Battery health state estimation method and system based on pre-trained large language model

Abstract

The present invention belongs to the technical field of lithium-ion batteries, and provides a battery health state estimation method and system based on a pre-trained large language model. The technical solution is to fuse multi-cycle charging information to perform candidate voltage combinations to efficiently obtain candidate features. By establishing a voltage-charge capacity curve in a constant current charging stage, effective candidate features can be extracted without the need for complete charging and discharging data or the need to establish a complex mathematical model or electrochemical model. At the same time, with the help of a pre-trained large language model and model fine-tuning technology, continuous, reliable and accurate estimation of the battery health state is achieved.

Description

Technical Field

[0001] The present invention relates to the field of lithium-ion battery technology, and in particular to a battery health state estimation method and system based on a pre-trained large language model. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, with the vigorous development of transportation electrification, the continuous innovation of research and technology in the field of renewable energy has attracted unprecedented attention. Lithium-ion batteries have unique advantages such as high energy density, long cycle life, low self-discharge rate, and low weight. At the same time, they have no memory effect and are widely used in electric vehicles, portable electronic devices, energy storage systems, medical equipment, military applications and other fields. Lithium-ion batteries can be divided into many types according to multiple angles such as positive electrode materials, battery structure, electrolyte type, and energy storage capacity. These different types of batteries are used in different fields and equipment according to their characteristics to meet diverse needs. However, under the interference of factors such as material degradation, extreme working conditions, overcharge and overdischarge, high temperature, and abuse, the performance of lithium-ion batteries will gradually decline, and there are many safety hazards.

[0004] The performance degradation of lithium-ion batteries is mainly reflected in the decrease in available capacity and the increase in internal resistance. The decrease in available capacity will reduce the battery life of electric vehicles or electronic devices, increase the charging burden, and thus cause problems such as unstable battery power output and reduced user experience. The increase in internal resistance will cause the battery to generate more heat during the charging and discharging process, which may accelerate battery aging and even cause safety accidents such as overheating, fire or explosion. Therefore, continuous, stable and accurate battery state of health (SOH) estimation is extremely important to ensure the reliability and safety of the battery, and is also a key indicator for evaluating the battery aging level.

[0005] At present, the methods for estimating the health status of lithium-ion batteries include: definition method, model-based estimation method and data-driven estimation method. In the definition method, the health status is defined as the percentage of the ratio of the remaining available capacity to the rated capacity. The health status is obtained by direct calculation. The defect of this method is that the remaining available capacity cannot be accurately obtained in engineering practice; the model-based method estimates the health status by constructing a mathematical model that can reflect the microscopic physical and chemical reaction mechanism of the battery, mainly including electrochemical model, empirical model and equivalent circuit model. The defect of this method is that it is completely dependent on accurate mechanisms and mathematical models, which makes it poorly adaptable in actual working conditions and has large fluctuations in estimation accuracy; without the need for in-depth understanding of the battery electrochemical reaction mechanism, the data-driven method can effectively establish the mapping relationship between the health status and the aging characteristics and achieve accurate estimation. This method is widely used and has high accuracy. However, the problem of feature engineering with strict prerequisites, training overfitting, high computational cost and other issues limit its further development. Therefore, the technical means and scope of application of this method still need to be strengthened. Summary of the invention

[0006] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a battery health state estimation method and system based on a pre-trained large language model, which breaks the barrier of performing feature engineering only in a single cycle, and instead integrates multi-cycle charging information to construct a feature combination, and establishes a mapping relationship between aging features and health status by fine-tuning the open source pre-trained large language model, thereby solving the key problem of poor flexibility of feature engineering in existing methods, and realizing continuous, stable and accurate estimation of battery health status with the help of large model prior knowledge and model fine-tuning technology.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A first aspect of the present invention provides a method for estimating a battery health state based on a pre-trained large language model, comprising the following steps:

[0009] Conduct lithium-ion battery cycle aging experiments and obtain lithium-ion battery cycle aging data;

[0010] Based on the cycle aging data of lithium-ion batteries, the voltage-charge capacity data of the constant current charging stage in each cycle is screened, and the voltage-charge capacity curve of the constant current charging stage is established;

[0011] Based on the voltage-charge capacity data and curves in the constant current charging stage, candidate voltage intervals and corresponding candidate features under different preset fixed charge capacities are extracted;

[0012] The candidate voltage range and the corresponding candidate features are used as input, and the battery health status estimation value is used as output, and the loaded pre-trained large language model is optimized to obtain an optimized large language model;

[0013] In actual charging data, multiple charging cycles for combining candidate voltage intervals are selected, corresponding candidate voltage intervals are combined, and corresponding candidate features are extracted;

[0014] Based on the optimized large language model, the candidate voltage ranges and corresponding candidate features of the charging data to be predicted are predicted to obtain the battery health status.

[0015] Furthermore, the lithium-ion battery cycle aging experiment is performed to obtain lithium-ion battery cycle aging data, including:

[0016] Step 101: performing a charging experiment on a lithium-ion battery at a certain fixed charging rate using a constant current and constant voltage charging protocol, so that the voltage in the constant current charging stage and the current in the constant voltage stage both reach a specified threshold;

[0017] Step 102: using a constant current discharge protocol to discharge the lithium-ion battery at a fixed discharge rate to make the voltage reach a specified threshold; wherein there is a set period of rest during the charge and discharge period;

[0018] Step 103: Record the total discharge capacity during the constant current discharge process and use it as a reference value of the actual capacity;

[0019] Step 104: Repeat steps 101-103 until the actual capacity of the lithium-ion battery decays to the set rated capacity, and each discharge ends with a set rest time;

[0020] Step 105: extracting long-term cycle aging data to obtain lithium-ion battery cycle aging data.

[0021] Furthermore, based on the voltage-charge capacity data and curve in the constant current charging stage, extracting the candidate voltage intervals and corresponding candidate features under different preset fixed charge capacities includes:

[0022] Based on the voltage-charge data and curves in the constant current charging stage, a sliding window technique is used on a certain cycle to charge the battery at a fixed charge level. is the window size, with the rated capacity of the battery as the step size, along x Move in the axis direction, extract the data coordinate information of each window, and define each window in y The dimension in the axial direction is the voltage range , to obtain the voltage interval set corresponding to the cycle;

[0023] Based on the voltage interval set extraction process, traverse all loops, and perform coordinate average calculation on multiple voltage intervals in all loops to obtain an average voltage interval set;

[0024] Based on the average voltage interval set, traverse all cycles and calculate the charge amount corresponding to each average voltage interval to obtain a charge amount sequence corresponding to each average voltage interval;

[0025] Based on the charge amount sequence corresponding to the average voltage interval, the Pearson correlation coefficient method is used to select the sequence with the strongest correlation with the health state, and the voltage interval corresponding to the sequence is defined as the candidate voltage interval , the charge amount corresponding to the candidate voltage interval is the candidate feature .

[0026] Furthermore, the method takes the candidate voltage interval and the corresponding candidate feature as input, takes the battery health status estimation value as output, and performs model optimization on the loaded pre-trained large language model to obtain the optimized large language model, including:

[0027] Based on the nonlinear characteristics of lithium-ion batteries, determine the pre-trained large language model and initialize the model parameters and word segmenter;

[0028] Based on the determined pre-trained large language model, adjust the input and output types, modify the type of the model output layer, and load it onto the GPU;

[0029] Based on the loaded pre-trained large language model, fine-tune the model and find the optimal hyperparameters based on grid search. The hyperparameters include the output layer learning rate, batch size, and number of training rounds. The model corresponding to the optimal hyperparameters is the optimal model.

[0030] Furthermore, the preset fixed charge amount Including: 20%, 30%, 40%, 50%, 60% and 70%, each cycle can collect p Voltage range , p Defined as:

[0031] ,

[0032] when When it is 20%, the corresponding candidate voltage interval is (3.40 V, 3.54 V), The candidate voltage range corresponding to 30% is (3.30 V, 3.38 V), The candidate voltage range corresponding to 40% is (3.36 V, 3.41 V), The candidate voltage range corresponding to 50% is (3.36 V, 3.48 V), The candidate voltage range corresponding to 60% is (3.33 V, 3.54 V), The candidate voltage range corresponding to 70% is (3.33 V, 3.63 V).

[0033] Furthermore, in a candidate voltage interval In the example, the initial charge node value is defined as , the terminal charge node value is defined as , candidate features Defined as: .

[0034] Furthermore, the BERT-base model based on the Transformer encoder architecture is used to train the large language model.

[0035] A second aspect of the present invention provides a battery health state estimation system based on a pre-trained large language model, comprising:

[0036] The cycle experiment module is used to perform a cycle aging experiment on a lithium-ion battery and obtain cycle aging data of the lithium-ion battery; based on the cycle aging data of the lithium-ion battery, the voltage-charge amount data in the constant current charging stage in each cycle is screened, and a voltage-charge amount curve in the constant current charging stage is established;

[0037] A model optimization module is used to extract candidate voltage intervals and corresponding candidate features under different preset fixed charge amounts based on the voltage-charge amount data and curves in the constant current charging stage; the candidate voltage intervals and corresponding candidate features are used as inputs, and the battery health status estimation value is used as output, and the loaded pre-trained large language model is optimized to obtain an optimized large language model;

[0038] A feature combination module, which is used to select multiple charging cycles for combining candidate voltage intervals in actual charging data, combine corresponding candidate voltage intervals, and extract corresponding candidate features;

[0039] The state estimation module predicts the candidate voltage ranges and corresponding candidate features of the charging data to be predicted based on the optimized large language model to obtain the battery health status.

[0040] Furthermore, in the model optimization module, the voltage-charge capacity data and curve in the constant current charging stage are used to extract candidate voltage intervals and corresponding candidate features under different preset fixed charge capacities, including:

[0041] Based on the voltage-charge data and curves in the constant current charging stage, a sliding window technique is used on a certain cycle to charge the battery at a fixed charge level. is the window size, with the rated capacity of the battery as the step size, alongx Move in the axis direction, extract the data coordinate information of each window, and define each window in y The dimension in the axial direction is the voltage range , to obtain the voltage interval set corresponding to the cycle;

[0042] Based on the voltage interval set extraction process, traverse all loops, and perform coordinate average calculation on multiple voltage intervals in all loops to obtain an average voltage interval set;

[0043] Based on the average voltage interval set, traverse all cycles and calculate the charge amount corresponding to each average voltage interval to obtain a charge amount sequence corresponding to each average voltage interval;

[0044] Based on the charge amount sequence corresponding to the average voltage interval, the Pearson correlation coefficient method is used to select the sequence with the strongest correlation with the health state, and the voltage interval corresponding to the sequence is defined as the candidate voltage interval , the charge amount corresponding to the candidate voltage interval is the candidate feature .

[0045] Furthermore, in the state estimation module, the BERT-base model based on the Transformer encoder architecture is used to train the large language model.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] 1. Compared with the prior art that performs feature engineering for aging feature extraction in a single cycle, the present invention integrates multi-cycle charging information to perform candidate voltage combinations to efficiently obtain candidate features, solving the problem of discontinuity and instability in feature engineering in practice, breaking the barrier of performing feature engineering only in a single cycle, and does not require additional data processing and complex calculation processes.

[0048] 2. Compared with the existing battery health status definition and estimation methods based on complete charging and discharging data, the present invention does not require complete charging and discharging data, nor does it require the establishment of complex mathematical models or electrochemical models. It can extract effective candidate features for battery health status estimation based only on a limited voltage-charge curve.

[0049] 3. Compared with the existing data-driven battery health status estimation method, the present invention achieves continuous, reliable and accurate estimation of battery health status with the help of pre-trained large language model and model fine-tuning technology. The pre-trained large language model provides prior knowledge, accelerates training convergence, and improves estimation efficiency; the model fine-tuning technology reduces data requirements and ensures estimation accuracy.

[0050] 4. Compared with the existing battery health status estimation technology, the feature engineering in the present invention is easy to implement, the estimation model is more stable, and it does not involve the deep electrochemical reaction mechanism and aging mechanism of lithium-ion batteries. It can adapt to the complex working conditions caused by fragmented charging behavior in actual engineering, and has good development prospects and research value.

[0051] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0053] Figure 1 is a flow chart of a method for estimating a battery health state based on a pre-trained large language model provided by an embodiment of the present invention;

[0054] Figure 2 It is a battery health state decay curve in the battery cycle aging data provided by an embodiment of the present invention;

[0055] Figure 3 is the voltage of the lithium-ion battery Cell#1 during the constant current charging stage provided by the embodiment of the present invention. V -Charge capacity Q curve;

[0056] Figure 4 The candidate features provided by the embodiment of the present invention are Q 50% The battery health status estimation result at ;

[0057] Figure 5 The candidate features provided by the embodiment of the present invention are Q 60% The battery health status estimation result at ;

[0058] Figure 6 It is a structural diagram of a battery health state estimation system based on a pre-trained large language model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0060] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0061] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0062] As mentioned in the background technology, the defects of the existing lithium-ion battery health status estimation method, the present invention breaks the barrier of only carrying out feature engineering in a single cycle, and instead integrates multi-cycle charging information to construct a feature combination, and establishes the mapping relationship between aging features and health status by fine-tuning the open source pre-trained large language model. The present invention solves the key problem of poor flexibility of feature engineering in existing methods, and achieves continuous, stable and accurate estimation of battery health status with the help of large model prior knowledge and model fine-tuning technology.

[0063] Embodiment 1

[0064] like Figure 1 As shown, this embodiment provides a battery health state estimation method based on a pre-trained large language model, comprising the following steps:

[0065] Step 1: Conduct a lithium-ion battery cycle aging experiment to obtain lithium-ion battery cycle aging data;

[0066] Wherein, in step 1, the lithium-ion battery cycle aging experiment is performed to obtain lithium-ion battery cycle aging data, including:

[0067] Step 101: performing a charging experiment on a lithium-ion battery at a certain fixed charging rate using a constant current and constant voltage charging protocol, so that the voltage in the constant current charging stage and the current in the constant voltage stage both reach a specified threshold;

[0068] Step 102: using a constant current discharge protocol to discharge the lithium-ion battery at a fixed discharge rate to make the voltage reach a specified threshold; wherein there is a set period of rest during the charge and discharge period;

[0069] In this embodiment, at room temperature, a long-cycle aging experiment was carried out on four lithium-ion battery cells using a constant current and constant voltage charging protocol and a constant current and discharge protocol. The cells were named Cell #1, Cell #2, Cell #3, and Cell #4. The rated capacity of the cell is 2.5Ah. In the constant current and constant voltage charging stage, the lithium-ion battery is charged with a constant current at a rate of 1C (current of 2.5A). When the voltage reaches 3.65V, constant voltage charging is performed until the current drops to 0.125A. In the constant current discharge stage, the lithium-ion battery is discharged with a constant current at a rate of 4C (current of 10A) until the voltage reaches the specified threshold of 2V.

[0070] Step 103: Record the total discharge capacity during the constant current discharge process and use it as a reference value of the actual capacity;

[0071] Step 104: Repeat the above steps 101-103 until the actual capacity of the lithium-ion battery decays to 80% of the rated capacity, and each discharge requires a period of rest;

[0072] Step 105: According to the above aging experiment, extract the long-term cycle aging data, including parameters such as voltage, current, charge capacity, and discharge capacity, and perform relevant preprocessing on the aging data. The battery health state decay curve obtained by the above long-term cycle aging experiment is as follows: Figure 2 shown.

[0073] Step 2: Based on the cycle aging data of lithium-ion batteries, the voltage-charge capacity data of the constant current charging stage in each cycle is screened, and the voltage-charge capacity curve of the constant current charging stage is established;

[0074] Specifically include:

[0075] Step 201: Based on the lithium-ion battery cycle aging data obtained in step 1, irrelevant parameter data are eliminated, and the voltage and charge capacity data of the constant current charging stage in each cycle are retained;

[0076] Step 202: Based on the voltage and charge data, combine the voltage-charge two-dimensional data points to specify the charge as x Axis, voltage is y axis, establishes the voltage during the constant current charging phase V -Charge capacity Q The voltage V-charge amount Q curve of the lithium-ion battery Cell #1 during the constant current charging stage is as follows: Figure 3 shown.

[0077] Step 3: Based on the voltage-charge capacity data and curves in the constant current charging stage, extract candidate voltage intervals and candidate features under different fixed charge capacities, including:

[0078] Step 301: Based on the voltage-charge data and curve, a sliding window technique is used on a certain cycle to is the window size, with the set value of the battery rated capacity, such as 5%, as the step length. x Move in the axis direction, extract the data coordinate information of each window, and define each window in y The dimension in the axial direction is the voltage range , to obtain the voltage interval set corresponding to the cycle;

[0079] Step 302: based on the voltage interval set extraction process, traverse all loops, and perform coordinate average calculation on multiple voltage intervals in all loops to obtain an average voltage interval set;

[0080] Step 303: based on the average voltage interval set, traverse all loops, calculate the charge amount corresponding to each average voltage interval, so as to obtain a charge amount sequence corresponding to each average voltage interval;

[0081] Step 304: Based on the charge amount sequence corresponding to the average voltage interval, the Pearson correlation coefficient method is used to select the sequence with the strongest correlation with the health state, and the voltage interval corresponding to the sequence is defined as the candidate voltage interval , the charge amount corresponding to the candidate voltage interval is the candidate feature ;

[0082] In this embodiment, the preset fixed charge amount Including: 20%, 30%, 40%, 50%, 60% and 70%, each cycle can collect p Voltage range , p Defined as:

[0083] ,

[0084] Among them, when When it is 20%, the corresponding candidate voltage interval is (3.40 V, 3.54 V), The candidate voltage range corresponding to 30% is (3.30 V, 3.38 V), The candidate voltage range corresponding to 40% is (3.36 V, 3.41 V), The candidate voltage range corresponding to 50% is (3.36 V, 3.48 V), The candidate voltage range corresponding to 60% is (3.33 V, 3.54 V), The candidate voltage range corresponding to 70% is (3.33 V, 3.63 V).

[0085] In a candidate voltage range In the example, the initial charge node value is defined as , the terminal charge node value is defined as , candidate features Defined as:

[0086] .

[0087] Step 4: Load the pre-trained large language model, find the optimal hyperparameters based on grid search, fine-tune the model, and build the optimal large language model. The specific steps include the following:

[0088] Step 401: Based on the nonlinear characteristics of lithium-ion batteries, determine a pre-trained large language model, and initialize the parameters and word segmenter of the model;

[0089] In this embodiment, the BERT-base model based on the Transformer encoder architecture is selected as the pre-trained large language model in the present invention; it consists of 12 layers of Transformer encoder layers, with a total parameter volume of 115 million. Each Transformer encoder layer contains a multi-head attention mechanism and a feedforward neural network. Each multi-head attention mechanism contains 12 self-attention mechanism heads.

[0090] Based on the BERT-base model, the model parameters and word segmenter pre-trained on a large corpus are loaded for model initialization. The BERT-base model is pre-trained on two large-scale corpora, BooksCorpus and English Wikipedia, and has strong prior knowledge. The pre-training of the BERT-base model includes two main tasks: masked language model and next sentence prediction, based on which the model can achieve bidirectional training. The pre-trained parameters and word segmenter are directly used as the initialization parameters of the pre-trained large language model in this case.

[0091] Step 402: Based on the determined pre-trained large language model, adjust the input and output types, modify the type of the model output layer, and load it onto the GPU;

[0092] In this embodiment, based on the BERT-base model, the input type is adjusted to a string type, the output type is adjusted to a floating point type, the output layer is modified to a linear layer, the internal parameters are initialized, and it is loaded onto the GPU.

[0093] Specifically: the original input is the charge amount corresponding to the candidate voltage interval , the original input type is a floating point number type, which should be adjusted to a string type . Adjust the output type to a floating point type, and output the estimated value of the battery health status under this cycle. Change the output layer to a linear layer, set the output dimension to 1, and initialize the internal parameters. To improve the efficiency of model fine-tuning, load the pre-trained large language model onto the GPU.

[0094] Step 403: Based on the loaded pre-trained large language model and lithium-ion battery cycle aging data, fine-tune the model and find the optimal hyperparameters based on grid search to obtain the optimal model; specifically, the following steps are included:

[0095] Step 4031: Divide the battery cycle aging data into a training set and a test set, and the training set is used for model fine-tuning; for example, the data of Cell #1, Cell #2, and Cell #3 are set as the training set, and the data of Cell #4 is set as the test set, and the training set is used for model fine-tuning.

[0096] Step 4032: Differentiate the learning rates on the base layer and output layer of the BERT-base model to preserve the prior knowledge in the base layer. The base layer of the BERT-base model refers to all Transformer encoder layers except the output layer. Set the learning rates of these base layers to smaller values ​​to preserve the influence of pre-trained prior knowledge on the health state estimation task. In this embodiment, the base layer learning rate is set to 3e-5;

[0097] Step 4033: Find the optimal hyperparameters based on grid search, where the hyperparameters include output layer learning rate, batch size, and number of training rounds. The model corresponding to the optimal hyperparameters is the optimal model.

[0098] In this embodiment, the output layer learning rate search set based on grid search optimization is {1e-4, 2e-4, 3e-4, 4e-4, 5e-4}, the batch size search set is {16, 32, 64, 128, 256}, and the training round number search set is {50, 60, 70, 80, 90, 100};

[0099] Step 4034: Based on the candidate features and battery health status labels, K-fold cross validation is applied to perform model fine-tuning to construct a mapping relationship between aging features and health status.

[0100] In this embodiment, the number of folds in K-fold cross validation is set to 5.

[0101] The candidate features corresponding to each candidate voltage range are used as input, and the estimated value of the battery health status is used as output to fine-tune the model. Based on the K-fold cross-validation results, the optimal hyperparameters are determined to be: {1e-4, 32, 60}. Based on the optimal hyperparameters, the optimal model under each candidate feature is obtained.

[0102] Step 5: In the actual charging data, multiple charging cycles that can be used to combine candidate voltage intervals are selected, the corresponding candidate voltage intervals are combined, and the corresponding candidate features are extracted;

[0103] Based on the candidate voltage intervals and the candidate features, actual charging data is observed and a plurality of charging cycles that can be used to combine the candidate voltage intervals are selected.

[0104] In this embodiment, the maximum interval between multiple charging cycles is required to be no more than 20 cycles.

[0105] Based on multiple charging cycles, the corresponding candidate voltage intervals are combined , and extract the corresponding candidate features There are multiple ways to combine candidate features obtained from multiple cycles, which are combined according to actual charging conditions. The combined candidate features can be identified as aging features under the current cycle.

[0106] Step 6: Based on the obtained optimal model and the lithium-ion battery cycle aging data, a reliable estimation of the battery health status is achieved, including:

[0107] The lithium-ion battery cycle aging data is divided into a training set and a test set, and the test set is used to verify the model estimation performance;

[0108] Based on the optimal model obtained by grid search, the test set is used to evaluate the performance of the model, and the root mean square error (RMSE) is selected as an indicator to measure the model estimation performance.

[0109] When the complete candidate features without multi-cycle combination are used for estimation, better estimation performance is achieved. and For example, corresponding to (3.36 V, 3.48 V) and (3.33 V, 3.54 V), the candidate features are and . Figure 4 and Figure 5 The candidate features are and The battery health status estimation results at 100 and 100 are RMSE 0.0070 and 0.0057 respectively. When using the candidate features of multi-cycle combination for estimation, in this case, the charging information from two cycles (with an interval of 20 cycles) is combined as an example. , respectively correspond to (3.36 V, 3.42 V) and (3.42 V, 3.48 V), and the combination of (3.36 V, 3.48 V) is in the candidate voltage range. When , they correspond to (3.33 V, 3.43 V) and (3.43 V, 3.54 V), respectively, and the combined candidate features are (3.33 V, 3.54 V). The RMSEs corresponding to the combined candidate features are 0.0075 and 0.0058, respectively.

[0110] Embodiment 2

[0111] like Figure 6 As shown, this embodiment provides a battery health state estimation system based on a pre-trained large language model, including:

[0112] The cycle experiment module is used to perform a cycle aging experiment on a lithium-ion battery and obtain cycle aging data of the lithium-ion battery; based on the cycle aging data of the lithium-ion battery, the voltage-charge amount data in the constant current charging stage in each cycle is screened, and a voltage-charge amount curve in the constant current charging stage is established;

[0113] A model optimization module is used to extract candidate voltage intervals and corresponding candidate features under different preset fixed charge amounts based on the voltage-charge amount data and curves in the constant current charging stage; the candidate voltage intervals and corresponding candidate features are used as inputs, and the battery health status estimation value is used as output, and the loaded pre-trained large language model is optimized to obtain an optimized large language model;

[0114] A feature combination module, which is used to select multiple charging cycles for combining candidate voltage intervals in actual charging data, combine corresponding candidate voltage intervals, and extract corresponding candidate features;

[0115] The state estimation module predicts the candidate voltage ranges and corresponding candidate features of the charging data to be predicted based on the optimized large language model to obtain the battery health status.

[0116] The implementation process of each module in the second embodiment is consistent with that in the first embodiment.

[0117] Embodiment 3

[0118] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the battery health state estimation method based on a pre-trained large language model as described above are implemented.

[0119] Embodiment 4

[0120] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the method for estimating the battery health status based on a pre-trained large language model as described above are implemented.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A battery health state estimation method based on a pre-trained large language model, characterized in that: The steps include: Conduct lithium-ion battery cycle aging experiments and obtain lithium-ion battery cycle aging data, including: Step 101: performing a charging experiment on a lithium-ion battery at a certain fixed charging rate using a constant current and constant voltage charging protocol, so that the voltage in the constant current charging stage and the current in the constant voltage stage both reach a specified threshold; Step 102: using a constant current discharge protocol to discharge the lithium-ion battery at a fixed discharge rate to make the voltage reach a specified threshold; wherein there is a set period of rest during the charge and discharge period; Step 103: Record the total discharge capacity during the constant current discharge process and use it as a reference value of the actual capacity; Step 104: Repeat steps 101-103 until the actual capacity of the lithium-ion battery decays to the set rated capacity, and each discharge ends with a set rest time; Step 105: extracting long-term cycle aging data to obtain lithium-ion battery cycle aging data; Based on the cycle aging data of lithium-ion batteries, the voltage-charge capacity data of the constant current charging stage in each cycle is screened, and the voltage-charge capacity curve of the constant current charging stage is established; Based on the voltage-charge capacity data and curves in the constant current charging stage, candidate voltage intervals and corresponding candidate features under different preset fixed charge capacities are extracted, including: Based on the voltage-charge data and curves in the constant current charging stage, a sliding window technique is used on a certain cycle to charge the battery at a fixed charge level. is the window size, with the rated capacity of the battery as the step size, along x Move in the axis direction, extract the data coordinate information of each window, and define each window in y The dimension in the axial direction is the voltage range , to obtain the voltage interval set corresponding to the cycle; Based on the voltage interval set extraction process, traverse all loops, and perform coordinate average calculation on multiple voltage intervals in all loops to obtain an average voltage interval set; Based on the average voltage interval set, traverse all cycles and calculate the charge amount corresponding to each average voltage interval to obtain a charge amount sequence corresponding to each average voltage interval; Based on the charge amount sequence corresponding to the average voltage interval, the Pearson correlation coefficient method is used to select the sequence with the strongest correlation with the health state, and the voltage interval corresponding to the sequence is defined as the candidate voltage interval , the charge amount corresponding to the candidate voltage interval is the candidate feature ; Preset fixed charge level Including: 20%, 30%, 40%, 50%, 60% and 70%, each cycle can collect p Voltage range , p Defined as: ， when When it is 20%, the corresponding candidate voltage interval is (3.40 V, 3.54 V), The candidate voltage range corresponding to 30% is (3.30 V, 3.38 V), The candidate voltage range corresponding to 40% is (3.36 V, 3.41 V), The candidate voltage range corresponding to 50% is (3.36 V, 3.48 V), The candidate voltage range corresponding to 60% is (3.33 V, 3.54 V), The candidate voltage range corresponding to 70% is (3.33 V, 3.63 V); the candidate voltage range and the corresponding candidate features are used as input, and the battery health status estimation value is used as output. The loaded pre-trained large language model is optimized to obtain the optimized large language model, including: Based on the nonlinear characteristics of lithium-ion batteries, determine the pre-trained large language model and initialize the model parameters and word segmenter; Based on the determined pre-trained large language model, adjust the input and output types, modify the type of the model output layer, and load it onto the GPU; Based on the loaded pre-trained large language model, fine-tune the model and find the optimal hyperparameters based on grid search. The hyperparameters include the output layer learning rate, batch size, and number of training rounds. The model corresponding to the optimal hyperparameters is the optimal model. In the actual charging data, multiple charging cycles for combining candidate voltage intervals are selected, the corresponding candidate voltage intervals are combined, and the corresponding candidate features are extracted, including Based on the candidate voltage intervals and the candidate features, observing actual charging data, selecting a plurality of charging cycles that can be used to combine the candidate voltage intervals; Based on multiple charging cycles, the corresponding candidate voltage intervals are combined , and extract the corresponding candidate features ; Based on the optimized large language model, the candidate voltage ranges and corresponding candidate features of the charging data to be predicted are predicted to obtain the battery health status.

2. The method for estimating the battery health state based on a pre-trained large language model according to claim 1, characterized in that: In a candidate voltage range In the example, the initial charge node value is defined as , the terminal charge node value is defined as , candidate features Defined as: .

3. The method for estimating the battery health state based on a pre-trained large language model according to claim 1, characterized in that: The BERT-base model based on the Transformer encoder architecture is used to train the large language model.

4. Battery health status estimation system based on pre-trained large language model, characterized in that: The method for estimating the battery health state based on the pre-trained large language model according to any one of claims 1 to 3 comprises: The cycle experiment module is used to perform a cycle aging experiment on a lithium-ion battery and obtain cycle aging data of the lithium-ion battery; based on the cycle aging data of the lithium-ion battery, the voltage-charge amount data in the constant current charging stage in each cycle is screened, and a voltage-charge amount curve in the constant current charging stage is established; A model optimization module is used to extract candidate voltage intervals and corresponding candidate features under different preset fixed charge amounts based on the voltage-charge amount data and curves in the constant current charging stage; the candidate voltage intervals and corresponding candidate features are used as inputs, and the battery health status estimation value is used as output, and the loaded pre-trained large language model is optimized to obtain an optimized large language model; A feature combination module, which is used to select multiple charging cycles for combining candidate voltage intervals in actual charging data, combine corresponding candidate voltage intervals, and extract corresponding candidate features; The state estimation module predicts the candidate voltage ranges and corresponding candidate features of the charging data to be predicted based on the optimized large language model to obtain the battery health status.

5. The battery health state estimation system based on the pre-trained large language model according to claim 4, characterized in that: In the model optimization module, the voltage-charge data and curve in the constant current charging stage are used to extract candidate voltage intervals and corresponding candidate features under different preset fixed charge capacities, including: Based on the voltage-charge data and curves in the constant current charging stage, a sliding window technique is used on a certain cycle to charge the battery at a fixed charge level. is the window size, with the rated capacity of the battery as the step size, along x Move in the axis direction, extract the data coordinate information of each window, and define each window in y The dimension in the axial direction is the voltage range , to obtain the voltage interval set corresponding to the cycle; Based on the voltage interval set extraction process, traverse all loops, and perform coordinate average calculation on multiple voltage intervals in all loops to obtain an average voltage interval set; Based on the average voltage interval set, traverse all cycles and calculate the charge amount corresponding to each average voltage interval to obtain a charge amount sequence corresponding to each average voltage interval; Based on the charge amount sequence corresponding to the average voltage interval, the Pearson correlation coefficient method is used to select the sequence with the strongest correlation with the health state, and the voltage interval corresponding to the sequence is defined as the candidate voltage interval , the charge amount corresponding to the candidate voltage interval is the candidate feature .

6. The battery health state estimation system based on a pre-trained large language model according to claim 4, characterized in that: In the state estimation module, the BERT-base model based on the Transformer encoder architecture is used to train the large language model.

Images