Lithium battery health state prediction method and device based on dynamic coupling mechanism

Abstract

The invention provides a lithium battery health state prediction method and device based on a dynamic coupling mechanism, and the method comprises the steps: collecting and preprocessing the multi-dimensional operation data of a lithium battery through a BMS, extracting multi-scale features, constructing a weight adaptive enhancement prediction model, and applying time sequence attention weighting to the multi-scale features, a feature calibration branch for coupling small sample learning and feature-level data enhancement and a parameter updating branch for online incremental learning are embedded, data sufficiency, prediction errors and distribution change indexes are calculated in real time, and double-branch weights are adaptively adjusted through a dynamic weight fusion layer. And carrying out weighted fusion to obtain an initial SOH predicted value. And through built-in physical constraint of the model and optimized adaptive noise robust Kalman filtering error correction, outputting a final SOH predicted value, and combining a threshold value to realize health assessment and early warning.

Description

Technical Field

[0001] This invention relates to the field of lithium battery health management technology, and in particular to a method and apparatus for predicting the health status of lithium batteries based on a dynamic coupling mechanism. Background Technology

[0002] In the rapid development of the new energy industry, lithium batteries, as core energy storage components, play an irreplaceable and crucial role. Lithium batteries are widely used in new energy vehicles, energy storage power stations, and portable electronic devices, and their state of health directly determines the operational stability, range, and safety of these devices. Traditional methods for predicting the State of Health (SOH) of lithium batteries largely rely on electrochemical mechanism modeling or single data-driven models. The former requires complex internal reaction kinetic analysis, resulting in difficulties in parameter identification and poor adaptability to complex operating conditions; the latter relies on a large amount of full-lifecycle aging data, leading to high computational costs, insufficient generalization ability, and susceptibility to increased prediction errors due to data noise. These problems not only reduce the accuracy and real-time performance of SOH prediction but also limit its engineering application in small-sample, dynamic operating conditions. In early technologies, lithium battery SOH monitoring relied heavily on offline detection or simple threshold judgment, exhibiting technical shortcomings such as delayed warnings and the inability to predict degradation trends in advance. In addition, traditional early warning models for safety hazards such as thermal runaway and sudden capacity drop in lithium batteries mostly focus on monitoring single thresholds such as temperature and voltage, lacking comprehensive consideration of multi-dimensional influencing factors such as charge / discharge rate, cycle number, and ambient humidity. This makes it difficult to fully capture the nonlinear and time-varying characteristics of the battery aging process, thus limiting the practicality and reliability of the early warning system and failing to provide full-cycle protection for the safe operation of equipment.

[0003] With the deep integration of big data and artificial intelligence technologies, modern lithium battery health status prediction methods have made breakthrough progress. Current research and practice are gradually shifting towards prediction schemes driven by multi-source data fusion. This involves deploying multiple types of sensors in the battery management system (BMS) to collect multi-dimensional operational data such as voltage, current, temperature, and cycle count in real time, and combining this with advanced algorithms such as deep learning and machine learning for feature mining and model building. These technological innovations have significantly improved the accuracy and adaptability of lithium battery SOH prediction, providing stronger technical support for the safe operation and maintenance of new energy equipment. However, existing methods still have significant shortcomings: insufficient generalization ability in small sample scenarios, with prediction accuracy dropping sharply in new battery models or niche application scenarios due to data scarcity; limited dynamic adaptability, making it difficult to adapt to complex operating conditions such as fluctuations in charge / discharge rates and sudden changes in ambient temperature in real time; low efficiency of hyperparameter optimization, long model training cycles, and susceptibility to getting trapped in local optima; and the lack of a collaborative mechanism between data augmentation and small sample learning / incremental learning, resulting in insufficient stability of prediction accuracy throughout the entire battery lifecycle. Summary of the Invention

[0004] Purpose of the Invention: This invention addresses the problems of poor generalization with small samples, weak dynamic adaptability, insufficient prediction rationality, high deployment difficulty, and lack of collaborative mechanisms in existing lithium battery SOH prediction technologies. It provides a lithium battery health state prediction method and device based on a dynamic coupling mechanism. Through an integrated design of "small sample + data enhancement coupling, incremental learning online update, dynamic weight adjustment, and physical constraint correction," it achieves accuracy, rationality, real-time performance, and embedded deployment adaptability in lithium battery SOH prediction throughout its entire life cycle, providing technical assurance for the safe and stable operation of lithium batteries.

[0005] Technical solution: The present invention provides a method for predicting the health status of lithium batteries based on a dynamic coupling mechanism, comprising the following steps:

[0006] Step 1: Collect multi-dimensional operating data of lithium batteries through the battery management system (BMS), preprocess the data, and construct a lithium battery health monitoring dataset;

[0007] Step 2: Use Attention-ConvTransformer to extract local temporal features, and combine it with improved Temporal Global Discriminant Analysis (T-GLDA) to extract global temporal features. Then, horizontally concatenate the local feature matrix and the global feature matrix to form a multi-scale feature matrix.

[0008] Step 3: Construct the weighted adaptive augmentation prediction model WA-TA-LightGRU-PC. The weighted adaptive augmentation prediction model embeds a feature calibration branch coupled with few-shot learning and feature-level data augmentation, a parameter update branch of online incremental learning, and a lightweight multilayer perceptron Light-MLP dynamic weight fusion layer. It calculates data sufficiency, prediction error, and distribution change indicators in real time, dynamically outputs the weights of the two branches, and performs weighted fusion (the feature calibration branch dominates in the few-shot and transition stages, and the parameter update branch dominates in the data sufficiency stage) to obtain the initial prediction value.

[0009] Step 4: The initial predicted value is given a reasonable constraint by the physical constraint layer built into the weighted adaptive enhancement prediction model. Then, the error is corrected by the optimized adaptive noise robust Kalman filter (ANR-KF) to output the final predicted value of the State of Health (SOH).

[0010] Step 5: Based on the predicted SOH (State of Health) value and the preset threshold, assess the health level of the lithium battery and generate a corresponding warning signal.

[0011] In step 1, the multi-dimensional operating data of the lithium battery includes charging curve data, discharging curve data, environmental data, and cycle aging data;

[0012] The preprocessing includes outlier removal based on the 3σ criterion, missing value imputation based on K-nearest neighbor interpolation, and max-min normalization. The normalization formula is as follows: , in, The original data, , These are the minimum and maximum values ​​of the features in the dataset, respectively. The data is after normalization;

[0013] The dataset is divided into training, validation, and test sets in a ratio (e.g., 7:2:1). The training set is used for learning model parameters, the validation set is used for hyperparameter optimization, and the test set is used for model performance evaluation.

[0014] Step 2 includes:

[0015] Attention-ConvTransformer extracts local temporal features: It uses three convolutional layers with kernel sizes of 3×3, 5×5, and 3×3 to extract local features including voltage inflection points and charging plateaus. The stride of each convolutional layer is 1, and the padding method is the same. A two-head attention mechanism is introduced to dynamically adjust the feature weights. The formula is as follows:

[0016] ,

[0017] in, , , , The convolution outputs a feature matrix. , , For learnable weight matrix, , , These are the query matrix, key matrix, and value matrix, all of which are obtained by linear transformation of the convolutional output features. For the attention head dimension, To query the product of the transpose of the key matrix and the key matrix, Scaling factor This is the activation function used to normalize the weights. This is the output local feature matrix; This indicates attention calculation; Represents convolution;

[0018] Improved Temporal Global Discriminant Analysis (T-GLDA) extracts global temporal features: It introduces a time weighting factor to optimize the intra-class and inter-class scatter matrices, capturing global aging patterns in the temporal dimension. The formula is as follows:

[0019] ,

[0020] ,

[0021] ,

[0022] ,

[0023] in, The global mean of the time series. Let i be the time weight factor for the i-th sample. For the current moment, Let be the time of collection for the i-th sample. The time decay coefficient, Let be the feature vector of the i-th sample. The scatter matrix within the time series class. The scatter matrix between time series classes, This represents the total number of aging stage categories. For the time weight of category c, Let c be the number of samples in category c. Let i be the feature vector of the i-th sample in category c. Let c be the time series mean of the c-th category, with superscript. Let e ​​denote the matrix transpose, and e denote the natural constant. Solve for... ( The inverse matrix and The eigenvectors corresponding to the first 64 largest eigenvalues ​​of the product of the original data are used to project the original data to obtain the global feature matrix. (dimension) );

[0024] Feature fusion: horizontal stitching and To form a multi-scale feature matrix Dimension .

[0025] In step 3, the weight adaptive enhancement prediction model WA-TA-LightGRU-PC includes a temporal attention layer, a feature calibration branch, a parameter update branch, and a Light-MLP dynamic weight fusion layer;

[0026] The temporal attention layer applies temporal attention weights to the multi-scale feature matrix F to highlight key temporal features. The formula is as follows:

[0027] ,

[0028] in, This is an operation to encode the temporal position of the multi-scale feature matrix F, in order to preserve temporal sequence information. for and Matrix product, for and The transpose of the product For feature dimension, The feature matrix is ​​a time-weighted matrix;

[0029] The feature calibration branch (few-shot learning + feature-level data augmentation coupling) embeds ProtoNet few-shot prototype learning to construct aging stage prototype features based on a small number of real few samples. Feature-level data augmentation is achieved through feature perturbation and interpolation. Then, the similarity between the input features and the prototype features is calculated, and the feature weights are dynamically adjusted. The formula is as follows:

[0030] ,

[0031] ,

[0032] ,

[0033] in, This represents the prototype characteristics of the k-th aging stage. For the k-th aging stage small sample set The number of samples, For samples from a small sample set, For feature vectors, For SOH labeling, for Feature mapping results after the temporal attention layer For summation operations; To enhance the feature matrix, For feature enhancement operations, The disturbance coefficient is... To L2 normalization operation, To and Linear interpolation operation for (the (k+1)th class of prototype features); For the calibrated feature matrix, for and The transpose product; obtained by mapping through two lightweight fully connected layers (64 hidden dimensions, GELU activation function). (SOH prediction value of the feature calibration branch);

[0034] The parameter update branch (online incremental learning) adopts an online sequence update mechanism, updating only the output layer weights in real time. It embeds Mini-Bayes optimization (Mini-BO) to dynamically adjust the update rate, eliminating the need for full retraining. The formula is:

[0035] ,

[0036] ,

[0037] in, Let be the output layer weight matrix at time t. For update rate, The actual SOH value at time t-1. The predicted SOH value for the parameter update branch at time t-1 is... Let be the time-weighted feature matrix at time t. for The transpose of the matrix, Update the SOH prediction value of the branch for the parameter at time t;

[0038] Light-MLP dynamic weight fusion layer input data sufficiency Prediction error and distribution changes Output the weights of the two branches. and ,satisfy We obtain the initial predicted value by weighting: ,

[0039] ,

[0040] in, For the mapping operation of a lightweight multilayer perceptron, the input is three indicators, and the output is a 2D vector. After activation, the normalized feature calibration branch weights are obtained. Update branch weights with parameters , These are the initial predicted values;

[0041] Weighting rules:

[0042] Small sample phase (number of iterations) ): The feature calibration branch is dominant;

[0043] Small sample transition phase ( ): It gradually increased to 0.8. Gradually reduced to 0.3 Gradually rising to 0.7, with dual-branch synergy;

[0044] Data Sufficient Phase ( ): 9. Parameter update branch is dominant.

[0045] In step 3, the data sufficiency The formula is calculated by multiplying the percentage of valid data by the distribution uniformity score:

[0046] ,

[0047] Among them, the number of valid samples is the number of qualified samples after preprocessing at the current moment, the number of theoretical samples for the entire life cycle is the total number of samples expected within the complete life cycle of the lithium battery, and the KL divergence score is obtained by KL divergence conversion between the current data distribution and the standard aging data distribution (range [0,1]).

[0048] Prediction error The formula for the normalized absolute error of SOH prediction is:

[0049] ,

[0050] in, This represents the final predicted SOH value at time t-1.

[0051] Data distribution changes The Wasserstein distance is used to measure the difference in distribution between new and historical data. The formula is as follows:

[0052] ,

[0053] in, This is the joint distribution set of the newly added data distribution and the historical data distribution. For in set The operation of finding the minimum target value in the middle. For joint distribution under the variable Seeking expectations, For the newly added data samples, For historical data samples, for and The L2 norm.

[0054] In step 4, the constraints of the physical constraint layer include: SOH ∈ [0,1], SOH decay in a single cycle ≤ 0.5%, and capacity decay constraints based on the SEI film growth mechanism.

[0055] ,

[0056] in, The rate of change of SOH over time. The attenuation coefficient is... The charging and discharging current (unit: A) is the current used for charging and discharging. For current influence factors, For activation energy, The gas constant is... Absolute temperature (unit: K). For Arrhenius factor;

[0057] Constraints are incorporated into the model's loss function through regularization terms:

[0058] ,

[0059] in, For the total loss function, To predict SOH value Compared with the actual SOH value The mean square error loss, For balance coefficient, This represents the physical constraint loss (1 if the constraint is violated, 0 otherwise).

[0060] In step 4, the calculation formula for the error correction is:

[0061] ,

[0062] ,

[0063] ,

[0064] ,

[0065] in, Let be the adaptive observation noise covariance matrix at time t. To adaptively update the coefficients, Let be the observation noise covariance matrix at time t-1. To convert the attention weight vector The operation of converting to a diagonal matrix, These are the initial predicted values. This is the error vector between the actual value and the initial predicted value. Let be the Kalman gain at time t. Let be the state covariance matrix at time t. The observation matrix (with a value of 1). for transpose, To constrain the penalty coefficient, The matrix within the parentheses is the inverse of the matrix. This is the final predicted SOH value. These are the predicted values ​​after physical constraints. It is an identity matrix.

[0066] In step 5, the lithium battery health level includes:

[0067] Health level );

[0068] Attenuation stage, ;

[0069] Warning level, ); ;

[0070] when A Level 1 warning will be issued (requiring immediate battery replacement). A level-two warning will be issued (reminding users to increase monitoring frequency) when... If no warning is issued, the warning signal will be output through the BMS system in two ways: audible and visual alarm and remote push.

[0071] This invention also provides a lithium battery health status prediction device based on a dynamic coupling mechanism, including a memory, a processor, a sensor module, a data transmission module, and an early warning module:

[0072] The sensor module includes a temperature sensor, a voltage sensor, a current sensor, a humidity sensor, and a cycle counter, which are arranged on the surface of the lithium battery cell and inside the battery pack to collect multi-dimensional operating data in real time.

[0073] The data transmission module uses CAN bus and wireless Bluetooth dual-mode transmission to transmit the collected data to the processor with a transmission delay of <10ms.

[0074] The memory uses an SD card (capacity ≥32GB) to store computer programs that can run on the processor, lithium battery operating data, model parameters, and historical prediction results;

[0075] The processor uses a core microcontroller that supports floating-point operations and executes the steps of the method described when running the computer program.

[0076] The warning module includes a buzzer, an LED indicator, and a wireless communication module, which are used to receive warning signals output by the processor and perform corresponding warning operations.

[0077] The present invention also provides an electronic device, including a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method.

[0078] Beneficial Effects: The innovative coupling of small-sample data augmentation significantly improves generalization ability: The feature calibration branch deeply couples small-sample prototype learning with feature-level data augmentation. Only ≤10% of the full lifecycle small-sample data is needed to generate high-fidelity augmented features through "prototype guidance + perturbation interpolation + similarity screening," solving the problems of insufficient small-sample data and poor adaptability of data augmentation to small samples in traditional methods. Experimental verification shows that in scenarios with only 100 iterations of small-sample data, the prediction error of this invention is ≤4.5%, a reduction of more than 55% compared to traditional methods (error ≥10%). Strong Incremental Learning Applicability and Outstanding Dynamic Adaptability: The parameter update branch is clearly defined as online incremental learning. Through "lightweight weight update + Mini-BO hyperparameter optimization," an update is triggered every 50-100 new iterations of data, with a latency of <20ms. After multiple updates, the error drift is ≤0.5%, avoiding the waste of computational power in full retraining. Meanwhile, combined with Light-MLP dynamic weight adjustment, it accurately adapts to the dynamic changes of "data scarcity to data abundance" throughout the entire lifecycle, improving dynamic adaptability by 60% compared to traditional static models. The model is integrated and collaborative, with a simple and efficient architecture: A unified WA-TA-LightGRU-PC model is constructed, integrating five core functions: temporal attention layer, dual-function branches, dynamic weight fusion, physical constraints, and error correction. This achieves deep collaboration in small-sample adaptation, data augmentation, incremental updates, dynamic weight adjustment, and accurate prediction, solving the problems of poor collaboration and complex deployment caused by traditional multi-module independent models. The model boasts impressive features: parameter size compressed to below 350KB, a 70% lightweight reduction compared to existing complex models (parameters ≥ 1MB); high prediction rationality and accuracy with robust reliability: the physical constraint layer applies multiple constraints based on electrochemical mechanisms, reducing unreasonable prediction rates by over 90%; the optimized ANR-KF dynamically adjusts noise covariance by combining attention weights and physical constraint losses, achieving a prediction error ≤ 2.5%, a 50% improvement in accuracy compared to traditional error correction methods (error ≥ 5%); excellent deployment adaptability and high engineering application value: the lightweight design results in inference latency < 30ms, supporting low-cost MCUs such as ARM Cortex-M7 without GPU computing power, making it fully compatible with embedded scenarios such as automotive BMS and energy storage systems. Furthermore, a three-level early warning mechanism combining audible and visual alarms and remote push notifications allows users to promptly monitor battery status, reducing maintenance costs and demonstrating broad engineering application prospects. Attached Figure Description

[0079] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention above and in other aspects will become clearer.

[0080] Figure 1 This is an overall flowchart of the lithium battery health status prediction method based on dynamic coupling mechanism of the present invention.

[0081] Figure 2 This is a schematic diagram of the internal structure of the weighted adaptive enhancement prediction model of the present invention.

[0082] Figure 3 This is a schematic diagram illustrating the working principle of the feature calibration branch (small sample + data augmentation) of the present invention.

[0083] Figure 4 This is a flowchart of the weight update process for the parameter update branch (incremental learning) of this invention. Detailed Implementation

[0084] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0085] like Figure 1 As shown, this invention discloses a method and apparatus for predicting the health status of lithium batteries based on a dynamic coupling mechanism, comprising the following steps:

[0086] I. Data Collection

[0087] 1.1 Experimental Data

[0088] The test subjects were selected as "ternary lithium battery (20Ah, nominal 3.7V)" and "lithium iron phosphate battery (30Ah, nominal 3.2V)". The full life cycle operation data was collected by the battery cycle test system (CT-4008-5V6A-S1), covering complex operating conditions such as normal temperature, high and low temperature (-20℃~60℃) and different charge and discharge rates (0.5C / 1C / 2C). The data sampling frequency was 1Hz, and the total number of cycles was 2000 (ternary lithium) and 3000 (lithium iron phosphate) respectively.

[0089] The collected multi-dimensional operational data includes:

[0090] Charging curve data: voltage-time curve, current-time curve, SOC-voltage curve;

[0091] Discharge curve data: discharge capacity, average discharge voltage, discharge efficiency;

[0092] Environmental data: internal temperature and humidity of the battery pack;

[0093] Cyclic aging data: number of cycles, cumulative charge / discharge capacity, actual SOH label (calculated from remaining capacity / rated capacity).

[0094] 1.2 Dataset Partitioning and Preprocessing Parameters

[0095] The preprocessing steps are: outlier removal according to the 3σ criterion → missing value imputation using K-nearest neighbor interpolation (K=5) → max-min normalization, where normalization maps all features to the [0,1] interval; after preprocessing, the dataset is divided into training, validation, and test sets in a 7:1:2 ratio.

[0096] Ternary lithium batteries: 1400 iterations for the training set, 200 iterations for the validation set, and 400 iterations for the test set, used for model training and parameter optimization.

[0097] Lithium iron phosphate battery: 3000 full cycles were used as a cross-model test set to verify the model's cross-scenario generalization ability.

[0098] II. Setting Core Model Parameters

[0099] like Figure 2 As shown, the core parameters of each module in the weighted adaptive augmentation prediction model (WA-TA-LightGRU-PC) of this invention are uniformly set as follows, taking into account both prediction accuracy and the lightweight requirements of embedded deployment:

[0100] 2.1 Multi-scale feature extraction parameters

[0101] Attention-ConvTransformer: Outputs a local feature matrix Dimension .

[0102] Improved T-GLDA: Outputs a global feature matrix Dimension .

[0103] Feature fusion: horizontal stitching and The multi-scale feature matrix is ​​obtained. Dimension .

[0104] 2.2 Weighted Adaptive Enhancement of Prediction Parameters

[0105] Temporal attention layer: Outputs a temporally weighted feature matrix .

[0106] like Figure 3 As shown, the feature calibration branch completes feature calibration and selection through few-sample learning.

[0107] like Figure 4 As shown, the parameter update branch dynamically optimizes the update rate and lightweight updates the model weights.

[0108] Light-MLP Dynamic Weight Fusion Layer: Achieves adaptive weight fusion based on multi-dimensional indicators.

[0109] 2.3 Physical Constraints and ANR-KF Error Correction Parameters

[0110] Physical constraint layer: Apply physical rules and constraints based on the battery aging mechanism.

[0111] Optimized ANR-KF: Adaptively corrects noise and constrains prediction bias.

[0112] 2.4 Health Assessment and Early Warning Parameters

[0113] Health level thresholds: Healthy level (SOH≥80%), Decay level (60%≤SOH<80%), Warning level (SOH<60%).

[0114] Warning triggers: Level 1 warning (SOH < 60%, continuous audible and visual alarm + remote push), Level 2 warning (60% ≤ SOH < 80%, short alarm + yellow light + remote prompt), no warning (SOH ≥ 80%, green light always on).

[0115] III. Experimental Results and Analysis

[0116] 3.1 Evaluation Indicators

[0117] Two metrics, prediction accuracy and small-sample generalization ability, were selected to comprehensively evaluate the performance of the model in this invention. The specific metrics are as follows:

[0118] 1. Prediction accuracy indicators: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), calculated using the following formulas:

[0119] ,

[0120] ,

[0121] 2. Small sample generalization metric: MAE in 100 / 200 iterations of small sample scenarios to evaluate the predictive ability under small sample data;

[0122] 3.2 Comparative Experimental Design

[0123] To verify the superiority of this invention, three sets of comparative models were set up, and parallel experiments were conducted based on the same dataset. The comparative models are all mainstream methods in the field of lithium battery SOH prediction:

[0124] 1. Comparison Model 1: Traditional single LSTM model (no few-shot learning, no incremental updates, trained only on the full dataset).

[0125] 2. Comparison Model 2: CNN+GRU fusion model (with simple data augmentation, fixed weights, no dynamic adjustment);

[0126] 3. Comparison Model 3: Existing hybrid-driven model (electrochemical mechanism + basic data driven, without physical constraint correction).

[0127] 4.3 Results Analysis

[0128] 4.3.1 Comparison of Overall Prediction Accuracy

[0129] Table 1. Performance comparison of different models on the ternary lithium battery test set.

[0130]

[0131] As shown in Table 1, the MAE of the model of this invention is only 1.8% and the RMSE is only 2.3% on the ternary lithium battery test set, which is 68.9% lower than the MAE of the traditional LSTM model and 37.9% lower than the MAE of the existing hybrid drive model.

[0132] 4.3.2 Comparison of Generalization Ability with Small Samples

[0133] Table 2. Comparison of MAE of different models in small sample scenarios

[0134]

[0135] As shown in Table 2, in scenarios with only 100 iterations of small samples, the MAE of the model of this invention is only 4.2%, which is 59.1% lower than the traditional LSTM model and 17.6% lower than the existing hybrid-driven model. Even with small sample data, the coupling of small sample prototype learning and feature-level data enhancement can still ensure high prediction accuracy, effectively solving the industry pain points of difficult data collection and poor generalization with small samples throughout the lithium battery life cycle.

[0136] In summary, the lithium battery health status prediction method and device of the present invention effectively solve the core problems of traditional methods, such as poor generalization with small samples, weak dynamic adaptability, insufficient prediction rationality, and high deployment difficulty. It takes into account both the prediction accuracy throughout the entire life cycle and the needs of embedded deployment, and has broad engineering application prospects in the fields of new energy vehicles, distributed energy storage systems, and portable electronic devices.

[0137] The above description is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention. All components not explicitly stated in this embodiment can be implemented using existing technology.

Claims

1. A method for predicting the health status of lithium batteries based on a dynamic coupling mechanism, characterized in that, Includes the following steps: Step 1: Collect multi-dimensional operating data of lithium batteries through the battery management system (BMS), preprocess the data, and construct a lithium battery health monitoring dataset. Step 2: Attention-ConvTransformer is used to extract local temporal features, and improved Temporal Global Discriminant Analysis (T-GLDA) is used to extract global temporal features. The local feature matrix and the global feature matrix are horizontally concatenated to form a multi-scale feature matrix. Step 3: Construct the weighted adaptive augmentation prediction model WA-TA-LightGRU-PC. The weighted adaptive augmentation prediction model embeds a feature calibration branch coupled with few-shot learning and feature-level data augmentation, a parameter update branch of online incremental learning, and a lightweight multilayer perceptron Light-MLP dynamic weight fusion layer. It calculates data sufficiency, prediction error, and distribution change indicators in real time, dynamically outputs the weights of the two branches, and weights them together to obtain the initial prediction value. Step 4: The initial predicted value is reasonably constrained by the physical constraint layer built into the weighted adaptive enhancement prediction model, and then the error is corrected by the optimized adaptive noise robust Kalman filter (ANR-KF) to output the final health state (SOH) predicted value. Step 5: Based on the predicted SOH (State of Health) value and the preset threshold, assess the health level of the lithium battery and generate a corresponding warning signal.

2. The method according to claim 1, characterized in that, In step 1, the multi-dimensional operating data of the lithium battery includes charging curve data, discharging curve data, environmental data, and cycle aging data; The preprocessing includes outlier removal based on the 3σ criterion, missing value imputation based on K-nearest neighbor interpolation, and max-min normalization. The normalization formula is as follows: ， in, This is the original data. , These are the minimum and maximum values ​​of the features in the dataset, respectively. The data is after normalization; The dataset is divided into training, validation, and test sets according to a set ratio. The training set is used for learning model parameters, the validation set is used for hyperparameter optimization, and the test set is used for model performance evaluation.

3. The method according to claim 2, characterized in that, Step 2 includes: Attention-ConvTransformer extracts local temporal features: It uses three convolutional layers with kernel sizes of 3×3, 5×5, and 3×3 to extract local features including voltage inflection points and charging plateaus. The stride of each convolutional layer is 1, and the padding method is the same. A two-head attention mechanism is introduced to dynamically adjust the feature weights. The formula is as follows: ， in, , , , The convolution outputs a feature matrix. , , For learnable weight matrix, , , These are the query matrix, key matrix, and value matrix, all of which are obtained by linear transformation of the convolutional output features. For the attention head dimension, To query the product of the transpose of the key matrix and the key matrix, Scaling factor This is the activation function used to normalize the weights. This is the output local feature matrix; This indicates attention calculation; Represents convolution; Improved Temporal Global Discriminant Analysis (T-GLDA) extracts global temporal features: It introduces a time weighting factor to optimize the intra-class and inter-class scatter matrices, capturing global aging patterns in the temporal dimension. The formula is as follows: ， ， ， ， in, The global mean of the time series. Let i be the time weight factor for the i-th sample. For the current moment, Let be the time of collection for the i-th sample. The time decay coefficient, Let be the feature vector of the i-th sample. The scatter matrix within the time series class. The scatter matrix between time series classes, This represents the total number of aging stage categories. For the time weight of category c, Let c be the number of samples in category c. Let i be the feature vector of the i-th sample in category c. Let c be the time series mean of the c-th category, with superscript. Let e ​​denote the matrix transpose, and e denote the natural constant. Solve for... The eigenvectors corresponding to the first 64 largest eigenvalues ​​are used to project the original data to obtain the global feature matrix. ; Feature fusion: horizontal stitching and To form a multi-scale feature matrix Dimension .

4. The method according to claim 3, characterized in that, In step 3, the weight adaptive enhancement prediction model WA-TA-LightGRU-PC includes a temporal attention layer, a feature calibration branch, a parameter update branch, and a Light-MLP dynamic weight fusion layer; The temporal attention layer applies temporal attention weights to the multi-scale feature matrix F to highlight key temporal features. The formula is as follows: ， in, This is an operation for temporal position encoding of the multi-scale feature matrix F. for and Matrix product, for and The transpose of the product For feature dimension, The feature matrix is ​​a time-weighted matrix; Feature calibration branch embedding ProtoNet few-shot prototype learning, constructing aging stage prototype features based on real few samples. Feature-level data augmentation is achieved through feature perturbation and interpolation. Then, the similarity between the input features and the prototype features is calculated, and the feature weights are dynamically adjusted. The formula is as follows: ， ， ， in, This represents the prototype characteristics of the k-th aging stage. For the k-th aging stage small sample set The number of samples, For samples from a small sample set, For feature vectors, For SOH labeling, for Feature mapping results after the temporal attention layer For summation operations; To enhance the feature matrix, For feature enhancement operations, The disturbance coefficient is... To L2 normalization operation, To and Linear interpolation operation; The calibrated feature matrix is ​​used; the SOH prediction value of the feature calibration branch is obtained by mapping through two lightweight fully connected layers. ; The parameter update branch adopts an online sequential update mechanism, updating only the output layer weights in real time, and embedding a micro Bayesian optimization to dynamically adjust the update rate. The formula is as follows: ， ， in, Let be the output layer weight matrix at time t. For update rate, The actual SOH value at time t-1. The predicted SOH value for the parameter update branch at time t-1 is... Let be the time-weighted feature matrix at time t. for The transpose of the matrix, Update the SOH prediction value of the branch for the parameter at time t; Light-MLP dynamic weight fusion layer input data sufficiency Prediction error and distribution changes Output the weights of the two branches. and ,satisfy We obtain the initial predicted value by weighting: , ， in, For the mapping operation of a lightweight multilayer perceptron, via After activation, the normalized feature calibration branch weights are obtained. Update branch weights with parameters , These are the initial predicted values.

5. The method according to claim 4, characterized in that, In step 3, the data sufficiency The formula is calculated by multiplying the percentage of valid data by the distribution uniformity score: ， Among them, the number of valid samples is the number of qualified samples after preprocessing at the current moment, and the theoretical number of samples for the entire life cycle is the total number of samples expected during the complete life cycle of the lithium battery. Prediction error The formula for the normalized absolute error of SOH prediction is: ， in, This represents the final predicted SOH value at time t-1. Data distribution changes The Wasserstein distance is used to measure the difference in distribution between new and historical data. The formula is as follows: ， in, This is the joint distribution set of the newly added data distribution and the historical data distribution. For in set The operation of finding the minimum target value in the middle. For joint distribution under the variable Seeking expectations, For the newly added data samples, For historical data samples, for and The L2 norm.

6. The method according to claim 5, characterized in that, In step 4, the constraints of the physical constraint layer include: SOH ∈ [0,1], SOH decay in a single cycle ≤ 0.5%, and capacity decay constraints based on the SEI film growth mechanism. ， in, The change rate of SOH over time. The attenuation coefficient is... This is the charging and discharging current. For current influence factors, Let R be the activation energy, R be the gas constant, and T be the absolute temperature. For Arrhenius factor; Constraints are incorporated into the model's loss function through regularization terms: ， in, For the total loss function, To predict SOH value Compared with the actual SOH value The mean square error loss, For balance coefficient, This represents the loss due to physical constraints.

7. The method according to claim 6, characterized in that, In step 4, the calculation formula for the error correction is: ， ， ， ， in, Let be the adaptive observation noise covariance matrix at time t. To adaptively update the coefficients, Let be the observation noise covariance matrix at time t-1. To convert the attention weight vector The operation of converting to a diagonal matrix, These are the initial predicted values. This is the error vector between the actual value and the initial predicted value. Let be the Kalman gain at time t. Let be the state covariance matrix at time t. For the observation matrix, To constrain the penalty coefficient, The matrix within the parentheses is the inverse of the matrix. This is the final predicted SOH value. is the predicted value after physical constraints, and E is the identity matrix.

8. The method according to claim 7, characterized in that, In step 5, the lithium battery health level includes: Health level ); Attenuation stage, ; Warning level, ); ; when A level-one warning will be issued at any time. A level-two warning was issued at that time. If no warning is issued, the warning signal will be output through the BMS system in two ways: audible and visual alarm and remote push.

9. A lithium battery health state prediction device based on a dynamic coupling mechanism, characterized in that, Includes memory, processor, sensor module, data transmission module, and early warning module: The sensor module includes a temperature sensor, a voltage sensor, a current sensor, a humidity sensor, and a cycle counter, which are arranged on the surface of the lithium battery cell and inside the battery pack to collect multi-dimensional operating data in real time. The data transmission module uses a dual-mode transmission of CAN bus and wireless Bluetooth to transmit the collected data to the processor; The memory uses an SD card to store computer programs that can run on the processor, lithium battery operating data, model parameters, and historical prediction results. The processor uses a core microcontroller that supports floating-point operations, and when running the computer program, it performs the steps of the method as described in any one of claims 1 to 9; The warning module includes a buzzer, an LED indicator, and a wireless communication module, which are used to receive warning signals output by the processor and perform corresponding warning operations.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 8.

Images