Lithium ion battery health state estimation method based on image recognition

Abstract

The invention discloses a lithium ion battery health state estimation method based on image recognition, and belongs to the field of image recognition and battery energy storage, and the method comprises the following steps: obtaining voltage data of a lithium ion battery in a constant-current and constant-voltage charging process, and generating an original two-dimensional image based on the voltage data; preprocessing the original two-dimensional image to obtain a single-channel gray feature map; inputting the feature map into a lightweight model fusing a residual network module and an SE network module to obtain a high-order feature related to the state of health of the battery; the high-order feature is mapped into a battery state of health estimate via a fully connected layer.

Description

Technical Field

[0001] This invention belongs to the fields of image recognition and battery energy storage technology, and particularly relates to a method for estimating the health status of lithium-ion batteries based on image recognition. Background Technology

[0002] Lithium-ion batteries have advantages such as large energy storage capacity, long cycle life, and low manufacturing cost. They are widely used in electric vehicles, grid energy storage, and other power systems.

[0003] A Battery Management System (BMS) continuously monitors the battery's internal resistance and cycle count to assess and track its health status over time. When battery performance degrades beyond a threshold, it provides diagnostics and alerts to the user or system controller for proactive maintenance or replacement. As batteries experience more usage cycles, battery life degradation exhibits non-linear, time-varying characteristics, making real-time and accurate assessment of battery state health difficult. Currently, existing technologies generally include measurement-based methods, model-based methods, and data-driven methods. Measurement-based methods for estimating battery state of health (SOH) can detect battery state in real time, providing direct and reliable state assessment results. This is mainly achieved by measuring the battery's electrical impedance tomography (EIS) and analyzing impedance spectra at different frequencies. However, measurement-based methods are affected by internal and external environmental factors, and accurate SOH estimation requires specialized equipment and sensors, increasing costs and demands. Model-based methods primarily include SOH estimation based on electrochemical models (EM) or equivalent circuit models (ECM). Electrochemical models are mathematical models that describe the battery's operation using inherent electrochemical principles, offering strong physical interpretability. However, they typically involve many equations and parameters, leading to high computational complexity. ECM-based models are easy to implement and can be applied to real-time SOH estimation. However, this model still contains some complex components, and these model parameters may drift as the battery ages, requiring higher sensitivity to ensure the accuracy and robustness of SOH estimation. Data-driven methods do not require knowledge of battery principles and models and can estimate battery degradation by establishing a correlation between capacity degradation and features extracted from measurement signals.

[0004] Data-driven methods for estimating state of health (SOH) typically involve cumbersome processes, such as data sampling and extracting features related to battery capacity. Furthermore, the choice of features can significantly impact the accuracy of the estimation. Therefore, this invention proposes an image recognition-based method for estimating the state of health of lithium-ion batteries. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes an image recognition-based method for estimating the health status of lithium-ion batteries, thereby resolving the issues present in the prior art.

[0006] To achieve the above objectives, this invention provides a method for estimating the health status of lithium-ion batteries based on image recognition, comprising:

[0007] Obtain voltage data during the constant current and constant voltage charging process of a lithium-ion battery, and generate an original two-dimensional image based on the voltage data;

[0008] The original two-dimensional image is preprocessed to obtain a single-channel grayscale feature map;

[0009] The feature map is input into a lightweight model that fuses the residual network module and the SE network module to obtain high-order features related to the battery health status.

[0010] The higher-order features are mapped to battery health state estimates through a fully connected layer.

[0011] Optionally, the process of preprocessing the original two-dimensional image to obtain a single-channel grayscale feature map includes:

[0012] Convert the original two-dimensional image into a grayscale image;

[0013] The grayscale image is scaled using the resize function, and the scaled image is then cropped from the center to obtain a fixed-size image.

[0014] The fixed-size image is subjected to image standardization processing to obtain a single-channel grayscale feature map.

[0015] Optionally, the process of inputting the feature map into a lightweight model that fuses the residual network module and the SE network module to obtain high-order features related to the battery health state includes:

[0016] The single-channel grayscale feature map is input into the residual network module for feature extraction to obtain preliminary feature information;

[0017] The preliminary feature information is input into the SE network module to obtain channel features related to battery health status;

[0018] The residual network module and the SE network module are alternately connected to form a hierarchical feature extraction structure, and high-order features related to the battery health status are obtained based on the hierarchical feature extraction structure.

[0019] Optionally, the residual network module includes: a two-dimensional convolutional layer, a max pooling layer, and a residual block; the residual block includes: a first convolutional sub-layer, a second convolutional sub-layer, and a skip connection layer.

[0020] Optionally, the expression for feature extraction based on the residual network module is:

[0021] y=F(x,{W i})+x

[0022] In the formula, x is the input layer, W i Let y be the weight layer of the i-th convolutional layer, y be the output layer, and F(.) be the intermediate features.

[0023] Optionally, the process of inputting the preliminary feature information into the SE network module to obtain channel features related to battery health status includes:

[0024] The preliminary feature information is compressed into one-dimensional data features with a global receptive field based on the Squeeze operation;

[0025] The channel weight coefficients are obtained from the one-dimensional data features based on the Excitation operation;

[0026] Based on the channel weight coefficients and preliminary feature information, channel features related to battery health status are obtained.

[0027] The present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a lithium-ion battery health status estimation method based on image recognition.

[0028] The present invention provides a storage medium storing a computer program that, when executed by a processor, implements a method for estimating the health status of a lithium-ion battery based on image recognition.

[0029] Compared with the prior art, the present invention has the following advantages and technical effects:

[0030] This invention presents an image recognition-based method for estimating the health status of lithium-ion batteries. It generates a raw two-dimensional image by acquiring voltage data during constant-current, constant-voltage charging and preprocesses it to obtain a single-channel grayscale feature map. This method inputs the feature map into a lightweight model that integrates a residual network module and an SE network module, effectively extracting high-order features related to the battery's health status. The residual network module alleviates the degradation problem during deep network training, enabling stable training while maintaining high network depth and capturing complex nonlinear relationships. The SE network module focuses on image features highly correlated with the battery's health status through an attention mechanism, reducing interference from irrelevant factors. Finally, the high-order features are mapped to an estimated battery health status value through a fully connected layer. This method eliminates the need for manual extraction of health features, simplifying the estimation process while maintaining high estimation accuracy and computational efficiency, providing an effective technical means for the rapid and accurate assessment of the health status of lithium-ion batteries. Attached Figure Description

[0031] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0032] Figure 1 This is a flowchart of the image recognition-based lithium-ion battery health status estimation method according to an embodiment of the present invention;

[0033] Figure 2 This is a flowchart of image preprocessing according to an embodiment of the present invention;

[0034] Figure 3 This refers to the sub-residual module in the SOH estimation model of this invention embodiment;

[0035] Figure 4 This refers to the sub-SE module in the SOH estimation model of this invention embodiment;

[0036] Figure 5 This is a structural diagram of the SOH estimation model according to an embodiment of the present invention;

[0037] Figure 6 This is a box plot of experimental error from an embodiment of the present invention. Detailed Implementation

[0038] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0039] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0040] Example 1

[0041] In the field of computer vision, useful details in images can be extracted and processed by recognizing them. Using computer vision for battery SOH estimation is a relatively new method that can greatly simplify the battery SOH estimation process.

[0042] The purpose of this invention is to automatically extract hidden features from images using image recognition and construct a correspondence between feature images and capacity degradation. No manual feature extraction is required, which simplifies the entire estimation process. Furthermore, the constructed lightweight image recognition model maintains high estimation accuracy while ensuring process simplification.

[0043] This invention obtains constant current and constant voltage charging voltage data from a battery aging test platform, exports original RGB three-channel images, preprocesses the exported images to generate grayscale images of the same size and format as feature maps, and then inputs them into an image recognition model with an integrated lightweight attention mechanism. This model automatically extracts hidden aging features from the images, establishes a correspondence between the feature maps and capacity degradation, and directly outputs the State of Health (SOH) value. Compared to existing technologies, this invention eliminates the need for manual extraction of any health features, ensuring a simplified estimation process while maintaining high accuracy and low computational resource consumption.

[0044] like Figure 1 As shown, this embodiment provides a method for estimating the health status of a lithium-ion battery based on image recognition, including the following steps:

[0045] Voltage data during the constant current and constant voltage charging process of a lithium-ion battery is acquired, and an original two-dimensional image is generated based on the voltage data. The original two-dimensional image is preprocessed to obtain a single-channel grayscale feature map. The feature map is input into a lightweight model that fuses a residual network module and an SE network module to obtain high-order features related to the battery health state. The high-order features are mapped to the battery health state estimate through a fully connected layer.

[0046] S1. Obtain the corresponding constant current and constant voltage charging voltage data, and then export the original image based on each set of constant current and constant voltage charging data. Specifically, the original charging data includes equally spaced time intervals and the voltage value at that time. The time value is used as the horizontal axis and the voltage value is used as the vertical axis to output the original two-dimensional image.

[0047] S2. Image preprocessing mainly consists of image grayscale conversion, image scaling, center cropping, and normalization. When an image is used as input to a model, the model primarily identifies the edges, textures, and shapes of curves in the input image. Grayscale images can better highlight these features, and since grayscale images have only one color channel, compared to the everyday three-channel RGB images, they can effectively reduce the computational load of the model and accelerate model training and inference. Therefore, preprocessing converts the input two-dimensional image from a three-channel color image to a grayscale image. Typically, the input image size for a convolutional neural network model is fixed. The image is scaled using the resize function to fit the model's input, and then the resulting image is center-cropped to obtain an image of size 224×224.

[0048] Image standardization ensures that the data distribution conforms to a normal distribution, which helps accelerate model convergence and improves the model's generalization ability. The specific formula is as follows:

[0049]

[0050] In the formula, x is the image before normalization, X normalize The image is normalized, and μ is the mean of the image, calculated as follows:

[0051]

[0052] Where w and h represent the width and height of the image, respectively, and p i,j It is the pixel value at the corresponding (i, j) position;

[0053] σ is the standard deviation, calculated as follows:

[0054]

[0055] The above steps output a single-channel grayscale image of size 224×224; the image processing flow is as follows. Figure 2 As shown.

[0056] S3. The single-channel grayscale feature map is input into the residual network for feature extraction to obtain preliminary feature information. The grayscale image passes through a two-dimensional convolutional layer, then is processed by the ReLU activation function, and then input into a max pooling layer, followed by input into the residual block. For example... Figure 3 and Figure 4 As shown, in the residual block, residual learning is achieved through skip connections and direct mappings, enabling the model's internal structure to possess constant mapping capabilities. This allows the network to be trained more stably while maintaining high depth. The residual block can be represented as:

[0057] y=F(x,{W i})+x

[0058] In the formula, x is the input layer, W i Let y be the weight layer of the i-th convolutional layer, and y be the output layer.

[0059] The input in the residual block passes through a convolutional layer, a ReLU activation function layer, and another convolutional layer in sequence. The result is then added to the input feature vector to obtain a preliminary result. This preliminary result is used as input and passes through a ReLU activation function layer to obtain the result after processing by the residual network.

[0060] In a specific implementation of this embodiment, the grayscale image is passed through a two-dimensional convolutional layer, then processed by a ReLU activation function, and then input into a max pooling layer, followed by a residual block. In the residual block, the input passes through a convolutional layer, a ReLU activation function layer, and another convolutional layer in sequence. The result is then added to the input feature vector and passed through a ReLU activation function layer to obtain the result processed by the residual network.

[0061] S4. The output of the module in S3 is used as the input to the SE network module to amplify the feature information. By applying the SE module, the attention mechanism is focused on the channels, making the network pay more attention to image features highly correlated with the state of battery (SOH), thereby reducing the influence of irrelevant factors. The output of the residual network in S3 is used as the input to the SE module. This input is processed through a Squeeze operation, followed by an Excitation operation. The result is then scaled with the input to obtain the output of the SE module.

[0062] The Squeeze operation compresses global feature information (i.e., preliminary feature information) into one-dimensional data with a global receptive field, thereby reducing computational cost. This can be represented as:

[0063]

[0064] W and H represent the width and height of the input feature map, respectively, where z c ∈R C It is for feature u c The result of performing global average pooling, u c (i,j) is its pixel value at the corresponding (i,j) position;

[0065] The excitation operation will assign corresponding weights through the fully connected layer and the non-linear activation function, calculated as follows:

[0066] S = F ex (Z,W)=σ(g(Z,W))=σ(W2δ(W 1z ))

[0067] In the formula, σ represents the sigma function, and δ represents the ReLU function. r is a hyperparameter, which is set to 16, and C represents the number of channels in the input model.

[0068] The weights of each channel obtained through the scale operation are applied to the original feature map to recalibrate the original features, calculated as follows:

[0069] x c =F scale (u c ,s c ) = s c u c

[0070] X = [x1, x2, ..., x C ],F scale The result of multiplying the channels, u c Indicates the input, s cThis represents the output obtained through the Excitation operation, i.e., the channel features related to the battery health status.

[0071] S5. Repeat steps S3 and S4 to build modules with the same structure. The whole module consists of two residual network modules and two SE modules. The input passes through the residual network module and the two SE modules, and the output is used as the input of the residual network module, and then input into the SE module.

[0072] S6. Use the output of step S5 as the input to the fully connected neural network, and output the estimated SOH. The residual network model has a local receptive field, which can extract detailed information from the charging image of the lithium-ion battery, while providing the stability and robustness of deep networks. The SE module can amplify the detailed information in the image and enhance the sensitivity of feature information. The fusion model can combine the advantages of both, so as to more accurately explore the potential relationship between battery image data and state variables (SOH).

[0073] Figure 5 The structure of the fusion model is shown. The preprocessed image serves as the model's input, and the final output is the SOH estimate of the battery. To reduce model complexity, the constructed network mainly consists of two convolutional layers, two pooling layers, two non-linear activation layers, and one fully connected layer. Residual blocks and channel attention modules are added after each pooling layer. These network modules are connected in series. Finally, the extracted image information is stretched and fully connected to output the SOH estimate corresponding to the battery.

[0074] S7. Use MSE as the loss function for model training and employ the Adam optimization algorithm. Specifically, the initial learning rate for Adam is lr = 2 × 10⁻⁶. -4 The first-order and second-order exponential decay factors are 0.9 and 0.999, respectively, betas = (0.9, 0.999), and weight decay = 0. The training time is typically 100 epochs, with a batch size of 1.

[0075] S8. The feature map obtained from step S2 of the image to be tested is input into the network constructed in S3-S7 to directly obtain the SOH estimation result. The error box plots for the test set on different experimental datasets (selecting the first 70% as the training set and the remaining data as the test set) are shown below. Figure 6 As shown, three other image recognition methods (DCNN, VIT, and ResNet) were selected for comparison. It can be seen that the proposed model's error can be reduced to below 1%, and its estimation performance is better than the other models. This demonstrates that the lightweight convolutional neural network and channel attention mechanism fusion model can maintain high estimation accuracy while ensuring process simplification.

[0076] Therefore, this invention adopts the above-mentioned image recognition-based lithium-ion battery SOH estimation method, which simplifies the cumbersome manual feature extraction and health feature selection process required for data-driven SOH estimation. It eliminates the need to manually extract any health features, ensuring the simplification of the entire estimation process. At the same time, the attention mechanism module is embedded into the residual network to construct a lightweight convolutional neural network and channel attention mechanism fusion model, which can maintain high estimation accuracy while ensuring process simplification.

[0077] This invention employs an image-based state of health estimation method, which uses constant current and constant voltage charging voltage curve images as input to estimate the state of health (SOH) of the battery, thus simplifying the SOH estimation process for lithium-ion batteries.

[0078] To more accurately uncover the potential relationship between battery image data and SOH, this invention embeds an attention mechanism module (SEBlock) into a residual network (ResNet), constructing a lightweight fusion model of convolutional neural network and channel attention mechanism, which can maintain high computational efficiency while ensuring high accuracy of estimation results.

[0079] The present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a lithium-ion battery health status estimation method based on image recognition.

[0080] The present invention provides a storage medium storing a computer program that, when executed by a processor, implements a method for estimating the health status of a lithium-ion battery based on image recognition.

[0081] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for estimating the state of health of a lithium-ion battery based on image recognition, characterized in that, Includes the following steps: Obtain voltage data during the constant current and constant voltage charging process of a lithium-ion battery, and generate an original two-dimensional image based on the voltage data; The original two-dimensional image is preprocessed to obtain a single-channel grayscale feature map; The feature map is input into a lightweight model that fuses the residual network module and the SE network module to obtain high-order features related to the battery health status. The higher-order features are mapped to battery health state estimates through a fully connected layer.

2. The lithium-ion battery health status estimation method based on image recognition according to claim 1, characterized in that, The process of preprocessing the original two-dimensional image to obtain a single-channel grayscale feature map includes: Convert the original two-dimensional image into a grayscale image; The grayscale image is scaled using the resize function, and the scaled image is then cropped from the center to obtain a fixed-size image. The fixed-size image is subjected to image standardization processing to obtain a single-channel grayscale feature map.

3. The lithium-ion battery health status estimation method based on image recognition according to claim 1, characterized in that, The process of inputting the feature map into a lightweight model that fuses the residual network module and the SE network module to obtain high-order features related to battery health status includes: The single-channel grayscale feature map is input into the residual network module for feature extraction to obtain preliminary feature information; The preliminary feature information is input into the SE network module to obtain channel features related to battery health status; The residual network module and the SE network module are alternately connected to form a hierarchical feature extraction structure, and high-order features related to the battery health status are obtained based on the hierarchical feature extraction structure.

4. The lithium-ion battery health status estimation method based on image recognition according to claim 3, characterized in that, The residual network module includes: a two-dimensional convolutional layer, a max pooling layer, and a residual block; the residual block includes: a first convolutional sub-layer, a second convolutional sub-layer, and a skip connection layer.

5. The lithium-ion battery health status estimation method based on image recognition according to claim 4, characterized in that, The expression for feature extraction based on the residual network module is: y=F(x,{W i })+x In the formula, x is the input layer, W i Let y be the weight layer of the i-th convolutional layer, y be the output layer, and F(.) be the intermediate features.

6. The lithium-ion battery health status estimation method based on image recognition according to claim 3, characterized in that, The process of inputting the preliminary feature information into the SE network module to obtain channel features related to battery health status includes: The preliminary feature information is compressed into one-dimensional data features with a global receptive field based on the Squeeze operation; The channel weight coefficients are obtained from the one-dimensional data features based on the Excitation operation; Based on the channel weight coefficients and preliminary feature information, channel features related to battery health status are obtained.

7. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the image recognition-based lithium-ion battery health status estimation method as described in claim 1.

8. A storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the image recognition-based lithium-ion battery health status estimation method as described in claim 1.

Images