A Health Status Estimation Method for Lithium-ion Batteries Based on Aging Multi-channel Feature Enhancement

Abstract

This invention relates to the field of battery monitoring technology and discloses a method for estimating the health status of lithium-ion batteries based on multi-channel aging feature enhancement. An aging feature enhancement module is constructed based on MCNN and cross-channel interactive aggregation attention. A battery feature model is then built based on this module and combined with a convolutional neural network. Charging data of the battery under test is acquired, and the charging data is preprocessed to obtain charging micro-product change features. The charging data and charging micro-product change features are input into the battery feature model. The battery feature model enhances the charging data and charging micro-product change features through cross-channel interactive aggregation attention to obtain charging depth features and charging micro-product change depth features. Based on these features, the battery health status is obtained, solving the problem that existing methods cannot extract effective battery aging features for battery prediction under different operating conditions and temperature conditions.

Description

Technical Field

[0001] This invention relates to the field of battery monitoring technology, and in particular to a method for estimating the health status of lithium-ion batteries based on aging multi-channel feature enhancement. Background Technology

[0002] The promotion and application of new energy storage systems can effectively mitigate the volatility caused by the grid connection of new energy power generation and promote the balance of power supply and load in the power system. As the core energy storage component of energy storage systems, lithium-ion batteries (mainly including lithium iron phosphate batteries and ternary lithium batteries) are widely used due to their advantages such as high energy density, low self-discharge rate, and long cycle life. During continuous system operation, lithium-ion batteries gradually age, and their remaining capacity decreases accordingly. The State of Health (SOH) index of a battery is defined as the ratio of the battery's maximum remaining usable capacity under current operating conditions to its rated capacity at the time of manufacture, reflecting the battery's current health status.

[0003] Currently, researchers primarily develop SOH estimation methods for lithium-ion batteries from the perspectives of laboratory research and practical online applications. In laboratory conditions, new batteries are cycled under fixed ambient temperatures and predefined protocols to study their aging behavior and capacity degradation. In practical online applications, researchers widely utilize cloud-acquired operational data to extract health indicators and design machine learning algorithms for SOH estimation. However, energy storage batteries may operate under various conditions, such as peak shaving, frequency control, and power control, and their charge and discharge behavior may rarely adhere to fixed charging protocols. Furthermore, the operating temperature conditions of energy storage batteries can vary significantly with environmental or operational conditions. These factors lead to significant differences in the aging trajectory of lithium-ion batteries under different temperatures and energy storage operating conditions, resulting in unclear aging characteristics in battery operating data and variations with the aging trajectory. This makes the extraction and modeling of health status features based on operating data extremely difficult, thus affecting accurate health status assessment and effective maintenance and management of battery systems.

[0004] Therefore, accurately extracting effective and robust battery aging characteristics based on the operating data of energy storage batteries under different operating conditions and temperatures, and integrating big data and deep learning models to achieve accurate estimation of the health status of energy storage lithium-ion batteries has become an urgent need in this field. Summary of the Invention

[0005] This invention provides a lithium-ion battery health status estimation method based on aging multi-channel feature enhancement, in order to solve the problem that existing methods cannot extract effective battery aging features for battery prediction under different operating conditions and temperature conditions.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] This invention provides a method for estimating the health status of lithium-ion batteries based on aging multi-channel feature enhancement, comprising the following steps:

[0008] Step 1: Construct an aging feature enhancement module based on MCNN (Multi-Column Convolutional Neural Network) and cross-channel interactive aggregation attention; and construct a battery feature model based on the aging feature enhancement module combined with the convolutional neural network.

[0009] Step 2: Obtain the charging data of the battery to be tested, preprocess the charging data to obtain the charging micro-product change features, input the charging data and the charging micro-product change features into the battery feature model, the battery feature model performs feature enhancement on the charging data and the charging micro-product change features through cross-channel interactive aggregation attention to obtain charging depth features and charging micro-product change depth features, and obtain the battery health status based on the charging depth features and charging micro-product change depth features.

[0010] The preprocessing consists of integration and difference operations.

[0011] This invention captures rich deep features from raw charging data using a multi-channel convolutional neural network, and then guides the neural network to enhance the features of the charging data using integral and differential data features, thus solving the problem of unclear aging features in raw charging data. Channel selection is performed on the multi-channel aging features, selecting the features most important for health status estimation and removing features irrelevant to health status estimation. This further eliminates the interference of redundant information on health status estimation under random charging conditions, improving the accuracy of health status estimation.

[0012] This invention effectively preprocesses the input charging data, obtaining the depth features of charging micro-product changes through integration and difference operations, thus providing a basis for feature extraction from another perspective. Furthermore, the charging depth features and the depth features of charging micro-product changes can more accurately represent the battery aging status.

[0013] The further described aging feature enhancement module based on MCNN and cross-channel interactive aggregated attention includes: constructing a multi-channel structure based on MCNN; constructing a cross-channel interactive aggregated attention structure for feature enhancement based on channel attention and temporal attention; and constructing the aging feature enhancement module by combining the multi-channel structure and the cross-channel interactive aggregated attention structure with batch normalization and max pooling. Multiple cascaded aging feature enhancement modules are used to extract aging multi-channel enhancement features with rich aging information.

[0014] Furthermore, the construction of the battery feature model based on the aging feature enhancement module combined with the convolutional neural network includes: connecting several stacked and connected aging feature enhancement modules, convolutional neural networks and stacked fully connected layers in series to construct the battery feature model.

[0015] Furthermore, in several stacked and connected aging feature enhancement modules, the input data of the first aging feature enhancement module is the charging data and the charging micro-product change feature, the input data of other subsequent aging feature enhancement modules are the charging depth feature and the charging micro-product change depth feature, and the output data of the aging feature enhancement module is the enhanced charging depth feature and the charging micro-product change depth feature.

[0016] In the multi-channel structure, one channel uses MCNN to process the input charging micro-product change features or the charging micro-product change depth features output by the previous aging feature enhancement module to obtain the first feature, and the other channel uses MCNN to process the input charging data or the charging depth features output by the previous aging feature enhancement module to obtain the second feature.

[0017] The cross-channel interactive aggregation attention structure performs a global average on the first feature to obtain a channel descriptor vector of the global average information of each channel of the first feature. Then, it uses one-dimensional convolution and sigmoid operation to adaptively weight the channel descriptor vector to obtain the third feature. The third feature is then transformed to the same time step as the second feature using linear transformation and sigmoid operation. Finally, it performs multiplication processing step by step to obtain the fourth feature.

[0018] The aging feature enhancement module performs max pooling and batch normalization on the third feature to obtain the enhanced charging depth feature.

[0019] The aging feature enhancement module performs max pooling and batch normalization on the fourth feature to obtain the enhanced micro-product change depth feature.

[0020] Furthermore, the convolutional neural network has more input channels than output channels, and channel selection is performed on the charging depth features and charging differential change depth features output by the stacked and cascaded aging feature enhancement modules. By employing a convolutional neural network with fewer output channels than input channels to perform channel selection on the multi-channel aging enhancement features, the channel features most relevant to SOH estimation are selected, ultimately achieving a more effective representation of battery aging.

[0021] Furthermore, the aging feature enhancement module also employs batch normalization and max pooling operations to perform convergence and dimensionality reduction processing on the enhanced charging depth features and charging micro-product change depth features.

[0022] Furthermore, the charging data includes the battery's total voltage, total current, temperature, and state of charge;

[0023] The charging differential change features and charging data are both input into the battery feature model in vector form.

[0024] Furthermore, the charging differential change characteristics include integral data and differential data. The integral data includes a charging capacity sequence and a charging energy sequence, and the differential data includes a first-order voltage difference sequence, a second-order voltage difference sequence, and a first-order temperature difference sequence.

[0025] Furthermore, the battery feature model uses the SmoothL1 loss function to calculate the loss value between the output label and the target label, and then uses the Adam optimization algorithm for backpropagation;

[0026] The output labels are obtained based on several stacked fully connected layers.

[0027] Furthermore, the method for obtaining the target tag includes: calculating the capacity corresponding to each charging segment using the reverse ampere-hour integration method, dividing it by the factory capacity to obtain a coarse estimate of the state of charge (SOH) for that charging segment; smoothing all the coarse estimates of SOH using a seasonal-trend decomposition algorithm, and using the smoothed SOH value as the target tag. By processing the charge and discharge segment data in historical data, multiple health indicators reflecting the battery aging mechanism are extracted from the energy storage battery data.

[0028] Beneficial effects:

[0029] The battery feature model designed in this invention can accurately extract effective and robust aging-related features from historical charging data, thereby achieving accurate health status estimation. This helps energy storage system operators to maintain and replace aging batteries in a timely manner, and has practical application value. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the network structure of the aging feature enhancement module in an embodiment of the present invention;

[0031] Figure 2 This is a schematic diagram of the network structure of the battery feature model according to an embodiment of the present invention;

[0032] exist Figure 1 and Figure 2 middle, This represents a multi-channel one-dimensional convolutional neural network. Indicates global average pooling. This represents a fully connected neural network. express function, The expression is multiplied at each time step. This represents max pooling and batch normalization;

[0033] Figure 3 This is a schematic diagram comparing the SOH estimation performance and error of 90% cross-validation on the actual operation dataset of the energy storage cabinet according to an embodiment of the present invention.

[0034] Figure 4 This is a bar chart showing the SOH estimation error on the 4-fold, 6-fold, and 9-fold cross-validation sets of the actual operation dataset of the energy storage cabinet in this embodiment of the invention. Detailed Implementation

[0035] The technical solution of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "connected" or "linked" and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up," "down," "left," "right," etc., are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship also changes accordingly.

[0037] This application provides a method for estimating the health status of lithium-ion batteries based on aging multi-channel feature enhancement, including the following steps:

[0038] Step 1: Construct an aging feature enhancement module based on MCNN and cross-channel interactive aggregation attention, and construct a battery feature model based on the aging feature enhancement module combined with a convolutional neural network;

[0039] Specifically, a multi-channel structure is constructed based on MCNN, a cross-channel interactive aggregation attention structure for feature enhancement is constructed based on channel attention and temporal attention, and an aging feature enhancement module is constructed based on the multi-channel structure and the cross-channel interactive aggregation attention structure combined with batch normalization and max pooling.

[0040] After constructing the aging feature enhancement module, several stacked and connected aging feature enhancement modules, convolutional neural networks, stacked fully connected layers, and blocks for performing batch normalization and max pooling operations are connected in series to construct the battery feature model.

[0041] In a series of stacked aging feature enhancement modules, the input data for the first aging feature enhancement module is charging data. Characteristics of changes in the differential product of charging The output data of the aging feature enhancement module is the enhanced charging depth feature. and depth characteristics of charging differential changes The input data for other subsequent aging feature enhancement modules is the charging depth feature output by the previous aging feature enhancement module. and depth characteristics of charging differential changes ;

[0042] The aging feature enhancement module first analyzes the input charging differential change features. Or the depth feature of the charge micro-product change output from the previous aging feature enhancement module. After processing using MCNN, the first feature x1 is obtained. This is applied to the input charging data. Or the charging depth feature from the previous aging feature enhancement module. After processing using MCNN, the second feature x2 is obtained.

[0043] In cross-channel interactive aggregation attention, the first feature calculated by MCNN will be used. Second feature Cross-channel interactive attention aggregation is performed. The process of cross-channel interactive attention aggregation can be described as follows: by applying the first feature... Perform a global average to obtain the first feature. The channel descriptor vectors of the global average information for each channel are adaptively weighted using one-dimensional convolution and sigmoid operations to obtain the first output third feature. The third feature is obtained by using linear transformation and sigmoid operation. Transformation to the second feature For the same time step, perform multiplication by each time step to obtain the fourth feature. The calculation of cross-channel interaction aggregation attention can be expressed as the following formula:

[0044]

[0045] in, for Total length of time For about The channel descriptor vector of the global average information for each channel; This is a one-dimensional convolution operation. It is the sigmoid function; This indicates point-by-point multiplication. This indicates the characteristics of the charging differential change after enhanced interaction; and These are weights and biases, respectively. The time step importance matrix is ​​represented by... pass After function processing and By performing time-point multiplication, the charging depth features after interaction enhancement are calculated. .

[0046] Finally, the aging feature enhancement module uses batch normalization and max pooling operations to... and After performing convergence and dimensionality reduction processing, the charging differential change depth feature is obtained after feature enhancement by the aging feature enhancement module. and charging depth characteristics .

[0047] The aging feature enhancement module has a large number of channels, some of which are redundant for SOH estimation. Therefore, in the battery feature model, a convolutional neural network with more input channels than output channels is used. After receiving the charging depth features and charging micro-product change depth features output from several stacked and cascaded aging feature enhancement modules, the convolutional neural network performs channel selection on the charging depth features and charging micro-product change depth features, selecting the channel most relevant to SOH estimation. The corresponding feature is... Finally, the charging depth features and charging micro-product change depth features enhanced by batch normalization and max pooling operations are converged and dimensionality reduced.

[0048] Here, the multi-channel convolutional neural network extracts discriminative deep features of the battery through different convolutional filters. It can effectively extract features from the original charging data, more effectively express the aging condition of the battery, and ultimately predict the health status of the battery.

[0049] Batch normalization eliminates internal variable shifting, accelerates training convergence, and reduces sensitivity to parameter initialization, while max pooling reduces the feature dimension of the data and enhances the learning of high-level abstract representations of the original input.

[0050] The battery feature model uses the SmoothL1 loss function to calculate the loss value between the output label and the target label, which can be expressed as the following formula:

[0051] ;

[0052] in, Indicates the loss value; Represented as target label; This represents the output label calculated by the battery feature model.

[0053] The output labels are obtained based on several stacked fully connected layers;

[0054] The method for obtaining the target label includes: calculating the capacity corresponding to each charging segment using the reverse ampere-hour integration method, dividing it by the factory capacity to obtain a rough estimate of the SOH corresponding to that charging segment; smoothing all the rough estimates of SOH using a seasonal-trend decomposition algorithm, and using the smoothed SOH value as the target label.

[0055] For training the battery feature model, the Adam optimization algorithm is used to perform backpropagation on the network parameters. During the optimization process, the initial learning rate is set to 0.01. It is stipulated that after 200 iterations, the learning rate will decay to half of the previous stage. The entire training is carried out for 2000 iterations, and the network parameters are fixed after training is completed.

[0056] Step 2: Obtain the charging data of the battery to be tested, preprocess the charging data to obtain the charging micro-product change features, input the charging data and the charging micro-product change features into the battery feature model, and the battery feature model performs feature enhancement on the charging data and the charging micro-product change features through cross-channel interactive aggregation of attention to obtain the charging depth features and the charging micro-product change depth features, and obtain the battery health status based on the charging depth features and the charging micro-product change depth features.

[0057] Specifically, the charging data during the battery charging process is acquired through the battery management system. The charging data includes the battery's total voltage, total current, temperature, and state of charge. After the charging data is acquired, preprocessing is performed using integral and differential operations to obtain the charging micro-product change characteristics. Specifically, this includes integral data and differential data. The integral data includes the charging capacity sequence and the charging energy sequence, while the differential data includes the first-order voltage difference sequence, the second-order voltage difference sequence, and the first-order temperature difference sequence. Finally, the charging micro-product change characteristics and the charging data are both input into the battery feature model in vector form.

[0058] The vector form of charging data is represented by the following formula:

[0059] ;

[0060] in, Represents charging data in vector form; Indicates the total battery voltage; Indicates the total current; Indicates average temperature; Indicates the state of charge; , , as well as These represent the total battery voltage, total current, average temperature, and state of charge in vector form, respectively. This indicates the dimension, which is 4 in this case; Indicates the length of time;

[0061] The charging capacity sequence in the integral data is calculated using the following formula:

[0062] ;

[0063] in, Represents a sequence of charging capacities; Indicates the first At that moment; Indicates the first The charging current at any given moment; Indicates the sampling time interval;

[0064] The charging energy sequence is calculated using the following formula:

[0065] ;

[0066] in, Indicates the charging energy sequence;

[0067] For differential data, the first-order voltage difference sequence, the second-order voltage difference sequence, and the first-order temperature difference sequence are calculated using the following three formulas:

[0068] ;

[0069] ;

[0070] ;

[0071] in, Represents a first-order voltage difference sequence; Represents a second-order difference sequence of voltages; This represents a first-order difference sequence of temperatures. and They represent the first Time and the The charging voltage; and They represent the first Time and the Temperature;

[0072] The concatenation of differential and integral data into a vector form is represented by the following formula:

[0073] ;

[0074] in, , , as well as These represent the vector forms of the charging capacity sequence, charging energy sequence, first-order voltage difference sequence, second-order voltage difference sequence, and first-order temperature difference sequence, respectively.

[0075] Finally, to verify the effectiveness of the present invention, the lithium-ion battery health state estimation method based on aging multi-channel feature enhancement provided by the present invention was applied to 36 energy storage cabinets, and the SOH was verified using 9-fold cross-validation. The results are as follows: Figure 3 , Figure 4 As shown, Figure 3 The estimated SOH values ​​are clustered near the red label values, with few outliers, indicating that the algorithm effectively found the mapping relationship between irregular data and SOH label values. Figure 4 The mean absolute error, mean mean square error, and mean relative error were as low as 1.59%, 2.09%, and 1.69%, respectively.

[0076] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for estimating the health status of lithium-ion batteries based on aging multi-channel feature enhancement, characterized in that, Includes the following steps: Step 1: Construct an aging feature enhancement module based on MCNN and cross-channel interactive aggregation attention, and construct a battery feature model based on the aging feature enhancement module combined with a convolutional neural network; The aging feature enhancement module based on MCNN and cross-channel interactive aggregated attention includes: constructing a multi-channel structure based on MCNN, constructing a cross-channel interactive aggregated attention structure for feature enhancement based on channel attention and temporal attention, and constructing an aging feature enhancement module based on the multi-channel structure and cross-channel interactive aggregated attention structure combined with batch normalization and max pooling. Step 2: Obtain the charging data of the battery to be tested, preprocess the charging data to obtain the charging micro-product change features, input the charging data and the charging micro-product change features into the battery feature model, the battery feature model performs feature enhancement on the charging data and the charging micro-product change features through cross-channel interactive aggregation attention to obtain charging depth features and charging micro-product change depth features, and obtain the battery health status based on the charging depth features and charging micro-product change depth features. The preprocessing includes integration and difference operations; The charging data includes the battery's total voltage, total current, temperature, and state of charge. Both the charging micro-product change characteristics and the charging data are input into the battery feature model in vector form. The charging micro-product change characteristics include integral data and differential data. The integral data includes a charging capacity sequence and a charging energy sequence. The differential data includes a first-order voltage difference sequence, a second-order voltage difference sequence, and a first-order temperature difference sequence. In the multi-channel structure, one channel uses MCNN to process the input charging micro-product change features or the charging micro-product change depth features output by the previous aging feature enhancement module to obtain the first feature, and the other channel uses MCNN to process the input charging data or the charging depth features output by the previous aging feature enhancement module to obtain the second feature. The cross-channel interactive aggregation attention structure performs a global average on the first feature to obtain a channel descriptor vector of the global average information of each channel of the first feature. Then, it uses one-dimensional convolution and sigmoid operation to adaptively weight the channel descriptor vector to obtain the third feature. The third feature is then transformed to the same time step as the second feature using linear transformation and sigmoid operation. Finally, it performs multiplication processing step by step to obtain the fourth feature. The aging feature enhancement module performs max pooling and batch normalization on the third feature to obtain the enhanced charging depth feature. The aging feature enhancement module performs max pooling and batch normalization on the fourth feature to obtain the enhanced micro-product change depth feature.

2. The lithium-ion battery health status estimation method based on aging multi-channel feature enhancement according to claim 1, characterized in that, The method of constructing a battery feature model based on aging feature enhancement modules combined with convolutional neural networks includes: connecting several stacked and connected aging feature enhancement modules, convolutional neural networks and stacked fully connected layers in series to construct a battery feature model.

3. The lithium-ion battery health status estimation method based on aging multi-channel feature enhancement according to claim 2, characterized in that, In a series of stacked aging feature enhancement modules, the input data of the first aging feature enhancement module is the charging data and the charging micro-product change feature, and the input data of the other subsequent aging feature enhancement modules are the charging depth feature and the charging micro-product change depth feature. The output data of the aging feature enhancement module are the enhanced charging depth feature and the charging micro-product change depth feature.

4. The lithium-ion battery health status estimation method based on aging multi-channel feature enhancement according to claim 3, characterized in that, The input channel of the convolutional neural network is larger than the output channel, and channel selection is performed on the charging depth feature and charging micro-product change depth feature output by the stacked and cascaded aging feature enhancement modules.

5. The lithium-ion battery health status estimation method based on aging multi-channel feature enhancement according to claim 3, characterized in that, The aging feature enhancement module also employs batch normalization and max pooling operations to perform convergence and dimensionality reduction processing on the enhanced charging depth features and charging micro-product change depth features.

6. The lithium-ion battery health status estimation method based on aging multi-channel feature enhancement according to claim 3, characterized in that, The battery feature model uses the SmoothL1 loss function to calculate the loss value between the output label and the target label, and then uses the Adam optimization algorithm for backpropagation. The output labels are obtained based on several stacked fully connected layers.

7. The lithium-ion battery health status estimation method based on aging multi-channel feature enhancement according to claim 6, characterized in that, The method for obtaining the target tag includes: calculating the capacity corresponding to each charging segment using the reverse ampere-hour integration method, dividing it by the factory capacity to obtain a rough estimate of the SOH corresponding to the charging segment; smoothing all the rough estimates of SOH using a seasonal-trend decomposition algorithm, and using the smoothed SOH value as the target tag.

Images