Lithium battery SOH estimation method and system fused with large language model

Abstract

The invention provides a lithium battery SOH estimation method and system fused with a large language model, and belongs to the technical field of battery management. The method comprises the following steps: S1, collecting multi-rate charging data, rated information and charging habit data of a lithium battery; s2, performing statistical analysis on each multi-rate charging data to obtain statistical characteristics; s3, generating cue words of a large language model based on the statistical characteristics, the rated information and the charging habit data; s4, inputting the cue word into a pre-trained large language model to obtain a first feature vector; s5, inputting the multi-rate charging data into a pre-trained time sequence model to obtain a second feature vector; and S6, fusing the first feature vector and the second feature vector based on a cross-modal attention mechanism to obtain a fused feature, and performing lithium battery SOH estimation based on the fused feature. The method has the advantage that the lithium battery SOH estimation precision is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, and particularly refers to a lithium battery SOH estimation method and system fusing a large language model. BACKGROUND

[0002] In the battery management system (BMS), the accurate estimation of the state of health (SOH) of the battery is crucial. The SOH refers to the ratio between the current actual capacity of the battery and its initial rated capacity, which directly affects the evaluation of the performance of the battery, the prediction of the remaining service life, and the safety and reliability of the system operation. With the wide application of lithium-ion batteries in electric vehicles, energy storage systems and other fields, higher requirements are put forward for the accuracy and reliability of the SOH estimation method.

[0003] Traditional SOH estimation methods are usually based on the charging time series data of the battery voltage, current, temperature and state of charge (SOC), and are analyzed and inferred through electrochemical models or data-driven models (such as neural networks, support vector machines, etc.). However, such methods have the following limitations: (1) Poor adaptability to complex charging modes: In the face of multi-rate charging scenarios, traditional methods are difficult to effectively capture the subtle differences in the dynamic characteristics of the battery under different rates. The multi-rate charging process involves complex electrochemical reactions and physical changes, and the voltage change rate, temperature rise trend, SOC adjustment process and the duration of each rate are coupled with each other, and traditional models often cannot fully integrate such multi-dimensional time series information, resulting in a significant decrease in the accuracy of SOH estimation under complex charging conditions.

[0004] (2) Insufficient use of comprehensive battery information: Traditional methods are mostly limited to basic electrical and thermal parameters, and do not introduce important information such as battery rated parameters (rated capacity, rated voltage, etc.) and usage behavior characteristics (charging and discharging frequency, deep discharge history, environmental conditions, etc.). These information has important reference significance for the overall evaluation of the health status of the battery, but traditional methods have not been able to systematically incorporate them into the modeling process, thereby limiting the overall representation ability of the true aging state of the battery.

[0005] (3) Weak feature extraction and fusion mechanism: Traditional time series models can only extract local or shallow features, and lack the ability to mine deep associations and cross-modal characteristics hidden in multi-modal charging data. In addition, in the feature fusion stage, traditional methods mostly rely on simple concatenation or linear weighting, which makes it difficult to achieve efficient integration of information from different sources, thereby restricting the comprehensive description of the health status of the battery and affecting the accuracy and robustness of the final SOH estimation.

[0006] Therefore, how to provide a lithium battery SOH estimation method and system fusing a large language model to improve the accuracy of lithium battery SOH estimation has become a technical problem to be solved. SUMMARY

[0007] The technical problem to be solved by the present application is to provide a lithium battery SOH estimation method and system fusing a large language model to improve the accuracy of lithium battery SOH estimation.

[0008] In a first aspect, the present application provides a lithium battery SOH estimation method fusing a large language model, comprising the following steps: Step S1, collecting multi-rate charging data, rated information and charging habit data of a lithium battery, and preprocessing each of the multi-rate charging data; Step S2, statistically analyzing each of the preprocessed multi-rate charging data to obtain statistical features; Step S3, generating prompt words of a large language model based on the statistical features, rated information and charging habit data; Step S4, inputting the prompt words into a pre-trained large language model to obtain a first feature vector; Step S5, inputting each of the preprocessed multi-rate charging data into a pre-trained time series model to obtain a second feature vector; Step S6, fusing the first feature vector and the second feature vector based on a cross-modal attention mechanism to obtain a fusion feature, and estimating the SOH of the lithium battery based on the fusion feature.

[0009] Further, the step S1 is specifically: Collecting multi-rate charging data, rated information and charging habit data of a lithium battery; the multi-rate charging data at least includes charging voltage, charging current, charging temperature, SOC and rate charging duration under each preset charging rate; the rated information at least includes rated capacity and rated voltage; the charging habit data at least includes charge-discharge cycle number, average charge-discharge depth and environmental temperature range; Dividing each of the multi-rate charging data into corresponding charging steps based on the charging rate, respectively counting charging features including at least charging duration, voltage variation range, temperature variation range and SOC variation range for each of the charging steps, and preprocessing each of the multi-rate charging data including at least outlier elimination and noise data elimination.

[0010] Further, the step S2 is specifically: Statistical analysis is performed on each of the pretreated multiple-rate charging data to extract statistical characteristics of charging voltage, charging current, charging temperature, SOC and rate charging duration under different charging rates; The step S3 is specifically: The statistical characteristics, rated information and charging habit data are combined to generate prompt words of the large language model in natural language form.

[0011] Further, in the step S4, the large language model is used to condense the input prompt words, capture potential correlations between different data dimensions, and convert them into a first feature vector in an abstract semantic representation form. In the step S5, the time series model is used to extract time series characteristics and time series patterns from the multiple-rate charging data, and then output a second feature vector.

[0012] Further, in the step S6, the formula of the cross-modal attention mechanism is: ; Wherein, represents the first feature vector; represents the second feature vector; represents the fusion feature; represents the normalized exponential function; represents the weight matrix of the query vector; represents the weight matrix of the key vector; represents the weight matrix of the value vector; represents the dimension of the query vector and the key vector; represents the transpose.

[0013] In a second aspect, the present application provides a lithium battery SOH estimation system fused with a large language model, comprising the following modules: A data acquisition module is used to acquire multiple-rate charging data, rated information and charging habit data of a lithium battery, and pretreat each of the multiple-rate charging data; A data statistical analysis module is used to perform statistical analysis on each of the pretreated multiple-rate charging data to obtain statistical characteristics; A prompt word generation module is used to generate prompt words of a large language model based on the statistical characteristics, rated information and charging habit data; A first feature vector extraction module is used to input the prompt words into a pre-trained large language model to obtain a first feature vector; A second feature vector extraction module is used to input each of the pretreated multiple-rate charging data into a pre-trained time series model to obtain a second feature vector; The lithium battery SOH estimation module is configured to fuse the first feature vector and the second feature vector based on a cross-modal attention mechanism to obtain a fused feature, and estimate the SOH of the lithium battery based on the fused feature.

[0014] Further, the data acquisition module is specifically configured to: acquire multi-rate charging data, rated information, and charging habit data of the lithium battery; the multi-rate charging data at least includes charging voltage, charging current, charging temperature, SOC, and rate charging duration under each preset charging rate; the rated information at least includes rated capacity and rated voltage; and the charging habit data at least includes charging and discharging cycle number, average charging and discharging depth, and environmental temperature range; divide each of the multi-rate charging data into a corresponding charging step based on a charging rate, and statistically count charging characteristics of each of the charging steps, including charging duration, voltage variation range, temperature variation range, and SOC variation range, and perform preprocessing on each of the multi-rate charging data, including outlier elimination and noise data elimination.

[0015] Further, the data statistical analysis module is specifically configured to: statistically analyze each of the preprocessed multi-rate charging data, and extract statistical characteristics of charging voltage, charging current, charging temperature, SOC, and rate charging duration under different charging rates; The prompt word generation module is specifically configured to: combine the statistical characteristics, rated information, and charging habit data to generate prompt words of the large language model in a natural language form.

[0016] Further, in the first feature vector extraction module, the large language model is used to condense the input prompt words, capture potential correlations between different data dimensions, and convert them into first feature vectors in an abstract semantic representation form; In the second feature vector extraction module, the time series model is used to extract time series characteristics and time series patterns from the multi-rate charging data, and then output second feature vectors.

[0017] Further, in the lithium battery SOH estimation module, the formula of the cross-modal attention mechanism is: ; wherein, represents the first feature vector; represents the second feature vector; represents the fused feature; represents a normalized exponential function; represents a weight matrix of a query vector; represents a weight matrix of a key vector; a weight matrix representing a value vector; denotes the dimension of a query vector and a key vector; denotes a transpose.

[0018] The present application has the advantages of: 1、By collecting multiple rate charging data of lithium batteries, rated information and charging habit data, each multiple rate charging data is preprocessed; then statistical characteristics are obtained by statistical analysis of each preprocessed multiple rate charging data; then the prompt words of the large language model are generated based on the statistical characteristics, rated information and charging habit data, the prompt words are input into the pre-trained large language model to obtain the first feature vector, and the preprocessed each multiple rate charging data is input into the pre-trained time sequence model to obtain the second feature vector, the first feature vector and the second feature vector are fused based on the cross-modal attention mechanism to obtain the fusion feature, and finally the lithium battery SOH estimation is performed based on the fusion feature; that is, first, collect multiple rate charging data (voltage / current / temperature / SOC), rated information and charging habit data, extract statistical characteristics after preprocessing; the statistical characteristics, rated information and charging habit data are converted into semantic feature vectors (first feature vectors) by the large language model, and the dynamic mode of multiple rate charging (second feature vector) is captured by using the time sequence model; finally, the first feature vector and the second feature vector are dynamically fused by using the cross-modal attention mechanism, wherein the attention weight automatically learns the complex correlation between the electrochemical characteristics and the battery historical behavior under different charging rates, realizes efficient collaboration of multi-source heterogeneous data, reduces the average error of SOH estimation under complex working conditions, especially improves the prediction robustness in extreme scenarios such as high-rate charging and deep cycling, and finally greatly improves the accuracy of lithium battery SOH estimation.

[0019] 2、By introducing a large language model into the field of lithium battery SOH estimation, battery data is processed by generating natural language prompt words, which is a breakthrough interdisciplinary innovation; the large language model can capture the potential correlation between data dimensions and abstract semantics, thereby revealing complex patterns that traditional methods cannot discover, improving the advanced nature and uniqueness of the technology.

[0020] 3、By collecting multiple rate charging data (including charging voltage, current, temperature, SOC and rate charging time), rated information (such as rated capacity and voltage) and charging habit data (such as charge and discharge cycle times, average charge and discharge depth and environmental temperature range), this multi-dimensional data collection ensures the richness and comprehensiveness of the input information, covering the static characteristics, dynamic behavior and use environment of the battery, providing a solid foundation for accurate SOH estimation.

[0021] 4. By preprocessing the data (such as outlier removal and noise removal) and performing statistical analysis (such as extracting statistical features), the data quality was effectively improved and noise interference was reduced. At the same time, by dividing the charging process into steps and statistically analyzing charging characteristics (such as charging time and voltage variation range), the data was efficiently organized and characterized, providing optimized information for subsequent model input.

[0022] 5. The prompt words generated by using large language models to process statistical features, rated information and charging habit data can condense information and capture potential correlations. This allows the model to output the first feature vector in an abstract semantic representation, enhancing the depth and flexibility of feature expression and improving the ability to understand complex data relationships.

[0023] 6. A pre-trained time series model is used to specifically handle the time series characteristics of multi-rate charging data, effectively extracting time-dependent features and patterns (such as the trend and periodic changes of the charging curve), ensuring that the second feature vector can accurately reflect the dynamic behavior of the battery, and making up for the shortcomings of traditional methods in time dimension analysis.

[0024] 7. A cross-modal attention mechanism is adopted to fuse the first feature vector (from the large language model) and the second feature vector (from the time series model), realizing the organic combination of semantic information and time series information. This fusion mechanism can adaptively weight the importance of different features, avoid information loss, and generate more comprehensive and accurate fused features, thereby significantly improving the accuracy and robustness of SOH estimation.

[0025] 8. By innovatively integrating large language models and time series models, and utilizing cross-modal attention mechanisms for feature fusion, high-precision estimation of lithium battery state of health (SOH) is achieved. Its advantages include: using multi-dimensional data (such as multi-rate charging data, rated information, and charging habit data) to ensure comprehensive input; improving data quality through data preprocessing and statistical analysis; using large language models to capture semantic associations and time series models to extract time patterns, thereby enhancing the depth and accuracy of feature representation; and the entire process is highly automated and scalable, suitable for real-time battery management systems, significantly improving the reliability and practicality of estimation. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Fig. 1 This is a flowchart of a lithium battery SOH estimation method that integrates a large language model according to the present invention.

[0028] Fig. 2 This is a schematic diagram of the structure of a lithium battery SOH estimation system that integrates a large language model according to the present invention.

[0029] Fig. 3 This is a flowchart of the architecture of the present invention. Detailed Implementation

[0030] The technical solution in this application embodiment follows the following general approach: First, multi-rate charging data, rated information, and charging habit data are collected and preprocessed to extract statistical features. Then, a large language model is used to transform the statistical features, rated information, and charging habit data into a first feature vector, while a time-series model is used to capture a second feature vector for multi-rate charging. Finally, a cross-modal attention mechanism is used to dynamically fuse the first and second feature vectors. The attention weights automatically learn the complex correlation between electrochemical characteristics and battery historical behavior at different charging rates, achieving efficient collaboration of multi-source heterogeneous data, reducing the average error of SOH estimation under complex operating conditions, and particularly improving the prediction robustness under extreme scenarios such as high-rate charging and deep cycling, thereby enhancing the accuracy of lithium battery SOH estimation.

[0031] Please refer to Figs. 1 to 3 As shown, a preferred embodiment of the lithium battery SOH estimation method integrating a large language model according to the present invention includes the following steps: Step S1: Collect multi-rate charging data, rated information and charging habit data of lithium battery, and preprocess each multi-rate charging data. Step S2: Perform statistical analysis on the preprocessed multi-rate charging data to obtain statistical characteristics; Step S3: Generate prompts for the large language model based on the statistical features, rated information, and charging habit data. Step S4: Input the prompt word into a pre-trained large language model (such as the Qwen series model) to obtain the first feature vector (embedding). That is, to leverage the powerful language understanding and semantic analysis capabilities of large language models to perform in-depth processing on the input prompt words; Step S5: Input the preprocessed multi-rate charging data into a pre-trained time series model (such as a recurrent neural network RNN, a long short-term memory network LSTM, a transformer, etc.) to obtain the second feature vector (embedding). Step S6: The first feature vector and the second feature vector are fused based on the cross attention mechanism to obtain fused features, and the SOH of the lithium battery is estimated based on the fused features.

[0032] In practice, a large amount of historical battery data with real SOH labels is used to supervise the training of each model. By optimizing the model parameters, the SOH of the battery can be accurately predicted based on the fused features. The large language model is in a frozen state, only performing inference and not participating in training, while the other models participate in training.

[0033] This invention introduces a large language model, which can fully explore the complex characteristics of batteries under multi-rate charging, as well as battery rating information, usage habits, and other aspects, to generate embeddings with rich semantics. After cross-attention fusion with the embeddings of traditional time-series models, the resulting features more comprehensively and accurately reflect the fused features of battery health status, thereby significantly improving the accuracy of SOH estimation. It can also more accurately assess the health status of batteries under complex charging modes and diverse usage scenarios.

[0034] This invention can effectively handle complex charging modes such as multi-rate charging and fully considers the characteristic changes of batteries under different usage conditions. Whether in electric vehicle applications with frequent variable-rate charging or energy storage system scenarios with specific usage habits, it can achieve reliable estimation of battery SOH through in-depth analysis and fusion of relevant information, greatly enhancing the adaptability of the battery management system to complex application scenarios.

[0035] By incorporating additional descriptive information such as battery ratings and charging habits into the SOH estimation process, a more comprehensive understanding of the battery's health status can be achieved. This information complements traditional charging timing data, providing richer evidence for accurate SOH estimation and helping to more realistically reflect the battery's aging and health status in actual use.

[0036] Step S1 specifically involves: Collect multi-rate charging data, rated information, and charging habit data of lithium batteries; the multi-rate charging data includes at least the charging voltage, charging current, charging temperature, SOC, and charging time at each preset charging rate; the rated information includes at least the rated capacity and rated voltage; the charging habit data includes at least the number of charge-discharge cycles, average depth of charge-discharge, and ambient temperature range. Based on the charging rate, the multi-rate charging data is divided into corresponding charging steps. For each charging step, charging characteristics including at least charging time, voltage variation range, temperature variation range, and SOC variation range are statistically analyzed. The multi-rate charging data is preprocessed, including at least outlier removal and noise data removal, to ensure data quality and usability.

[0037] Step S2 specifically involves: Statistical analysis was performed on the preprocessed multi-rate charging data to extract statistical characteristics of charging voltage, charging current, charging temperature, SOC, and charging time at different charging rates. Step S3 specifically involves: The statistical features, rated information, and charging habit data are combined to generate prompt words for a large language model in natural language form.

[0038] Example of the prompt: "A certain lithium battery, with a rated capacity of XAh and a rated voltage of YV, has undergone Z charge-discharge cycles, and its recent average depth of charge is D%. The charging occurred in a certain month in city A, with an ambient temperature typically between [Ta, Tb]. During this charging process, the battery was charged at a 0.5C rate for a duration of T1 minutes, during which the voltage rose from V1 to V2 (if the voltage initially decreased and then increased due to a reduction in the current rate, the relaxation time and magnitude would be described separately), the temperature rose from Tem1 to Tem2, and the state of charge (SOC) increased from S1 to S2; the battery was then charged at a 1C rate for a duration of T2 minutes..."

[0039] In step S4, the large language model is used to condense the input prompt words, capture the potential correlations between different data dimensions, and convert them into a first feature vector in an abstract semantic representation form. After processing by the large language model, the output is a first feature vector that highly corresponds to the input information (prompt words). The first feature vector is a high-dimensional numerical sequence that integrates rated information such as the battery's multi-rate charging characteristics, rated capacity, and rated voltage, as well as actual usage information such as the user's daily usage habits, charging and discharging frequency, and usage environment, in an abstract semantic representation. It not only condenses the key features of the battery's operating state but also captures the potential correlations between different data dimensions, forming a comprehensive and accurate high-dimensional feature description of battery health status information, providing support for subsequent battery health status estimation.

[0040] In step S5, the time series model is used to extract time series features and time series patterns from multi-rate charging data, and then output a second feature vector.

[0041] The time series model learns the time series features and patterns in the time series data and outputs a second feature vector that reflects the dynamic changes in the battery charging process. The second feature vector mainly focuses on capturing the time series characteristics of the battery charging process.

[0042] In step S6, the formula for the cross-modal attention mechanism is: ; in, This represents the first eigenvector; This represents the second eigenvector; Indicates fusion characteristics; This represents the normalized exponential function; The weight matrix representing the query vector; The weight matrix representing the key vector (query); The weight matrix representing the value vector (key); Indicates the dimensions of the query vector (value) and key vector; This indicates transpose.

[0043] The first feature vector is used as the query, and the second feature vector is used as the key and value for cross-modal information fusion.

[0044] A preferred embodiment of the lithium battery SOH estimation system integrating a large language model according to the present invention includes the following modules: The data acquisition module is used to collect multi-rate charging data, rated information and charging habit data of lithium batteries, and to preprocess the multi-rate charging data. The data statistical analysis module is used to perform statistical analysis on the preprocessed multi-rate charging data to obtain statistical characteristics; The prompt word generation module is used to generate prompt words for the large language model based on the statistical features, rated information, and charging habit data. The first feature vector extraction module is used to input the prompt word into a pre-trained large language model (such as the Qwen series model) to obtain the first feature vector (embedding). That is, to leverage the powerful language understanding and semantic analysis capabilities of large language models to perform in-depth processing on the input prompt words; The second feature vector extraction module is used to input the preprocessed multi-rate charging data into a pre-trained time series model (such as a recurrent neural network RNN, a long short-term memory network LSTM, a transformer, etc.) to obtain the second feature vector (embedding). The lithium battery SOH estimation module is used to fuse the first feature vector and the second feature vector based on the cross attention mechanism to obtain fused features, and to estimate the lithium battery SOH based on the fused features.

[0045] In practice, a large amount of historical battery data with real SOH labels is used to supervise the training of each model. By optimizing the model parameters, the SOH of the battery can be accurately predicted based on the fused features. The large language model is in a frozen state, only performing inference and not participating in training, while the other models participate in training.

[0046] This invention introduces a large language model, which can fully explore the complex characteristics of batteries under multi-rate charging, as well as battery rating information, usage habits, and other aspects, to generate embeddings with rich semantics. After cross-attention fusion with the embeddings of traditional time-series models, the resulting features more comprehensively and accurately reflect the fused features of battery health status, thereby significantly improving the accuracy of SOH estimation. It can also more accurately assess the health status of batteries under complex charging modes and diverse usage scenarios.

[0047] This invention can effectively handle complex charging modes such as multi-rate charging and fully considers the characteristic changes of batteries under different usage conditions. Whether in electric vehicle applications with frequent variable-rate charging or energy storage system scenarios with specific usage habits, it can achieve reliable estimation of battery SOH through in-depth analysis and fusion of relevant information, greatly enhancing the adaptability of the battery management system to complex application scenarios.

[0048] By incorporating additional descriptive information such as battery ratings and charging habits into the SOH estimation process, a more comprehensive understanding of the battery's health status can be achieved. This information complements traditional charging timing data, providing richer evidence for accurate SOH estimation and helping to more realistically reflect the battery's aging and health status in actual use.

[0049] The data acquisition module is specifically used for: Collect multi-rate charging data, rated information, and charging habit data of lithium batteries; the multi-rate charging data includes at least the charging voltage, charging current, charging temperature, SOC, and charging time at each preset charging rate; the rated information includes at least the rated capacity and rated voltage; the charging habit data includes at least the number of charge-discharge cycles, average depth of charge-discharge, and ambient temperature range. Based on the charging rate, the multi-rate charging data is divided into corresponding charging steps. For each charging step, charging characteristics including at least charging time, voltage variation range, temperature variation range, and SOC variation range are statistically analyzed. The multi-rate charging data is preprocessed, including at least outlier removal and noise data removal, to ensure data quality and usability.

[0050] The data statistical analysis module is specifically used for: Statistical analysis was performed on the preprocessed multi-rate charging data to extract statistical characteristics of charging voltage, charging current, charging temperature, SOC, and charging time at different charging rates. The prompt word generation module is specifically used for: The statistical features, rated information, and charging habit data are combined to generate prompt words for a large language model in natural language form.

[0051] Example of the prompt: "A certain lithium battery, with a rated capacity of XAh and a rated voltage of YV, has undergone Z charge-discharge cycles, and its recent average depth of charge is D%. The charging occurred in a certain month in city A, with an ambient temperature typically between [Ta, Tb]. During this charging process, the battery was charged at a 0.5C rate for a duration of T1 minutes, during which the voltage rose from V1 to V2 (if the voltage initially decreased and then increased due to a reduction in the current rate, the relaxation time and magnitude would be described separately), the temperature rose from Tem1 to Tem2, and the state of charge (SOC) increased from S1 to S2; the battery was then charged at a 1C rate for a duration of T2 minutes..."

[0052] In the first feature vector extraction module, the large language model is used to condense the input prompt words, capture the potential correlation between different data dimensions, and convert them into a first feature vector in an abstract semantic representation form; After processing by the large language model, the output is a first feature vector that highly corresponds to the input information (prompt words). The first feature vector is a high-dimensional numerical sequence that integrates rated information such as the battery's multi-rate charging characteristics, rated capacity, and rated voltage, as well as actual usage information such as the user's daily usage habits, charging and discharging frequency, and usage environment, in an abstract semantic representation. It not only condenses the key features of the battery's operating state but also captures the potential correlations between different data dimensions, forming a comprehensive and accurate high-dimensional feature description of battery health status information, providing support for subsequent battery health status estimation.

[0053] In the second feature vector extraction module, the time series model is used to extract time series features and time series patterns from multi-rate charging data, and then output the second feature vector.

[0054] The time series model learns the time series features and patterns in the time series data and outputs a second feature vector that reflects the dynamic changes in the battery charging process. The second feature vector mainly focuses on capturing the time series characteristics of the battery charging process.

[0055] In the lithium battery SOH estimation module, the formula for the cross-modal attention mechanism is: ; in, This represents the first eigenvector; This represents the second eigenvector; Indicates fusion characteristics; This represents the normalized exponential function; The weight matrix representing the query vector; The weight matrix representing the key vector (query); The weight matrix representing the value vector (key); Indicates the dimensions of the query vector (value) and key vector; This indicates transpose.

[0056] The first feature vector is used as the query, and the second feature vector is used as the key and value for cross-modal information fusion.

[0057] In summary, the advantages of this invention are: 1. By collecting multi-rate charging data, rated information, and charging habit data of lithium batteries, the multi-rate charging data is preprocessed. Then, statistical analysis is performed on the preprocessed multi-rate charging data to obtain statistical features. Next, based on the statistical features, rated information, and charging habit data, prompt words for a large language model are generated. The prompt words are input into a pre-trained large language model to obtain the first feature vector. The preprocessed multi-rate charging data is input into a pre-trained time-series model to obtain the second feature vector. The first and second feature vectors are fused using a cross-modal attention mechanism to obtain fused features. Finally, the SOH of the lithium battery is estimated based on the fused features. In other words, first, multi-rate charging data (voltage / current / temperature / SO) is collected. C) Rated information and charging habit data are preprocessed to extract statistical features. A large language model is used to transform these statistical features, rated information, and charging habit data into semantic feature vectors (first feature vector). Simultaneously, a time-series model is used to capture the dynamic patterns of multi-rate charging (second feature vector). Finally, a cross-modal attention mechanism is used to dynamically fuse the first and second feature vectors. The attention weights automatically learn the complex correlation between electrochemical characteristics and battery historical behavior at different charging rates, achieving efficient collaboration of multi-source heterogeneous data. This reduces the average error of SOH estimation under complex operating conditions, particularly improving the prediction robustness in extreme scenarios such as high-rate charging and deep cycling, ultimately greatly enhancing the accuracy of lithium battery SOH estimation.

[0058] 2. By introducing large language models into the field of lithium battery SOH estimation and processing battery data by generating natural language prompts, this is a groundbreaking interdisciplinary innovation. Large language models can capture the potential correlations and abstract semantics between data dimensions, thereby revealing complex patterns that are difficult to discover by traditional methods, thus enhancing the advancement and uniqueness of the technology.

[0059] 3. By collecting multi-rate charging data (including charging voltage, current, temperature, SOC, and multi-rate charging duration), rated information (such as rated capacity and voltage), and charging habit data (such as charge-discharge cycle count, average charge-discharge depth, and ambient temperature range), this multi-dimensional data acquisition ensures the richness and comprehensiveness of the input information, covering the battery's static characteristics, dynamic behavior, and usage environment, providing a solid foundation for accurately estimating SOH.

[0060] 4. By preprocessing the data (such as outlier removal and noise removal) and performing statistical analysis (such as extracting statistical features), the data quality was effectively improved and noise interference was reduced. At the same time, by dividing the charging process into steps and statistically analyzing charging characteristics (such as charging time and voltage variation range), the data was efficiently organized and characterized, providing optimized information for subsequent model input.

[0061] 5. The prompt words generated by using large language models to process statistical features, rated information and charging habit data can condense information and capture potential correlations. This allows the model to output the first feature vector in an abstract semantic representation, enhancing the depth and flexibility of feature expression and improving the ability to understand complex data relationships.

[0062] 6. A pre-trained time series model is used to specifically handle the time series characteristics of multi-rate charging data, effectively extracting time-dependent features and patterns (such as the trend and periodic changes of the charging curve), ensuring that the second feature vector can accurately reflect the dynamic behavior of the battery, and making up for the shortcomings of traditional methods in time dimension analysis.

[0063] 7. A cross-modal attention mechanism is adopted to fuse the first feature vector (from the large language model) and the second feature vector (from the time series model), realizing the organic combination of semantic information and time series information. This fusion mechanism can adaptively weight the importance of different features, avoid information loss, and generate more comprehensive and accurate fused features, thereby significantly improving the accuracy and robustness of SOH estimation.

[0064] 8. By innovatively integrating large language models and time series models, and utilizing cross-modal attention mechanisms for feature fusion, high-precision estimation of lithium battery state of health (SOH) is achieved. Its advantages include: using multi-dimensional data (such as multi-rate charging data, rated information, and charging habit data) to ensure comprehensive input; improving data quality through data preprocessing and statistical analysis; using large language models to capture semantic associations and time series models to extract time patterns, thereby enhancing the depth and accuracy of feature representation; and the entire process is highly automated and scalable, suitable for real-time battery management systems, significantly improving the reliability and practicality of estimation.

[0065] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A lithium battery SOH estimation method of a fusion large language model, characterized by: Includes the following steps: Step S1: Collect multi-rate charging data, rated information and charging habit data of lithium battery, and preprocess each multi-rate charging data. Step S2: Perform statistical analysis on the preprocessed multi-rate charging data to obtain statistical characteristics; Step S3: Generate prompt words for a large language model based on the statistical features, rated information, and charging habit data; Step S4: Input the prompt words into a pre-trained large language model to obtain the first feature vector; Step S5: Input the preprocessed multi-rate charging data into the pre-trained time series model to obtain the second feature vector; Step S6: The first feature vector and the second feature vector are fused based on the cross-modal attention mechanism to obtain fused features, and the SOH of the lithium battery is estimated based on the fused features.

2. The lithium battery SOH estimation method of fusing large language models according to claim 1, wherein: Step S1 specifically involves: Collect multi-rate charging data, rated information, and charging habit data of lithium batteries; the multi-rate charging data includes at least the charging voltage, charging current, charging temperature, SOC, and charging time at each preset charging rate. The rated information includes at least the rated capacity and rated voltage; the charging habit data includes at least the number of charge-discharge cycles, average depth of charge-discharge, and ambient temperature range. Based on the charging rate, the multi-rate charging data is divided into corresponding charging steps. For each charging step, charging characteristics including at least charging time, voltage variation range, temperature variation range, and SOC variation range are statistically analyzed. The multi-rate charging data is then preprocessed, including at least outlier removal and noise data removal.

3. The lithium battery SOH estimation method of fusing large language models according to claim 1, wherein: Step S2 specifically involves: Statistical analysis was performed on the preprocessed multi-rate charging data to extract statistical characteristics of charging voltage, charging current, charging temperature, SOC, and charging time at different charging rates. Step S3 specifically involves: The statistical features, rated information, and charging habit data are combined to generate prompt words for a large language model in natural language form.

4. The lithium battery SOH estimation method of fusing large language models according to claim 1, wherein: In step S4, the large language model is used to condense the input prompt words, capture the potential correlations between different data dimensions, and convert them into a first feature vector in an abstract semantic representation form. In step S5, the time series model is used to extract time series features and time series patterns from multi-rate charging data, and then output a second feature vector.

5. The lithium battery SOH estimation method of fusing large language models according to claim 1, wherein: In the step S6, the formula of the cross-modal attention mechanism is: ; wherein, denotes a first feature vector; denotes a second feature vector; denotes a fused feature; denotes a normalized exponential function; denotes a weight matrix of a query vector; denotes a weight matrix of a key vector; denotes a weight matrix of a value vector; denotes a dimension of a query vector and a key vector; denotes a transpose.

6. A lithium battery SOH estimation system fused with a large language model, characterized by: Includes the following modules: The data acquisition module is used to collect multi-rate charging data, rated information and charging habit data of lithium batteries, and to preprocess the multi-rate charging data. The data statistical analysis module is used to perform statistical analysis on the preprocessed multi-rate charging data to obtain statistical characteristics; The prompt word generation module is used to generate prompt words for a large language model based on the statistical features, rated information, and charging habit data. The first feature vector extraction module is used to input the prompt word into a pre-trained large language model to obtain the first feature vector; a second feature vector extraction module configured to input the preprocessed multi-rate charging data into a pre-trained time sequence model to obtain a second feature vector; a lithium battery SOH estimation module configured to fuse the first feature vector and the second feature vector based on a cross-modal attention mechanism to obtain fused features, and estimate the SOH of the lithium battery based on the fused features.

7. The lithium battery SOH estimation system of a fused large language model of claim 6, wherein: The data collection module is specifically configured to: collect multi-rate charging data, rated information and charging habit data of the lithium battery; the multi-rate charging data at least includes charging voltage, charging current, charging temperature, SOC and rate charging duration under each preset charging rate; the rated information at least includes rated capacity and rated voltage; the charging habit data at least includes charging and discharging cycle number, average charging and discharging depth and environmental temperature range; divide each of the multi-rate charging data into corresponding charging steps based on the charging rate, respectively count charging characteristics of each of the charging steps, and pre-process each of the multi-rate charging data, the charging characteristics at least including abnormal value elimination and noise data elimination.

8. The lithium battery SOH estimation system of a fused large language model of claim 6, wherein: The data statistical analysis module is specifically configured to: statistically analyze each of the preprocessed multi-rate charging data, and extract statistical characteristics of charging voltage, charging current, charging temperature, SOC and rate charging duration under different charging rates; The prompt word generation module is specifically configured to: combine the statistical characteristics, rated information and charging habit data to generate prompt words of the large language model in the form of natural language.

9. The lithium battery SOH estimation system of a fused large language model of claim 6, wherein: In the first feature vector extraction module, the large language model is used to condense the input prompt words, capture potential correlations between different data dimensions, and convert them into first feature vectors in the form of abstract semantic representation; In the second feature vector extraction module, the time sequence model is used to extract time sequence characteristics and time sequence patterns from the multi-rate charging data, and then output second feature vectors.

10. The lithium battery SOH estimation system fused with a large language model of claim 6, wherein: In the lithium battery SOH estimation module, the formula of the cross-modal attention mechanism is: ； wherein, denotes a first feature vector; denotes a second feature vector; denotes a fused feature; denotes a normalized exponential function; denotes a weight matrix of the query vector; denotes a weight matrix of the key vector; denotes a weight matrix of the value vector; denotes a dimension of the query vector and the key vector; denotes a transpose.

Images