A battery system multi-performance index joint estimation method based on CNN-BiLSTM-SAM
By combining a CNN-BiLSTM-SAM hybrid deep learning model with KPCA and multi-level regularization techniques, the problem of joint estimation of multiple performance indicators in lithium-ion battery systems was solved, achieving high-precision SOC and SOE prediction and improving the model's generalization ability and prediction stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANGTZE DELTA REGION INST OF UNIV OF ELECTRONICS SCI & TECH OF CHINE (HUZHOU)
- Filing Date
- 2026-04-30
- Publication Date
- 2026-05-29
AI Technical Summary
In existing technologies, the coupling relationship between multiple performance parameters of lithium-ion battery systems is not fully utilized. The features have high dimensionality and nonlinear correlations. Deep learning models have insufficient generalization ability and low efficiency in hyperparameter tuning, making it difficult to achieve efficient joint estimation of multiple performance indicators.
A hybrid deep learning model of CNN-BiLSTM-SAM is adopted, combined with KPCA for nonlinear dimensionality reduction, and L2 regularization, Dropout mechanism and data augmentation are introduced. Genetic algorithm is used to optimize hyperparameters and construct a multi-level regularization system to achieve joint estimation of multiple performance indicators.
It significantly improves the model's generalization performance and prediction accuracy, accurately tracks the changing trends of SOC and SOE at different aging stages, reduces overfitting, and provides an efficient battery state monitoring and lifetime prediction solution.
Smart Images
Figure CN122109856A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of battery management system technology, and in particular to a method for joint estimation of multiple performance indicators of battery systems based on CNN-BiLSTM-SAM. Background Technology
[0002] Lithium-ion battery systems are widely used in electric vehicles, energy storage systems, and portable electronic devices due to their high energy density and long cycle life. A Battery Management System (BMS) needs to monitor multiple performance parameters of the battery system in real time to ensure safe and efficient battery operation. These performance parameters include State of Health (SOH), reflecting the battery's aging level; State of Charge (SOC), reflecting the remaining charge; and State of Energy (SOE), reflecting the remaining energy. These three are the main monitoring targets of the battery management system and together constitute a complete description of the battery's state.
[0003] However, SOH, SOC, and SOE cannot be directly measured and must be indirectly estimated using measurable parameters (such as voltage, current, and temperature). Existing technologies have the following shortcomings:
[0004] Limitations of single-state estimation: Most methods only estimate a single performance parameter (such as only SOH or only SOC), ignoring the coupling relationship between multiple performance parameters, making it difficult to achieve a comprehensive evaluation of battery system performance.
[0005] High feature dimensionality and nonlinear correlation: Health features extracted from battery charging and discharging behavior are often high in dimensionality, and there are complex nonlinear correlations between the features. If the original features are used directly for modeling, it will not only increase the computational complexity of the model, but may also lead to overfitting and reduce the model's generalization ability.
[0006] Insufficient generalization ability of the model: Deep learning models are prone to overfitting in battery state prediction, especially when the training samples are limited, the prediction accuracy of the model on the test set drops significantly.
[0007] Hyperparameters rely on manual tuning: The predictive performance of deep learning models is highly dependent on the selection of hyperparameters. Traditional manual tuning methods are inefficient and it is difficult to find the global optimal solution.
[0008] Therefore, there is an urgent need for a method for joint estimation of multiple performance indicators of battery systems that can effectively handle characteristic nonlinear correlations, achieve joint estimation of multiple state parameters, have good generalization ability, and adaptively optimize hyperparameters. Summary of the Invention
[0009] To overcome the limitations of existing single-state estimation techniques, such as high feature dimensionality, nonlinear correlation, insufficient model generalization ability, and reliance on manual hyperparameter tuning, this invention provides a joint estimation method for multiple performance indicators of battery systems based on CNN-BiLSTM-SAM, comprising the following steps:
[0010] S1. Collect timing data during the battery's cyclic charging and discharging process;
[0011] S2. Based on time-series data, the actual discharge capacity of the battery is calculated using the ampere-hour integration method, and then the battery SOH is calculated.
[0012] S3. Extract multi-dimensional health characteristic parameters related to SOH during battery cycle charging and discharging;
[0013] S4. Use KPCA to perform nonlinear dimensionality reduction on the multidimensional health feature parameters to obtain the dimensionality-reduced feature vector;
[0014] S5. Integrate time series data, SOH, and dimensionality-reduced feature vectors to construct the input feature matrix;
[0015] S6. Construct a CNN-BiLSTM-SAM hybrid deep learning model for battery state prediction using the input feature matrix;
[0016] S7. Introduce a regularization system to train the CNN-BiLSTM-SAM hybrid deep learning model to obtain the regularized CNN-BiLSTM-SAM hybrid deep learning model.
[0017] S8. Use a genetic algorithm to globally optimize the key hyperparameters of the regularized CNN-BiLSTM-SAM hybrid deep learning model to obtain the optimal hyperparameters.
[0018] S9. Retrain the regularized CNN-BiLSTM-SAM hybrid deep learning model using the optimal hyperparameters, and use the trained model to predict the battery SOC and SOE, outputting the joint estimation results of multiple battery performance indicators.
[0019] Preferably, in step S1, the time series data includes voltage, current and temperature, and the time series data is preprocessed and abnormal loop data is removed.
[0020] Preferably, in step S2, the formula for calculating the battery SOH is: ;
[0021] in, The number of charge / discharge cycles for the battery. For the first Battery health status under one cycle For the battery in the first The actual maximum discharge capacity under one cycle This refers to the rated discharge capacity of the new battery.
[0022] Preferably, in step S3, the multi-dimensional health characteristic parameters include constant current charging time (CCCT), constant current discharging time (CCDT), constant voltage charging time (CVCT), constant voltage rising charging time (CVRT), and constant voltage falling discharging time (CVDT).
[0023] Preferably, in step S4, the nonlinear dimensionality reduction using KPCA specifically involves:
[0024] set up This represents a set of data with zero mean. Given the total number of data samples, the kernel function maps... Implementation point arrive The mapping, the feature space center data is: ;
[0025] in, For kernel function on the 1st The nonlinear mapping result of a data sample , This is the eigenvector space after nonlinear mapping;
[0026] The covariance of the mapped data in the feature space is: ;
[0027] in, Let covariance be the data mapped in the feature space. For kernel function on the 1st The nonlinear mapping result of a data sample;
[0028] Diagonalize the covariance matrix through eigenvalue decomposition: ;
[0029] in, These are the eigenvalues of the covariance matrix. These are the eigenvectors of the covariance matrix. ;get: ;
[0030] in, These are the eigenvectors obtained by eigenvalue decomposition of the covariance matrix.
[0031] The linear representation is: ; ;
[0032] in, The constant coefficient, , For kernel function on the 1st , The nonlinear mapping result of a data sample , The values are 1, 2, ..., N;
[0033] definition Kernel matrix for: ;
[0034] The kernel matrix is corrected by centering: ;
[0035] in, The kernel matrix after centering correction is the first... Line number Column elements, The first kernel in the original kernel matrix Line number Column elements, The first kernel in the original kernel matrix Line number Column elements, The first kernel in the original kernel matrix Line number The elements of the column.
[0036] Preferably, in step S5, the input feature is: ;
[0037] in, for Time-time feature vector for Voltage at any moment for Current at any moment for Temperature at all times This is the SOH value for the current cycle.
[0038] Preferably, in step S6, the model includes an input layer, a CNN module, a SAM module, a BiLSTM module, and an output layer;
[0039] The operating principle of the CNN module is as follows: ; ;
[0040] in, , The first Convolutional layers output features and input features. , The first Feature weights and bias coefficients after convolution, * denotes the convolution symbol. For activation function, To Max pooling result, A function to perform max pooling operations;
[0041] The operating principle of the SAM module is as follows: ; ; ; ; ;
[0042] in, For the first Attention energy value of each feature For time steps, The hyperbolic tangent activation function is used. Train the parameter matrix for the SAM module. The bias training parameters for the SAM module. To train the parameter vector for the SAM module, h i It is a bidirectional feature sequence. For the first Attention weights for each feature These are intermediate mapping features for the SAM module. For activation function, For mapping parameters, For output gate coefficients;
[0043] The operating principle of the BiLSTM module is as follows: ; ; ;
[0044] in, , The forward LSTM at time step , The hidden state, , The backward LSTM at time steps , The hidden state, In time step The input feature vector, This is a function for the bidirectional long short-term memory network layer. For bidirectional LSTM at time steps The final output state.
[0045] Preferably, in step S7, the L2 regularization operation is as follows:
[0046] Let the original loss function be... ,in For all trainable parameters of the model, the total loss function after introducing L2 regularization is: ;
[0047] in, The L2 regularization coefficient is... It is the sum of squares of all weight parameters;
[0048] The formula for the weight gradient is: ;
[0049] in, The total loss function with respect to the weights gradient, The original loss function with respect to the weights The gradient;
[0050] The weight update formula is: ;
[0051] Where η is the learning rate. For the updated weights, The weights before the update. The total loss function corresponding to the weights before the update. The total loss function before update with respect to the weights The gradient;
[0052] Output of the Dropout layer for: ;
[0053] Where a(x) is the activation output of the previous layer network. is a binary mask vector with the same dimension as a(x), and ⊙ is the element-wise multiplication operator;
[0054] The data augmentation mechanism is as follows:
[0055] Noise injection: A small amount of Gaussian noise is added to the original voltage, current, and temperature sequence;
[0056] Time scaling: Performs slight time-axis scaling transformations on time-series data;
[0057] Fragment sampling: Randomly sampling subsequences from different starting points from a long sequence.
[0058] Preferably, in step S8, the genetic algorithm aims to minimize the MAE of the test set, and the hyperparameters include the number of CNN convolutional kernels, the number of BiLSTM hidden units, the initial learning rate, the L2 regularization coefficient, the Dropout ratio, and the batch size.
[0059] Preferably, in step S9, the formulas for predicting SOC and SOE are as follows: ; ;
[0060] in, , , The battery number , , One sample data, For the first In the loop, the th The state of charge of the battery in each sample For the first In the loop, the th Battery energy state of a sample; For the first In the loop, the first... Electricity data for each sample; For the first In the loop, the first... Current data for each sample; For the first In the loop, the first... Voltage data for each sample; For the first In the loop, the first... Energy data for each sample; For the battery in the first The maximum energy under each cycle The sampling time interval between two adjacent samples.
[0061] The beneficial effects of this invention are as follows:
[0062] This invention addresses the problem of nonlinear correlation among five health features by innovatively introducing KPCA (Kernel Principal Component Analysis) for dimensionality reduction. By mapping the original features to a high-dimensional feature space using radial basis function kernels, the nonlinearly separable problem is transformed into a linearly separable one. Principal components are then extracted using PCA. Dimensionality reduction results show that the cumulative contribution rate of the first three principal components reaches 90.366%, effectively reducing the feature dimensionality (from five dimensions to three) while preserving the effective information of the original data to the maximum extent. Comparative experiments show that compared to linear PCA dimensionality reduction, KPCA dimensionality reduction reduces the test set MAE (Mean Absolute Error) by 18.9%, significantly improving the expressive power of the features and the generalization performance of the model.
[0063] This invention overcomes the limitations of traditional methods that estimate only a single state. It uses the nonlinear health characteristics after KPCA dimensionality reduction, the battery aging information SOH, and instantaneous measurements (voltage, current, temperature) as model inputs. This allows the model to simultaneously utilize the battery's nonlinear degradation characteristics, overall health status, and local temporal characteristics to achieve high-precision joint prediction of SOC and SOE. Experimental results show that this method can accurately track the changing trends of SOC and SOE at different aging stages (initial cycle, 50 cycles, 100 cycles, and 150 cycles), with prediction errors controlled within a small range.
[0064] This invention addresses the overfitting problem that often occurs in deep learning models for battery state prediction by innovatively constructing a multi-layered regularization system that includes L2 regularization, Dropout mechanism, early stopping strategy, and data augmentation. These four regularization techniques work together at the parameter, structural, training process, and data levels, creating a synergistic effect. Comparative experiments show that the multi-layered regularization system reduces the error difference between the training and test sets from 0.0137 to 0.0018, effectively suppressing overfitting and enabling the model to maintain stable predictive ability even on complex, high-noise battery time-series data.
[0065] This invention introduces a genetic algorithm into the training process of the CNN-BiLSTM-SAM model, aiming to minimize the MAE on the test set and globally optimize key hyperparameters. Experimental results show that the absolute value of the SOH prediction error is reduced from less than 0.16 to less than 0.02, and the SOC prediction error is reduced from 0.1577 to 0.0294, verifying the effectiveness of the genetic algorithm in hyperparameter optimization. This method avoids the inefficiency and local optima problems of traditional manual parameter tuning, providing an efficient and reliable solution for the engineering application of the model.
[0066] This invention constructs a complete battery state prediction method system encompassing data preprocessing, SOH calculation, multi-dimensional health feature extraction, KPCA nonlinear dimensionality reduction, input feature construction integrating dimensionality-reduced features and SOH information, CNN-BiLSTM-SAM modeling, multi-level regularization, genetic algorithm optimization, and accuracy verification. This method system possesses advantages such as strong feature physical interpretability, superior nonlinear feature representation ability, high prediction accuracy, good generalization ability, and adaptive optimization. It can provide theoretical basis and technical support for battery management system state monitoring and lifetime prediction, and has significant engineering application prospects. Attached Figure Description
[0067] Figure 1 This is a schematic diagram of the overall process of an embodiment of the present invention;
[0068] Figure 2 This is a schematic diagram illustrating the correlation analysis results between battery health characteristic parameters and SOH in an embodiment of the present invention;
[0069] Figure 3 This is a schematic diagram of the KPCA analysis results of battery health characteristic parameters according to an embodiment of the present invention;
[0070] Figure 4 This is a schematic diagram of the training set prediction results of SOH in an embodiment of the present invention;
[0071] Figure 5 This is a schematic diagram of the training set prediction error results of SOH in an embodiment of the present invention;
[0072] Figure 6 This is a schematic diagram of the SOH test set prediction results before and after the introduction of the optimization algorithm in an embodiment of the present invention;
[0073] Figure 7 This is a schematic diagram of the training set prediction results of SOC in an embodiment of the present invention;
[0074] Figure 8 This is a schematic diagram of the SOC test set prediction results before and after the introduction of the optimization algorithm in an embodiment of the present invention;
[0075] Figure 9 This is a schematic diagram of the training set prediction results of SOE in an embodiment of the present invention;
[0076] Figure 10 This is a schematic diagram of the SOE test set prediction results before and after the introduction of the optimization algorithm in an embodiment of the present invention. Detailed Implementation
[0077] Embodiments of the present invention provide a method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM, such as... Figure 1 As shown, it includes the following steps:
[0078] S1. Collect time-series data during the battery's cyclic charging and discharging process, preprocess the time-series data and remove abnormal cyclic data to obtain preprocessed time-series data.
[0079] Timing data includes voltage, current, and temperature.
[0080] S2. Based on the preprocessed time-series data, the actual discharge capacity of the battery is calculated using the ampere-hour integration method to obtain the actual maximum discharge capacity, and then the battery SOH is calculated using the following formula: ;
[0081] in, The number of charge / discharge cycles for the battery. For the first Battery health status under one cycle For the battery in the first The actual maximum discharge capacity under one cycle This refers to the rated discharge capacity of the new battery.
[0082] S3. Extract multi-dimensional health characteristic parameters that are strongly correlated with SOH during the battery cycle charging and discharging process, including constant current charging time CCCT, constant current discharging time CCDT, constant voltage charging time CVCT, constant voltage rising charging time CVRT, and constant voltage falling discharging time CVDT.
[0083] Calculate the Pearson correlation coefficients between the constant current charging time (CCCT), constant current discharging time (CCDT), constant voltage charging time (CVCT), constant voltage rise charging time (CVRT), and constant voltage fall discharging time (CVDT) and the state of equilibrium (SOH), such as... Figure 2 As shown, this is to verify that these five health characteristic parameters are strongly correlated with SOH.
[0084] S4. Based on the problem of nonlinear correlation between multidimensional health feature parameters, KPCA is used for nonlinear dimensionality reduction, and the first few principal components are extracted as the feature vector after dimensionality reduction.
[0085] Nonlinear dimensionality reduction using KPCA specifically involves:
[0086] set up Given a set of zero-mean data samples, Given the total number of data samples, the kernel function maps... Implementation point arrive The mapping can yield feature space center data that must satisfy: ;
[0087] in, For kernel function on the 1st The nonlinear mapping result of a data sample , This is the eigenvector space after nonlinear mapping;
[0088] The covariance of the mapped data in the feature space is: ;
[0089] in, Let covariance be the data mapped in the feature space. For kernel function on the 1st The nonlinear mapping result of a data sample;
[0090] The covariance matrix can be diagonalized through eigenvalue decomposition, and the decomposition formula is as follows: ;
[0091] in, These are the eigenvalues of the covariance matrix. These are the eigenvectors of the covariance matrix. ;get: ;
[0092] in, The eigenvectors obtained by eigenvalue decomposition of the covariance matrix are the eigenvalues. correspond.
[0093] Considering all All exist The eigenvectors are linearly represented as: ; ;
[0094] in, The constant coefficient, , For kernel function on the 1st , The nonlinear mapping result of a data sample , The values are 1, 2, ..., N;
[0095] definition Kernel matrix for: ;
[0096] get: ;
[0097] Further simplification yields: ;
[0098] Training samples In the eigenvector space middle The projection of the direction is: ;
[0099] in, The feature vector index;
[0100] Replacing the inner product with a kernel function yields: ;
[0101] If the above formula is not true, it needs to be adjusted to: ;
[0102] The kernel matrix is corrected by centering: ;
[0103] in, The kernel matrix after centering correction is the first... Line number Column elements, The first kernel in the original kernel matrix Line number Column elements, The first kernel in the original kernel matrix Line number Column elements, The first kernel in the original kernel matrix Line number The elements of the column.
[0104] This embodiment uses the radial basis function (RBF) as the kernel function. The kernel width parameter is automatically determined using the median heuristic to construct the kernel matrix, which is then centered. Eigenvalue decomposition is performed on the centered kernel matrix to obtain the eigenvalues and eigenvectors of the covariance matrix. The top eigenvalues are selected based on the cumulative contribution rate. Principal components, ≤ Original feature dimension (the dimension of the 5 health feature vectors in step S3), such as Figure 3 As shown; based on the dimensionality reduction calculation results, features with a cumulative contribution rate of over 90% are selected as the dimensionality-reduced feature vectors and used as input for subsequent SOH prediction. The comparison results of different dimensionality reduction strategies are shown below:
[0105] Table 1: ;
[0106] S5. By fusing the preprocessed time-series data, SOH, and dimensionality-reduced feature vectors, an input feature matrix containing battery nonlinear health characteristics and aging information is constructed.
[0107] The input feature matrix is: ;
[0108] in, for Voltage at any moment for Current at any moment for Temperature at all times This is the SOH value for the current cycle.
[0109] S6. Construct a CNN-BiLSTM-SAM hybrid deep learning model for battery state prediction using the input feature matrix. The model includes an input layer, a CNN module, a SAM module, a BiLSTM module, and an output layer.
[0110] CNN module: Contains at least one convolutional layer and one activation layer, used to extract local feature patterns from the input data. The operating principle of this module is as follows: ; ;
[0111] in, , The first Convolutional layers output features and input features. , The first Feature weights and bias coefficients after convolution, * denotes the convolution symbol. For activation function, To Max pooling result, A function to implement max pooling operations.
[0112] SAM module: Self-attention mechanism layer, which dynamically allocates weights by calculating the correlation between features, enhancing the attention to key time steps. The operating principle of this module is as follows: ; ; ; ; ;
[0113] in, For the first Attention energy value of each feature For time steps, The hyperbolic tangent activation function is used. Train the parameter matrix for the SAM module. The bias training parameters for the SAM module. To train the parameter vector for the SAM module, h i It is a bidirectional feature sequence. For the first Attention weights for each feature The intermediate mapping features of the SAM module are used to represent the bidirectional feature sequence h. i The result after linear transformation For activation function, For mapping parameters, This is the output gate coefficient.
[0114] BiLSTM module: A bidirectional long short-term memory network layer that processes time-series data in both forward and backward directions, comprehensively capturing the contextual dependencies of time series data. The operating principle of this module is as follows: ; ; ;
[0115] in, , The forward LSTM at time step , The hidden state is used to capture forward temporal dependencies; , The backward LSTM at time steps , The hidden state is used to capture backward temporal dependencies; for Time-time feature vector; This is a bidirectional long short-term memory network layer operation function used to extract temporally dependent features; For bidirectional LSTM at time steps The final output state.
[0116] S7. Train the CNN-BiLSTM-SAM hybrid deep learning model. During the training process, a multi-level regularization system including L2 regularization, Dropout mechanism, early stopping strategy and data augmentation is introduced to suppress overfitting and improve the model's generalization ability, so as to obtain the regularized CNN-BiLSTM-SAM hybrid deep learning model.
[0117] The L2 regularization operation process is as follows:
[0118] Let the original loss function be... ,in For all trainable parameters of the model (including weights) and bias parameters The total loss function after introducing L2 regularization. It can be represented as: ;
[0119] in, This is the L2 regularization coefficient, used to control the intensity of the penalty; It is the sum of squares of all weight parameters.
[0120] From the perspective of gradient descent, the mechanism of L2 regularization is more intuitive. During each parameter update, the weights... The gradient includes not only the gradient of the original loss function but also the contribution of the regularization term. The formula for the weight gradient is: ;
[0121] in, The total loss function with respect to the weights gradient, The original loss function with respect to the weights The gradient.
[0122] The weight update formula is: ;
[0123] in, The learning rate for model training. For the updated weights, The weights before the update. The total loss function corresponding to the weights before the update. The total loss function before update with respect to the weights The gradient.
[0124] The Dropout mechanism works as follows:
[0125] During forward propagation, the Dropout layer randomly sets the output of some neurons to zero with probability p (retention probability) or drop rate q (drop rate, typically q = 1 - p). Neurons with their output set to zero do not participate in forward or backward propagation in the current iteration, effectively removing them from the network temporarily. In each iteration, the dropped neurons are randomly and independently selected, ensuring that the network structure is different for each training iteration.
[0126] From a mathematical perspective, the output of the Dropout layer for: ;
[0127] Where a(x) is the activation output of the previous layer network. is a binary mask vector of the same dimension as a(x), where each element independently follows a Bernoulli distribution (p); ⊙ is the element-wise multiplication operator.
[0128] During the testing phase, all neurons participate in the computation, but the weights need to be scaled. Typically, the training weights are multiplied by the retention probabilities. This is to keep the expected output consistent between the training and testing phases.
[0129] How data augmentation mechanisms work:
[0130] Noise injection: A small amount of Gaussian noise is added to the original voltage, current and temperature sequences to simulate random interference in actual measurements and enhance the model's robustness to noise.
[0131] Time scaling: Performs slight time-axis scaling transformations on time-series data to simulate behavioral changes at different charge and discharge rates.
[0132] Fragment sampling: Randomly sample subsequences from different starting points in a long sequence to increase the model's adaptability to data at different stages.
[0133] The comparison results of different regularization strategies are shown below:
[0134] Table 2: ;
[0135] S8. With the goal of minimizing the MAE of the test set, a genetic algorithm is used to globally optimize the hyperparameters of the regularized CNN-BiLSTM-SAM hybrid deep learning model to obtain the optimal hyperparameters.
[0136] Hyperparameters include the number of CNN convolutional kernels, the number of BiLSTM hidden units, the initial learning rate, the L2 regularization coefficient, the dropout ratio, and the batch size. The range of hyperparameter optimization variations is shown below:
[0137] Table 3: ;
[0138] S9. Retrain the regularized CNN-BiLSTM-SAM hybrid deep learning model using optimal hyperparameters. Utilize the trained model to jointly predict battery SOC and SOE, outputting multi-performance index estimation results, including predictions of SOC and SOE time-series changes within a single charge-discharge cycle, covering the complete aging process of the battery from initial cycle to end of life. Figures 4-10The diagrams show the prediction results for SOH, SOC, and SOE, the error results, and the prediction results before and after the introduction of the optimization algorithm.
[0139] During the battery's charge and discharge cycle, the first Under each cycle, the formulas for predicting SOC and SOE are as follows: ; ;
[0140] in, , , The battery number , , One sample data, For the first In the loop, the th The state of charge of the battery in each sample For the first In the loop, the th Battery energy state of a sample The value range is 0 to ; For the first In the loop, the first... Electricity data for each sample; For the first In the loop, the first... Current data for each sample; For the first In the loop, the first... Voltage data for each sample; For the first In the loop, the first... Energy data for each sample; For the battery in the first The maximum energy under each cycle The sampling time interval between two adjacent samples.
[0141] This invention has been described through embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of this invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, this invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of this invention.
Claims
1. A method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM, characterized in that, Includes the following steps: S1. Collect timing data during the battery's cyclic charging and discharging process; S2. Based on time-series data, the actual discharge capacity of the battery is calculated using the ampere-hour integration method, and then the battery SOH is calculated. S3. Extract multi-dimensional health characteristic parameters related to SOH during battery cycle charging and discharging; S4. Use KPCA to perform nonlinear dimensionality reduction on the multidimensional health feature parameters to obtain the dimensionality-reduced feature vector; S5. Integrate time series data, SOH, and dimensionality-reduced feature vectors to construct the input feature matrix; S6. Construct a CNN-BiLSTM-SAM hybrid deep learning model for battery state prediction using the input feature matrix; S7. Introduce a regularization system to train the CNN-BiLSTM-SAM hybrid deep learning model to obtain the regularized CNN-BiLSTM-SAM hybrid deep learning model. S8. Use a genetic algorithm to globally optimize the key hyperparameters of the regularized CNN-BiLSTM-SAM hybrid deep learning model to obtain the optimal hyperparameters. S9. Retrain the regularized CNN-BiLSTM-SAM hybrid deep learning model using the optimal hyperparameters, and use the trained model to predict the battery SOC and SOE, outputting the joint estimation results of multiple battery performance indicators.
2. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S1, the time series data includes voltage, current and temperature. The time series data is preprocessed and abnormal loop data is removed.
3. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S2, the formula for calculating the battery SOH is: ； in, The number of charge / discharge cycles for the battery. For the first Battery health status under one cycle For the battery in the first The actual maximum discharge capacity under one cycle This refers to the rated discharge capacity of the new battery.
4. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S3, the multi-dimensional health characteristic parameters include constant current charging time (CCCT), constant current discharging time (CCDT), constant voltage charging time (CVCT), constant voltage rising charging time (CVRT), and constant voltage falling discharging time (CVDT).
5. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S4, the nonlinear dimensionality reduction using KPCA is specifically performed as follows: set up This represents a set of data with zero mean. Given the total number of data samples, the kernel function maps... Implementation point arrive The mapping, the feature space center data is: ； in, For kernel function on the 1st The nonlinear mapping result of a data sample , This is the eigenvector space after nonlinear mapping; The covariance of the mapped data in the feature space is: ； in, Let covariance be the data mapped in the feature space. For kernel function on the 1st The nonlinear mapping result of a data sample; Diagonalize the covariance matrix through eigenvalue decomposition: ； in, These are the eigenvalues of the covariance matrix. These are the eigenvectors of the covariance matrix. ;get: ； in, These are the eigenvectors obtained by eigenvalue decomposition of the covariance matrix. The linear representation is: ； ； in, The constant coefficient, , For kernel function on the 1st , The nonlinear mapping result of a data sample , The values are 1, 2, ..., N; definition Kernel matrix for: ； The kernel matrix is corrected by centering: ； in, The kernel matrix after centering correction is the first... Line number Column elements, The first kernel in the original kernel matrix Line number Column elements, The first kernel in the original kernel matrix Line number Column elements, The first kernel in the original kernel matrix Line number The elements of the column.
6. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S5, the input features are: ； in, for Time-time feature vector for Voltage at any moment for Current at any moment for Temperature at all times This is the SOH value for the current cycle.
7. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S6, the model includes an input layer, a CNN module, a SAM module, a BiLSTM module, and an output layer; The operating principle of a CNN module is as follows: ； ； in, , The first Convolutional layers output features and input features. , The first Feature weights and bias coefficients after convolution, * denotes the convolution symbol. For activation function, To Max pooling result, A function to perform max pooling operations; The operating principle of the SAM module is as follows: ； ； ； ； ； in, For the first Attention energy value of each feature For time steps, The hyperbolic tangent activation function is used. Train the parameter matrix for the SAM module. The bias training parameters for the SAM module. To train the parameter vector for the SAM module, h i It is a bidirectional feature sequence. For the first Attention weights for each feature These are intermediate mapping features for the SAM module. For activation function, For mapping parameters, For output gate coefficients; The operating principle of the BiLSTM module is as follows: ； ； ； in, , The forward LSTM at time step , The hidden state, , The backward LSTM at time steps , The hidden state, In time step The input feature vector, This is a function for the bidirectional long short-term memory network layer. For bidirectional LSTM at time steps The final output state.
8. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S7, the L2 regularization operation is as follows: Let the original loss function be... ,in For all trainable parameters of the model, the total loss function after introducing L2 regularization is: ； in, The L2 regularization coefficient is... It is the sum of squares of all weight parameters; The formula for the weight gradient is: ； in, The total loss function with respect to the weights gradient, The original loss function with respect to the weights The gradient; The weight update formula is: ； Where η is the learning rate. For the updated weights, The weights before the update. The total loss function corresponding to the weights before the update. The total loss function before update with respect to the weights The gradient; Output of the Dropout layer for: ； Where a(x) is the activation output of the previous layer network. is a binary mask vector with the same dimension as a(x), and ⊙ is the element-wise multiplication operator; The data augmentation mechanism is as follows: Noise injection: A small amount of Gaussian noise is added to the original voltage, current, and temperature sequence; Time scaling: Performs slight time-axis scaling transformations on time-series data; Fragment sampling: Randomly sampling subsequences from different starting points from a long sequence.
9. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S8, the genetic algorithm aims to minimize the MAE of the test set. The hyperparameters include the number of CNN convolutional kernels, the number of BiLSTM hidden units, the initial learning rate, the L2 regularization coefficient, the Dropout ratio, and the batch size.
10. The method for joint estimation of multiple performance indicators of a battery system based on CNN-BiLSTM-SAM according to claim 1, characterized in that, In step S9, the formulas for predicting SOC and SOE are as follows: ； ； in, , , The battery number , , One sample data, For the first In the loop, the th The state of charge of the battery in each sample For the first In the loop, the th Battery energy state of a sample; For the first In the loop, the first... Electricity data for each sample; For the first In the loop, the first... Current data for each sample; For the first In the loop, the first... Voltage data for each sample; For the first In the loop, the first... Energy data for each sample; For the battery in the first The maximum energy under each cycle The sampling time interval between two adjacent samples.