High-magnification lithium battery health assessment method and system based on deep learning

Abstract

The invention relates to the technical field of lithium battery health assessment, in particular to a high-magnification lithium battery health assessment method and system based on deep learning, and the system comprises a data preprocessing module, a dynamic characteristic extraction module, a health factor generation module, a health state assessment module, a real-time optimization module and a consistency verification module. According to the method, the operation data is subjected to standardization and noise reduction processing, dynamic characteristics are extracted in combination with an improved LSTM model, health factors are generated through PCA dimension reduction and SVR, the health state is evaluated through an MLP model, meanwhile, the real-time performance is optimized through pruning and quantification technologies, and finally the reliability is ensured through consistency verification. According to the invention, the accuracy and real-time performance of high-rate lithium battery health assessment can be effectively improved, and reliable support is provided for battery management.

Description

Technical Field

[0001] This invention belongs to the field of battery health management and artificial intelligence technology, specifically a method and system for health assessment of high-rate lithium batteries based on deep learning. Background Technology

[0002] With the increasing demand for lithium batteries in high-rate applications, accurate assessment of their state of health (SOH) has become particularly important. Existing lithium battery SOH assessment methods mainly rely on data-driven models, equivalent circuit models, or electrochemical-thermal coupling models, but their applicability and prediction accuracy under high-rate conditions are still insufficient, making it difficult to meet the needs of efficient, accurate, and real-time assessment in practical applications.

[0003] A search revealed a method and apparatus for predicting the health status of lithium batteries, published on September 26, 2025, with publication number CN119619893B. This patent achieves accurate prediction of lithium battery health status by preprocessing health factor data and combining convolutional layers, bidirectional long short-term memory networks, attention mechanisms, and encoder-decoder structures. However, this technical solution mainly targets the assessment of lithium battery health status under normal operating conditions and does not fully consider the drastic changes in the internal dynamic characteristics of the battery under high-rate charge and discharge conditions, which may lead to a decrease in prediction accuracy in high-rate scenarios. In addition, this solution has high requirements for data quality, and the robustness of the model may be affected by high noise levels.

[0004] A search revealed a method for predicting the health status of lithium batteries, published on July 12, 2024, with publication number CN116679211B. This patent improves the accuracy of lithium battery health status prediction by establishing a second-order equivalent circuit model and combining an improved supply-demand optimization algorithm (ISDO), a Gaussian process regression model, and a Transformer deep learning model. However, this technical solution has limited applicability under high-rate operating conditions, and its equivalent circuit model may not accurately reflect the nonlinear characteristics of the battery under high-rate conditions. Furthermore, while the method of fusing multiple models improves prediction accuracy, it increases computational complexity and system implementation difficulty, making it unsuitable for scenarios with high real-time requirements.

[0005] The aforementioned problems indicate that existing lithium battery health status assessment methods still have certain shortcomings in terms of applicability, prediction accuracy, model robustness, and real-time performance under high-rate operating conditions. Therefore, this invention provides a high-rate lithium battery health assessment method and system based on deep learning. The aim is to improve assessment accuracy and robustness by introducing a deep learning model optimized for high-rate operating conditions, while simplifying the calculation process to meet real-time requirements, thereby better adapting to the actual needs of high-rate application scenarios. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a high-rate lithium battery health assessment method and system based on deep learning, which at least partially solves the problems existing in the prior art.

[0007] In a first aspect of the present invention, a method and system for health assessment of high-rate lithium batteries based on deep learning are provided, the system comprising:

[0008] The data preprocessing module is used to standardize and denoise the collected lithium battery operating data. The standardization process uses a normalization formula to unify data with different dimensions to the same scale, while the denoising process uses wavelet transform to decompose the signal and reconstruct high-frequency noise components to remove interference. The operating data includes multi-dimensional features such as voltage, current, temperature, and internal resistance.

[0009] The dynamic characteristic extraction module is used to extract the dynamic characteristics of the battery under high-rate operation based on preprocessed runtime data and an improved Long Short-Term Memory (LSTM) network model. The improved LSTM model introduces an adaptive forgetting gate mechanism, which adjusts the weights of the forgetting gate by calculating the correlation coefficient between the current input and historical states, as shown in the following formula: , in, This represents the output of the forget gate. For activation function, and These are the weight matrix and bias term of the forget gate, respectively. This is the hidden state from the previous moment. For the current input, The adaptive weighting coefficients are determined by the following formula: , Through the above mechanism, the model can more accurately capture the drastic changes in the dynamic characteristics inside the battery under high-rate conditions.

[0010] The health factor generation module is used to generate health factors based on the extracted dynamic characteristics. The generation process includes two steps: First, principal component analysis (PCA) is used to reduce the dimensionality of the dynamic characteristics, retaining principal components with a cumulative contribution rate exceeding 95%. Second, a support vector regression (SVR) model is used to fit the dimensionality-reduced principal components to obtain the predicted values ​​of the health factors. The kernel function of the SVR model is the radial basis function (RBF), and its expression is: , in, The kernel width parameter was determined through cross-validation optimization.

[0011] The health status assessment module is used to evaluate the health status of lithium batteries based on generated health factors and a multilayer perceptron (MLP) model. The MLP model takes health factors and their corresponding statistical characteristics, including mean, variance, and skewness, as input. The output is a health status score, ranging from 0 to 1, representing the battery's degree of health from complete failure to complete failure. The MLP model is trained using the Adaptive Moment Estimation (Adam) algorithm, with mean squared error (MSE) as the loss function. The expression for MLP is: , in, For the sample size, For the true value, These are predicted values.

[0012] The real-time optimization module reduces computational complexity through model pruning and quantization techniques to meet real-time requirements. Model pruning is achieved by removing neurons whose absolute weight value is less than a preset threshold, and quantization converts floating-point weights into low-precision integer representations. The pruning threshold and quantization bit width are determined experimentally to ensure that the model accuracy loss does not exceed 1%.

[0013] The consistency verification module is used to determine the reliability of the evaluation system based on consistency indicators of multiple evaluation results. These consistency indicators include the Pearson correlation coefficient and the Kendall rank correlation coefficient, which measure linear correlation and ordination consistency, respectively. The formula for calculating the Pearson correlation coefficient is as follows: , in, and These are the values ​​of the two evaluation results. and Its mean; the formula for calculating the Kendall rank correlation coefficient is: , in, and These represent the number of consistent pairs and inconsistent pairs, respectively; when the consistency index falls below a preset threshold, a reassessment process is triggered.

[0014] The present invention has at least the following beneficial effects: The high-rate lithium battery health assessment method and system based on deep learning provided by the present invention, through the data preprocessing module, eliminates dimensional differences and noise interference by standardizing and denoising the raw data, laying a reliable data foundation for subsequent dynamic characteristic extraction. The dynamic characteristic extraction module introduces an adaptive forgetting gate mechanism through an improved LSTM model, significantly improving the model's ability to capture the dynamic characteristics of the battery under high-rate conditions, and solving the problem of insufficient response of traditional models in drastically changing scenarios. The health factor generation module combines PCA and SVR technologies, which not only reduces the data dimensionality but also improves the accuracy of health factor generation, avoiding the interference of redundant features on the assessment results. The health status assessment module integrates multiple statistical features through an MLP model to achieve a comprehensive assessment of the battery health status, and further improves the convergence speed and prediction accuracy of the model by using the Adam optimization algorithm and the MSE loss function. The real-time optimization module significantly reduces the computational complexity while ensuring assessment accuracy through model pruning and quantization techniques, meeting the strict real-time requirements of high-rate application scenarios. Finally, the consistency verification module quantifies the consistency of the evaluation results using Pearson correlation coefficient and Kendall rank correlation coefficient, providing an intuitive basis for the reliability of the system and effectively improving the overall performance of high-rate lithium battery health assessment. Attached Figure Description

[0015] Figure 1 The schematic diagram of the module structure of the high-rate lithium battery health assessment system based on deep learning provided in the embodiments of the present invention shows the connection relationship between the data preprocessing module, dynamic characteristic extraction module, health factor generation module, health status assessment module, real-time optimization module and consistency verification module.

[0016] Figure 2 This diagram illustrates the principle of the adaptive forget gate mechanism of the improved Long Short-Term Memory (LSTM) network model in this embodiment of the invention. It details the calculation process of the forget gate output and the adaptive weight coefficients, as well as their role in dynamic feature extraction.

[0017] Figure 3 This is a flowchart illustrating the workflow of the health factor generation module in this embodiment of the invention, describing the steps of principal component analysis (PCA) dimensionality reduction and support vector regression (SVR) model fitting, and their application in health factor generation.

[0018] The attached diagram is labeled as follows: 1. Data preprocessing module; 2. Dynamic feature extraction module; 3. Health factor generation module; 4. Health status assessment module; 5. Real-time optimization module; 6. Consistency verification module. Detailed Implementation

[0019] This invention provides a method and system for health assessment of high-rate lithium batteries based on deep learning, the module structure of which is as follows: Figure 1 As shown, the system includes a data preprocessing module 1, a dynamic characteristic extraction module 2, a health factor generation module 3, a health status assessment module 4, a real-time optimization module 5, and a consistency verification module 6. These modules are logically connected to form a complete system for comprehensively assessing the health status of high-rate lithium batteries. The following section, in conjunction with the appendix... Figure 1 To be continued Figure 3 The specific embodiments of the present invention are described in detail below.

[0020] Data preprocessing module 1 is the starting point of the entire system, responsible for standardizing and denoising the collected lithium battery operating data. In practical applications, lithium battery operating data typically includes multi-dimensional features such as voltage, current, temperature, and internal resistance. Since these data may contain dimensional differences and noise interference, preprocessing is necessary to improve the accuracy of subsequent analysis. Standardization uses a normalization formula to unify data with different dimensions to the same scale; for example, mapping the voltage range from 0 to 4.2 volts to 0 to 1. Denoising uses wavelet transform to decompose the signal and reconstruct high-frequency noise components to remove interference. Specifically, wavelet transform separates high-frequency noise components by performing multi-scale decomposition of the signal and retains the effective low-frequency signal through reconstruction. The output of data preprocessing module 1 is directly passed to dynamic characteristic extraction module 2 as the basis for subsequent analysis.

[0021] The core of the dynamic characteristic extraction module 2 is an improved Long Short-Term Memory (LSTM) network model. This model significantly improves the ability to capture the dynamic characteristics of batteries under high-rate conditions by introducing an adaptive forgetting gate mechanism. For example... Figure 2 As shown, the improved LSTM model's adaptive forget gate mechanism adjusts the forget gate weights by calculating the correlation coefficient between the current input and historical states. Specifically, the output of the forget gate... From the formula Confirmed, among which For activation function, and These are the weight matrix and bias term of the forget gate, respectively. This is the hidden state from the previous moment. For the current input, These are adaptive weighting coefficients. The value is determined by the formula The calculation shows that the formula dynamically adjusts the weights of the forget gate by measuring the distance between the current input and the historical state. The output of the dynamic feature extraction module 2 is a set of feature vectors that reflect the dynamic characteristics inside the battery, and these feature vectors are then passed to the health factor generation module 3.

[0022] The workflow of health factor generation module 3 is as follows: Figure 3 As shown, the process mainly includes two steps: Principal Component Analysis (PCA) dimensionality reduction and Support Vector Regression (SVR) model fitting. First, PCA is used to reduce the dimensionality of the feature vectors output by the dynamic characteristic extraction module 2, retaining principal components with a cumulative contribution rate exceeding 95%. This process is achieved by calculating the eigenvalues ​​and eigenvectors of the covariance matrix, ensuring that the reduced data retains the information of the original data to the greatest extent possible. Second, the SVR model is used to fit the reduced principal components to obtain the predicted values ​​of the health factors. The radial basis function (RBF) is used as the kernel function for the SVR model, and its expression is: ,in The kernel width parameter is determined through cross-validation optimization. The output of the health factor generation module 3 is a set of health factors, which are then passed to the health status assessment module 4.

[0023] The health status assessment module 4 uses a multilayer perceptron (MLP) model to fuse and analyze health factors and their corresponding statistical characteristics to assess the health status of the lithium battery. The MLP model's input includes health factors and their corresponding statistical characteristics such as mean, variance, and skewness. The output is a health status score, ranging from 0 to 1, representing the battery's degree of health from complete failure to complete failure. The MLP model is trained using the Adaptive Moment Estimation (Adam) algorithm, with the mean squared error (MSE) loss function, expressed as follows: ,in For the sample size, For the true value, These are predicted values. The output of the health status assessment module 4 is a set of health status scores, which are then passed to the real-time optimization module 5 and the consistency verification module 6.

[0024] The real-time optimization module 5 reduces computational complexity through model pruning and quantization techniques to meet real-time requirements. Model pruning is achieved by removing neurons whose absolute weights are less than a preset threshold, and quantization converts floating-point weights into low-precision integer representations. The pruning threshold and quantization bit width were determined experimentally to ensure that the model accuracy loss does not exceed 1%. The output of the real-time optimization module 5 is an optimized model that significantly reduces computational complexity while maintaining evaluation accuracy, meeting the stringent real-time requirements of high-rate applications.

[0025] The consistency verification module 6 quantifies the consistency of the evaluation results using the Pearson correlation coefficient and Kendall's rank correlation coefficient, thereby determining the reliability of the evaluation system. The formula for calculating the Pearson correlation coefficient is as follows: ,in and These are the values ​​of the two evaluation results. and Its mean. The formula for calculating the Kendall rank correlation coefficient is: ,in and These represent the number of consistent and inconsistent pairs, respectively. When the consistency index falls below a preset threshold, a re-evaluation process is triggered. The output of the consistency verification module 6 is a judgment on the reliability of the evaluation system, providing an intuitive basis for the overall performance of the system.

[0026] In practical applications, the aforementioned modules work collaboratively through logical connections to jointly assess the health status of high-rate lithium batteries. For example, in the electric vehicle field, the health status of lithium batteries directly affects the vehicle's range and safety. The system provided by this invention can monitor the operating status of lithium batteries in real time and accurately assess their health level, thereby ensuring the safe operation of the vehicle. Furthermore, this invention can also be applied to energy storage systems, drones, and other fields, providing technical support for the health management of high-rate lithium batteries.

[0027] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0028] In the field of electric vehicles, lithium battery health assessment is a crucial step in ensuring vehicle range and operational safety. The high-rate lithium battery health assessment system based on deep learning provided by this invention can monitor battery operating status in real time and accurately assess its health level. The following, in conjunction with the appendix... Figure 1 To be continued Figure 3 The system's operating principle is explained in detail, including specific implementation steps.

[0029] First, in data preprocessing module 1, the collected lithium battery operating data includes multi-dimensional features such as voltage, current, temperature, and internal resistance. These data may contain dimensional differences and noise interference, therefore standardization and noise reduction are necessary. Standardization uses a normalization formula to unify data with different dimensions to the same scale; for example, mapping the voltage range from 0 to 4.2 volts to 0 to 1. Noise reduction decomposes the signal using wavelet transform, separating high-frequency noise components and reconstructing the effective low-frequency signal. This process ensures the data quality for subsequent analysis, providing a reliable foundation for dynamic characteristic extraction module 2.

[0030] Subsequently, the dynamic feature extraction module 2 analyzes the preprocessed data using an improved Long Short-Term Memory (LSTM) network model. For example... Figure 2As shown, the improved LSTM model introduces an adaptive forget gate mechanism, dynamically adjusting the weights of the forget gate by calculating the correlation coefficient between the current input and historical states. Specifically, the output of the forget gate... From the formula Confirmed, among which These are adaptive weighting coefficients, and their values ​​are obtained through the formula... The calculations show that this mechanism can more accurately capture the drastic changes in the dynamic characteristics inside the battery under high-rate operating conditions, thereby generating a set of feature vectors that reflect the dynamic characteristics of the battery.

[0031] In module 4 of the health status assessment, the Multilayer Perceptron (MLP) model fuses and analyzes health factors and their corresponding statistical characteristics. The input to the MLP model includes health factors and their statistical characteristics such as mean, variance, and skewness. The output is a health status score, ranging from 0 to 1, representing the degree of battery health from complete failure to complete health. The training process of the MLP model uses the Adaptive Moment Estimation Optimization (Adam) algorithm, with the Mean Squared Error (MSE) loss function, expressed as follows: Through this process, the system can comprehensively assess the health status of lithium batteries and generate a set of health status scores.

[0032] To meet the stringent real-time requirements of high-rate applications, the real-time optimization module 5 reduces computational complexity through model pruning and quantization techniques. Model pruning is achieved by removing neurons whose absolute weights are less than a preset threshold, while quantization converts floating-point weights into low-precision integer representations. The pruning threshold and quantization bit width were determined experimentally to ensure that the model's accuracy loss does not exceed 1%. The optimized model significantly reduces computational complexity while maintaining evaluation accuracy, thus meeting real-time requirements.

[0033] Finally, the consistency verification module 6 quantifies the consistency of the evaluation results using the Pearson correlation coefficient and the Kendall rank correlation coefficient. The formula for calculating the Pearson correlation coefficient is as follows: The formula for calculating the Kendall rank correlation coefficient is: When the consistency index falls below a preset threshold, a reassessment process is triggered. This process provides a direct indicator of the system's reliability.

[0034] Through the steps described above, the system provided by this invention can monitor the operating status of lithium batteries in real time and accurately assess their health level. For example, in practical applications of electric vehicles, the system can generate a health status score based on real-time collected battery data and determine the reliability of the assessment results through a consistency verification module. This process not only improves assessment accuracy but also significantly reduces computational complexity, meeting the stringent requirements of real-time performance and reliability in high-rate application scenarios. Furthermore, this invention can also be applied to energy storage systems and drones, providing technical support for the health management of high-rate lithium batteries.

[0035] All content not described in detail in this specification is prior art known to those skilled in the art, and the model parameters of each electrical appliance are not specifically limited; conventional equipment can be used. Electrical control components not mentioned in this technical solution are not shown in the figures because they are prior art, and will not be described further here.

[0036] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-rate lithium battery health assessment system based on deep learning, characterized in that, The system includes: The data preprocessing module (1) is used to standardize and denoise the collected lithium battery operating data. The standardization process uses a normalization formula to unify data of different dimensions to the same scale, and the denoising process uses wavelet transform to decompose the signal and reconstruct high-frequency noise components to remove interference. The operating data includes voltage, current, temperature and internal resistance. The dynamic characteristic extraction module (2) is used to extract the dynamic characteristics of the battery under high-rate conditions based on the preprocessed running data and the improved long short-term memory network model. The improved long short-term memory network model introduces an adaptive forget gate mechanism, which adjusts the weight of the forget gate by calculating the correlation coefficient between the current input and the historical state. The health factor generation module (3) is used to generate health factors based on the extracted dynamic characteristics. The health factor generation process includes two steps: First, the dynamic characteristics are reduced in dimensionality through principal component analysis, and the principal components with a cumulative contribution rate of more than 95% are retained. Second, the principal components after dimensionality reduction are fitted using the support vector regression model to obtain the predicted value of the health factor. The radial basis function is selected as the kernel function of the support vector regression model. The health status assessment module (4) is used to assess the health status of lithium batteries based on the generated health factors and combined with the multilayer perceptron model. The input of the multilayer perceptron model is the health factors and their corresponding statistical features, including the mean, variance and skewness. The output is the health status score, which ranges from zero to one. The training process of the multilayer perceptron model adopts the adaptive moment estimation optimization algorithm, and the loss function is the mean square error. The real-time optimization module (5) is used to reduce computational complexity through model pruning and quantization techniques. Model pruning is achieved by removing neurons whose absolute weights are less than a preset threshold, and quantization techniques convert floating-point weights into low-precision integer representations. The consistency verification module (6) is used to determine the reliability of the evaluation system based on the consistency index of multiple evaluation results; the consistency index includes Pearson correlation coefficient and Kendall rank correlation coefficient.

2. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, In the adaptive forgetting gate mechanism of the improved Long Short-Term Memory network model, the output of the forgetting gate... Determined by the following formula: f t = σ W f × h t - 1 x t + b f × α t in, For activation function, and These are the weight matrix and bias term of the forget gate, respectively. This is the hidden state from the previous moment. For the current input, For adaptive weighting coefficients; adaptive weighting coefficients The value is determined by the following formula:

3. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, The radial basis function expression of the support vector regression model is: in, The kernel width parameter is determined through cross-validation optimization.

4. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, The mean squared error loss function expression for the multilayer perceptron model is as follows: in, For the sample size, For the true value, These are predicted values.

5. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, The formula for calculating the Pearson correlation coefficient is as follows: in, and These are the values ​​of the two evaluation results. and Let it be its mean.

6. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, The formula for calculating the Kendall rank correlation coefficient is as follows: in, and These represent the number of consistent pairs and inconsistent pairs, respectively.

7. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, In the real-time optimization module (5), the pruning threshold and quantization bit width are determined experimentally to ensure that the model accuracy loss does not exceed one percent.

8. The high-rate lithium battery health assessment system based on deep learning according to claim 1, characterized in that, When the consistency index falls below a preset threshold, a reassessment process is triggered.

Images