A Battery Capacity Estimation Method and System Based on Domain Adversarial Network
By building a feature extractor and regression predictor of the domain adversarial network, the problem of insufficient accuracy and generalization capabilities of battery capacity estimation is solved, and high-precision battery capacity prediction under label-free data is achieved.
Patent Information
- Application Number
- CN202211639361.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-20
- Publication Date
- 2025-07-08
- Estimated Expiration
- 2042-12-20
AI Technical Summary
The existing battery capacity estimation methods have problems with insufficient accuracy and poor generalization capabilities, especially in the case of no label data, and traditional methods cannot effectively improve the accuracy of battery capacity estimation.
Using a domain adversarial network method, by constructing a feature extractor, regression predictor and domain classifier, characterization features are automatically extracted and battery capacity estimation is carried out to form a domain adversarial network, which improves the accuracy and generalization ability of battery capacity estimation.
It realizes high-precision prediction of battery capacity without label data, improves the accuracy and applicability of battery capacity estimation, and is suitable for battery data from different sources.
Smart Images
Figure CN116224101B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of battery detection, and particularly to a battery capacity estimation method and system based on a domain adversarial network. Background Art
[0002] Batteries have been widely used in fields such as electronic devices, new energy vehicles, and energy storage. Due to the complex mechanism of internal aging reactions in batteries and the significant influence of external environments and operating conditions, accurately estimating the state of health (SOH) of batteries is a difficult problem in battery management.
[0003] The state of health of a battery is generally characterized by capacity decay. Therefore, it is necessary to estimate the battery capacity. Traditional battery capacity estimation methods include model-based prediction methods and big data-based prediction methods. However, traditional methods have the following disadvantages:
[0004] 1. Manually extracting features from the collected battery data (such as voltage, current, temperature, etc.), such as constant current charging duration, maximum value, minimum value, mean value, etc., cannot guarantee the final accuracy of the model, and the prediction effects for battery data from different sources will vary, with poor generalization ability; 2. Using supervised learning requires labeled battery data for training. For unlabeled battery data, due to the distribution differences between the training data and the prediction data, the accuracy is poor; 3. Only estimating the battery capacity based on unlabeled battery data is not applicable to the transfer learning method of pre-training + fine-tuning, and the accuracy based on the maximum mean discrepancy (MMD) for matching the marginal distribution of features is insufficient.
[0005] Therefore, how to provide a battery capacity estimation method and system based on a domain adversarial network to improve the accuracy and generalization ability of battery capacity estimation has become an urgent technical problem to be solved. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to provide a battery capacity estimation method and system based on a domain adversarial network to improve the accuracy and generalization ability of battery capacity estimation.
[0007] In a first aspect, the present invention provides a battery capacity estimation method based on a domain adversarial network, including the following steps:
[0008] Step S10: Obtain source domain data and target domain data of the battery;
[0009] Step S20: Mix and splice the source domain data and the target domain data and shuffle them to obtain mixed data;
[0010] Step S30: Label and segment the mixed data to obtain a battery data set;
[0011] Step S40: Construct a feature extractor, and use the feature extractor to extract characterization features from the battery dataset;
[0012] Step S50: Construct a domain adversarial network including a regression predictor and a domain classifier; use the regression predictor and the characterization features to estimate the battery capacity and obtain a predicted value of the battery capacity; use the domain classifier to perform domain classification on each characterization feature.
[0013] Further, in the step S10, the source domain data at least includes voltage, current, temperature, and a label, where the label is the battery capacity; the target domain data at least includes voltage, current, and temperature.
[0014] Further, the step S20 is specifically as follows:
[0015] Perform alternating mixed splicing and shuffling on each piece of source domain data and target domain data by row to obtain mixed data;
[0016] The step S30 is specifically as follows:
[0017] Perform annotation of the data domain on the mixed data, where the data domain is source domain data or target domain data;
[0018] Segment the annotated mixed data into battery data of equal length according to the time series, fill the battery data with insufficient length with 0, and construct a battery dataset based on each piece of battery data.
[0019] Further, in the step S40, the feature extractor is constructed based on Conv2D, BatchNorm, Relu, and MaxPool, and is used to perform 5 times of non-linear transformation on the battery data in the battery dataset, and flatten the result of the non-linear transformation to extract characterization features.
[0020] Further, in the step S50, the regression predictor is constructed based on Linear and Relu, and is used to perform 2 times of non-linear transformation on the characterization features and then convert them into one-dimensional data, and use the one-dimensional data as the predicted value of the battery capacity;
[0021] The domain classifier is constructed based on Linear, Relu, and BatchNorm, and is used to perform 3 times of non-linear transformation on the characterization features and then convert them into two-dimensional data, and perform domain classification on each characterization feature based on the two-dimensional data to judge the data source of the characterization features.
[0022] In a second aspect, the present invention provides a battery capacity estimation system based on a domain adversarial network, including the following modules:
[0023] A data acquisition module, configured to acquire source domain data and target domain data of a battery;
[0024] A data mixing and splicing module, configured to mix and splice the source domain data and the target domain data and shuffle them to obtain mixed data;
[0025] A battery dataset construction module, configured to label and segment the mixed data to obtain a battery dataset;
[0026] A characterization feature extraction module, configured to construct a feature extractor, and use the feature extractor to extract characterization features from the battery dataset;
[0027] A battery capacity estimation module, configured to construct a domain adversarial network including a regression predictor and a domain classifier; use the regression predictor and the characterization features to estimate the battery capacity to obtain a predicted value of the battery capacity; use the domain classifier to perform domain classification on each characterization feature.
[0028] Further, in the data acquisition module, the source domain data at least includes voltage, current, temperature, and a label, where the label is the battery capacity; the target domain data at least includes voltage, current, and temperature.
[0029] Further, the data mixing and splicing module is specifically configured to:
[0030] Perform alternating mixing and splicing of each of the source domain data and the target domain data by row and shuffle them to obtain mixed data;
[0031] The battery dataset construction module is specifically configured to:
[0032] Perform annotation of the data domain on the mixed data, where the data domain is source domain data or target domain data;
[0033] Segment the annotated mixed data into equal-length battery data according to the time series, fill the battery data with insufficient length with 0, and construct a battery dataset based on each battery data.
[0034] Further, in the characterization feature extraction module, the feature extractor is constructed based on Conv2D, BatchNorm, Relu, and MaxPool, and is configured to perform 5 times of non-linear transformation on the battery data of the battery dataset, and flatten the result of the non-linear transformation to extract characterization features.
[0035] Further, in the battery capacity estimation module, the regression predictor is constructed based on Linear and Relu, and is configured to perform 2 times of non-linear transformation on the characterization features and then convert them into one-dimensional data, and use the one-dimensional data as the predicted value of the battery capacity;
[0036] The domain classifier is constructed based on Linear, Relu, and BatchNorm, and is used to perform three non-linear transformations on the representation features and then convert them into two-dimensional data. Based on the two-dimensional data, domain classification is performed on each representation feature to determine the data source of the representation features.
[0037] The advantages of the present invention are as follows:
[0038] A feature extractor is constructed through Conv2D, BatchNorm, Relu, and MaxPool to extract representation features from the battery dataset, that is, a deep neural network is used to automatically extract representation features instead of traditional manual extraction, which has better accuracy and generalization ability; a regression predictor and a domain classifier are constructed to be used for battery capacity estimation and representation feature classification respectively, forming a domain adversarial network, so that the final feature extractor obtains information independent of the domain (data source), thereby improving the capacity prediction performance for target domain data, and does not require the target domain data to have labels, and is applicable to the capacity estimation of battery data with only unlabeled data, to improve the accuracy and generalization ability of battery capacity estimation. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The present invention will be further described below with reference to the accompanying drawings in conjunction with embodiments.
[0040] Figure 1 is a flowchart of a battery capacity estimation method based on a domain adversarial network according to the present invention.
[0041] Figure 2 is a schematic structural diagram of a battery capacity estimation system based on a domain adversarial network according to the present invention.
[0042] Figure 3 is a schematic flow diagram of the present invention.
[0043] Figure 4 is a schematic diagram of a feature extractor according to the present invention.
[0044] Figure 5 is a schematic diagram of a regression predictor according to the present invention.
[0045] Figure 6 is a schematic diagram of a domain classifier according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0046] The overall idea of the technical solution in the embodiments of this application is as follows: A feature extractor, a regression predictor, and a domain classifier are constructed through a deep neural network. The feature extractor automatically extracts representative features to improve accuracy and generalization ability; a domain adversarial network is formed through the regression predictor and the domain classifier, so that the final feature extractor obtains information independent of the domain (data source), thereby improving the capacity prediction performance for target domain data and being applicable to capacity estimation of battery data with only unlabeled data.
[0047] Please refer to Figures 1 to 6 As shown, a preferred embodiment of a battery capacity estimation method based on a domain adversarial network according to the present invention includes the following steps:
[0048] Step S10: Obtain source domain data and target domain data of the battery; the source domain data represents labeled data that can be used for training; the target domain data is the data for which battery capacity estimation is required.
[0049] Step S20: Mix and splice the source domain data and the target domain data and shuffle them to obtain mixed data.
[0050] Step S30: Label and segment the mixed data to obtain a battery data set.
[0051] Step S40: Construct a feature extractor, and use the feature extractor to extract representative features from the battery data set.
[0052] Step S50: Construct a domain adversarial network including a regression predictor and a domain classifier; use the regression predictor and the representative features to estimate the battery capacity to obtain a predicted value of the battery capacity; use the domain classifier to classify each representative feature by domain.
[0053] In step S10, the source domain data includes at least voltage, current, temperature, and a label, and the label is the battery capacity; the target domain data includes at least voltage, current, and temperature.
[0054] The specific content of step S20 is as follows:
[0055] Mix and splice each source domain data and target domain data alternately by row and shuffle them to obtain mixed data.
[0056] The specific content of step S30 is as follows:
[0057] Label the mixed data by data domain, and the data domain is source domain data or target domain data.
[0058] The labeled mixed data is segmented into battery data of equal length according to the time series, and the battery data with insufficient length is filled with 0s, and a battery data set is constructed based on each battery data. For example, the mixed data is segmented in units of 360 seconds, and the data with a length less than 360 seconds is filled with 0s.
[0059] In step S40, the feature extractor is constructed based on Conv2D, BatchNorm, Relu, and MaxPool, and is used to perform 5 non-linear transformations on the battery data in the battery data set, and flatten the results of the non-linear transformations to extract representative features.
[0060] The input data is time series data such as voltage, current, and temperature, which is transformed into high-dimensional representative features through mathematical operations. The input dimension is: [360, 3], and the output dimension is: [512, 1].
[0061] The purpose of Conv is to increase the dimension of the data while adding non-linear information, and at the same time use a 2D convolution kernel to extract the interaction information between channels for efficient feature extraction. The formula is as follows:
[0062]
[0063] Among them, H represents the length of the input matrix; W represents the width of the output matrix; j represents the step size of the sliding window;
[0064] BatchNorm is used to adjust the distribution of the output data of each layer of the network and reduce the possibility of gradient dispersion. The formula is as follows:
[0065]
[0066] Among them, represents the value after being standardized; x i represents the i-th element; μ β represents the average value of the vector; represents the variance of the vector; ε represents a very small value to prevent the denominator from being 0;
[0067] Relu is used to obtain a non-linear mapping relationship. The formula is as follows:
[0068] RELU(x) = max(0, f(x));
[0069] Among them, f(x) represents the output of BatchNorm of the previous network layer;
[0070] MaxPool is used for downsampling to filter redundant information. When the calculation step size is 1 and the convolution kernel is 2*2, the formula is as follows:
[0071]
[0072] Among them, m represents the length of the matrix; n represents the width of the matrix.
[0073] In the step S50, the regression predictor is constructed based on Linear and Relu, and is used to perform two non-linear transformations on the characterization features and then convert them into one-dimensional data, and use the one-dimensional data as the predicted value of the battery capacity;
[0074] The formula of Linear is: f(x) = w T x + b;
[0075] The domain classifier is constructed based on Linear, Relu, and BatchNorm, and is used to perform three non-linear transformations on the characterization features and then convert them into two-dimensional data, and perform domain classification on each characterization feature based on the two-dimensional data to determine the data source of the characterization features; the formula of the domain classifier is:
[0076] G d =(G f (x); u, z) = sigm(u T G f (x) + z);
[0077]
[0078] Among them, u T and z represent a set of parameters to be learned and optimized; G f (x) represents the output result of the feature extractor.
[0079] A preferred embodiment of the battery capacity estimation system based on the domain adversarial network of the present invention includes the following modules:
[0080] A data acquisition module, configured to acquire source domain data and target domain data of the battery; the source domain data represents labeled data available for training; the target domain data is the data for which battery capacity estimation is required;
[0081] A data mixing and splicing module, configured to mix and splice the source domain data and the target domain data and shuffle them to obtain mixed data;
[0082] A battery data set construction module, configured to label and segment the mixed data to obtain a battery data set;
[0083] A characterization feature extraction module, configured to construct a feature extractor, and use the feature extractor to extract characterization features from the battery data set;
[0084] The battery capacity estimation module is used to construct a domain adversarial network including a regression predictor and a domain classifier; estimate the battery capacity by using the regression predictor and the characterization features to obtain a predicted value of the battery capacity; classify the domain of each characterization feature by using the domain classifier.
[0085] In the data acquisition module, the source domain data at least includes voltage, current, temperature and a label, and the label is the battery capacity; the target domain data at least includes voltage, current and temperature.
[0086] The data hybrid splicing module is specifically used for:
[0087] Perform alternating hybrid splicing and shuffle of each source domain data and target domain data by row to obtain hybrid data;
[0088] The battery data set construction module is specifically used for:
[0089] Label the data domain of the hybrid data, and the data domain is source domain data or target domain data;
[0090] Divide the labeled hybrid data into equal-length battery data according to the time series, fill the battery data with insufficient length with 0, and construct a battery data set based on each battery data. For example, divide the hybrid data in units of 360 seconds, and fill the data with a length of less than 360 seconds with 0.
[0091] In the characterization feature extraction module, the feature extractor is constructed based on Conv2D, BatchNorm, Relu and MaxPool, and is used to perform 5 non-linear transformations on the battery data of the battery data set, and flatten the results of the non-linear transformations to extract characterization features.
[0092] The input data is time series data such as voltage, current, and temperature, which is converted into high-dimensional characterization features through mathematical operations. The input dimension is: [360, 3], and the output dimension is: [512, 1].
[0093] The purpose of Conv is to increase the dimension of the data while adding non-linear information, and at the same time use a 2D convolution kernel to extract the interaction information between channels for efficient feature extraction. The formula is as follows:
[0094]
[0095] Among them, H represents the length of the input matrix; W represents the width of the output matrix; j represents the step size of the sliding window;
[0096] BatchNorm is used to adjust the distribution of the output data of each layer of the network and reduce the possibility of gradient dispersion. The formula is as follows:
[0097]
[0098] Among them, represents the value after standardization; x i represents the i-th element; μ β represents the average value of the vector; represents the variance of the vector; ε represents an extremely small value to prevent the denominator from being zero;
[0099] Relu is used to obtain a non-linear mapping relationship, and the formula is as follows:
[0100] RELU(x) = max(0, f(x));
[0101] Among them, f(x) represents the output of the previous network layer BatchNorm;
[0102] MaxPool is used for downsampling to filter redundant information. When the calculation step size is 1 and the convolutional kernel is 2*2, the formula is as follows:
[0103]
[0104] Among them, m represents the length of the matrix; n represents the width of the matrix.
[0105] In the battery capacity estimation module, the regression predictor is constructed based on Linear and Relu, and is used to perform 2 non-linear transformations on the characterization features and then convert them into one-dimensional data, and use the one-dimensional data as the predicted value of the battery capacity;
[0106] The formula of Linear is: f(x) = w T x + b;
[0107] The domain classifier is constructed based on Linear, Relu, and BatchNorm, and is used to perform 3 non-linear transformations on the characterization features and then convert them into two-dimensional data, and perform domain classification on each characterization feature based on the two-dimensional data to determine the data source of the characterization features; the formula of the domain classifier is:
[0108] G d = (G f (x); u, z) = sigm(u T G f (x) + z);
[0109]
[0110] Among them, u T and z represent a set of parameters that need to be learned and optimized; G f(x) represents the output result of the feature extractor.
[0111] In summary, the advantages of the present invention are as follows:
[0112] The feature extractor is constructed by Conv2D, BatchNorm, Relu, and MaxPool to extract the characterization features from the battery dataset, that is, the deep neural network is used to automatically extract the characterization features to replace the traditional manual extraction, which has better accuracy and generalization ability; by constructing a regression predictor and a domain classifier for battery capacity estimation and characterization feature classification respectively, a domain adversarial network is formed, so that the final feature extractor obtains information independent of the domain (data source), thereby improving the capacity prediction performance for the target domain data, and no labels are required for the target domain data, which is applicable to the capacity estimation of battery data with only unlabeled data, and improves the accuracy and generalization ability of battery capacity estimation.
[0113] Although the specific implementation manners of the present invention have been described above, those skilled in the art of this technology should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be covered by the scope protected by the claims of the present invention.
Claims
1. A battery capacity estimation method based on a domain adversarial network, characterized in that: It includes the following steps: Step S10: Obtain the source domain data and target domain data of the battery; Step S20: Perform hybrid splicing and shuffle on the source domain data and target domain data to obtain hybrid data; Step S30: Label and segment the hybrid data to obtain a battery data set; Step S40: Construct a feature extractor, and use the feature extractor to extract characterization features from the battery data set; the feature extractor is constructed based on Conv2D, BatchNorm, Relu, and MaxPool, and is used to perform 5 times of non-linear transformation on the battery data in the battery data set, and flatten the results of the non-linear transformation to extract characterization features; Step S50: Construct a domain adversarial network including a regression predictor and a domain classifier; Use the regression predictor and the characterization features to estimate the battery capacity to obtain a predicted value of the battery capacity; use the domain classifier to perform domain classification on each characterization feature; The regression predictor is constructed based on Linear and Relu, and is used to perform 2 times of non-linear transformation on the characterization features and then convert them into one-dimensional data, and use the one-dimensional data as the predicted value of the battery capacity; The domain classifier is constructed based on Linear, Relu, and BatchNorm, and is used to perform 3 times of non-linear transformation on the characterization features and then convert them into two-dimensional data, and perform domain classification on each characterization feature based on the two-dimensional data to determine the data source of the characterization features.
2. The battery capacity estimation method based on the domain adversarial network according to claim 1, characterized in that: In the step S10, the source domain data at least includes voltage, current, temperature, and a label, and the label is the battery capacity; the target domain data at least includes voltage, current, and temperature.
3. The battery capacity estimation method based on a domain adversarial network according to claim 1, wherein: The step S20 is specifically: Perform alternating hybrid splicing and shuffle on each of the source domain data and target domain data by row to obtain hybrid data; The step S30 is specifically: Label the hybrid data in the data domain, and the data domain is the source domain data or the target domain data; Segment the labeled hybrid data into battery data of equal length according to the time series, fill the battery data with insufficient length with 0, and construct a battery data set based on each battery data.
4. A battery capacity estimation system based on a domain adversarial network, characterized in that: It includes the following modules: A data acquisition module, which is used to acquire the source domain data and target domain data of the battery; A data hybrid splicing module, which is used to perform hybrid splicing and shuffle on the source domain data and target domain data to obtain hybrid data; A battery data set construction module, which is used to label and segment the hybrid data to obtain a battery data set; A characterization feature extraction module, which is used to construct a feature extractor, and use the feature extractor to extract characterization features from the battery data set; the feature extractor is constructed based on Conv2D, BatchNorm, Relu, and MaxPool, and is used to perform 5 times of non-linear transformation on the battery data in the battery data set, and flatten the results of the non-linear transformation to extract characterization features; A battery capacity estimation module, which is used to construct a domain adversarial network including a regression predictor and a domain classifier; Estimate the battery capacity using the regression predictor and the characterization features to obtain a predicted value of the battery capacity; classify the domains of each characterization feature using the domain classifier; The regression predictor is constructed based on Linear and Relu, and is used to perform a second non-linear transformation on the characterization features and then convert them into one-dimensional data, and use the one-dimensional data as the predicted value of the battery capacity; The domain classifier is constructed based on Linear, Relu, and BatchNorm, and is used to perform a third non-linear transformation on the characterization features and then convert them into two-dimensional data, and classify the domains of each characterization feature based on the two-dimensional data to determine the data source of the characterization features.
5. The battery capacity estimation system based on a domain adversarial network according to claim 4, characterized in that: In the data acquisition module, the source domain data at least includes voltage, current, temperature, and a label, where the label is the battery capacity; the target domain data at least includes voltage, current, and temperature.
6. The battery capacity estimation system based on the domain adversarial network according to claim 4, characterized in that: The data mixing and splicing module is specifically used for: alternately mixing and splicing each source domain data and target domain data by row and shuffling to obtain mixed data; The battery dataset construction module is specifically used for: labeling the data domains of the mixed data, where the data domain is source domain data or target domain data; dividing the labeled mixed data into equal-length battery data according to the time series, padding the battery data with insufficient length with 0, and constructing a battery dataset based on each battery data.
Citation Information
Patent Citations
Method and device for predicting remaining service life of PEMFC
CN114814589A