Lithium ion battery health state prediction method based on large model

Abstract

The invention relates to the technical field of battery management systems, in particular to a lithium ion battery SOH prediction method based on a large model, and mainly solves the technical problem that an existing lithium ion battery SOH prediction method is low in prediction accuracy. The method comprises the following steps: firstly, converting numerical data of a data set into text codes, and extracting text features by using a large language model; meanwhile, multi-time-scale historical information is used, a multi-time-scale information interval Transform module is built, and numerical value features are extracted from numerical value data of the data set; and finally, predicting a future SOH value by using two kinds of modal information of the extracted text features and numerical features. According to the method, training is carried out through multi-modal data, the data advantages of different modals can be fully utilized, the prediction precision and robustness of the model are improved through complementary information, lithium ion battery SOH estimation is carried out through data of different modals and time scales, and the prediction result has higher precision.

Description

Technical Field

[0001] The present invention relates to the technical field of battery management systems, and in particular to a method for predicting the health status of lithium-ion batteries based on a large model. Background Art

[0002] As energy shortages and climate change become increasingly serious, the development of new energy sources has also ushered in important opportunities. Batteries, as core energy storage and supply components, play a key role in this process. During the development process, lithium-ion batteries have become the focus of research due to their advantages such as high energy density and long life. However, as the use time increases, lithium-ion batteries will gradually age due to factors such as charge and discharge cycles and temperature fluctuations, resulting in problems such as capacity decay and increased internal resistance, which not only reduces battery performance but also brings safety hazards. Therefore, accurately predicting the health state (S0H) of lithium-ion batteries is crucial to extending their life, ensuring safety, and optimizing battery management.

[0003] In the prior art, the SOH prediction method for lithium-ion batteries only uses numerical data as a data set to train the model, but the charging and discharging of lithium-ion batteries is a very complex chemical reaction process. Only using the data collected by the sensor when the lithium-ion battery is charged and discharged cannot accurately reflect the aging process of the battery; at the same time, these methods only consider historical information of a single time scale, so that the information in the historical data cannot be fully utilized, and the model cannot fully learn the relationship between historical data and SOH; therefore, the existing SOH prediction method for lithium-ion batteries has a low prediction accuracy and is difficult to meet actual needs. Summary of the invention

[0004] In order to overcome the technical defect of low prediction accuracy in the existing SOH prediction method of lithium-ion batteries, the present invention provides a lithium-ion battery health state prediction method based on a large model.

[0005] The method for predicting the health status of a lithium-ion battery based on a large model provided by the present invention comprises the following steps:

[0006] S1. Collect initial data of lithium-ion battery under charging and discharging status;

[0007] S2. performing noise reduction processing on the initial data to form noise reduction data;

[0008] Draw an incremental capacity curve according to the initial data, extract a health factor from the incremental capacity curve, decompose the incremental capacity curve into different intrinsic mode functions by using variational mode decomposition, and perform dimension reduction on the different intrinsic mode functions to obtain dimension-reduced data;

[0009] S3. The noise reduction data, health factor and dimension reduction data are used as a data set, the label SOH value is calculated according to the data set, and the data set is divided into a training set and a test set;

[0010] S4. Use multi-head attention to convert the numerical data in the data set into text encoding, build a large language model, write prompt words, and use the large language model to extract text features;

[0011] S5. Build a multi-time scale information interval Transformer module, extract numerical data of different time scales from the dataset, and use multivariate attention to extract numerical features;

[0012] S6. using cross attention to obtain the correlation between the tag SOH value and the text feature and the numerical feature, and reconstructing the text feature and the numerical feature according to the calculated correlation to obtain a reconstructed feature;

[0013] S7. Utilize the reconstructed features to predict the SOH values ​​during multiple future charge and discharge cycles through a fully connected neural network.

[0014] Optionally, in step S1, the initial data includes current data, voltage data, capacity data and charging time data.

[0015] Optionally, in step S2, discrete wavelet is used to perform noise reduction processing on the collected lithium-ion batteries.

[0016] Optionally, in step S2, the expression of variational mode decomposition is:

[0017] ;

[0018] in, is the number of modes, For the modal components, For the The center frequency, is the Dirac function, is the convolution operation, is the input incremental capacity curve, is the intrinsic mode function.

[0019] Optionally, in step S2, principal component analysis is used to reduce the dimensions of different intrinsic mode functions.

[0020] Optionally, in step S4, the expression of the multi-head attention is:

[0021] ;

[0022] ;

[0023] in, is the numerical data in the dataset, is the pre-trained token vector, and for The linear mapping of for The linear mapping of , and Respectively indicate that , and After being divided into eight parts Head vector, It is The dimension size of the head vector, is the output of each head after multi-head attention, The text encoding associated with the numerical data is obtained by weighted summation of the pre-trained token vectors.

[0024] Optionally, in step S4, the expression of the large language model is:

[0025] ;

[0026] in, represents the extracted text features, represents a pre-trained large language model, Indicates prompt words, is the text encoding.

[0027] Optionally, in step S5, numerical data of different time scales are extracted from the data set by the following steps:

[0028] S51. Encoding the multivariate features in the data set;

[0029] S52. Use 2D-TCN with convolution kernels of different sizes to extract features of different time scales from the data set. The expression of the 2D-TCN is:

[0030] ;

[0031] in, represents the dilation factor of the convolution kernel, The expansion factor is The convolution kernel, is the convolution kernel at position The value of Indicates that the data set is The value at is the value after convolution.

[0032] Optionally, in step S5, the expression of the multivariate attention is:

[0033] ;

[0034] ;

[0035] in, , and are the linear mappings of 2D-TCN outputs, is the extracted numerical feature.

[0036] Optionally, in step S6, the expression of the cross attention is:

[0037] ;

[0038] ;

[0039] ;

[0040] in, It represents the features after the extracted text features and numerical features are concatenated. Indicates the historical tag SOH value, Represents the reconstruction feature.

[0041] The technical solution provided by the present invention has the following advantages compared with the prior art:

[0042] The method for predicting the health status of lithium-ion batteries based on a large model provided by the present invention first converts the numerical data of a data set into text encoding, uses a large language model to extract text features, and provides appropriate prompt words to the large language model, thereby improving the large language model's understanding of the data and making the information extracted by the large language model more reasonable; at the same time, the method uses historical information at multiple time scales, which helps the model to better understand and capture the temporal dependencies in the data, and extracts numerical features from the numerical data of the data set by building a multi-time scale information interval Transformer module; finally, the method uses the extracted text features and numerical features to predict future SOH values. The method can fully utilize the advantages of data in different modes by training with multi-modal data, improve the prediction accuracy and robustness of the model through complementary information, and use data of different modes and time scales to estimate the SOH of lithium-ion batteries, and its prediction results have higher accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0045] Figure 1 A flow chart showing a method for predicting the health status of a lithium-ion battery based on a large model in an embodiment of the present invention;

[0046] Figure 2 It shows an incremental capacity curve diagram in an embodiment of the present invention;

[0047] Figure 3 represents a variational mode decomposition diagram in an embodiment of the present invention;

[0048] Figure 4 Represents a multivariate feature coding diagram in an embodiment of the present invention;

[0049] Figure 5 A schematic diagram showing the structure of multivariate attention in an embodiment of the present invention;

[0050] Figure 6 A diagram showing the prediction results of SOH in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the scheme of the present invention will be further described below. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0052] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein; it is obvious that the embodiments in the specification are only part of the embodiments of the present invention, rather than all of the embodiments.

[0053] Combine the following Figures 1 to 6 The specific embodiments of the present invention are described in detail.

[0054] This embodiment provides a method for predicting the health status of a lithium-ion battery based on a large model, including steps S1 to S7.

[0055] S1. Collect initial data of lithium-ion battery under charging and discharging status.

[0056] Specifically, the initial data includes current data, voltage data, capacity data, and charging time data. For example, this embodiment uses the commercial model 21700 lithium-ion battery provided by Imperial College London and the Faraday Institute as the research object, and selects the data collected under the working conditions of 40°C and full discharge depth (0~100% SoC cycle) at 0.3C constant current charging and 1C constant current discharging. The lithium-ion battery charging and discharging experiment process is: first, charge at a constant current of 0.3C until the voltage reaches 4.2V, and then charge at a constant voltage of 4.2V until the current drops below 0.01C to activate the battery; then let the battery stand to eliminate the influence of battery polarization; after meeting the requirements of the cycle experiment, the lithium-ion battery is cycled and discharged 77 times, then the battery is left standing, and then the above process is repeated to obtain the initial data of the lithium-ion battery under charging and discharging conditions.

[0057] S2. Perform denoising on the initial data to form denoised data; draw an incremental capacity curve based on the initial data, extract the health factor from the incremental capacity curve, and use variational mode decomposition to decompose the incremental capacity curve into different intrinsic mode functions, and perform dimensionality reduction on the different intrinsic mode functions to obtain dimensionality reduced data.

[0058] Specifically, discrete wavelet is used to perform noise reduction processing on the collected lithium-ion battery. The noise reduction processing can remove the noise interference in the data collection process under the charging and discharging state of the lithium-ion battery.

[0059] It is easy to understand that the incremental capacity curve is a graphical representation of the relationship between the battery voltage (V) and the incremental capacity (dQ / dV, i.e., the change in capacity per unit voltage change) during the charging or discharging process, reflecting the capacity change provided by the battery at different voltages. For example, this embodiment uses the voltage and capacity (integral of current) of a lithium-ion battery during charging to draw an incremental capacity curve, and the calculation formula for the incremental capacity is:

[0060] ;

[0061] in, The amount of electricity charged into the lithium-ion battery, is the current during charging, and for The start and end time of the change interval. By differentiating the capacity-voltage (QV) data, a curve showing the change of incremental capacity with voltage can be obtained, which is called the incremental capacity curve (ICC for short). In order to better reflect the law of ICC reaction, this embodiment uses a low-pass filter to filter ICC. The filtered IC curve is as follows Figure 2 As shown, ICC can provide information about the charge capacity released by a battery at different voltages, thereby revealing the electrochemical behavior and characteristics of the battery.

[0062] It should be noted that the health factor is the data information corresponding to the peak and trough positions in ICC. The peaks and troughs of ICC can reflect the intercalation or diffusion process of lithium ions inside the battery, reflecting the changes in the electrochemical reaction process or material aging.

[0063] It should be noted that in order to further understand the internal structure of ICC, this method uses variational mode decomposition (VMD) to decompose ICC into several modes with different frequencies and characteristics.

[0064] Specifically, the expression of variational mode decomposition is:

[0065] ;

[0066] in, is the number of modes, For the modal components, For the The center frequency, is the Dirac function, is the convolution operation, is the input incremental capacity curve, is the intrinsic mode function.

[0067] Specifically, principal component analysis is used to reduce the dimensionality of different intrinsic mode functions. Dimensionality reduction can remove redundant information in each modal data, reduce data storage requirements and computational complexity, reduce computational complexity, thereby speeding up model training and reasoning, preventing model overfitting, and also improving model performance.

[0068] S3. The denoised data, health factor and dimension reduction data are taken as a data set, the label SOH value is calculated according to the data set, and the data set is divided into a training set and a test set.

[0069] It is easy to understand that the tag SOH value is calculated based on the maximum capacity of each charging cycle of the lithium-ion battery in the data set and the rated capacity of the lithium-ion battery.

[0070] S4. Use multi-head attention to convert the numerical data in the dataset into text encoding, build a large language model, write prompt words, and use the large language model to extract text features.

[0071] Specifically, the expression of multi-head attention is:

[0072] ;

[0073] ;

[0074] in, is the numerical data in the dataset, is the pre-trained token vector, and for The linear mapping of for The linear mapping of , and Respectively indicate that , and After being divided into eight parts Head vector, It is The dimension size of the head vector, is the output of each head after multi-head attention, The text encoding associated with the numerical data is obtained by weighted summation of the pre-trained token vectors.

[0075] Specifically, the expression of the large language model is:

[0076] ;

[0077] in, represents the extracted text features, represents a pre-trained large language model, Indicates prompt words, is the text encoding.

[0078] It is easy to understand that the prompt words include task description and data description. The data description is a text description of the maximum, minimum, median and trend of the data. For example, the minimum value of the j-th feature is {min}, the maximum value of the j-th feature is {max}, the median of the j-th feature is {median}, and the trend of the j-th feature is {"increase"} or {"decrease"}. The text information is then converted into text encoding.

[0079] S5. Build a multi-time scale information interval Transformer module, extract numerical data of different time scales from the dataset, and use multivariate attention to extract numerical features.

[0080] It should be noted that the data set divides the time series data into several continuous time windows, each window contains m+n continuous historical charge and discharge data points, m is the historical time step used for training, and n is the time span of the future SOH data to be predicted.

[0081] Specifically, numerical data of different time scales are extracted from the data set in the following steps:

[0082] S51. Encode the multivariate features in the data set;

[0083] S52. Use 2D-TCN with convolution kernels of different sizes to extract features of different time scales from the data set. The expression of 2D-TCN is:

[0084] ;

[0085] in, represents the dilation factor of the convolution kernel, The expansion factor is The convolution kernel, is the convolution kernel at position The value of Indicates that the data set is The value at is the value after convolution.

[0086] Specifically, the expression of multivariate attention is:

[0087] ;

[0088] ;

[0089] in, , and are the linear mappings of 2D-TCN outputs, is the extracted numerical feature.

[0090] S6. Cross attention is used to obtain the correlation between the label SOH value and the text features and numerical features, and the text features and numerical features are reconstructed according to the calculated correlation to obtain the reconstructed features. By reconstructing text features and numerical features using cross attention, data of different modalities can form complementary information, thereby improving the prediction accuracy and robustness of the model.

[0091] Specifically, the expression of cross attention is:

[0092] ;

[0093] ;

[0094] ;

[0095] in, It represents the features after the extracted text features and numerical features are concatenated. Indicates the historical tag SOH value, Represents the reconstruction feature.

[0096] S7. Using the reconstructed features, the SOH values ​​during multiple future charge and discharge cycles are predicted through a fully connected neural network.

[0097] The large-model-based lithium-ion battery health status prediction method of this embodiment makes full use of the data information collected during the charging and discharging state of the lithium-ion battery, and through the complementary information of different modes, effectively improves the accuracy and reliability of the health status assessment of the lithium battery during multiple future charging and discharging, and provides an effective method for the battery management system, which has important guiding significance for battery safety and service life extension in key application fields such as automobiles and industry.

[0098] The above is only a specific implementation of the present invention, which enables those skilled in the art to understand or implement the present invention. Although detailed descriptions are given with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments, and they should all be covered by the protection scope of the claims.

Claims

1. A method for predicting the health status of lithium-ion batteries based on a large model, characterized in that: The steps include: S1. Collect initial data of lithium-ion battery under charging and discharging status; S2. performing noise reduction processing on the initial data to form noise reduction data; Draw an incremental capacity curve according to the initial data, extract a health factor from the incremental capacity curve, decompose the incremental capacity curve into different intrinsic mode functions by using variational mode decomposition, and perform dimension reduction on the different intrinsic mode functions to obtain dimension-reduced data; S3. The noise reduction data, health factor and dimension reduction data are used as a data set, the label SOH value is calculated according to the data set, and the data set is divided into a training set and a test set; S4. Use multi-head attention to convert the numerical data in the data set into text encoding, build a large language model, write prompt words, and use the large language model to extract text features; S5. Build a multi-time scale information interval Transformer module, extract numerical data of different time scales from the dataset, and use multivariate attention to extract numerical features; S6. using cross attention to obtain the correlation between the tag SOH value and the text feature and the numerical feature, and reconstructing the text feature and the numerical feature according to the calculated correlation to obtain a reconstructed feature; S7. Utilize the reconstructed features to predict the SOH values ​​during multiple future charge and discharge cycles through a fully connected neural network.

2. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 1, characterized in that: In step S1, the initial data includes current data, voltage data, capacity data and charging time data.

3. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 1, characterized in that: In step S2, discrete wavelet is used to perform noise reduction processing on the collected lithium-ion batteries.

4. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 3, characterized in that: In step S2, the expression of variational mode decomposition is: ； in, is the number of modes, For the modal components, For the The center frequency, is the Dirac function, is the convolution operation, is the input incremental capacity curve, is the intrinsic mode function.

5. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 4, characterized in that: In step S2, principal component analysis is used to reduce the dimensions of different intrinsic mode functions.

6. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 1, characterized in that: In step S4, the expression of the multi-head attention is: ； ； in, is the numerical data in the dataset, is the pre-trained token vector, and for The linear mapping of for The linear mapping of , and Respectively indicate , and After being divided into eight parts Head vector, It is The dimension size of the head vector, is the output of each head after multi-head attention, The text encoding associated with the numerical data is obtained by weighted summation of the pre-trained token vectors.

7. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 6, characterized in that: In step S4, the expression of the large language model is: ； in, represents the extracted text features, represents a pre-trained large language model, Indicates prompt words, is the text encoding.

8. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 1, characterized in that: In step S5, numerical data of different time scales are extracted from the data set by the following steps: S51. Encoding the multivariate features in the data set; S52. Use 2D-TCN with convolution kernels of different sizes to extract features of different time scales from the data set. The expression of the 2D-TCN is: ； in, represents the dilation factor of the convolution kernel, The expansion factor is The convolution kernel, is the convolution kernel at position The value of Indicates that the data set is The value at is the value after convolution.

9. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 8, characterized in that: In step S5, the expression of the multivariate attention is: ； ； in, , and are the linear mappings of 2D-TCN outputs, is the extracted numerical feature.

10. The method for predicting the health status of a lithium-ion battery based on a large model according to claim 1, characterized in that: In step S6, the expression of the cross attention is: ； ； ； in, It represents the features after the extracted text features and numerical features are concatenated. Indicates the historical tag SOH value, Represents the reconstruction feature.

Images