SoH estimation method based on bayesian optimization
By constructing a CNN+LSTM model based on Bayesian optimization, performing dual feature extraction of the battery health feature matrix and fusion of multiple LSTM sub-models, the problems of high computational cost and insufficient model generalization ability in existing technologies are solved, and efficient and accurate SOH estimation is achieved.
Patent Information
- Application Number
- CN202610449174.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-03
AI Technical Summary
Existing SOH estimation methods rely on complex physical models or statistical analysis, which are computationally expensive and lack accuracy. Single deep learning models are difficult to cope with complex and ever-changing battery conditions and data characteristics, feature extraction is incomplete, and the model's generalization ability and parameter optimization efficiency are insufficient.
A CNN+LSTM model is constructed using a battery health feature matrix, and multiple iterations are performed using Bayesian optimization. A precise objective function is defined, and dual extraction of local features and long-term temporal features is performed. A multi-LSTM sub-model fusion structure is adopted, and the generalization ability of the model is improved through Bayesian optimization, and parameter configuration is intelligently selected.
It significantly reduces computational costs, decreases reliance on massive amounts of experimental data, improves the accuracy of SOH estimation results and the generalization ability of the model, optimizes efficiency, and reduces the bias of a single model.
Smart Images
Figure CN122330710A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data processing technology, specifically a Bayesian optimization-based SOH estimation method. Background Technology
[0002] Battery state of health (SOH) is a core indicator for evaluating battery performance, reflecting the degree of battery aging and remaining lifespan, and is of great significance for battery management and maintenance. In recent years, deep learning technologies (CNN, LSTM, etc.) have been gradually applied to the field of SOH estimation. These technologies have natural advantages in feature extraction and time series data processing, and have become an important research direction for SOH estimation.
[0003] However, existing SOH estimation methods suffer from several technical problems: they rely heavily on complex physical models or statistical analyses, require extensive experimental data for implementation, have high computational costs, and struggle to guarantee the accuracy of SOH estimation results. Furthermore, while deep learning models such as CNN and LSTM have been applied in this field, single models are insufficient to handle complex and variable battery conditions and data characteristics, exhibiting significant deficiencies in the comprehensiveness of feature extraction, the generalization ability of the model, and the efficiency of parameter optimization. Summary of the Invention
[0004] To address the aforementioned issues and overcome the shortcomings of existing technologies, this invention provides a Bayesian optimization-based SOH estimation method. Addressing the technical problem of insufficient accuracy in SOH estimation due to the reliance on complex physical models or statistical analyses, the need for extensive experimental data, and high computational costs, this invention constructs a CNN+LSTM model using a battery health feature matrix, defines a precise objective function, and employs multiple rounds of Bayesian optimization iteration to achieve intelligent parameter search, significantly reducing computational costs and dependence on massive amounts of experimental data. Furthermore, addressing the technical problem that while deep learning models such as CNN and LSTM have been applied in this field, single models struggle to cope with complex and variable battery conditions and data characteristics, exhibiting significant deficiencies in feature extraction comprehensiveness, model generalization ability, and parameter optimization efficiency, this invention uses a CNN+LSTM model for dual extraction of local and long-term temporal features, covering all dimensions of battery aging characteristics. A multi-LSTM sub-model fusion structure is used to calculate the final SOH estimation result, reducing the bias of a single model. Bayesian optimization enhances the model's generalization ability, intelligently selecting the parameter configuration most likely to improve performance based on the expected improvement value, avoiding ineffective training, and improving optimization efficiency.
[0005] The technical solution adopted by this invention is as follows: The SOH estimation method based on Bayesian optimization provided by this invention includes the following steps:
[0006] Step S1: Obtain the battery dataset;
[0007] Step S2: Construct the battery health feature matrix;
[0008] Step S3: Initialize Bayesian optimization;
[0009] Step S4: Multi-round Bayesian optimization iteration;
[0010] Step S5: Determine the convergence condition;
[0011] Step S6: Full training of the optimal configuration model.
[0012] Furthermore, in step S1, obtaining the battery dataset specifically involves selecting an experimental battery dataset and filtering out battery health characteristics that are strongly correlated with battery aging from the data.
[0013] Furthermore, in step S2, constructing the battery health feature matrix specifically involves constructing the battery health feature matrix by arranging the extracted battery health features in chronological order.
[0014] Further, in step S3, the initialization of Bayesian optimization includes the following steps:
[0015] Define the set of hyperparameters to be optimized, determine the hyperparameters to be optimized for CNN and LSTM neural networks, and delineate the search space for Bayesian optimization;
[0016] Initial parameter sampling and training: An active learning strategy is adopted, and n sets of initial parameter configurations are randomly sampled. For each set of initial parameter configurations, a CNN+LSTM model is trained based on the battery health feature matrix.
[0017] Determine the initial optimal value, use the mean squared error of the validation set as the evaluation index, calculate the model error for each set of initial parameters, and select the parameter corresponding to the minimum MSE as the initial current optimal parameter. The MSE corresponding to the initial current optimal parameter is the first objective mean squared error. ;
[0018] The objective function is calculated using the mean squared error, resulting in an objective function with the mean squared error as its core.
[0019] Further, in step S4, the multi-round Bayesian optimization iteration includes the following steps:
[0020] A Gaussian process proxy model and a desired improvement acquisition model are built to obtain the desired improvement value of the parameter configuration;
[0021] The target parameter configuration is selected, and the target parameter configuration for this round is determined by maximizing the expected improvement value: with the hyperparameter search space X1 as the constraint range, the L-BFGS gradient optimizer is used to solve for the expected improvement value. Maximize parameter configuration;
[0022] Configuration based on target parameters Train the CNN+LSTM model, and use the validation set data to calculate the mean squared error (MSE) of the model, denoted as the second objective MSE. If the second objective MSE is less than the first objective MSE, then update the current optimal parameters to... At the same time Update to the second objective mean square error; if the second objective mean square error is greater than or equal to the first objective mean square error, then keep the current optimal parameters unchanged.
[0023] Repeat the above steps to complete the parameter optimization and update for each iteration.
[0024] Further, in step S5, determining the convergence condition includes the following steps:
[0025] After each Bayesian optimization iteration, check whether any convergence condition is met:
[0026] (1) The mean square error of the second objective in N consecutive iterations is less than the error fluctuation threshold;
[0027] (2) The total number of iterations in Bayesian optimization is greater than or equal to the maximum number of iterations threshold;
[0028] (3) The calculated maximum expected improvement value is less than or equal to the expected improvement value threshold;
[0029] If any of the above convergence conditions are met, the Bayesian optimization iteration stops, and the SOH estimated optimal configuration model is generated using the final current optimal parameters as the standard.
[0030] Further, in step S6, the full training of the optimal configuration model includes the following steps:
[0031] The optimal configuration model for SOH estimation is built based on the optimal parameters. The CNN layer constructs a feature extraction layer according to the optimal number and size of convolutional kernels, performs convolution operations, and outputs local time features of battery aging. The LSTM layer constructs a time series modeling layer according to the optimal number of hidden layer units and layers, receives the local time features of battery aging and captures long-term dependencies through a gate control mechanism, and adopts a multi-LSTM sub-model fusion structure to calculate the final SOH estimation result.
[0032] The entire battery dataset is input into SOH to estimate the optimal configuration model, and training is performed using the optimal learning rate, Dropout rate, and batch size.
[0033] The test set is input into the trained SOH estimation optimal configuration model, and the model performance is evaluated from four dimensions: mean absolute error, mean square error, coefficient of determination, and inference time. After the evaluation is passed, the SOH estimation optimal configuration model is deployed in the actual battery management system for real-time estimation of battery SOH.
[0034] The beneficial results achieved by the present invention using the above solution are as follows:
[0035] (1) To address the technical problem that the implementation process requires a large amount of experimental data to support the complex physical model or statistical analysis with multiple dependencies, resulting in high computational costs and insufficient accuracy of SOH estimation results, a CNN+LSTM model is constructed using the battery health feature matrix, a precise objective function is defined, and multiple rounds of Bayesian optimization iteration are used to achieve intelligent parameter search, which greatly reduces computational costs and reduces dependence on massive experimental data.
[0036] (2) To address the technical problem that although deep learning models such as CNN and LSTM have been applied in this field, a single model is difficult to cope with the complex and ever-changing battery conditions and data characteristics. There are obvious deficiencies in the comprehensiveness of feature extraction, the generalization ability of the model, and the efficiency of parameter optimization. The CNN+LSTM model is adopted to extract both local features and long-term time-series features, covering the full-dimensional features of battery aging. The final SOH estimation result is calculated by using a multi-LSTM sub-model fusion structure to reduce the bias of a single model. Bayesian optimization is used to improve the generalization ability of the model. The parameter configuration that is most likely to improve performance is intelligently selected by the expected improvement value to avoid ineffective training and improve optimization efficiency. Attached Figure Description
[0037] Figure 1 This is a flowchart illustrating the SOH estimation method based on Bayesian optimization provided by the present invention.
[0038] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation
[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0040] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0041] Example 1, see Figure 1 The present invention provides a Bayesian optimization-based SOH estimation method, which includes the following steps:
[0042] Step S1: Obtain the battery dataset, specifically by selecting the experimental battery dataset and extracting battery health features;
[0043] Step S2: Construct a battery health feature matrix, specifically by constructing a battery health feature matrix according to the extracted battery health features in chronological order;
[0044] Step S3: Initialize Bayesian optimization, specifically by defining the set of hyperparameters to be optimized, training a CNN+LSTM model based on the battery health feature matrix for each set of initial parameter configurations, determining the initial optimal parameters and the first objective mean square error, and using the mean square error to calculate the objective function;
[0045] Step S4: Multi-round Bayesian optimization iteration, specifically, building a Gaussian process surrogate model and an expected improvement acquisition model to obtain the expected improvement value of the parameter configuration, determining the target parameter configuration for this round by maximizing the expected improvement value, and obtaining the current optimal parameters;
[0046] Step S5: Determine the convergence condition. Specifically, if the convergence condition is met, stop the Bayesian optimization iteration and generate the SOH estimate optimal configuration model based on the final current optimal parameters.
[0047] Step S6: Full training of the optimal configuration model, specifically, calculating the final SOH estimation result based on the SOH estimation optimal configuration model.
[0048] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, obtaining the battery dataset specifically involves selecting an experimental battery dataset (University of Maryland battery dataset) and filtering out 7 battery health features that are strongly correlated with battery aging from the data: constant current discharge time, constant voltage charging time, constant current charging time, incremental capacity peak value, voltage corresponding to the incremental capacity peak value, internal discharge time, and internal charging time, which provides a basis for subsequent feature modeling.
[0049] Example 3, see Figure 1This embodiment is based on the above embodiment. In step S2, the construction of the battery health feature matrix specifically involves constructing the battery health feature matrix by arranging the extracted battery health features in chronological order, thus transforming the unstructured battery data into a structured input that can be directly processed by the CNN+LSTM model. The formula used is as follows:
[0050] ;
[0051] In the formula, X represents the battery health feature matrix. This represents the health feature vector at the corresponding time point, where T represents the time step of battery aging. Represents the space of real numbers.
[0052] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the initialization of Bayesian optimization includes the following steps:
[0053] Define a set of hyperparameters to be optimized, determine the hyperparameters to be optimized for CNN and LSTM neural networks, and delineate the search space for Bayesian optimization. The hyperparameters specifically include the number of convolutional kernels and the kernel size of the CNN model, the number of hidden layer units of the LSTM model, and general training parameters, including the network learning rate, Dropout rate, and batch size.
[0054] Initial parameter sampling and training employ an active learning strategy, randomly sampling n sets of initial parameter configurations. For each set of initial parameter configurations, a CNN+LSTM model is trained based on the battery health feature matrix.
[0055] Determine the initial optimal value, use the mean squared error of the validation set as the evaluation index, calculate the model error for each set of initial parameters, and select the parameter corresponding to the minimum MSE as the initial current optimal parameter. The MSE corresponding to the initial current optimal parameter is the first objective mean squared error. ;
[0056] The objective function is calculated using the mean squared error, resulting in an objective function with the mean squared error as its core. The formula used is as follows:
[0057] ;
[0058] In the formula, Indicates parameter configuration The corresponding objective function is based on the mean squared error. Indicates the number of samples in the validation set. This represents the true SOH value of the i-th sample. Indicates parameter configuration The predicted SOH value of the model.
[0059] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the multi-round Bayesian optimization iteration includes the following steps:
[0060] A Gaussian process proxy model and a desired improvement acquisition model are built to obtain the desired improvement value of the parameter configuration;
[0061] The formula for the Gaussian process proxy model is as follows:
[0062] ;
[0063] In the formula, Represents the Gaussian distribution function. Represents the mean function, Represents the covariance function. This indicates a comparison of parameter configurations;
[0064] The formula for the desired improvement in the data acquisition model is as follows:
[0065] ;
[0066] In the formula, Represents the mathematical expectation. This represents the mean square error of the first objective. Indicates parameter configuration Expected improvement value;
[0067] The target parameter configuration is selected, and the target parameter configuration for this round is determined by maximizing the expected improvement value: with the hyperparameter search space X1 as the constraint range, the L-BFGS gradient optimizer is used to solve for the parameter configuration that maximizes the expected improvement value, and the formula used is as follows:
[0068] ;
[0069] In the formula, For the target parameter configuration of this iteration, X1 is the hyperparameter search space;
[0070] Configuration based on target parameters Train the CNN+LSTM model, and use the validation set data to calculate the mean squared error (MSE) of the model, denoted as the second objective MSE. If the second objective MSE is less than the first objective MSE, then update the current optimal parameters to... At the same time Update to the second objective mean square error; if the second objective mean square error is greater than or equal to the first objective mean square error, then keep the current optimal parameters unchanged.
[0071] Repeat the above steps to complete the parameter optimization and update for each iteration.
[0072] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, determining the convergence condition includes the following steps:
[0073] After each Bayesian optimization iteration, check whether any convergence condition is met:
[0074] (1) The mean square error of the second objective in N consecutive iterations is less than the error fluctuation threshold;
[0075] (2) The total number of iterations in Bayesian optimization is greater than or equal to the maximum number of iterations threshold;
[0076] (3) The calculated maximum expected improvement value is less than or equal to the expected improvement value threshold;
[0077] If any of the above convergence conditions are met, the Bayesian optimization iteration stops, and the SOH estimation optimal configuration model is generated using the final current optimal parameters as the standard.
[0078] The optimal configuration model for SOH estimation is still a CNN+LSTM fusion structure, where the convolution operation formula of the CNN model is: ,in, Here, k represents the convolution kernel, k represents the kernel length, and d represents the number of output channels. This represents a one-dimensional convolution operation. This represents the bias term of the CNN convolutional layer. This represents the activation function. This represents the CNN feature vector at time t. Represents the battery health feature vector at time t;
[0079] The formula for the local temporal feature matrix extracted by the CNN model is as follows: ,in, This represents the local time characteristics of battery aging and serves as the input to the LSTM layer.
[0080] By performing the above operations, a CNN+LSTM model is constructed using the battery health feature matrix, a precise objective function is defined, and multiple rounds of Bayesian optimization iteration are used to achieve intelligent parameter search, which significantly reduces computational costs and dependence on massive experimental data. This solves the technical problem that the existence of complex physical models or statistical analyses with multiple dependencies requires a large amount of experimental data to support the implementation process, resulting in high computational costs and insufficient accuracy of SOH estimation results.
[0081] Example 7, see Figure 1 This embodiment is based on the above embodiment. In step S6, the full training of the optimal configuration model includes the following steps:
[0082] Based on the optimal parameters, a SOH estimation optimal configuration model is built. The CNN layer is used to construct a feature extraction layer with the optimal number and size of convolutional kernels, and convolution operations are performed to output local temporal features of battery aging. The LSTM layer is constructed using the optimal number of hidden layer units and layers to create a time-series modeling layer, receiving local time features of battery aging. Long-term dependencies are captured through a gate control mechanism, and a multi-LSTM sub-model fusion structure is used to calculate the final SOH estimation result. The formula used is as follows:
[0083] ;
[0084] In the formula, K represents the number of LSTM sub-models. This represents the weighting coefficient of the k-th sub-model. This represents the output of the k-th LSTM sub-model at time t. This represents the fused SOH estimate at time t;
[0085] Input the entire battery dataset (training set + validation set) into the optimal configuration model, and perform training (including forward propagation and back propagation) using the optimal learning rate, Dropout rate, and batch size.
[0086] The test set is input into the trained model, and the model performance is evaluated from four dimensions: mean absolute error, mean squared error, coefficient of determination, and inference time. Once the evaluation is satisfactory, the model is deployed in the actual battery management system for real-time estimation of battery state of harm (SOH).
[0087] By performing the above operations, a CNN+LSTM model is used to extract both local and long-term temporal features, covering all dimensions of battery aging. A multi-LSTM sub-model fusion structure is used to calculate the final SOH estimation result, reducing the bias of a single model. Bayesian optimization is used to improve the model's generalization ability. The expected improvement value is used to intelligently select the parameter configuration most likely to improve performance, avoiding ineffective training and improving optimization efficiency. This solves the technical problem that although deep learning models such as CNN and LSTM have been applied in this field, a single model is difficult to cope with the complex and ever-changing battery conditions and data characteristics, and there are obvious deficiencies in the comprehensiveness of feature extraction, the generalization ability of the model, and the efficiency of parameter optimization.
[0088] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
[0090] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A SOH estimation method based on Bayesian optimization, characterized in that: The method includes the following steps: Step S1: Obtain the battery dataset, specifically by selecting the experimental battery dataset and extracting battery health features; Step S2: Construct a battery health feature matrix, specifically by constructing a battery health feature matrix according to the extracted battery health features in chronological order; Step S3: Initialize Bayesian optimization, specifically by defining the set of hyperparameters to be optimized, training a CNN+LSTM model based on the battery health feature matrix for each set of initial parameter configurations, determining the initial optimal parameters and the first objective mean square error, and using the mean square error to calculate the objective function; Step S4: Multi-round Bayesian optimization iteration, specifically, building a Gaussian process surrogate model and an expected improvement acquisition model to obtain the expected improvement value of the parameter configuration, determining the target parameter configuration for this round by maximizing the expected improvement value, and obtaining the current optimal parameters; Step S5: Determine the convergence condition. Specifically, if the convergence condition is met, stop the Bayesian optimization iteration and generate the SOH estimate optimal configuration model based on the final current optimal parameters. Step S6: Full training of the optimal configuration model, specifically, calculating the final SOH estimation result based on the SOH estimation optimal configuration model.
2. The SOH estimation method based on Bayesian optimization according to claim 1, characterized in that: In step S3, the initialization of Bayesian optimization includes the following steps: Define the set of hyperparameters to be optimized, determine the hyperparameters to be optimized for CNN and LSTM neural networks, and delineate the search space for Bayesian optimization; Initial parameter sampling and training: An active learning strategy is adopted, and n sets of initial parameter configurations are randomly sampled. For each set of initial parameter configurations, a CNN+LSTM model is trained based on the battery health feature matrix. Determine the initial optimal value, use the mean squared error of the validation set as the evaluation index, calculate the model error for each set of initial parameters, and select the parameter corresponding to the minimum MSE as the initial current optimal parameter. The MSE corresponding to the initial current optimal parameter is the first objective mean squared error. ; The objective function is calculated using the mean squared error, resulting in an objective function with the mean squared error as its core.
3. The SOH estimation method based on Bayesian optimization according to claim 1, characterized in that: In step S4, the multi-round Bayesian optimization iteration includes the following steps: A Gaussian process proxy model and a desired improvement acquisition model are built to obtain the desired improvement value of the parameter configuration; Select the target parameter configuration and determine the target parameter configuration for this round by maximizing the expected improvement value: with the hyperparameter search space X1 as the constraint range, use the L-BFGS gradient optimizer to solve for the parameter configuration that maximizes the expected improvement value; Configuration based on target parameters Train the CNN+LSTM model, and use the validation set data to calculate the mean squared error (MSE) of the model, denoted as the second objective MSE. If the second objective MSE is less than the first objective MSE, then update the current optimal parameters to... At the same time Update to the second objective mean square error; if the second objective mean square error is greater than or equal to the first objective mean square error, then keep the current optimal parameters unchanged. Repeat the above steps to complete the parameter optimization and update for each iteration.
4. The SOH estimation method based on Bayesian optimization according to claim 1, characterized in that: In step S5, determining the convergence condition includes the following steps: After each Bayesian optimization iteration, check whether any convergence condition is met: (1) The mean square error of the second objective in N consecutive iterations is less than the error fluctuation threshold; (2) The total number of iterations in Bayesian optimization is greater than or equal to the maximum number of iterations threshold; (3) The calculated maximum expected improvement value is less than or equal to the expected improvement value threshold; If any of the above convergence conditions are met, the Bayesian optimization iteration stops, and the SOH estimated optimal configuration model is generated using the final current optimal parameters as the standard.
5. The SOH estimation method based on Bayesian optimization according to claim 1, characterized in that: In step S6, the full training of the optimal configuration model includes the following steps: The final SOH estimation optimal configuration model is built based on the optimal parameters. The CNN layer is constructed with the optimal number and size of convolutional kernels to build a feature extraction layer, perform convolution operations, and output the local time features of battery aging. The LSTM layer is constructed with the optimal number of hidden layer units and layers to build a time series modeling layer, receive the local time features of battery aging and capture long-term dependencies through a gate control mechanism, and adopt a multi-LSTM sub-model fusion structure to calculate the final SOH estimation result. The entire battery dataset is input into SOH to estimate the optimal configuration model, and training is performed using the optimal learning rate, Dropout rate, and batch size. The test set is input into the trained SOH estimation optimal configuration model, and the model performance is evaluated from four dimensions: mean absolute error, mean square error, coefficient of determination, and inference time. After the evaluation is passed, the SOH estimation optimal configuration model is deployed in the actual battery management system for real-time estimation of battery SOH.
6. The SOH estimation method based on Bayesian optimization according to claim 1, characterized in that: In step S1, obtaining the battery dataset specifically involves selecting an experimental battery dataset and filtering out battery health characteristics that are strongly correlated with battery aging from the data.
7. The SOH estimation method based on Bayesian optimization according to claim 1, characterized in that: In step S2, constructing the battery health feature matrix specifically involves constructing the battery health feature matrix by arranging the extracted battery health features in chronological order.