A lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features

Abstract

This application discloses a lithium battery SOH estimation method based on multi-source degradation characterization maps and dynamic feature fusion. The method includes: synchronously collecting voltage, current, and temperature data within a specific state of charge range during lithium battery discharge cycling; generating a gray trajectory map based on voltage and current data, generating a temperature thermogram based on temperature data, generating a degradation state map based on cycle number and aging characteristics, and forming a multi-source fusion map through weighted fusion; constructing a dynamic feature fusion network model and a multi-source feature constraint loss function; and deploying the trained dynamic feature fusion network model to a battery management system for online SOH estimation. This invention constructs a multi-source degradation characterization map, comprehensively quantifies the battery aging state, and proposes to effectively fuse heterogeneous feature information through a dynamic feature fusion network model and a multi-source feature constraint loss function, thereby improving the accuracy of lithium battery SOH estimation.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and in particular to a lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features. Background Technology

[0002] As the core power source for electric vehicles and energy storage systems, lithium batteries' state of health (SOH) is closely related to the operational safety and lifespan of these devices. With increasing battery usage, the internal resistance gradually increases, leading to localized overheating and potentially causing thermal runaway and other safety incidents. Therefore, accurately estimating the SOH of lithium batteries has significant engineering value in reducing the incidence of safety accidents and minimizing economic losses.

[0003] In existing research, most scholars tend to directly use raw battery operating data (such as voltage curves, current curves, and electrochemical impedance spectroscopy curves) as model input, or preprocess it using basic feature engineering techniques (such as generating differential voltage curves and capacity increment curves). Other studies employ algorithms such as Grom's angle field (GAF) or Markov transfer field (MTF) to convert one-dimensional electrochemical impedance spectroscopy data into two-dimensional image features, mining the spatiotemporal correlation features within the data. However, converting only a single data source (such as voltage or EIS) makes it difficult to capture the coupling effects of multiple physical variables in the aging process, especially neglecting the synergistic trends of temperature and voltage / current changes and the synergistic effects of cycle decay. Furthermore, existing fusion methods (such as feature stitching and early fusion) lack interaction mechanisms between multiple data sources, resulting in insufficient exploration of the complementarity between physical states such as voltage, current, and temperature and the aging process information, making it impossible to comprehensively quantify the battery aging state. Summary of the Invention

[0004] To overcome the shortcomings of existing SOH estimation methods based on single-source image transformation and simple feature fusion, this invention proposes a lithium battery SOH estimation method based on multi-source degradation characterization maps and dynamic feature fusion.

[0005] The technical solution of the present invention is as follows:

[0006] S1: Collect battery operation data. Specific steps include: setting the maximum number of battery cycles, performing cyclic charge and discharge tests on the lithium battery, selecting the discharge stage in each discharge cycle where the SOC drops from 80% to 30%, synchronously collecting voltage, current and temperature data, and recording the current number of cycles.

[0007] S2: Generate four types of degradation characterization maps for lithium batteries based on battery operation data, including a gray trajectory map, a temperature thermogram, a degradation state map, and a multi-source fusion map. Specific steps include: first, normalizing the collected lithium battery data; then, generating a gray trajectory map reflecting the coordinated change trend of voltage and current based on the voltage and current data corresponding to each cycle; next, generating a temperature thermogram reflecting the coordinated change trend of battery temperature, voltage, and current based on the gray trajectory map and the temperature data of the corresponding cycle; then, generating a degradation state map reflecting the impact of cycle number on battery degradation based on the number of cycles and aging characteristics; finally, fusing the temperature thermogram and the degradation state map to generate a multi-source fusion map comprehensively reflecting the battery aging state.

[0008] S3: Construct a dynamic feature fusion network model and a multi-source feature constraint loss function. Specific steps include: First, based on the grayscale trajectory map, temperature heatmap, degradation state map, and multi-source fusion map, feature extraction is performed using an encoder, followed by feature transformation to obtain key features for each type of image. Then, based on the extracted features, dynamic weights are calculated and weighted to generate fusion features, adaptively fusing complementary information from different images. Next, the fusion features are input into a fully connected layer to estimate the battery's state of equilibrium (SOH). Finally, the model parameters are jointly optimized using a multi-source feature constraint loss function that includes regression loss, feature diversity loss, and orthogonality constraint loss, thereby suppressing feature redundancy and improving the comprehensiveness of aging characterization.

[0009] S4: Dynamic Feature Fusion Network Model Training and Online SOH Estimation. The specific steps include: First, based on the battery charge-discharge cycle data collected in S1, a degradation characterization graphic dataset containing grayscale trajectory maps, temperature heat maps, degradation state maps, and multi-source fusion maps is generated according to the method in S2. Then, the dataset is divided into training, testing, and validation sets. Next, the dynamic feature fusion network model constructed in S3 and the multi-source feature constraint loss function are used for training and testing. Finally, the trained model is deployed to the battery management system for online SOH estimation.

[0010] In this solution, step S2, which involves generating four types of degradation characterization maps for lithium batteries based on battery operating data, includes the following specific steps:

[0011] S201: Normalize the recorded voltage, current, and temperature data of the lithium battery during the c-th charge-discharge cycle, as follows:

[0012]

[0013] Among them, U c This represents the voltage data sequence collected in the c-th cycle, specifically U. c ={u1,…,u k ,…,u N};Ic This represents the current data sequence collected in the c-th cycle, specifically I. c ={i1,…,i k ,…,i N};T c This represents the temperature data sequence collected in the c-th cycle, specifically T. c ={t1,…,t k ,…,t N}; N represents the number of sampling points in the c-th iteration; u k u k t k Let u' represent the voltage, current, and temperature at sampling point k, respectively, k = 1, ..., N; k 、u' k ,t' k Let k represent the normalized voltage, current, and temperature at sampling point k, respectively; min(·) represents the minimum value in the sequence; max(·) represents the maximum value in the sequence.

[0014] S202: Based on the voltage u' normalized from S201 k and current i' k The pixel coordinates corresponding to the voltage and current trajectories are calculated, then adjacent points are connected to form the trajectory path Γ, and finally a complete grayscale trajectory image is generated. Specifically as follows:

[0015]

[0016]

[0017]

[0018] in, This represents the floor function, W represents the image width, H represents the image height, and m is the trajectory point index, corresponding to the sampling point k; p m The x-coordinate represents the pixel coordinates of trajectory point m. m The x-coordinate and y-coordinate of the trajectory point m are represented by... m Let Γ represent the ordinate of the trajectory point m, and let Γ represent the trajectory path. p m With p m+1 The line segments between them are denoted by ∪, where ∪ represents the union of line segments, d((x,y),Γ) represents the Euclidean distance from pixel (x,y) to trajectory Γ, and σ represents the Gaussian kernel width control parameter. Represents grayscale trajectory map The gray value at pixel (x,y), where exp(·) represents an exponential function;

[0019] S203: Temperature t' normalized according to S201k Define a hue conversion function for temperature, and use this function to generate a complete grayscale trajectory map for S202. Render the temperature heatmap. Specifically as follows:

[0020] H(t′ k )=240°·(1-t′ k )

[0021]

[0022] Wherein, H(t' k ) represents the normalized temperature t' k The hue angle of the map, H m This represents the RGB color vector corresponding to trajectory point m. Let f(·) represent the red, green, and blue channel values ​​in the RGB color vector corresponding to the trajectory point m, respectively; f(·) represents the conversion function from HSV to RGB standard color space; S represents saturation, which is a constant of 1; and V represents brightness, which is a constant of 1. Represents a temperature thermogram. This represents the RGB pixel value vector at pixel (x, y) in the temperature heatmap. These represent the red, green, and blue channel values ​​of the temperature heatmap at pixel (x, y), respectively, where ||·|| represents the Euclidean distance, and σ... T The temperature diffusion scale parameter is represented by τ, the dynamic temperature weighting coefficient is represented by exp(·), and e (·) Both represent exponential functions;

[0023] S204: The normalized voltage sequence U' based on the current cycle number c. c and current sequence I' c And the normalized voltage sequence U'1 and current sequence I'1 of the first discharge cycle, the current cycle number c, and the maximum cycle number c. max The aging characteristics are constructed, and finally a degradation state map is formed. Specifically as follows:

[0024]

[0025] Among them, U′ j Let I′ represent the normalized voltage sequence of the j-th cycle. j Let DTW(·) represent the normalized current sequence of the j-th cycle, and let DTW(·) represent the dynamic time warping function. This represents the maximum value of DTW during iterations 1 to c. U (c) represents the voltage aging characteristics at cycle number c, HF I(c) represents the current aging characteristics at cycle number c, HF cycle (c) represents the cyclic aging characteristics at the number of cycles c; Represents a degenerate state diagram. This represents the RGB vector at pixel (x, y) in the degenerate state map. These represent the red, green, and blue channel values ​​at pixel (x, y) in the degraded state map, respectively.

[0026] S205: Based on the temperature and thermogram of S203 Degradation state diagram of S204 Finally, a weighted fusion process is performed to generate a multi-source fusion graph. Specifically as follows:

[0027]

[0028] Where β is a measure of temperature in a thermogram. Degeneration state diagram The contribution ratio weighting coefficient.

[0029] In this solution, step S3, which involves constructing the dynamic feature fusion network model and the multi-source feature constraint loss function, comprises the following specific steps:

[0030] S301: Based on the grayscale trajectory map obtained in S2 Temperature thermogram Degeneration state diagram Multi-source fusion diagram A gray trajectory map encoder, a temperature heatmap encoder, a degradation state map encoder, and a multi-source fusion map encoder were constructed separately. Feature extraction was performed on each, and then feature transformation was performed on each to obtain their respective feature vectors, as detailed below:

[0031]

[0032] Among them, E g (·), E t (·), E a (·), E f (·) represent grayscale trajectory maps respectively. Temperature thermogram Degeneration state diagram Multi-source fusion diagram The corresponding CNN encoder used for feature extraction; θ g θ t θ a θ f Representing grayscale trajectory images Temperature thermogram Degeneration state diagram Multi-source fusion diagram The corresponding trainable parameters of the CNN encoder; Φ g (·), Φ t (·), Φ a (·), Φ f (·) represent grayscale trajectory maps respectively. Temperature thermogram Degeneration state diagram Multi-source fusion diagram A transformation function converts the feature map obtained after feature extraction into a feature vector, transforming the feature map output by the CNN encoder into a fixed-dimensional feature vector; f g f t f a f f They represent The feature vectors obtained after feature extraction and feature transformation have the same dimension.

[0033] S302: According to f in S301 g f t f a f f Perform fusion and output fused features f fuse The details are as follows:

[0034] s g =v g ·tanh(W g ·f g +b g )

[0035] s t =v t ·tanh(W t ·f t +b t )

[0036] s a =v a ·tanh(W a ·f a +b a )

[0037] s f =v f ·tanh(W f ·f f +b f )

[0038]

[0039]

[0040] ffuse =δ g ·f g +δ t ·f t +δ a ·f a +δ f ·f f

[0041] Among them, W g W t W a W f They represent f respectively g f t f a f f The corresponding transformation weights; b g b t b a b f They represent f respectively g f t f a f f Corresponding bias; v g v t v a v f They represent f respectively g f t f a f f The corresponding score weights, tanh(·) represents the hyperbolic tangent activation function, s g s t s a s f They represent f respectively g f t f a f f Corresponding feature score; e (·) Represents an exponential function; γ represents a scaling factor; δ g δ t δ a δ f They represent f respectively g f t f a f f The dynamic weight coefficients of the corresponding features; f fuse Indicates fusion characteristics;

[0042] S303: Based on the fusion feature f in S302 fuse Design a SOH estimation network to map fused features to SOH. pred The details are as follows:

[0043] SOH pred =MLP(f fuse ;θ p )

[0044] Where, θ p This indicates the parameters of the SOH-estimated network; MLP(·) represents a fully connected layer.

[0045] S304: A loss function was designed, comprising regression loss, feature diversity loss, orthogonality constraint loss, etc., as detailed below:

[0046]

[0047] in, λ1 and λ2 represent the regression loss function, the feature diversity loss function, and the orthogonality constraint loss function, respectively, and represent hyperparameters used to balance different optimization objectives.

[0048] The regression loss function The specific representation is as follows:

[0049]

[0050] Where, N batch Indicates the batch sample quantity; Let SOH represent the network estimate of sample i. This represents the true SOH of sample i;

[0051] The feature diversity loss function The specific representation is as follows:

[0052]

[0053] in, They represent f respectively g f t f a f f The transpose of , ||·|| denotes the Euclidean norm, and |·| denotes the absolute value;

[0054] The orthogonal constraint loss function The specific representation is as follows:

[0055] W gtaf =[W g W t W a W f ]

[0056]

[0057] Among them, W gtaf Indicates the fusion weight. W gtaf The transpose of I E Represents the identity matrix. This represents the square of the Frobenius norm.

[0058] In this scheme, the training of the dynamic feature fusion network model and the online estimation of SOH in step S4 are specifically carried out as follows:

[0059] S401: Based on the battery charge-discharge cycle data collected in S1, generate a grayscale trajectory map for each sample according to the method in S2. Temperature thermogram Degeneration state diagram Multi-source fusion diagram Multiple graph datasets are generated and divided into training, validation and test sets according to a preset ratio;

[0060] S402: Use the dynamic feature fusion network model and multi-source feature constraint loss function constructed in step S3 to train the model. After the validation set converges, save the trained model and deploy it to the battery management system.

[0061] S403: The battery management system collects the voltage, current, temperature and current cycle number of the battery in the current cycle, and generates a grayscale trajectory map, temperature heat map, degradation state map and multi-source fusion map according to the method in step S2; inputs the graphics into the pre-trained model and directly outputs the battery SOH.

[0062] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0063] This invention reveals the coupling relationship between temperature and voltage / current changes by generating gray trajectory maps, temperature thermograms, degradation state maps, and multi-source fusion maps, comprehensively quantifying the battery aging state. It also constructs a dynamic weighted feature fusion network model and designs a multi-source feature constraint loss function to effectively fuse heterogeneous feature information, overcoming the problems of one-sided representation, inefficient fusion, and high redundancy in existing technologies, thereby improving the accuracy of lithium battery SOH estimation. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0065] Figure 1 This is a flowchart of a lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features proposed in this invention.

[0066] Figure 2 This is a diagram illustrating the effect of constructing a gray trajectory diagram as proposed in this invention. Detailed Implementation

[0067] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0068] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0069] In a specific embodiment, such as Figure 1 As shown, a lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features includes the following steps:

[0070] S1: Collect battery operation data. Specific steps include: setting the maximum number of battery cycles, performing cyclic charge and discharge tests on the lithium battery, selecting the discharge stage in each discharge cycle where the SOC drops from 80% to 30%, synchronously collecting voltage, current and temperature data, and recording the current number of cycles.

[0071] S2: Generate four types of degradation characterization maps for lithium batteries based on battery operation data, including a gray trajectory map, a temperature thermogram, a degradation state map, and a multi-source fusion map. Specific steps include: first, normalizing the collected lithium battery data; then, generating a gray trajectory map reflecting the coordinated change trend of voltage and current based on the voltage and current data corresponding to each cycle; next, generating a temperature thermogram reflecting the coordinated change trend of battery temperature, voltage, and current based on the gray trajectory map and the temperature data of the corresponding cycle; then, generating a degradation state map reflecting the impact of cycle number on battery degradation based on the number of cycles and aging characteristics; finally, fusing the temperature thermogram and the degradation state map to generate a multi-source fusion map comprehensively reflecting the battery aging state.

[0072] S3: Construct a dynamic feature fusion network model and a multi-source feature constraint loss function. Specific steps include: First, based on the grayscale trajectory map, temperature heatmap, degradation state map, and multi-source fusion map, feature extraction is performed using an encoder, followed by feature transformation to obtain key features for each type of image. Then, based on the extracted features, dynamic weights are calculated and weighted to generate fusion features, adaptively fusing complementary information from different images. Next, the fusion features are input into a fully connected layer to estimate the battery's state of equilibrium (SOH). Finally, the model parameters are jointly optimized using a multi-source feature constraint loss function that includes regression loss, feature diversity loss, and orthogonality constraint loss, thereby suppressing feature redundancy and improving the comprehensiveness of aging characterization.

[0073] S4: Dynamic Feature Fusion Network Model Training and Online SOH Estimation. The specific steps include: First, based on the battery charge-discharge cycle data collected in S1, a degradation characterization graphic dataset containing grayscale trajectory maps, temperature heat maps, degradation state maps, and multi-source fusion maps is generated according to the method in S2. Then, the dataset is divided into training, testing, and validation sets. Next, the dynamic feature fusion network model constructed in S3 and the multi-source feature constraint loss function are used for training and testing. Finally, the trained model is deployed to the battery management system for online SOH estimation.

[0074] In this solution, step S2, which involves generating four types of degradation characterization maps for lithium batteries based on battery operating data, includes the following specific steps:

[0075] S201: Normalize the recorded voltage, current, and temperature data of the lithium battery during the c-th charge-discharge cycle, as follows:

[0076]

[0077] Among them, U c This represents the voltage data sequence collected in the c-th cycle, specifically U. c ={u1,…,u k ,…,u N};I c This represents the current data sequence collected in the c-th cycle, specifically I. c ={i1,…,i k ,…,i N};T c This represents the temperature data sequence collected in the c-th cycle, specifically T. c ={t1,…,t k ,…,t N}; N represents the number of sampling points in the c-th iteration; u k i k t k Let u' represent the voltage, current, and temperature at sampling point k, respectively, k = 1, ..., N;k 、i' k ,t' k Let k represent the normalized voltage, current, and temperature at sampling point k, respectively; min(·) represents the minimum value in the sequence; max(·) represents the maximum value in the sequence.

[0078] S202: Based on the voltage u normalized from S201 k and current i k The pixel coordinates corresponding to the voltage and current trajectories are calculated, then adjacent points are connected to form the trajectory path Γ, and finally a complete grayscale trajectory image is generated. Generated grayscale trajectory map The effect is as follows Figure 2 As shown, the details are as follows:

[0079]

[0080] in, This represents the floor function, W represents the image width, H represents the image height, and m is the trajectory point index, corresponding to the sampling point k; p m The x-coordinate represents the pixel coordinates of trajectory point m. m The x-coordinate and y-coordinate of the trajectory point m are represented by... m Let Γ represent the ordinate of the trajectory point m, and let Γ represent the trajectory path. p m With p m+1 The line segments between them are denoted by ∪, where ∪ represents the union of line segments, d((x,y),Γ) represents the Euclidean distance from pixel (x,y) to trajectory Γ, and σ represents the Gaussian kernel width control parameter. Represents grayscale trajectory map The gray value at pixel (x,y), where exp(·) represents an exponential function;

[0081] S203: Temperature t' normalized according to S201 k Define a hue conversion function for temperature, and use this function to generate a complete grayscale trajectory map for S202. Render the temperature heatmap. Specifically as follows:

[0082] H(t′ k )=240°·(1-t′ k )

[0083]

[0084] Among them, G(t' k ) represents the normalized temperature t' k The hue angle of the mapping, H m This represents the RGB color vector corresponding to trajectory point m. Let f(·) represent the red, green, and blue channel values ​​in the RGB color vector corresponding to the trajectory point m, respectively; f(·) represents the conversion function from HSV to RGB standard color space; S represents saturation, which is a constant of 1; and V represents brightness, which is a constant of 1. Represents a temperature thermogram. This represents the RGB pixel value vector at pixel (x, y) in the temperature heatmap. These represent the red, green, and blue channel values ​​of the temperature heatmap at pixel (x, y), respectively, where ||·|| represents the Euclidean distance, and σ... T The temperature diffusion scale parameter is represented by τ, the dynamic temperature weighting coefficient is represented by exp(·), and e (·) Both represent exponential functions;

[0085] S204: The normalized voltage sequence U' based on the current cycle number c. c and current sequence I' c And the normalized voltage sequence U'1 and current sequence I'1 of the first discharge cycle, the current cycle number c, and the maximum cycle number c. max The aging characteristics are constructed, and finally a degradation state map is formed. Specifically as follows:

[0086]

[0087]

[0088] Among them, U′ j Let I′ represent the normalized voltage sequence of the j-th cycle. j Let DTW(·) represent the normalized current sequence of the j-th cycle, and let DTW(·) represent the dynamic time warping function. This represents the maximum value of DTW during iterations 1 to c. U (c) represents the voltage aging characteristics at cycle number c, HF I (c) represents the current aging characteristics at cycle number c, HF cycle (c) represents the cyclic aging characteristics at the number of cycles c; Represents a degenerate state diagram. This represents the RGB vector at pixel (x, y) in the degenerate state map. These represent the red, green, and blue channel values ​​at pixel (x, y) in the degraded state map, respectively.

[0089] S205: Based on the temperature and thermogram of S203 Degradation state diagram of S204 Finally, a weighted fusion process is performed to generate a multi-source fusion graph. Specifically as follows:

[0090]

[0091] Where β is a measure of temperature in a thermogram. Degeneration state diagram The contribution ratio weighting coefficient.

[0092] In this solution, step S3, which involves constructing the dynamic feature fusion network model and the multi-source feature constraint loss function, comprises the following specific steps:

[0093] S301: Based on the grayscale trajectory map obtained in S2 Temperature thermogram Degeneration state diagram Multi-source fusion diagram A gray trajectory map encoder, a temperature heatmap encoder, a degradation state map encoder, and a multi-source fusion map encoder were constructed separately. Feature extraction was performed on each, and then feature transformation was performed on each to obtain their respective feature vectors, as detailed below:

[0094]

[0095] Among them, E g (·), E t (·), E a (·), E f (·) represent grayscale trajectory maps respectively. Temperature thermogram Degeneration state diagram Multi-source fusion diagram The corresponding CNN encoder used for feature extraction; θ g θ t θ a θ f Representing grayscale trajectory images Temperature thermogram Degeneration state diagram Multi-source fusion diagram The corresponding trainable parameters of the CNN encoder; Φ g (·), Φ t (·), Φ a (·), Φ f (·) represent grayscale trajectory maps respectively. Temperature thermogram Degeneration state diagram Multi-source fusion diagram A transformation function converts the feature map obtained after feature extraction into a feature vector, transforming the feature map output by the CNN encoder into a fixed-dimensional feature vector; f g f t f a ff They represent The feature vectors obtained after feature extraction and feature transformation have the same dimension.

[0096] S302: According to f in S301 g f t f a f f Perform fusion and output fused features f fuse The details are as follows:

[0097] s g =v g ·tanh(W g ·f g +b g )

[0098] s t =v t ·tanh(W t ·f t +b t )

[0099] s a =v a ·tanh(W a ·f a +b a )

[0100] s f =v f ·tanh(W f ·f f +b f )

[0101]

[0102] f fuse =δ g ·f g +δ t ·f t +δ a ·f a +δ f ·f f

[0103] Among them, W g W t W a W f They represent f respectively g f t f a f f The corresponding transformation weights; b g b t ba b f They represent f respectively g f t f a f f Corresponding bias; v g v t v a v f They represent f respectively g f t f a f f The corresponding score weights, tanh(·) represents the hyperbolic tangent activation function, s g s t s a s f They represent f respectively g f t f a f f Corresponding feature score; e (·) Represents an exponential function; γ represents a scaling factor; δ g δ t δ a δ f They represent f respectively g f t f a f f The dynamic weight coefficients of the corresponding features; f fuse Indicates fusion characteristics;

[0104] S303: Based on the fusion feature f in S302 fuse Design a SOH estimation network to map fused features to SOH. pred The details are as follows:

[0105] SOH pred =MLP(f fuse ;θ p )

[0106] Where, θ p This indicates the parameters of the SOH-estimated network; MLP(·) represents a fully connected layer.

[0107] S304: A loss function was designed, comprising regression loss, feature diversity loss, orthogonality constraint loss, etc., as detailed below:

[0108]

[0109] in, λ1 and λ2 represent the regression loss function, the feature diversity loss function, and the orthogonality constraint loss function, respectively, and represent hyperparameters used to balance different optimization objectives.

[0110] The regression loss function The specific representation is as follows:

[0111]

[0112] Where, N batch Indicates the batch sample quantity; Let SOH represent the network estimate of sample i. This represents the true SOH of sample i;

[0113] The feature diversity loss function The specific representation is as follows:

[0114]

[0115] in, They represent f respectively g f t f a f f The transpose of , ||·|| denotes the Euclidean norm, and |·| denotes the absolute value;

[0116] The orthogonal constraint loss function The specific representation is as follows:

[0117] W gtaf =[W g W t W a W f ]

[0118]

[0119] Among them, W gtaf Indicates the fusion weight. W gtaf The transpose of I E Represents the identity matrix. This represents the square of the Frobenius norm.

[0120] In this scheme, the training of the dynamic feature fusion network model and the online estimation of SOH in step S4 are specifically carried out as follows:

[0121] S401: Based on the battery charge-discharge cycle data collected in S1, generate a grayscale trajectory map for each sample according to the method in S2. Temperature thermogram Degeneration state diagram Multi-source fusion diagram Multiple graph datasets are generated and divided into training, validation and test sets according to a preset ratio;

[0122] S402: Use the dynamic feature fusion network model and multi-source feature constraint loss function constructed in step S3 to train the model. After the validation set converges, save the trained model and deploy it to the battery management system.

[0123] S403: The battery management system collects the voltage, current, temperature and current cycle number of the battery in the current cycle, and generates a grayscale trajectory map, temperature heat map, degradation state map and multi-source fusion map according to the method in step S2; inputs the graphics into the pre-trained model and directly outputs the battery SOH.

Claims

1. A lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features, characterized in that, The specific steps are as follows: S1: Collect battery operation data. Specific steps include: setting the maximum number of battery cycles, performing cyclic charge and discharge tests on the lithium battery, selecting the discharge stage in each discharge cycle where the SOC drops from 80% to 30%, synchronously collecting voltage, current and temperature data, and recording the current number of cycles. S2: Generate four types of degradation characterization maps for lithium batteries based on battery operation data, including a gray trajectory map, a temperature thermogram, a degradation state map, and a multi-source fusion map. Specific steps include: first, normalizing the collected lithium battery data; then, generating a gray trajectory map reflecting the coordinated change trend of voltage and current based on the voltage and current data corresponding to each cycle; next, generating a temperature thermogram reflecting the coordinated change trend of battery temperature, voltage, and current based on the gray trajectory map and the temperature data of the corresponding cycle; then, generating a degradation state map reflecting the impact of cycle number on battery degradation based on the number of cycles and aging characteristics; finally, fusing the temperature thermogram and the degradation state map to generate a multi-source fusion map comprehensively reflecting the battery aging state. S3: Construct a dynamic feature fusion network model and a multi-source feature constraint loss function. Specific steps include: First, based on the gray trajectory map, temperature heatmap, degradation state map, and multi-source fusion map, feature extraction is performed using an encoder, followed by feature transformation to obtain key features for each type of graph. Then, based on the extracted features, dynamic weights are calculated and weighted to generate fusion features, adaptively fusing complementary information from different graphs. Next, the fusion features are input into a fully connected layer to estimate the battery's state of equilibrium (SOH). Finally, the model parameters are jointly optimized using a multi-source feature constraint loss function that includes regression loss, feature diversity loss, and orthogonality constraint loss, thereby suppressing feature redundancy and improving the comprehensiveness of aging characterization. S4: Dynamic Feature Fusion Network Model Training and Online SOH Estimation. The specific steps include: First, based on the battery charge-discharge cycle data collected in S1, a degradation characterization graphic dataset containing gray trajectory maps, temperature heat maps, degradation state maps, and multi-source fusion maps is generated according to the method in S2. Then, the dataset is divided into training sets, test sets, and validation sets. Next, the dynamic feature fusion network model constructed in S3 and the multi-source feature constraint loss function are used for training and testing. Finally, the trained model is deployed to the battery management system for online SOH estimation.

2. The lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features according to claim 1, characterized in that, Step S2, which involves generating four types of degradation characterization diagrams for lithium batteries based on battery operating data, includes the following specific steps: S201: Record the lithium battery in the first... The voltage, current, and temperature data from each charge-discharge cycle were normalized as follows: in, Indicates the first The voltage data sequence collected in the next cycle is specifically as follows: ; Indicates the first The current data sequence collected in the next cycle is specifically as follows: ; Indicates the first The temperature data sequence collected in the next cycle is specifically as follows: ; Indicates the first Number of sampling points in the next loop; , , These represent the sampling points. Voltage, current and temperature, ; , , These represent the sampling points. Normalized voltage, current, and temperature; This represents the minimum value in the sequence; Represents the maximum value in the sequence; S202: Based on the voltage normalized from S201 and current Calculate the pixel coordinates corresponding to the voltage and current trajectories, and then connect adjacent points to form the trajectory path. Finally, a complete gray trajectory map is generated. The details are as follows: in, This represents the floor function. Indicates the width of the image. Indicates the height of the image. For trajectory point index, and sampling point correspond; Represents trajectory points pixel coordinates, Represents trajectory points x-coordinate Represents trajectory points The ordinate, Indicates the trajectory path. express and The straight line segment between them Represents the union of line segments. Represents pixels To trajectory Euclidean distance, This represents the Gaussian kernel width control parameter. Represents gray trajectory diagram At pixel grayscale value at that location Represents an exponential function; S203: Temperature normalized based on S201 Define a hue conversion function for temperature, and use this function to generate a complete grayscale trajectory map for S202. Render the temperature heatmap. The details are as follows: in, Represents the normalized temperature The hue angle of the map Represents trajectory points The corresponding RGB color vector, , , Representing trajectory points respectively The corresponding red, green, and blue channel values ​​in the RGB color vector. This represents the function for converting from the HSV to the RGB standard color space. This represents saturation, and is taken as a constant of 1. Indicates brightness, taken as a constant of 1; Represents a temperature thermogram. This indicates the temperature heatmap at the pixel level. RGB pixel value vector at that location These are temperature heatmaps at pixels. The red, green, and blue channel values ​​at that location. Represents Euclidean distance. This represents the temperature diffusion scale parameter. This represents the dynamic temperature weighting coefficient. and Both represent exponential functions; S204: Based on the current loop count Normalized voltage sequence and current sequence and the normalized voltage sequence of the first discharge cycle and current sequence Current loop count Maximum number of loops The aging characteristics are constructed, and finally a degradation state map is formed. The details are as follows: in, Indicates the first Normalized voltage sequence of the next cycle. Indicates the first Normalized current sequence of the next cycle This represents a dynamic time warping function. Indicates traversal arrive In the next loop Maximum value Indicates the number of loops Voltage aging characteristics under these conditions Indicates the number of loops Current aging characteristics under the following conditions Indicates the number of loops The cyclic aging characteristics under these conditions; Represents a degenerate state diagram. This indicates the degradation state map at the pixel level. RGB vector at that location, These represent the degradation state map at the pixel level. The red, green, and blue channel values ​​at the location; S205: Based on the temperature and thermogram of S203 Degradation state diagram of S204 Finally, a weighted fusion is performed to form a multi-source fusion graph. The details are as follows: in, It is a thermogram that measures temperature. Degeneration state diagram The contribution ratio weighting coefficient.

3. The lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features according to claim 1, characterized in that, Step S3, which involves constructing the dynamic feature fusion network model and the multi-source feature constraint loss function, includes the following specific steps: S301: Based on the gray trajectory diagram obtained in S2 Temperature thermogram Degeneration state diagram Multi-source fusion diagram A gray trajectory map encoder, a temperature heat map encoder, a degradation state map encoder, and a multi-source fusion map encoder were constructed respectively. Feature extraction was performed on each, and then feature transformation was performed to obtain their respective feature vectors, as follows: in, , , , Representing gray trajectory diagrams Temperature thermogram Degeneration state diagram Multi-source fusion diagram The corresponding CNN encoder used for feature extraction; , , , Representing gray trajectory diagrams Temperature thermogram Degeneration state diagram Multi-source fusion diagram The trainable parameters of the corresponding CNN encoder; , , , Representing gray trajectory diagrams Temperature thermogram Degeneration state diagram Multi-source fusion diagram A transformation function that converts the feature map obtained after feature extraction into a feature vector converts the feature map output by the CNN encoder into a fixed-dimensional feature vector. , , , They represent , , , The feature vectors obtained after feature extraction and feature transformation have the same dimension. S302: According to S301 , , , Perform fusion and output fused features The details are as follows: in, , , , They represent , , , The corresponding transformation weights; , , , They represent , , , Corresponding bias; , , , They represent , , , The corresponding score weights, This represents the hyperbolic tangent activation function. They represent , , , Corresponding feature scores; Represents an exponential function; Indicates the scaling factor; They represent , , , The dynamic weight coefficients of the corresponding features; Indicates fusion characteristics; S303: Based on the fusion features in S302 Design a SOH estimation network to map the fused features to The details are as follows: in, This indicates the parameters of the SOH-estimated network; Indicates a fully connected layer; S304: A loss function was designed, comprising regression loss, feature diversity loss, orthogonality constraint loss, etc., as detailed below: in, , , These represent the regression loss function, the feature diversity loss function, and the orthogonality constraint loss function, respectively. , This represents hyperparameters, used to balance different optimization objectives; The regression loss function The specific representation is as follows: in, Indicates the batch sample quantity; Indicates sample Network-estimated SOH Indicates sample The true SOH; The feature diversity loss function The specific representation is as follows: in, , , , They represent , , , transpose, Denotes the Euclidean norm. Represents absolute value; The orthogonal constraint loss function The specific representation is as follows: in, Indicates the fusion weight. express transpose, Represents the identity matrix. This represents the square of the Frobenius norm.

4. The lithium battery SOH estimation method based on the fusion of multi-source degradation characterization maps and dynamic features according to claim 1, characterized in that, The specific steps for training the dynamic feature fusion network model and estimating SOH online in step S4 are as follows: S401: Based on the battery charge-discharge cycle data collected in S1, generate a gray trajectory map for each sample according to the method in S2. Temperature thermogram Degeneration state diagram Multi-source fusion diagram This generates various graph datasets, which are then divided into training, validation, and test sets according to a preset ratio. S402: Use the dynamic feature fusion network model and multi-source feature constraint loss function constructed in step S3 to train the model. After the validation set converges, save the trained model and deploy it to the battery management system. S403: The battery management system collects the voltage, current, temperature and current cycle number of the battery in the current cycle, and generates a gray trajectory map, temperature heat map, degradation state map and multi-source fusion map according to the method in step S2; inputs the graphics into the pre-trained model and directly outputs the battery SOH.

Images