An artificial intelligence-based lithium battery health degree detection method and system

Abstract

The present application relates to a kind of lithium battery health degree detection method and system based on artificial intelligence, it is related to artificial intelligence technical field, lithium battery health degree detection method, comprising: according to vehicle source data evaluation vehicle basic characteristics, obtain battery characteristic data table;Battery charging state is as research object, analysis charging segment, determine sampling time threshold, according to battery residual charge judgment data missing and interpolation filling;Based on charge charging time, calculate battery capacity, analyze battery capacity in horizontal direction, obtain initial capacity value;And according to preset trip constraint condition, extract the effective capacity data in all initial capacity value;Train single package health degree evaluation matrix or double package health degree evaluation matrix, predict charging segment sequence, construct battery health degree evaluation model;Correct single package health degree evaluation matrix or double package health degree evaluation matrix, obtain optimal health degree evaluation weight matrix, complete lithium battery health degree detection in combination with physical compensation mechanism;The present application avoids SOH upward fluctuation, improves SOH estimation accuracy and model adaptability.

Description

Technical Field

[0001] This invention relates to the technical field of power battery testing, and in particular to an artificial intelligence-based method and system for testing the health of lithium batteries. Background Technology

[0002] As the global energy structure shifts towards cleaner and lower-carbon energy, lithium batteries, with their advantages of high energy density, long cycle life, and low self-discharge rate, have become the core energy carrier for electric vehicles, energy storage power stations, and portable electronic devices.

[0003] However, traditional SOH detection methods for lithium batteries (such as empirical SOH estimation methods and performance-based SOH estimation methods) have significant limitations. Among them, empirical SOH estimation methods, also known as statistical methods, mainly include cycle number method, ampere-hour method and weighted ampere-hour method, as well as event-oriented aging accumulation method. According to the different information sources used in lifetime prediction, performance-based SOH estimation methods are divided into three categories: mechanism-based, feature-based, and data-driven.

[0004] Among them, mechanism-based prediction analyzes and establishes the battery's operating mechanism model and aging model from the perspective of the battery's inherent mechanism, describes the battery's aging behavior from the perspective of electrochemical principles, and predicts the battery's lifespan through the analysis of the battery model.

[0005] Feature-based prediction utilizes the evolution of feature parameters exhibited during battery aging to establish a correspondence between feature quantities and battery life for life prediction.

[0006] Data-driven prediction utilizes battery performance test data to extract patterns in battery performance evolution for lifespan prediction. For example, analytical models derived from data fitting and artificial neural network models are both data-driven methods.

[0007] Existing application number CN202411807136.3 discloses a rapid testing instrument and method for detecting the state of health (SOH) of new energy vehicle batteries. It includes: connecting a testing probe to the charging port of a new energy vehicle battery, using a mobile phone program to control the testing instrument, acquiring battery-related data, analyzing and processing the data using an intelligent algorithm combined with the collected data, assessing the battery's SOH, generating a detailed analysis report, providing battery valuation and maintenance suggestions, and offering battery safety status information. This invention can quickly and accurately detect battery SOH values, providing battery safety status information and health status, etc. Compared to traditional testing methods, it does not require battery disassembly, is faster, more accurate, and simpler to operate. It can be widely applied to scenarios such as routine vehicle maintenance battery status checks, vehicle fault repair battery checks, and used car recycling value assessment.

[0008] Application number CN202410513892.9 discloses a real-vehicle lithium battery SOH estimation method based on big data and hybrid machine learning. By embedding CatBoost as the base learner of the NGBoost algorithm, a new N-CatBoost boosting algorithm framework is provided, which can combine the advantages of the two algorithms. This allows CatBoost's efficient processing of classification features and its ability to prevent overfitting to be fully utilized, as well as NGBoost's ability to evaluate prediction uncertainty, thereby significantly improving the estimation accuracy of lithium battery SOH.

[0009] The existing technical solutions mentioned above have the following drawbacks: 1. The battery data collected is usually uncertain and incomplete, resulting in low accuracy of battery health estimation; 2. Due to the different sources of battery data, the evolution of the pre-trained battery health assessment model is too limited, which in turn leads to an upward fluctuation in battery health status. Summary of the Invention

[0010] To address the shortcomings of existing technologies, the present invention aims to provide an artificial intelligence-based lithium battery health detection method and system. This method uses artificial intelligence to train a battery health assessment model based on historical vehicle source data. Then, it uses real-time vehicle source data combined with a physical compensation mechanism to correct the model and perform transfer learning. The physical compensation mechanism also constrains the model to return it to a specific range, effectively preventing upward fluctuations in state of health (SOH) and significantly improving SOH estimation accuracy and model adaptability.

[0011] The above-mentioned objective of this invention is achieved through the following technical solutions:

[0012] An artificial intelligence-based method for detecting the health of lithium batteries, comprising:

[0013] The information correlation between vehicle basic characteristics and battery charge and discharge characteristics is calculated based on vehicle source data. Based on the information correlation, battery candidate characteristics including sampling time, total current and battery remaining charge are determined from the battery charge and discharge characteristics. All battery candidate characteristics are summarized into a battery characteristic data table.

[0014] The battery charging state is determined based on the battery characteristic data table and the total current polarity, and a charging segment is extracted from the battery charging state. At the same time, the sampling time threshold of the charging segment is determined, and the data missing state is judged according to the battery remaining charge, and the corresponding missing value is filled in.

[0015] Based on the charge charging time of the charging segment, the battery capacity of each vehicle is calculated one by one. The total battery capacity is analyzed in a targeted manner to obtain the initial capacity values ​​corresponding to the battery capacity. The initial capacity values ​​are then cleaned and verified to obtain the effective capacity data.

[0016] The effective capacity data is divided into single-packet capacity datasets and dual-packet capacity datasets according to the capacity type. These datasets are then input into a neural network model for training to obtain single-packet health assessment matrices and dual-packet health assessment matrices, and the corresponding theoretical charging segment sequences are output.

[0017] Based on the battery feature data table, the single-pack health assessment matrix and the dual-pack health assessment matrix are corrected to obtain the corresponding optimal health assessment weight matrix, thereby generating a comprehensive battery health assessment model and completing the battery health detection.

[0018] By adopting the above technical solution, the correlation between each basic vehicle feature and the battery charging and discharging feature is calculated based on the vehicle source data. N key battery features are selected based on the correlation, and corresponding source data is extracted from the vehicle source data based on these key battery features to construct a battery feature data table. The battery charging state is determined based on the battery feature data table and the total current polarity, and charging segments are extracted from the battery charging state. The sampling time threshold is determined based on the sampling time proportion of the charging segments, and the source data within the sampling time is judged for missing data based on the remaining battery charge, and interpolation is performed to fill in the missing data using an interpolation strategy to obtain corrected historical source data. Simultaneously, based on the charging time of the charging segments, a two-stage capacity calculation method is used to calculate the battery capacity of each vehicle, and the battery capacity is analyzed horizontally and vertically to obtain several initial capacity values. Based on the travel constraints and the fitting residual cleaning and robust regression verification of all initial capacity values, effective capacity data is obtained. Then, based on the capacity type, the effective capacity data is divided into dual-pack capacity datasets and single-pack capacity datasets. By dividing the dual-pack capacity dataset and the single-pack capacity dataset according to data type and time series, historical array input information and historical metadata output information corresponding to the dual-pack capacity dataset and the single-pack capacity dataset are obtained, resulting in a charging segment sequence. Simultaneously, all historical array input information is input into a graph neural network for iterative training to initialize the battery health assessment model, yielding single-pack and dual-pack health assessment matrices. Based on the physical compensation mechanism of the battery feature data table set, the single-pack and dual-pack health assessment matrices are corrected to obtain the corresponding optimal health assessment weight matrix, thereby completing the lithium battery health detection, reducing the upward fluctuation trend of SOH, improving SOH estimation accuracy, and enhancing the adaptability of the battery health assessment model.

[0019] The present invention is further configured such that: the specific steps of calculating the information correlation degree between vehicle basic characteristics and battery charge / discharge characteristics based on vehicle source data, determining battery candidate characteristics including sampling time, total current and battery remaining charge from the battery charge / discharge characteristics based on the information correlation degree, and summarizing all the battery candidate characteristics into a battery characteristic data table include:

[0020] The vehicle status is monitored, and vehicle source data is extracted to obtain several sets of basic vehicle features and battery charging / discharging features. Based on the vehicle source data, the information correlation between each set of basic vehicle features and battery charging / discharging features is calculated to obtain a feature validity evaluation value, which is then compared with a preset first feature correlation threshold.

[0021] If the effective evaluation value of the feature is greater than or equal to the first feature-related threshold, then the current vehicle basic feature is used as a battery candidate feature;

[0022] If the effective evaluation value of the feature is less than the first feature-related threshold, then the current vehicle basic features are removed.

[0023] Arrange all the valid evaluation values ​​of the battery candidate features in descending order, and extract the source data corresponding to the battery candidate features according to the preset target number;

[0024] The extracted battery candidate features and their corresponding source data are summarized to generate a battery feature data table.

[0025] Furthermore, if the number of candidates for the battery candidate features is less than the target number, then the source data corresponding to all battery candidate features that currently exceed the first feature-related threshold are used to construct an initial battery-related table.

[0026] Simultaneously, the valid evaluation values ​​of all vehicle basic features that do not exceed the first feature-related threshold are compared with a preset second feature-related threshold for judgment:

[0027] If the effective evaluation value of the feature is less than the second feature-related threshold, then the current vehicle basic features will be removed.

[0028] If the effective evaluation value of the feature is greater than or equal to the second feature-related threshold, then according to the difference between the preset target number and the candidate number, the same number of source data corresponding to the vehicle basic features are selected and filled into the initial battery-related table to complete the construction of the battery feature data table.

[0029] By adopting the above technical solution, the correlation between each vehicle basic feature and battery charging and discharging feature is calculated based on the vehicle source data to obtain the effective feature evaluation value, and then compared with the preset first feature correlation threshold and the second feature correlation threshold in sequence. At the same time, all battery candidate features are obtained according to the preset target number, and the battery feature data table is constructed. This reduces feature data that is not related to battery charging and discharging features and reduces the impact of invalid features on the battery.

[0030] The present invention is further configured to: determine the battery charging state based on the battery characteristic data table and the total current polarity, extract charging segments from the battery charging state, determine the sampling time threshold of the charging segment, and determine the data missing state based on the remaining battery charge, and fill in the corresponding missing values. The specific steps include:

[0031] Determine the polarity of the total current in the battery characteristic data table:

[0032] If the polarity of the total current is irregular and alternates between positive and negative over a period of time, then the current battery is determined to be in a discharge state, and all discharge segments in the discharge state are eliminated.

[0033] If the polarity of the total current is concentrated and remains negative for a prolonged period, then the current battery is determined to be in a charging state, and all charging segments during the battery charging state are extracted.

[0034] Cluster the sampling times of all the charging segments to obtain the proportion of each sampling time, and take the sampling time with the largest proportion as the sampling time threshold;

[0035] If the current sampling time is less than or equal to the sampling time threshold, it is determined that there is no missing source data, and the current sampling time is corrected to the sampling time threshold.

[0036] If the current sampling time is greater than the sampling time threshold, it is determined that the current source data is missing, and the remaining battery charge is assessed.

[0037] If the remaining charge of the battery increases, the vehicle is determined to be in a charging state during the current sampling time, and the source data during the current sampling time is interpolated and filled using a preset interpolation strategy.

[0038] If the remaining charge of the battery remains unchanged and the sampling time is less than a preset time span threshold, the vehicle is determined to be in a non-charging state during the current sampling time. At the same time, the source data during the current sampling time is interpolated and filled using an interpolation strategy.

[0039] If the remaining charge of the battery remains unchanged, but the sampling time is greater than or equal to a preset time span threshold, the vehicle is determined to be in a non-charging state during the current sampling time, and no interpolation is performed on the source data during the current sampling time.

[0040] By adopting the above technical solution, the source data is used to determine the research object, correct the sampling time, and judge the remaining charge of the battery based on the total current polarity and battery remaining charge in the battery characteristic data table. This completes the data missing judgment and interpolation filling, ensuring the integrity and authenticity of the source data.

[0041] The present invention is further configured such that: the specific steps of calculating the battery capacity of each vehicle one by one according to the charge charging time of the charging segment, performing directional analysis on all the battery capacities, and obtaining initial capacity values ​​corresponding to several battery capacities include:

[0042] Feature extraction is performed on all the charging segments to obtain the charging start range, charging end range, and charge increase interval; based on the charging start range and the charging end range, the first segment charging capacity within the preset charge increase interval is calculated;

[0043] The battery capacity of the first segment of the first N charging increase intervals is calculated based on the first segment charging capacity, and the charging capacity of the second segment of the remaining charging increase intervals is calculated separately.

[0044] Calculate the battery capacity of all vehicles based on the first segment charging capacity and the second segment charging capacity.

[0045] A directional analysis is performed on all the battery capacities to obtain initial capacity values ​​corresponding to several battery capacities; the directional analysis includes lateral clustering analysis based on capacity values ​​and longitudinal clustering analysis based on current values.

[0046] By adopting the above technical solution, the battery capacity calculation is divided into two stages based on the charge charging time, and the total battery capacity is analyzed horizontally and vertically to obtain several initial capacity values; this reduces the probability of falsely labeled battery remaining charge.

[0047] The present invention is further configured such that the specific steps of cleaning and verifying all the initial capacity values ​​to obtain valid capacity data include:

[0048] Based on the travel constraint condition that the number of driving days is positively correlated with the driving mileage, all the initial capacity values ​​are linearly fitted and cleaned. The fitting residual value of each initial capacity value is calculated and compared with the preset outlier threshold.

[0049] If the fitting residual value is greater than or equal to the outlier threshold, the current battery capacity is determined to be an outlier, and all outliers are removed.

[0050] If the fitting residual value is less than the outlier threshold, the current battery capacity is determined to be a non-outlier, and all the non-outliers are retained to generate a non-outlier set.

[0051] Robust regression verification is performed on the set of non-outliers, and the corrected residual value of each non-outlier is calculated and compared with a preset correction threshold.

[0052] If the corrected residual value is greater than or equal to the correction threshold, then the current non-outlier point is removed.

[0053] If the corrected residual value is less than the corrected threshold, then the direction of change of the battery capacity of the current non-outlier point with driving mileage is verified:

[0054] If the battery capacity does not decrease linearly with the driving mileage, then the current non-outlier point will be removed.

[0055] If the battery capacity decreases linearly with the mileage, the current outlier will be retained in the set of outliers and thus determined as valid capacity data.

[0056] By adopting the above technical solution, and based on the preset travel constraints, all initial capacity values ​​are fitted with residual judgment and robust regression verification to obtain effective capacity data, thereby reducing the problem of excessive capacity deviation caused by time-missed sampling differences.

[0057] The present invention is further configured such that: the specific steps of dividing the effective capacity data into single-packet capacity datasets and dual-packet capacity datasets according to capacity type, inputting them into a neural network model for training to obtain single-packet health assessment matrices and dual-packet health assessment matrices, and outputting the corresponding theoretical charging segment sequences include:

[0058] The effective capacity data is divided into dual-packet capacity datasets and single-packet capacity datasets based on capacity type.

[0059] The dual-packet capacity dataset and the single-packet capacity dataset are divided according to data type and time series, respectively, to obtain the historical array input information and historical metadata output information corresponding to the dual-packet capacity dataset and the historical array input information and historical metadata output information corresponding to the single-packet capacity dataset.

[0060] The historical metadata output information is sorted according to the vehicle number and charging segment, and corresponding charging segment sequences are generated that are positively correlated with driving time and driving mileage.

[0061] The historical array input information is normalized according to the preset extreme value label to obtain the corresponding array training dataset; the array training dataset is input into the graph neural network for several iterations of training, and the historical metadata output information is used as the output feature of the graph neural network to obtain the single-packet health evaluation matrix and the double-packet health evaluation matrix, and then the theoretical charging segment sequence is output respectively.

[0062] By adopting the above technical solution, the dual-pack capacity dataset and the single-pack capacity dataset are divided separately to obtain historical array input information and historical metadata output information. At the same time, the historical array input information is normalized to obtain the array training dataset, which is then input into the graph neural network for several iterations of training to initialize the battery health assessment model, obtain the corresponding single-pack health assessment matrix and dual-pack health assessment matrix, and output the theoretical charging segment sequence respectively; thus improving the accuracy of the battery health assessment model.

[0063] The present invention is further configured such that the construction steps of the battery health assessment model include:

[0064] The feature extraction layer fills the data input matrix constructed from the array training dataset with zero vectors according to the convolution kernel dimension, and uses several stacked convolution blocks of dimension L to extract convolution features from the array training dataset, thereby obtaining the battery feature matrix.

[0065] The gate layer uses matrix multiplication to filter the battery feature matrices of different stacked convolutional blocks to obtain the battery weight matrix;

[0066] The linear layer performs a linear transformation on the battery weight matrix based on the weight vector matrix and bias matrix, combined with the pooling function and loss function, to obtain a single-packet health assessment matrix or a dual-packet health assessment matrix.

[0067] The fully connected layer maps the single-packet health assessment matrix or the double-packet health assessment matrix into the corresponding one-dimensional feature tensor, and matches it with the historical metadata output information to obtain the corresponding theoretical charging segment sequence.

[0068] By adopting the above technical solution, the array training dataset is sequentially passed through a feature extraction layer, a gating layer, a linear layer, and a fully connected layer for convolutional feature extraction, noise reduction and filtering, linear transformation, and matching prediction, thereby obtaining a single-packet health assessment matrix or a double-packet health assessment matrix and outputting the corresponding theoretical charging segment sequence; thus improving the model's evolution capability.

[0069] The present invention is further configured such that: the specific steps of correcting the single-pack health assessment matrix or the dual-pack health assessment matrix according to the battery feature data table to obtain the corresponding optimal health assessment weight matrix include:

[0070] The real-time array input information from the battery feature data table is loaded into the corresponding single-pack health assessment matrix and the dual-pack health assessment matrix for prediction, resulting in the corresponding theoretical charging segment sequence. This sequence is then matched with the corresponding real-time charging segment sequence to obtain the sequence similarity, which is then compared with a preset similarity threshold.

[0071] If the sequence similarity is greater than or equal to the similarity threshold, the learning rate of the current single-packet health assessment matrix or double-packet health assessment matrix remains unchanged.

[0072] If the sequence similarity is less than the similarity threshold, the current battery health assessment model is determined to be overfitted. The learning rate of the current single-pack health assessment matrix and / or the dual-pack health assessment matrix is ​​dynamically corrected and compared with the preset warning learning threshold.

[0073] If the corrected learning rate is greater than the warning learning threshold, the current battery health assessment model is deemed to have redundant structure. Random pruning and / or adding an attention mechanism are then performed on the feature extraction layer. At the same time, the convolution kernel dimension of the feature extraction layer is expanded and / or the loss function of the linear layer is replaced, thereby obtaining the corresponding optimal health assessment weight matrix.

[0074] By adopting the above technical solution, the single-pack health assessment matrix or the dual-pack health assessment matrix is ​​predicted based on the real-time array input information to obtain the charging segment prediction sequence, which is then matched with the real-time charging segment sequence. The learning rate is dynamically adjusted based on the matching result, thereby completing the correction of each layer of the model and obtaining the optimal health assessment weight matrix, which improves the prediction accuracy of the battery health assessment model.

[0075] The present invention is further configured such that the specific steps for completing the lithium battery health detection by combining the optimal health assessment weight matrix with a physical compensation mechanism include:

[0076] The prediction accuracy of the optimal health assessment weight matrix is ​​calculated and compared with the preset assessment accuracy threshold. If the prediction accuracy is greater than or equal to the assessment accuracy threshold, it indicates that the training performance of the current battery health assessment model meets the standard and the sampling time of the array training dataset is unified.

[0077] If the prediction accuracy is less than the evaluation accuracy threshold, it indicates that the training performance of the current battery health assessment model is not up to standard. The interpolation strategy of the array training dataset is optimized and / or a temperature correction coefficient is added to the interpolation strategy according to the temperature-capacity distribution relationship, thereby completing the detection of lithium battery health.

[0078] By adopting the above technical solution, based on the physical compensation mechanism, the prediction accuracy of the optimal health assessment weight matrix is ​​calculated using the temperature correction coefficient, and compared with the preset assessment accuracy threshold. Based on the comparison results, the temperature correction coefficient or interpolation strategy is optimized for the array training dataset, ensuring the accuracy of the source data and improving the model prediction accuracy.

[0079] Secondly, the present invention also provides an artificial intelligence-based lithium battery health detection system, which adopts the following technical solution:

[0080] An artificial intelligence-based lithium battery health detection system, applied to the aforementioned lithium battery health detection method, includes a feature screening module, a data correction module, a capacity extraction module, a model building module, and an evaluation and optimization module; wherein, the feature screening module is used to screen battery candidate features related to battery charging and discharging characteristics, including sampling time, total current, and battery remaining charge, from the vehicle's basic features based on vehicle source data, and to construct a battery feature data table;

[0081] The data correction module is used to determine the battery charging state based on the battery characteristic data table, extract the charging segment from the battery charging state, and combine it with the battery remaining charge correction source data.

[0082] The capacity extraction module is used to calculate and analyze the battery capacity of each vehicle one by one according to the charge charging time of the charging segment, and obtain the initial capacity value corresponding to several battery capacities.

[0083] The model building module is used to clean and verify all the initial capacity values ​​to obtain effective capacity data. According to the capacity type, the effective capacity data is divided into single-pack capacity dataset and dual-pack capacity dataset, which are then input into the neural network model for training to obtain single-pack health assessment matrix and dual-pack health assessment matrix, and then comprehensively generate battery health assessment model.

[0084] The evaluation and optimization module is used to modify the single-pack health evaluation matrix and the dual-pack health evaluation matrix according to the physical compensation mechanism, complete the optimization of the battery health evaluation model, and complete the battery health detection.

[0085] By adopting the above technical solution, the lithium battery health detection method is deployed on a cloud big data platform as a detection system composed of several functional modules. The detection system extracts battery-related feature information from the vehicle's basic features through the feature screening module, and extracts the corresponding source data, which is then transmitted to the data correction module for charging segment analysis and source data correction based on the battery's remaining charge. Based on the corrected source data, the vehicle's battery capacity is calculated. Then, the model building module builds and initializes the battery health assessment model based on the battery capacity, and the evaluation optimization module corrects the battery health assessment model based on the physical compensation mechanism to complete the lithium battery health detection.

[0086] The present invention is further configured such that: the capacity extraction module includes a capacity calculation submodule, a capacity analysis submodule, and a data extraction submodule; wherein,

[0087] The capacity calculation submodule is used to calculate the vehicle's battery capacity in segments based on the charge charging time of the charging segment; the capacity analysis submodule is used to analyze the battery capacity horizontally based on capacity value clustering and vertically based on total current value to obtain several initial capacity values.

[0088] The data extraction submodule is used to preset travel constraints including the number of driving days and the driving mileage, and to filter and extract effective capacity data from all the initial capacity values ​​through data fitting, cleaning, and robust regression verification.

[0089] By adopting the above technical solution, the battery capacity is calculated in segments based on the charging time of the charging segment through the capacity calculation submodule. The total battery capacity is analyzed horizontally and vertically to obtain several initial capacity values. Based on preset travel constraints including driving days and mileage, the initial capacity values ​​are subjected to data fitting, cleaning, and robust regression verification for secondary screening and extraction to obtain effective capacity data. This reduces the probability of falsely labeled battery remaining charge and also reduces the problem of excessive capacity deviation caused by time sampling differences.

[0090] In summary, the beneficial technical effects of the present invention are as follows:

[0091] 1. By combining real-time vehicle source data with a physical compensation mechanism, the model is corrected and transferred to learn. By imposing constraints, it regresses to a specific range, avoiding upward fluctuations in SOH and improving the accuracy of SOH estimation and the adaptability of the model.

[0092] 2. Based on the preset target number, all candidate battery features are obtained, and a battery feature data table is constructed. This reduces feature data that is irrelevant to the battery and reduces the impact of invalid features on the battery.

[0093] 3. Based on the preset travel constraints, the effective capacity data in all initial capacity values ​​are subjected to fitting residual judgment and robust regression verification to reduce the problem of excessive capacity deviation caused by time-missed sampling differences.

[0094] 4. By training the dataset using a graph neural network training array, the battery health assessment model was initialized, which improved the accuracy of the battery health assessment model. Attached Figure Description

[0095] Figure 1 This is a flowchart illustrating a lithium battery health assessment method according to one embodiment of the present invention.

[0096] Figure 2 This is a flowchart illustrating a lithium battery health assessment method according to one embodiment of the present invention.

[0097] Figure 3 This is a flowchart illustrating a lithium battery health assessment method according to one embodiment of the present invention.

[0098] Figure 4 This is a schematic diagram of the structure of a lithium battery health detection system according to one embodiment of the present invention. Detailed Implementation

[0099] The present invention will be further described in detail below with reference to the accompanying drawings.

[0100] Example 1:

[0101] Reference Figure 1 The present invention discloses an artificial intelligence-based method for detecting the health of lithium batteries, comprising:

[0102] S1: Calculate the information correlation degree between the vehicle's basic characteristics and the battery's charging and discharging characteristics based on the vehicle source data, and determine the battery candidate characteristics, including sampling time, total current and battery remaining charge, from the battery charging and discharging characteristics based on the information correlation degree, and summarize all the battery candidate characteristics into a battery feature data table;

[0103] In this embodiment, the vehicle source data is the real-time vehicle operation data stored in the vehicle processing module, including sensor data (such as speed, steering angle, tire pressure, maximum power, maximum torque, comprehensive driving range, and energy consumption per 100 kilometers), power system status (battery pack, battery management system BMS, motor controller, motor, charging system and controller), energy consumption information (battery charge, rated capacity, energy density), etc., reflecting the overall vehicle operation status.

[0104] Vehicle source data includes basic vehicle characteristics and battery charge / discharge characteristics;

[0105] Basic vehicle characteristics include vehicle status, total voltage, insulation resistance, and other vehicle-specific features unrelated to battery characteristics.

[0106] Battery charge and discharge characteristics include sampling time, total current, minimum temperature, maximum temperature, average voltage, battery remaining charge (SOC), maximum voltage, and minimum voltage.

[0107] S2: Based on the battery feature data table, the battery charging state is taken as the research object, charging segments are extracted and analyzed, the sampling time threshold is determined, and data missing and interpolation filling are determined based on the remaining battery charge.

[0108] S3: Determine the battery charging state based on the battery feature data table and the total current polarity, extract charging segments from the battery charging state, determine the sampling time threshold of the charging segments, and determine the data missing state based on the remaining charge of the battery, and fill in the corresponding missing values.

[0109] S4: Calculate the battery capacity of each vehicle one by one according to the charge charging time of the charging segment, perform directional analysis on all the battery capacities to obtain initial capacity values ​​corresponding to several battery capacities, clean and verify all the initial capacity values ​​to obtain effective capacity data.

[0110] S5: Based on the battery feature data table, correct the single-pack health assessment matrix and the dual-pack health assessment matrix to obtain the corresponding optimal health assessment weight matrix, thereby generating a comprehensive battery health assessment model and completing the battery health detection.

[0111] The implementation principle of this embodiment is as follows: Key features of battery health (such as charging voltage change rate, capacity decay rate, internal resistance change, etc.) are selected from vehicle source data (such as charging / discharging curves, temperature, current, and voltage). The duration characteristics of the charging voltage range and a multiple linear regression model are referenced. Features significantly affecting SOH can be screened out through statistical analysis or machine learning methods (such as LSTM networks), eliminating redundant or noisy data.

[0112] Next, based on the battery's remaining charge (SOC) state, missing values ​​are filled in using time series interpolation methods (such as linear interpolation or LSTM prediction), and normalization is used to time-align charging segments from different vehicles or at different times.

[0113] Then, based on the ampere-hour integration method or the relationship between charging time and current, the capacity value of a single charge can be calculated; the capacity differences of batteries from different vehicles can be analyzed and compared horizontally (e.g., classified by capacity size); the capacity changes of the same battery at different time points can also be tracked vertically to identify aging trends; and then abnormal data can be eliminated to retain the effective capacity value.

[0114] Subsequently, a data-driven approach is adopted, such as a hybrid model of LSTM and multilayer perceptron, or a multiple linear regression model. Then, effective capacity data and feature data are input, and the model parameters are optimized by mean square error. At the same time, the SOH value of the charging segment sequence is predicted, and an evaluation model is constructed. The prediction results are then adjusted in conjunction with an electrochemical model to improve accuracy.

[0115] Finally, the single-packet health assessment matrix or the dual-packet health assessment matrix is ​​corrected through iterative algorithms (such as gradient descent); and weights are assigned according to the importance of features (such as the greater impact of capacity decay on SOH) to form the optimal weight matrix; at the same time, battery aging mechanisms (such as the effect of temperature on capacity) are introduced to correct model prediction bias and ensure that the results conform to actual physical laws.

[0116] Example 2:

[0117] The specific steps of step S1 include:

[0118] The vehicle status is monitored, and vehicle source data is extracted to obtain several sets of basic vehicle features and battery charging / discharging features. Based on the vehicle source data, the information correlation between each set of basic vehicle features and battery charging / discharging features is calculated to obtain a feature validity evaluation value, which is then compared with a preset first feature correlation threshold.

[0119] If the effective evaluation value of the feature is greater than or equal to the first feature-related threshold, then the current vehicle basic feature is used as a battery candidate feature;

[0120] If the effective evaluation value of the feature is less than the first feature-related threshold, then the current vehicle basic features are removed.

[0121] Arrange all the valid evaluation values ​​of the battery candidate features in descending order, and extract the source data corresponding to the battery candidate features according to the preset target number;

[0122] The extracted battery candidate features and their corresponding source data are summarized to generate a battery feature data table.

[0123] If the number of candidates for the battery candidate features is less than the target number, then the source data corresponding to all battery candidate features that currently exceed the first feature-related threshold are used to construct an initial battery-related table.

[0124] Simultaneously, the valid evaluation values ​​of all vehicle basic features that do not exceed the first feature-related threshold are compared with a preset second feature-related threshold for judgment:

[0125] If the effective evaluation value of the feature is less than the second feature-related threshold, then the current vehicle basic features will be removed.

[0126] If the effective evaluation value of the feature is greater than or equal to the second feature-related threshold, then according to the difference between the preset target number and the candidate number, the same number of source data corresponding to the vehicle basic features are selected and filled into the initial battery-related table to complete the construction of the battery feature data table.

[0127] In this embodiment, the vehicle's central processing unit collects real-time vehicle operating status data detected by onboard sensors. Based on the correlation between the vehicle source data and battery charging / discharging characteristics, it filters all relevant vehicle features, extracts all features related to battery charging / discharging characteristics, and extracts the data corresponding to the current feature to construct a battery feature data table specific to the vehicle's battery status. The data table includes information such as sampling time, total current, minimum temperature, maximum temperature, average voltage, SOC, maximum voltage, and minimum voltage. The first feature correlation threshold can be 0.95. In this embodiment, considering the problem of incomplete extraction of battery-related feature information due to interruptions or missing values ​​during vehicle source data acquisition, a second feature correlation threshold is set to further extract battery-related features.

[0128] If the number of candidate battery features is less than the target number, then the source data corresponding to all current battery candidate features that exceed the first feature correlation threshold are used to construct an initial battery correlation table.

[0129] Simultaneously, the effective evaluation values ​​of all vehicle basic features that do not exceed the first feature correlation threshold are compared with the preset second feature correlation threshold for judgment:

[0130] If the effective evaluation value of the feature is less than the second feature-related threshold, then the current vehicle basic features will be removed.

[0131] If the effective evaluation value of the feature is greater than or equal to the second feature-related threshold, then based on the difference between the preset target number and the candidate number, the same number of source data corresponding to the vehicle basic features are selected to fill the initial battery-related table, thus completing the construction of the battery feature data table.

[0132] In this embodiment, the second feature correlation threshold can be 0.8.

[0133] The implementation principle of this embodiment is as follows: raw data (such as voltage, current, temperature, SOC, charging time, etc.) is extracted from the vehicle data processing module, and the correlation strength between each feature and battery health is calculated using the Pearson correlation coefficient or mutual information method to generate effective feature evaluation values. For example, the correlation coefficient between the duration of the inflection point of the charging voltage curve and capacity decay can reach 0.85, while ambient temperature fluctuations may only show a weak correlation (less than 0.3).

[0134] The effective evaluation values ​​of features are compared with a preset first threshold (usually set to 0.9 to 0.99). Features exceeding the threshold are added to the candidate set and sorted in descending order of evaluation value. When the number of candidates is less than the preset number of influencing elements, a secondary screening is initiated: features that do not reach the first threshold but exceed the second threshold (usually 0.8 to 0.89) are re-sorted, and features with the largest difference are added to the initial table. For example, if the preset requirement is 8 features but there are only 6 candidates, the 2 highest evaluation value items are added from the secondary features.

[0135] The algorithm prioritizes retaining strongly correlated features (such as capacity decay rate and internal resistance change), followed by indirectly correlated features (such as charging cycle frequency). A dynamic thresholding mechanism is used to avoid missing important features while preventing inefficient features from interfering with model training. The final output is a structured battery feature data table, providing a high signal-to-noise ratio input for subsequent health modeling.

[0136] Example 3:

[0137] Reference Figure 2 The specific steps in step S2 include:

[0138] Determine the polarity of the total current in the battery characteristic data table:

[0139] If the polarity of the total current is irregular and alternates between positive and negative over a period of time, then the current battery is determined to be in a discharge state, and all discharge segments in the discharge state are eliminated.

[0140] If the polarity of the total current is concentrated and remains negative for a prolonged period, then the current battery is determined to be in a charging state, and all charging segments during the battery charging state are extracted.

[0141] Cluster the sampling times of all the charging segments to obtain the proportion of each sampling time, and take the sampling time with the largest proportion as the sampling time threshold;

[0142] If the current sampling time is less than or equal to the sampling time threshold, it is determined that there is no missing source data, and the current sampling time is corrected to the sampling time threshold.

[0143] If the current sampling time is greater than the sampling time threshold, it is determined that the current source data is missing, and the remaining battery charge is assessed.

[0144] If the remaining charge of the battery increases, the vehicle is determined to be in a charging state during the current sampling time, and the source data during the current sampling time is interpolated and filled using a preset interpolation strategy.

[0145] If the remaining charge of the battery remains unchanged and the sampling time is less than a preset time span threshold, the vehicle is determined to be in a non-charging state during the current sampling time. At the same time, the source data during the current sampling time is interpolated and filled using an interpolation strategy.

[0146] If the remaining charge of the battery remains unchanged, but the sampling time is greater than or equal to a preset time span threshold, the vehicle is determined to be in a non-charging state during the current sampling time, and no interpolation is performed on the source data during the current sampling time.

[0147] The implementation principle of this embodiment is as follows: By analyzing the battery characteristic data table, it can be seen that compared with the irregular and complex discharge state, the charging state is more concentrated and continuous within a specific time period. This is a phenomenon where a large number of users charge within the same time period. Therefore, the charging state is selected as the research object. A continuous negative current is extracted as a charging segment, and most charging time is 3 hours, with 95% of the charging time being less than or equal to 6 hours. Furthermore, due to the different data sampling times, for data less than or equal to 10 seconds, there are 1, 3, and 10 seconds (10 seconds accounts for more than 95%), which are uniformly sampled to 10 seconds; for data greater than 10 seconds, the sampling time ranges from 100 seconds to 1000 seconds. This indicates that there may be missing data or no charging may have occurred: If the SOC changes, and if the SOC range is large and increases (indicating that charging has definitely occurred), regardless of the duration, direct interpolation is performed. It is also possible that the vehicle is charging, but the data has not been collected. If the SOC does not change and the time span is less than five minutes (statistically, the average charging time for one SOC is less than five minutes), interpolation is performed.

[0148] Example 4:

[0149] The specific steps in step S3 include:

[0150] Feature extraction is performed on all the charging segments to obtain the charging start range, charging end range, and charge increase interval; based on the charging start range and the charging end range, the first segment charging capacity within the preset charge increase interval is calculated;

[0151] The battery capacity of the first segment of the first N charging increase intervals is calculated based on the first segment charging capacity, and the charging capacity of the second segment of the remaining charging increase intervals is calculated separately.

[0152] Calculate the battery capacity of all vehicles based on the first segment charging capacity and the second segment charging capacity.

[0153] A directional analysis is performed on all the battery capacities to obtain initial capacity values ​​corresponding to several battery capacities; the directional analysis includes lateral clustering analysis based on capacity values ​​and longitudinal clustering analysis based on current values.

[0154] The implementation principle of this embodiment is as follows: Analysis of massive charging segments shows that the initial SOC charging value is concentrated between 40% and 70%, and the ending value is mostly above 90%, with the SOC charging increase range concentrated between 20% and 40%. A segmented improved ampere-hour integral calculation method is used to calculate the capacity (the reliability of the battery's state of charge (SOC) can be ensured through an open-circuit voltage (OCV) correction strategy, but since open-circuit voltage data is currently unavailable, a segmented capacity calculation method is used to address the issue of falsely labeled SOC).

[0155] Inflated SOC (State of Charge) ratings can lead to prolonged charging even with an SOC of 99%. Therefore, it's advisable to calculate in segments: use the SOC 75%-98% range to represent the first 99 segments, and calculate the last segment separately using an SOC of 99%.

[0156]

[0157] Among them, C i This indicates the battery capacity, and 'i' represents the sequence number of the charge increase interval. This indicates the start time when SOC reaches 75%. This indicates the end time when the SOC reaches 98%, and I represents the total current. and These represent the start and end times when SOC reaches 99%, respectively.

[0158] After horizontal and vertical capacity analysis, the capacity was divided into two levels: 275Ah and 135Ah. Due to inconsistent data sampling, as well as reasons such as missed sampling and charging interruptions, a small portion of the data capacity calculation may be biased when the difference is uniformly sampled. Therefore, abnormal data cleaning is required.

[0159] Example 5:

[0160] Reference Figure 3 The specific steps in step S3 also include:

[0161] Based on the travel constraint condition that the number of driving days is positively correlated with the driving mileage, all the initial capacity values ​​are linearly fitted and cleaned. The fitting residual value of each initial capacity value is calculated and compared with the preset outlier threshold.

[0162] In this embodiment, the travel constraints are that the number of days of travel does not exceed 30 days and the mileage does not exceed 20,000 km.

[0163] If the fitting residual value is greater than or equal to the outlier threshold, the current battery capacity is determined to be an outlier, and all outliers are removed.

[0164] If the fitting residual value is less than the outlier threshold, the current battery capacity is determined to be a non-outlier, and all the non-outliers are retained to generate a non-outlier set.

[0165] Robust regression verification is performed on the set of non-outliers, and the corrected residual value of each non-outlier is calculated and compared with a preset correction threshold.

[0166] If the corrected residual value is greater than or equal to the correction threshold, then the current non-outlier point is removed.

[0167] If the corrected residual value is less than the corrected threshold, then the direction of change of the battery capacity of the current non-outlier point with driving mileage is verified:

[0168] If the battery capacity does not decrease linearly with the driving mileage, then the current non-outlier point will be removed.

[0169] If the battery capacity decreases linearly with the mileage, the current outlier will be retained in the set of outliers and thus determined as valid capacity data.

[0170] The implementation principle of this embodiment is as follows: After data cleaning, it can be determined that the travel constraints for most vehicles are no more than 30 days of driving and no more than 20,000 km of mileage. Therefore, in order to balance the validity of data and capacity decay, data from 658 vehicles with more than 30 days of driving are extracted as valid data. At the same time, a linear regression model is constructed based on the cumulative mileage and capacity data to detect outliers in the valid data. The outlier threshold is set to 0.6 times the standard deviation of the residuals, and non-outliers that do not exceed the outlier threshold are retained. To further reduce the problem of overestimation of capacity due to time-related sampling differences, robust regression correction is used to extract non-outliers where the battery capacity of each vehicle decreases linearly with mileage. The current non-outliers are retained as valid capacity data through trend direction verification and dynamic thresholding.

[0171] Example 6:

[0172] The specific steps in step S4 include:

[0173] The effective capacity data is divided into dual-packet capacity datasets and single-packet capacity datasets based on capacity type.

[0174] The dual-packet capacity dataset and the single-packet capacity dataset are divided according to data type and time series, respectively, to obtain the historical array input information and historical metadata output information corresponding to the dual-packet capacity dataset and the historical array input information and historical metadata output information corresponding to the single-packet capacity dataset.

[0175] The historical metadata output information is sorted according to the vehicle number and charging segment, and corresponding charging segment sequences are generated that are positively correlated with driving time and driving mileage.

[0176] The historical array input information is normalized according to the preset extreme value label to obtain the corresponding array training dataset; the array training dataset is input into the graph neural network for several iterations of training, and the historical metadata output information is used as the output feature of the graph neural network to obtain the single-packet health evaluation matrix and the double-packet health evaluation matrix, and then the theoretical charging segment sequence is output respectively.

[0177] The implementation principle of this embodiment is as follows: Based on two capacity types, 275Ah and 135Ah, the effective capacity data is divided into dual-packet capacity datasets and single-packet capacity datasets. A window of size 128 with a step size of 128 is used to divide the charging segments, with insufficient portions filled using nearest-neighbor padding. Total current, average voltage, temperature, and SOC are taken as historical array input information (ArrayData information), while mileage, capacity, and charging segment information are taken as historical metadata output information (Metadata information).

[0178] The metadata information is sorted according to vehicle number and charging segment, and finally a continuous sequence of charging segments sorted by time and mileage is formed (due to the screening, removal of outliers and correction of SOC charging intervals, the interval between time and mileage is not the same, which is also one of the reasons affecting the accuracy of the model).

[0179] The source data is normalized using a minimax normalization method; the extreme value labels are normalized based on a custom capacity range with a minimum value of 0 and a maximum value of 135. However, this causes the RMSE, MAE and other indicator values ​​to depend on the order of magnitude of the labels, resulting in no comparability between different orders of magnitude.

[0180] Example 7:

[0181] The specific steps in step S4 also include:

[0182] The feature extraction layer fills the data input matrix constructed from the array training dataset with zero vectors according to the convolution kernel dimension, and uses several stacked convolution blocks of dimension L to extract convolution features from the array training dataset, thereby obtaining the battery feature matrix.

[0183] The gate layer uses matrix multiplication to filter the battery feature matrices of different stacked convolutional blocks to obtain the battery weight matrix;

[0184] The linear layer performs a linear transformation on the battery weight matrix based on the weight vector matrix and bias matrix, combined with the pooling function and loss function, to obtain a single-packet health assessment matrix or a dual-packet health assessment matrix.

[0185] The fully connected layer maps the single-packet health assessment matrix or the double-packet health assessment matrix into the corresponding one-dimensional feature tensor, and matches it with the historical metadata output information to obtain the corresponding theoretical charging segment sequence.

[0186] The implementation principle of this embodiment is as follows: A graph neural network is composed of a large number of interconnected neurons. A neuron consists of a basic input X, parameter weights W, bias b, activation function, and output Y, which can be represented as Y = XW + b. The operation of performing an inner product on different data windows and filter matrices (a set of fixed weight matrices, i.e., filters (neurons with a set of fixed weights; multiple filters stacked together form a convolution kernel)) is the convolution operation. By sampling and pooling, the feature map data is fused using the average or maximum value, further reducing the feature map dimension and extracting effective information. The dimension of the feature map can also be controlled by setting the number of convolution kernels and the convolution stride.

[0187] The input sequence is convolved with six stacked L-dimensional convolutional kernels to extract features. The input matrix is ​​padded with (L-1) / 2 zero vectors to match the dimension of the convolutional kernels, resulting in an expanded input matrix. ReLU and tanh are used as activation functions for the stacked convolutional blocks, respectively. A gating layer filters out redundant feature information, simplifying the model's complexity. A linear layer linearly transforms its input using parameters W and bias b, producing a max-pooled output matrix. A fully connected layer then maps the max-pooled output matrix into a one-dimensional feature tensor, representing the predicted regression result.

[0188] The overall model training loss was stable across all data sets, but some fluctuations were observed (indicating the presence of a few outliers). Training the model separately for single and double-label data significantly outperformed training with all data together. (The mixed single and double-label data, with two labels, is prone to overfitting). For single and double-label data: linear regression and robust regression yielded similar results, with the latter having approximately one-third the data volume of the former, indicating that trend correction can compensate for insufficient data. However, for the mixed data, robust regression significantly outperformed linear regression, demonstrating that data normalization can reduce model overfitting.

[0189] Example 8:

[0190] The specific steps in step S5 include:

[0191] The real-time array input information from the battery feature data table is loaded into the corresponding single-pack health assessment matrix and the dual-pack health assessment matrix for prediction, resulting in the corresponding theoretical charging segment sequence. This sequence is then matched with the corresponding real-time charging segment sequence to obtain the sequence similarity, which is then compared with a preset similarity threshold.

[0192] If the sequence similarity is greater than or equal to the similarity threshold, the learning rate of the current single-packet health assessment matrix or double-packet health assessment matrix remains unchanged.

[0193] If the sequence similarity is less than the similarity threshold, the current battery health assessment model is determined to be overfitted. The learning rate of the current single-pack health assessment matrix and / or the dual-pack health assessment matrix is ​​dynamically corrected and compared with the preset warning learning threshold.

[0194] If the corrected learning rate is greater than the warning learning threshold, the current battery health assessment model is deemed to have redundant structure. Random pruning and / or adding an attention mechanism are then performed on the feature extraction layer. At the same time, the convolution kernel dimension of the feature extraction layer is expanded and / or the loss function of the linear layer is replaced, thereby obtaining the corresponding optimal health assessment weight matrix.

[0195] The implementation principle of this embodiment is as follows: for the same batch of data, the higher the learning rate, the easier it is for the model to overfit, which is very likely caused by the complex model structure (2 convolutional layers, 6 parallel connections in each layer). The number of parallel connections in each layer will be reduced in the future.

[0196] The performance was low when training on all data, possibly due to the dataset having multiple label values, resulting in insufficient generalization ability of the model to extract multi-class features. Future attempts will attempt to add a one-dimensional convolutional attention mechanism: ECA has advantages such as high efficiency, lightweight design, and strong representational ability. Furthermore, during training, it was found that the larger the output feature dimension of the convolution kernel in the hyperparameters, the better the model's performance; the subsequent value range will be adjusted upwards from [3, 15]. This regression task uses the mean squared error (MSE) loss function, which is sensitive to outliers. A smoothed L2 loss function will be used later to reduce the impact of outliers. Early stopping and dynamic learning rate adjustment can also be added to help the model converge faster during training, while avoiding getting trapped in local optima and saving training time. The learning rate will be reduced when the performance on the validation set no longer improves; the learning rate will be dynamically adjusted when the performance on the validation set stagnates.

[0197] Example 9:

[0198] The specific steps in step S5 also include:

[0199] The prediction accuracy of the optimal health assessment weight matrix is ​​calculated and compared with the preset assessment accuracy threshold. If the prediction accuracy is greater than or equal to the assessment accuracy threshold, it indicates that the training performance of the current battery health assessment model meets the standard and the sampling time of the array training dataset is unified.

[0200] If the prediction accuracy is less than the evaluation accuracy threshold, it indicates that the training performance of the current battery health assessment model is not up to standard. The interpolation strategy of the array training dataset is optimized and / or a temperature correction coefficient is added to the interpolation strategy according to the temperature-capacity distribution relationship, thereby completing the detection of lithium battery health.

[0201] The implementation principle of this embodiment is as follows: the training performance of this model largely depends on the dataset, and time is crucial for calculating the capacity label. Therefore, the time sampling should be as uniform as possible when collecting the dataset. Furthermore, because some data samples have long sampling times and contain blind-box regions, the current sampling strategy may result in an excessively large capacity label, requiring optimization.

[0202] Correction factors can also be added to the calculation based on the temperature-capacity distribution relationship.

[0203] Double-packet calibration: Single-package calibration: Calculation of battery capacity after correction: in, The temperature compensation capacity of the interval where the i-th charge is charged increases, T i This represents the real-time temperature over the interval where the charge increases by i charges. This indicates the reference capacity during the charge charging increase range at 25°C.

[0204] Example 10:

[0205] Reference Figure 4 An artificial intelligence-based lithium battery health detection system, applied to the aforementioned lithium battery health detection method, includes a feature selection module, a data correction module, a capacity extraction module, a model construction module, and an evaluation and optimization module; wherein,

[0206] The feature filtering module is used to filter candidate battery features related to battery charging and discharging characteristics, including sampling time, total current and battery remaining charge, from the vehicle source data based on the vehicle's basic features, and to build a battery feature data table.

[0207] The data correction module is used to determine the battery charging state based on the battery characteristic data table, extract the charging segment from the battery charging state, and combine it with the battery remaining charge correction source data.

[0208] The capacity extraction module is used to calculate and analyze the battery capacity of each vehicle one by one according to the charge charging time of the charging segment, and obtain the initial capacity value corresponding to several battery capacities.

[0209] The model building module is used to clean and verify all the initial capacity values ​​to obtain effective capacity data. According to the capacity type, the effective capacity data is divided into single-pack capacity dataset and dual-pack capacity dataset, which are then input into the neural network model for training to obtain single-pack health assessment matrix and dual-pack health assessment matrix, and then comprehensively generate battery health assessment model.

[0210] The evaluation and optimization module is used to modify the single-pack health evaluation matrix and the dual-pack health evaluation matrix according to the physical compensation mechanism, complete the optimization of the battery health evaluation model, and complete the battery health detection.

[0211] The capacity extraction module includes a capacity calculation submodule, a capacity analysis submodule, and a data extraction submodule; wherein, the capacity calculation submodule is used to calculate the vehicle's battery capacity in segments based on the charge charging time of the charging segment; the capacity analysis submodule is used to analyze the battery capacity horizontally based on capacity value clustering and vertically based on total current value to obtain several initial capacity values;

[0212] The data extraction submodule is used to preset travel constraints including the number of driving days and the driving mileage, and to filter and extract effective capacity data from all the initial capacity values ​​through data fitting, cleaning, and robust regression verification.

[0213] The implementation principle of this embodiment is as follows: the feature selection module selects key features of battery health (such as charging voltage change rate, capacity decay rate, internal resistance change, etc.) from the vehicle source data, and filters out features that have a significant impact on SOH through statistical analysis or machine learning methods (such as LSTM network) to eliminate redundant or noisy data.

[0214] The data correction module fills in missing values ​​based on the battery's remaining charge (SOC) state using linear interpolation or LSTM prediction, and performs time alignment for charging segments from different vehicles or at different times.

[0215] The capacity calculation submodule of the capacity extraction module can calculate the capacity value of a single charge based on the ampere-hour integration method or the relationship between charging time and current; and through the capacity analysis submodule, it can perform horizontal analysis and comparison of the capacity differences of batteries from different vehicles (such as classifying them by capacity size); at the same time, it can track the capacity changes of the same battery at different time points and vertically analyze and identify aging trends; the data extraction submodule removes abnormal data and retains valid capacity values ​​according to preset conditions (such as charging completion and ambient temperature range).

[0216] Subsequently, the model building module employs a hybrid model of multilayer perceptron (MLP) or a multiple linear regression model; it inputs effective capacity data and feature data, optimizes model parameters through a cost function (such as mean square error), and constructs an evaluation model by predicting the SOH value of the charging segment sequence; and adjusts the prediction results by combining the battery aging rate formula to improve accuracy.

[0217] Finally, the evaluation and optimization module corrects the single-pack health assessment matrix or the dual-pack health assessment matrix using the gradient descent method; and assigns weights according to the impact of capacity decay on SOH to form the corresponding optimal weight matrix; at the same time, it introduces a temperature correction coefficient for battery capacity to correct model prediction bias and ensure that the results conform to actual physical laws.

[0218] The embodiments described herein are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for detecting the health of lithium batteries based on artificial intelligence, characterized in that, include: The information correlation between vehicle basic characteristics and battery charge and discharge characteristics is calculated based on vehicle source data. Based on the information correlation, battery candidate characteristics including sampling time, total current and battery remaining charge are determined from the battery charge and discharge characteristics. All battery candidate characteristics are summarized into a battery characteristic data table. The battery charging state is determined based on the battery characteristic data table and the total current polarity, and a charging segment is extracted from the battery charging state. At the same time, the sampling time threshold of the charging segment is determined, and the data missing state is judged according to the battery remaining charge, and the corresponding missing value is filled in. Based on the charging time of the charging segment, the battery capacity of each vehicle is calculated one by one. A targeted analysis is performed on all the battery capacities to obtain initial capacity values ​​corresponding to several battery capacities. All the initial capacity values ​​are then cleaned and verified to obtain valid capacity data, including: Based on the travel constraint condition that the number of driving days is positively correlated with the driving mileage, all the initial capacity values ​​are linearly fitted and cleaned. The fitting residual value of each initial capacity value is calculated and compared with the preset outlier threshold. If the fitting residual value is less than the outlier threshold, the current battery capacity is determined to be a non-outlier, and all the non-outliers are retained to generate a non-outlier set. Robust regression verification is performed on the set of non-outliers, and the corrected residual value of each non-outlier is calculated and compared with a preset correction threshold. If the corrected residual value is less than the corrected threshold, then the direction of change of the battery capacity of the current non-outlier point with driving mileage is verified: If the battery capacity decreases linearly with the driving mileage, the current non-outlier point is retained in the set of non-outlier points and thus determined as valid capacity data. The effective capacity data is divided into single-packet capacity datasets and dual-packet capacity datasets according to the capacity type. These datasets are then input into a neural network model for training to obtain single-packet health assessment matrices and dual-packet health assessment matrices, and the corresponding theoretical charging segment sequences are output. Based on the battery feature data table, the single-pack health assessment matrix and the dual-pack health assessment matrix are corrected to obtain the corresponding optimal health assessment weight matrix, thereby comprehensively generating a battery health assessment model and completing the battery health detection, including: The real-time array input information from the battery feature data table is loaded into the corresponding single-pack health assessment matrix and the dual-pack health assessment matrix for prediction, resulting in the corresponding theoretical charging segment sequence. This sequence is then matched with the corresponding real-time charging segment sequence to obtain the sequence similarity, which is then compared with a preset similarity threshold. If the sequence similarity is less than the similarity threshold, the current battery health assessment model is determined to be overfitted. The learning rate of the current single-pack health assessment matrix and / or the dual-pack health assessment matrix is ​​dynamically corrected and compared with the preset warning learning threshold. If the corrected learning rate is greater than the warning learning threshold, the current battery health assessment model is deemed to be structurally redundant. Random pruning and / or adding an attention mechanism are then performed on the feature extraction layer. At the same time, the convolution kernel dimension of the feature extraction layer is expanded and / or the loss function of the linear layer is replaced, thereby obtaining the corresponding optimal health assessment weight matrix. The prediction accuracy of the optimal health assessment weight matrix is ​​calculated and compared with a preset assessment accuracy threshold: If the prediction accuracy is less than the evaluation accuracy threshold, it indicates that the training performance of the current battery health assessment model is not up to standard. The interpolation strategy of the array training dataset is optimized and / or a temperature correction coefficient is added to the interpolation strategy according to the temperature-capacity distribution relationship, thereby completing the detection of lithium battery health.

2. The method for detecting the health of lithium batteries based on artificial intelligence according to claim 1, characterized in that, The specific steps of calculating the information correlation between vehicle basic characteristics and battery charge / discharge characteristics based on vehicle source data, determining battery candidate characteristics including sampling time, total current, and remaining battery charge from the battery charge / discharge characteristics based on the information correlation, and summarizing all the battery candidate characteristics into a battery characteristic data table include: The vehicle status is monitored, and vehicle source data is extracted to obtain several sets of basic vehicle characteristics and battery charging and discharging characteristics. Based on the vehicle source data, the information correlation degree between each group of vehicle basic features and battery charging and discharging features is calculated to obtain a feature validity evaluation value, which is then compared with a preset first feature correlation threshold for judgment. If the effective evaluation value of the feature is greater than or equal to the first feature-related threshold, then the current vehicle basic feature is used as a battery candidate feature; Arrange all the valid evaluation values ​​of the battery candidate features in descending order, and extract the source data corresponding to the battery candidate features according to the preset target number; The extracted battery candidate features and their corresponding source data are summarized to generate a battery feature data table.

3. The method for detecting the health of lithium batteries based on artificial intelligence according to claim 1, characterized in that, The specific steps of determining the battery charging state based on the battery characteristic data table and total current polarity, extracting charging segments from the battery charging state, determining the sampling time threshold of the charging segment, and judging the missing data state based on the remaining battery charge and filling in the corresponding missing values ​​include: Determine the polarity of the total current in the battery characteristic data table: If the polarity of the total current is concentrated and remains negative for a prolonged period, then the current battery is determined to be in a charging state, and all charging segments during the battery charging state are extracted. Cluster the sampling times of all the charging segments to obtain the proportion of each sampling time, and take the sampling time with the largest proportion as the sampling time threshold; If the current sampling time is greater than the sampling time threshold, it is determined that the current source data is missing, and the remaining battery charge within the current sampling time is determined: If the remaining charge of the battery increases, the vehicle is determined to be in a charging state during the current sampling time, and the source data during the current sampling time is interpolated and filled using a preset interpolation strategy. If the remaining charge of the battery remains unchanged and the sampling time is less than a preset time span threshold, the vehicle is determined to be in a non-charging state during the current sampling time. At the same time, the source data during the current sampling time is interpolated and filled using a preset interpolation strategy.

4. The method for detecting the health of lithium batteries based on artificial intelligence according to claim 1, characterized in that, The specific steps of calculating the battery capacity of each vehicle based on the charge charging time of the charging segment, performing targeted analysis on all battery capacities, and obtaining initial capacity values ​​corresponding to several battery capacities include: Feature extraction is performed on all the charging segments to obtain the charging start range, charging end range, and charge increase interval; Based on the charging start range and the charging end range, calculate the first segment of charging capacity within the preset charging increase interval; The battery capacity of the first segment of the first N charging increase intervals is calculated based on the first segment charging capacity, and the charging capacity of the second segment of the remaining charging increase intervals is calculated separately. Calculate the battery capacity of all vehicles based on the first segment charging capacity and the second segment charging capacity. A directional analysis is performed on all the battery capacities to obtain initial capacity values ​​corresponding to several battery capacities; the directional analysis includes lateral clustering analysis based on capacity values ​​and longitudinal clustering analysis based on current values.

5. The method for detecting the health of lithium batteries based on artificial intelligence according to claim 1, characterized in that, The specific steps of dividing the effective capacity data into single-packet capacity datasets and dual-packet capacity datasets according to capacity type, inputting them into a neural network model for training to obtain single-packet health assessment matrices and dual-packet health assessment matrices, and outputting the corresponding theoretical charging segment sequences include: The effective capacity data is divided into dual-packet capacity datasets and single-packet capacity datasets based on capacity type. The dual-packet capacity dataset and the single-packet capacity dataset are divided according to data type and time series, respectively, to obtain the historical array input information and historical metadata output information corresponding to the dual-packet capacity dataset and the historical array input information and historical metadata output information corresponding to the single-packet capacity dataset. The historical metadata output information is sorted according to the vehicle number and charging segment, and corresponding charging segment sequences are generated that are positively correlated with driving time and driving mileage. The historical array input information is normalized according to the preset extreme value labels to obtain the corresponding array training dataset; The array training dataset is input into the graph neural network for several iterations of training. At the same time, the historical metadata output information is used as the output feature of the graph neural network to obtain the single-packet health evaluation matrix and the double-packet health evaluation matrix, and then output the theoretical charging segment sequence respectively.

6. The method for detecting the health of lithium batteries based on artificial intelligence according to claim 5, characterized in that, The steps for constructing the battery health assessment model include: The feature extraction layer fills the data input matrix constructed from the array training dataset with zero vectors according to the convolution kernel dimension, and uses several stacked convolution blocks of dimension L to extract convolution features from the array training dataset, thereby obtaining the battery feature matrix. The gate layer uses matrix multiplication to filter the battery feature matrices of different stacked convolutional blocks to obtain the battery weight matrix; The linear layer performs a linear transformation on the battery weight matrix based on the weight vector matrix and bias matrix, combined with the pooling function and loss function, to obtain a single-packet health assessment matrix or a dual-packet health assessment matrix. The fully connected layer maps the single-packet health assessment matrix or the double-packet health assessment matrix into a one-dimensional feature tensor, and matches it with the historical metadata output information to obtain the corresponding theoretical charging segment sequence.

7. An artificial intelligence-based lithium battery health detection system, applied to the lithium battery health detection method according to any one of claims 1 to 6, characterized in that: in, The feature filtering module is used to filter candidate battery features related to battery charging and discharging characteristics, including sampling time, total current and battery remaining charge, from the vehicle source data based on the vehicle's basic features, and to build a battery feature data table. The data correction module is used to determine the battery charging state based on the battery characteristic data table, extract the charging segment from the battery charging state, and combine it with the battery remaining charge correction source data. The capacity extraction module is used to calculate and analyze the battery capacity of each vehicle one by one according to the charge charging time of the charging segment, and obtain the initial capacity value corresponding to several battery capacities. The model building module is used to clean and verify all the initial capacity values ​​to obtain effective capacity data. According to the capacity type, the effective capacity data is divided into single-pack capacity dataset and dual-pack capacity dataset, which are then input into the neural network model for training to obtain single-pack health assessment matrix and dual-pack health assessment matrix, and then comprehensively generate battery health assessment model. The evaluation and optimization module is used to modify the single-pack health evaluation matrix and the dual-pack health evaluation matrix according to the physical compensation mechanism, complete the optimization of the battery health evaluation model, and complete the battery health detection.

8. The artificial intelligence-based lithium battery health detection system according to claim 7, characterized in that, The capacity extraction module includes a capacity calculation submodule, a capacity analysis submodule, and a data extraction submodule; wherein... The capacity calculation submodule is used to calculate the vehicle's battery capacity in segments based on the charge charging time of the charging segment. The capacity analysis submodule is used to analyze the battery capacity based on capacity value clustering horizontally and based on total current value vertically to obtain several initial capacity values. The data extraction submodule is used to preset travel constraints including the number of driving days and the driving mileage, and to filter and extract effective capacity data from all the initial capacity values ​​through data fitting, cleaning, and robust regression verification.

Images