 0, it is approximately:

[0022]

[0023] S2.3, perform noise reduction on the IC curve by exponential sliding filtering;

[0024] Due to noise contamination, the extracted IC curve has data distortion, making it difficult to analyze the IC curve;

[0025] In the Exponential Moving Average (EMA), the current filter value is the weighted average of all historical data, where the weight decays exponentially over time. The closer the data is to the current moment, the higher the weight is, and the farther the data is, the lower the weight is. This method can effectively retain the trend of the data; therefore, exponential sliding filtering is used to smooth and denoise the data. The specific expression is as follows:

[0026] y t =α·x t +(1-α)·y t-1

[0027] In the formula, x t is the original data value at time t, which is the value of the IC curve; t is the filtered value at time t; y t-1 is the filtering value of the previous moment; α is the smoothing factor, and its value range is (0, 1);

[0028] S3. Fuse the IC curve and voltage curve information to extract the voltage curve tangent angle as a feature. The fusion steps are as follows:

[0029] S3.1. Extract two peak voltage intervals of the IC curve to obtain a voltage interval, and map the voltage interval information to a charging voltage curve;

[0030] Analysis of the IC curve shows that there are two peak voltage intervals, and the two peak voltage intervals of the eight single cells are both located in the two peak voltage intervals of voltage interval I (3.5V-3.7V) and voltage interval II (3.8V-3.85V);

[0031] S3.2. In the mapped voltage range, calculate the maximum tangent angle of the charging voltage curve and the corresponding time as the health factor; the calculation steps are as follows:

[0032] S3.2.1. Normalize each voltage value and time within the voltage range, using maximum and minimum normalization. The specific expression is as follows:

[0033]

[0034] In the formula, v' represents the standard value obtained after the original voltage value v is normalized, v min is the minimum voltage in the voltage range, v max is the maximum voltage in the voltage range. Similarly, t' represents the standard value obtained after normalization of the original time t, t min is the minimum time within the voltage interval, t max is the maximum value of time within the voltage interval;

[0035] S3.2.2. Calculate the voltage tangent slope. The specific expression is as follows:

[0036]

[0037] In the formula, k cv is the voltage tangent slope, v' p is the pth normalized voltage value in the voltage interval, v' p-1 is the p-1th normalized voltage value in the voltage interval, t' p is the pth normalized time value in the voltage interval, t' p-1 is the p-1th normalized time value in the voltage interval, where p represents the number of samples in the voltage interval;

[0038] S3.2.3. Calculate the voltage tangent angle. The specific expression is as follows:

[0039] α cv =arctan(k cv )

[0040] In the formula, α cv is the voltage tangent angle, arctan represents the inverse tangent function;

[0041] S3.2.4. Calculate the maximum voltage tangent angle and corresponding time within the voltage range. The specific expression is as follows:

[0042]

[0043] In the formula, α cvmax is the maximum voltage tangent angle within the voltage range, is the pth voltage tangent angle within the voltage range;

[0044] S3.3, calculate the Spearman correlation coefficient between the maximum value of the tangent angle extracted in S3.2 and the corresponding time feature and the battery SOH and establish a heat map to determine the health feature with the strongest correlation;

[0045] The expression of Spearman correlation coefficient is as follows:

[0046]

[0047] Where ρ is the Spearman correlation coefficient, n is the number of samples of voltage tangent in the voltage interval, and d q is the rank difference of each pair of samples; q here represents the number of voltage tangent samples;

[0048] S4. Establish a LSTM-Informer-SA lithium-ion battery health status estimation model;

[0049] The steps to build are as follows:

[0050] S4.1. Run the LSTM and Informer models in parallel to establish a LSTM-Informer-SA lithium-ion battery health status estimation model;

[0051] Parallel operation means that the input data is passed to both LSTM and Informer at the same time, and they run independently in their respective networks to extract local and global time series information respectively;

[0052] Specifically, LSTM processes the input sequence to capture local time series feature information, while Informer processes the input sequence to capture global time series feature information; the two outputs are concatenated (i.e., merged in parallel) in the feature dimension, and then the concatenated result is input into the self-attention mechanism layer to achieve the fusion of local time series information and global time series information, and finally the battery SOH is estimated through the fully connected layer;

[0053] However, directly splicing the time series feature information extracted by the two models will introduce feature redundancy or cause information imbalance. Therefore, a self-attention mechanism layer is connected after the models are connected in parallel, which can adaptively select features and fully integrate the diverse features from different models, thereby maximizing the efficiency of feature utilization and improving the overall performance of the model. The operation expression is as follows:

[0054]

[0055] Where Q is the matrix of all query vectors, K is the matrix of all key vectors, V is the matrix of all value vectors, and d k is the dimension of the key vector, and softmax is the normalization function;

[0056] The Informer model is a deep learning model designed for long time series prediction tasks. By introducing ProbSparse Attention, Self-attenion Distilling and generative decoder, the model significantly reduces the computational cost of the model in long time series tasks, while enhancing the model's ability to process long sequence data containing a large amount of redundant information. The specific calculation operations are as follows:

[0057] The expression of ProbSparseAttention is as follows:

[0058]

[0059] In the formula, represents the sparsity measure of the i-th vector of the query matrix and the key matrix, which is calculated by the neural network; d represents the length of the input vector is 2; L k Indicates that the length of the vector in the Key matrix is ​​30; q i represents the query vector; K is the set of key vectors, is one of the key vectors; the value range of i and j is (1,30);

[0060] The expression of Self-attenion Distilling is as follows:

[0061]

[0062] In the formula, Represents the sequence obtained after Self-attenion Distilling processing; Maxpool represents the maximum pooling operation; Elu represents the exponential linear unit activation function; Conv1d represents the one-dimensional convolution operation; AB represents the global average block; Indicates a sequence that has not been processed by Self-attenion Distilling;

[0063] The expression of the generative decoder is as follows:

[0064]

[0065] In the formula, Represents the global timing characteristics of the input decoder; Concat is a data concatenation operation; is the prior data; is the output sequence placeholder; L token Indicates the length of the prior data part; L y Indicates the number of time steps for prediction; d model represents the feature dimension; R represents a real number;

[0066] S4.2. The cycle data features of any seven batteries are used as the input of the LSTM-Informer-SA network, and the corresponding health status values ​​are used as the output of the LSTM-Informer-SA network. The battery health status values ​​are calculated using the LSTM-Informer-SA network.

[0067] Select any seven single-cell battery cycle data obtained in S3, and then divide the selected battery data set into a training set and a validation set at a ratio of 70% and 30%, and use the remaining battery cycle data as a test set. Set the time step of the divided data to 30; then use the most correlated feature as the input of the LSTM-Informer-SA network, and the corresponding health status value as the output of the LSTM-Informer-SA network;

[0068] S5. Using the features of step S3 as the model input of S4 to obtain the final result, and to estimate the battery health status; specifically as follows:

[0069] Based on the LSTM-Informer-SA network established in step S4, the tangent angle of the voltage curve in the voltage ranges of 3.5V-3.7V and 3.8V-3.85V is used as the feature input of the model to predict the SOH value.

[0070] Beneficial Effects of the Invention

[0071] The charging data of the lithium-ion battery of the present invention is read, the IC curve and voltage curve information during the battery charging process are integrated, the battery aging health characteristics that can characterize the entire life cycle are extracted, and the LSTM-Informer-SA hybrid neural network that can extract local dependencies and global information of time series data is used. In terms of feature extraction, the correlation of features is stronger. In terms of model construction, it is not easily restricted by the problem of gradient disappearance or explosion. It can not only capture short-term dependency information, but also excel in capturing long-distance dependency information, improving the accuracy of local and short-term information processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 is a flow chart of the present invention;

[0073] Figure 2 Schematic diagram of IC curve and voltage curve results of the present invention, wherein part (a) is a schematic diagram of the voltage range of the IC curve, and part (b) is a schematic diagram of the voltage range of the voltage curve;

[0074] Figure 3 This is a characteristic schematic diagram extracted by the present invention;

[0075] Figure 4 It is a characteristic correlation heat map of the present invention, wherein part (a) is a diagram of the maximum tangent angle of voltage interval I, and part (b) is a diagram of the maximum tangent angle corresponding to time of voltage interval I;

[0076] Figure 5 It is a schematic diagram of the model structure of the present invention;

[0077] Figure 6Schematic diagram of health status estimation result and error of the present invention, wherein part (a) is a schematic diagram of the true value and predicted value of the first battery, part (b) is a schematic diagram of the true value and predicted value of the second battery, part (c) is a schematic diagram of the true value and predicted value of the third battery, part (d) is a schematic diagram of the true value and predicted value of the fourth battery, part (e) is a schematic diagram of the true value and predicted value of the fifth battery, part (f) is a schematic diagram of the true value and predicted value of the sixth battery, part (g) is a schematic diagram of the true value and predicted value of the seventh battery, and part (h) is a schematic diagram of the true value and predicted value of the eighth battery; DETAILED DESCRIPTION

[0078] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0079] like Figure 1 As shown, a lithium battery health state estimation method integrating local and global dependency information includes:

[0080] S1. Collect battery power data during the battery charge and discharge cycle; the details are as follows:

[0081] The collected batteries are lithium-ion batteries of the same model;

[0082] The experimental process is as follows: each single cell is subjected to an aging cycle test until the overall capacity of each battery is less than 80% of the nominal capacity;

[0083] The collected battery power data include: real-time recording of battery current, voltage, capacity and data collection time;

[0084] The present invention records eight single-cell battery data in total;

[0085] S2. Extract the IC curve and voltage curve during the battery charging cycle and perform smoothing. The steps are as follows:

[0086] S2.1, dividing charging segment data for each battery according to charging current;

[0087] Specifically, since the directions of the current are different during charging and discharging, that is, when the battery is charging, the current is positive, and when it is discharging, the current is negative; therefore, when the current is positive, it is marked as charging data; when the current is negative, it is marked as discharging data;

[0088] In this embodiment, according to the change of positive and negative currents, the cycle data of eight single cells are divided into charging segments of 78, 73, 76, 47, 46, 46, 77 and 76 respectively;

[0089] S2.2, extracting the IC curve and voltage curve during the battery charging cycle according to the charging segment data;

[0090] Since the battery IC curve and voltage curve contain the battery aging information, the battery IC curve and voltage curve corresponding to each charging cycle segment of the battery are extracted and integrated to deeply analyze the impact of the charging and discharging process on the health status characteristics;

[0091] Here are the steps:

[0092] S2.2.1. Determine the battery capacity increment IC by the battery capacity and voltage value; the expression is as follows:

[0093]

[0094]

[0095] In the formula, Q c represents the charging capacity, V represents the voltage, I represents the current, T represents the charging time interval, dQ c represents the differential of charging capacity, dV represents the differential of voltage; in this embodiment, Q c , V, and I are obtained specifically according to the sampling time (charging time interval);

[0096] S2.2.2.2, introduce fitting error, and change dQ in S2.1.1 c and dV are replaced by ΔQ c and ΔV;

[0097] Introduce a fitting error into the expression of capacity increment, that is, dQ c and dV are replaced by ΔQ c and ΔV, replacing dV, ΔQ with a fixed voltage interval ΔV c To correspond to the capacity change in the ΔV range, when ΔV approaches 0, it is approximately:

[0098]

[0099] S2.3, perform noise reduction on the IC curve by exponential sliding filtering;

[0100] Due to noise contamination, the extracted IC curve has data distortion, making it difficult to analyze the IC curve;

[0101] In the Exponential Moving Average (EMA), the current filter value is the weighted average of all historical data, where the weight decays exponentially over time. The closer the data is to the current moment, the higher the weight is, and the farther the data is, the lower the weight is. This method can effectively retain the trend of the data; therefore, exponential sliding filtering is used to smooth and denoise the data. The specific expression is as follows:

[0102] y t =α·x t +(1-α)·y t-1

[0103] In the formula, x t is the original data value at time t, which is the value of the IC curve; t is the filtered value at time t; y t-1 is the filtering value of the previous moment; α is the smoothing factor, and its value range is (0, 1);

[0104] Finally, the filtered IC curves under different charging segments are obtained as follows: Figure 2 As shown in parts (a) and (b);

[0105] S3. Fuse the IC curve and voltage curve information to extract the voltage curve tangent angle as a feature. The fusion steps are as follows:

[0106] S3.1. Extract two peak voltage intervals of the IC curve to obtain a voltage interval, and map the voltage interval information to a charging voltage curve;

[0107] Specifically, analysis of the IC curve can yield two peak voltage intervals, and the two peak voltage intervals of the eight single-cell batteries are both located in the two peak voltage intervals of voltage interval I (3.5V-3.7V) and voltage interval II (3.8V-3.85V);

[0108] S3.2. In the mapped voltage range, calculate the maximum tangent angle of the charging voltage curve and the corresponding time as the health factor; the calculation steps are as follows:

[0109] S3.2.1. Normalize each voltage value and time within the voltage range, using maximum and minimum normalization. The specific expression is as follows:

[0110]

[0111] In the formula, v' represents the standard value obtained after the original voltage value v is normalized, v min is the minimum voltage in the voltage range, v max is the maximum voltage in the voltage range. Similarly, t' represents the standard value obtained after normalization of the original time t, tmin is the minimum time within the voltage interval, t max is the maximum value of time within the voltage interval;

[0112] S3.2.2. Calculate the voltage tangent slope. The specific expression is as follows:

[0113]

[0114] In the formula, k cv is the voltage tangent slope, v' p is the pth normalized voltage value in the voltage interval, v' p-1 is the p-1th normalized voltage value in the voltage interval, t' p is the pth normalized time value in the voltage interval, t' p-1 is the p-1th normalized time value in the voltage interval, where p represents the number of samples in the voltage interval;

[0115] S3.2.3. Calculate the voltage tangent angle. The specific expression is as follows:

[0116] α cv =arctan(k cv )

[0117] In the formula, α cv is the voltage tangent angle, arctan represents the inverse tangent function;

[0118] S3.2.4. Calculate the maximum voltage tangent angle and corresponding time within the voltage range. The specific expression is as follows:

[0119]

[0120] In the formula, α cvmax is the maximum voltage tangent angle within the voltage range, is the pth voltage tangent angle within the voltage range;

[0121] S3.3, calculate the Spearman correlation coefficient between the maximum value of the tangent angle extracted in S3.2 and the corresponding time feature and the battery SOH and establish a heat map to determine the health feature with the strongest correlation;

[0122] The expression of Spearman correlation coefficient is as follows:

[0123]

[0124] Where ρ is the Spearman correlation coefficient, n is the number of samples of voltage tangent in the voltage interval, and d q is the rank difference of each pair of samples; q here represents the number of voltage tangent samples;

[0125] The heat map created is as follows Figure 3 As shown, it can be found that when in voltage interval I (battery health status corresponding to maximum angle 1 and battery health status corresponding to maximum time 1), the Spearman correlation coefficient is -0.949 and 0.97, indicating that within voltage interval I, the Spearman correlation coefficient is greater than ±0.9, which can better reflect the battery health status;

[0126] Furthermore, the health feature with the strongest correlation is determined as the maximum tangent angle of voltage interval I and the corresponding time as follows Figure 4 As shown in parts (a) and (b);

[0127] S4. Establish a LSTM-Informer-SA lithium-ion battery health status estimation model;

[0128] The steps to build are as follows:

[0129] S4.1. Run the LSTM and Informer models in parallel to establish a LSTM-Informer-SA lithium-ion battery health status estimation model;

[0130] Parallel operation means that the input data is passed to both LSTM and Informer at the same time, and they run independently in their respective networks to extract local and global time series information respectively;

[0131] Specifically, LSTM processes the input sequence to capture local time series feature information, while Informer processes the input sequence to capture global time series feature information; the two outputs are concatenated (i.e., merged in parallel) in the feature dimension, and then the concatenated result is input into the self-attention mechanism layer to achieve the fusion of local time series information and global time series information, and finally the battery SOH is estimated through the fully connected layer;

[0132] However, directly splicing the time series feature information extracted by the two models will introduce feature redundancy or cause information imbalance. Therefore, a self-attention mechanism layer is connected after the models are connected in parallel, which can adaptively select features and fully integrate the diverse features from different models, thereby maximizing the efficiency of feature utilization and improving the overall performance of the model. The operation expression is as follows:

[0133]

[0134] Where Q is the matrix of all query vectors, K is the matrix of all key vectors, V is the matrix of all value vectors, and d k is the dimension of the key vector, and softmax is the normalization function;

[0135] Among them, LSTM includes: memory cells, gated memory cells, candidate memory cells and hidden states; the memory cells store the information that needs to be updated; the gated memory cells are used to control the memory cells, usually including input gates, forget gates and output gates, which are used to control the writing, retention and output of information; the candidate memory cells are responsible for generating new information for updating the memory cells; and the hidden states store the processed memory cell information, which can be passed to the next layer of the network or used as the final prediction result; these structural units work together to realize the extraction of local dependency information of the subsequent parallel model on the time series feature data; the specific calculation process is as follows:

[0136] F t =σ(W f ·[X t ,H t-1 ]+b f )

[0137] I t =σ(W i ·[X t ,H t-1 ]+b i )

[0138] O t =σ(W o ·[X t ,H t-1 ]+b o )

[0139]

[0140] H t =O t tanh(C t )

[0141] In the formula, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent activation function, and X t represents the input at time t, H t-1 represents the hidden state at the previous moment, W f , W i , W o , W c are the first, second, third, and fourth weight parameters, b f 、b i 、b o 、b c are the first, second, third, and fourth bias terms;

[0142] In this embodiment, the weight parameters and bias items are initial values ​​of the computer program, and the initial values ​​are 0;

[0143] The Informer model is a deep learning model designed for long time series prediction tasks. By introducing ProbSparse Attention, Self-attenion Distilling and generative decoder, the model significantly reduces the computational cost of the model in long time series tasks, while enhancing the model's ability to process long sequence data containing a large amount of redundant information. The specific calculation operations are as follows:

[0144] The expression of ProbSparseAttention is as follows:

[0145]

[0146] In the formula, represents the sparsity measure of the i-th vector of the query matrix and the key matrix, which is calculated by the neural network; d represents the length of the input vector is 2; L k Indicates that the length of the vector in the Key matrix is ​​30; q i represents the query vector; K is the set of key vectors, is one of the key vectors; the value range of i and j is (1,30);

[0147] The expression of Self-attenion Distilling is as follows:

[0148]

[0149] In the formula, Represents the sequence obtained after Self-attenion Distilling processing; Maxpool represents the maximum pooling operation; Elu represents the exponential linear unit activation function; Conv1d represents the one-dimensional convolution operation; AB represents the global average block; Indicates a sequence that has not been processed by Self-attenion Distilling;

[0150] The expression of the generative decoder is as follows:

[0151]

[0152] In the formula, Represents the global timing characteristics of the input decoder; Concat is a data concatenation operation; is the prior data; is the output sequence placeholder; L token Indicates the length of the prior data part; L y Indicates the number of time steps for prediction; d model represents the feature dimension; R represents a real number;

[0153] S4.2. The cycle data features of any seven batteries are used as the input of the LSTM-Informer-SA network, and the corresponding health status values ​​are used as the output of the LSTM-Informer-SA network. The battery health status values ​​are calculated using the LSTM-Informer-SA network.

[0154] Specifically, any seven single-cell battery cycle data obtained by S3 are selected, and then the selected battery data set is divided into a training set and a validation set at a ratio of 70% and 30%, and the remaining battery cycle data is used as a test set, and the time step of the divided data is set to 30; then the most correlated feature is used as the input of the LSTM-Informer-SA network, and the corresponding health status value is used as the output of the LSTM-Informer-SA network;

[0155] Based on the above model, the present invention proposes an innovative timing model fusion strategy. Specifically, two independent timing models, LSTM and Informer, are run in parallel to capture local and global timing features respectively; the LSTM model is good at extracting short-term local dependencies in the input sequence, while the Informer model efficiently models long-term global dependencies through a sparse attention mechanism, significantly reducing computational complexity and retaining key global information; the output feature vectors of the two models are then concatenated and passed to the self-attention mechanism layer as a unified input; the self-attention mechanism dynamically adjusts the weight of each feature based on the similarity between features, which can highlight the features that are most critical to the prediction task and effectively suppress redundant or irrelevant features; through this context-based adaptive weight allocation, the model can optimize the feature fusion process and improve the accuracy and efficiency of information extraction; finally, the features processed by the self-attention mechanism are input to the prediction layer to generate an accurate prediction of the battery state of health (SOH); the specific structure is as follows Figure 5 As shown in the figure, this innovative method combining local and global time series features can significantly improve the accurate assessment of battery health status through automatic feature selection and weight adjustment, which has high academic value and practical significance.

[0156] S5. Using the features of step S3 as the model input of S4 to obtain the final result, and to estimate the battery health status; specifically as follows:

[0157] Based on the LSTM-Informer-SA network established in step S4, the tangent angle of the voltage curve in the voltage ranges of 3.5V-3.7V and 3.8V-3.85V is used as the feature input of the model to predict the SOH value.

[0158] The results are as follows Figure 6As shown in parts (a), (b), (c), (d), (e), (f), (g), and (h), it can be seen that the errors between the actual values ​​and the predicted values ​​of the eight batteries are mostly between -2% and 2%, which is within an acceptable range.

[0159] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A lithium battery health status estimation method integrating local and global dependency information, characterized in that: The following steps are involved: S1, collecting battery power data during the battery charge and discharge cycle; The collected batteries are eight lithium-ion batteries of the same model; S2, extracting the capacity increment IC curve and voltage curve during the battery charging cycle and performing smoothing; S3, integrating IC curve and voltage curve information to extract the voltage curve tangent angle as a feature; S4. Establish a LSTM-Informer-SA lithium-ion battery health status estimation model; S5. Use the features of step S3 as the model input of S4 to obtain the final result.

2. A lithium battery health status estimation method integrating local and global dependency information according to claim 1, characterized in that: The steps of extracting the IC curve and the voltage curve during the battery charging cycle and performing smoothing are as follows: S2.1, dividing charging segment data for each battery according to charging current; The charging segment data is divided according to the charging current, that is, when the battery is charging, the current is positive, and when the battery is discharging, the current is negative; therefore, when the current is positive, it is marked as charging data; when the current is negative, it is marked as discharging data; S2.2, extracting the IC curve and voltage curve during the battery charging cycle according to the charging segment data; The steps of the extraction are as follows: S2.2.

1. Determine the battery capacity increment IC by the battery capacity and voltage value; the expression is as follows: In the formula, Q c represents the charging capacity, V represents the voltage, I represents the current, T represents the charging time interval, dQ c represents the differential of charge capacity, and dV represents the differential of voltage; S2.2.2.2, introduce fitting error, and change dQ in S2.1.1 c and dV are replaced by ΔQ c and ΔV; the expressions are as follows: S2.3, perform noise reduction on the IC curve by exponential sliding filtering; The expression is as follows: y t =α·x+(1-α)·y t-1 In the formula, x t is the original data value at time t, which is the value of the IC curve; t is the filtered value at time t; y t-1 is the filtering value at the previous moment; α is the smoothing factor.

3. The method for estimating the health status of a lithium battery by integrating local and global dependency information according to claim 1, characterized in that: The fusion steps of fusing the IC curve and the voltage curve information to extract the voltage curve tangent angle as a feature are as follows: S3.

1. Extract two peak voltage intervals of the IC curve to obtain a voltage interval, and map the voltage interval information to a charging voltage curve; The two peak voltage intervals are located in the two peak voltage intervals of 3.5V-3.7V and 3.8V-3.85V; S3.

2. In the mapped voltage range, the maximum tangent angle of the charging voltage curve and the corresponding time are calculated as the health factor; The calculation steps are as follows: S3.2.

1. Normalize each voltage value and time within the voltage range, using maximum and minimum normalization; S3.2.2, calculate the voltage tangent slope; S3.2.3, calculate the voltage tangent angle; S3.2.

4. Obtain the maximum voltage tangent angle and corresponding time within the voltage interval; S3.

3. Calculate the Spearman correlation coefficient between the maximum tangent angle extracted in S3.2 and the corresponding time feature and the battery SOH, and establish a heat map to determine the health feature with the strongest correlation.

4. The method for estimating the health status of a lithium battery by integrating local and global dependency information according to claim 1, characterized in that: The steps for establishing the LSTM-Informer-SA lithium-ion battery health status estimation model are as follows: S4.

1. Run the LSTM and Informer models in parallel to establish a LSTM-Informer-SA lithium-ion battery health status estimation model; The parallel operation is to pass the input data to the LSTM and Informer parts at the same time, and run them independently in their respective networks to extract local and global time series information respectively; S4.

2. The cycle data features of any seven batteries are used as the input of the LSTM-Informer-SA network, and the corresponding health status values ​​are used as the output of the LSTM-Informer-SA network. The battery health status values ​​are calculated using the LSTM-Informer-SA network.

Images