Battery state of health estimation method based on symmetric positive definite manifold and logarithmic euclidean registration and electronic device

Abstract

The application discloses a battery state of health estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration and electronic equipment, and relates to the field of battery state of health estimation; covariance matrices of each source domain battery sample are geometrically modeled on a symmetric positive definite manifold by using affine invariant Riemannian metrics, and are unified to a same registration center by adopting a logarithmic Euclidean registration mechanism, so that covariance structure heterogeneity caused by differences in production batches of battery cells, vehicle purposes or regional environment is eliminated, the adaptability of the model to unseen working conditions or new batches of batteries is significantly enhanced, and the SOH estimation generalization capability in cross-working-condition and cross-batch scenes is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of battery health state estimation, and in particular to a battery health state estimation method and electronic device based on symmetric positive definite manifold and logarithmic Euclidean registration. Background Technology

[0002] Battery State of Health (SOH) estimation is one of the core functions of a Battery Management System (BMS), and it is crucial for ensuring the safe operation of electric vehicles and energy storage power stations, extending battery life, and optimizing usage strategies. With the rapid deployment of electric vehicles and large-scale energy storage systems, the accumulation of massive amounts of battery operating data has provided new opportunities for data-driven SOH estimation methods. However, existing technologies still face the following key challenges in practical applications.

[0003] First, even within the same operating entity, the operational data collected from different battery packs often only covers limited operating conditions due to differences in cell production batches, vehicle usage (e.g., high-frequency shallow charging and discharging for ride-hailing vehicles, and low-frequency deep charging and discharging for official vehicles), and geographical environment (e.g., hot and humid southern regions vs. cold and dry northern regions). SOH estimation models trained on such localized data have significantly limited generalization ability when faced with new operating conditions or new battery batches.

[0004] Secondly, there is a significant mismatch in the covariance structure of health factors among different battery packs. Here, "health factor" refers to the feature quantity extracted from battery operation data that can effectively characterize its aging state (such as degradation index constructed based on charge-discharge curves). Due to differences in cell production batches, vehicle usage patterns, and regional environments, the covariance matrices of these health factors exhibit high heterogeneity. Traditional methods typically align the covariance matrix directly in Euclidean space, ignoring its non-Euclidean geometric properties, which are essentially symmetric positive definite (SPD) manifolds. This leads to feature alignment distortion and a significant decrease in SOH estimation accuracy when applied to new operating conditions or new battery batches.

[0005] Finally, new battery packs face the challenge of cold start. For newly put into operation battery systems or new batches of cells, there is a lack of sufficient aging process data in the early stages. Traditional data-driven models that rely on a large number of historical degradation samples are difficult to model effectively, resulting in the absence or unreliability of SOH estimation function in the critical early stages. Summary of the Invention

[0006] To overcome the limitations of existing State of Health (SOH) estimation methods in cross-operating condition and batch scenarios, such as weak generalization ability, covariance structure mismatch, and cold start problems, this invention proposes a battery health state estimation method based on symmetric positive definite manifold and log-Euclidean registration, including: Historical charge and discharge data of battery samples contained in each source domain are obtained from multiple source domains. For each battery sample, multiple health factor vectors are extracted based on its historical data from multiple charge-discharge cycles, and the corresponding covariance matrix is ​​calculated based on these multiple health factor vectors. The covariance matrix is ​​a symmetric positive definite matrix, which forms geometric points on a symmetric positive definite manifold and is used to characterize the statistical characteristics of the health factors of the battery sample. The geometric relationship between the covariance matrices on the symmetric positive definite manifold is characterized by the affine invariant Riemannian metric, and the covariance matrices are registered to a unified registration center by the log-Euclidean registration mechanism to obtain the set of registered covariance matrices. The target student model is obtained by training a meta-knowledge distillation model based on the registered covariance matrix set. The meta-knowledge distillation model includes a teacher model and a student model. Obtain the health factor vector of the battery under test and generate the corresponding covariance matrix. Use the registration center to perform a registration transformation on the covariance matrix to obtain the registered covariance matrix under test. Use the target student model to predict the covariance matrix under test and output the battery health status estimate of the battery under test.

[0007] Furthermore, the health factor vector includes at least one of the following: the peak position and amplitude of the voltage response curve, the peak position and peak area of ​​the differential voltage curve, the rate of change of internal resistance with the number of cycles, and the area of ​​the capacity increment curve.

[0008] Furthermore, the covariance matrix is ​​obtained by averaging the deviations between the multiple health factor vectors corresponding to the battery sample and the mean of the multiple health factor vectors.

[0009] Furthermore, the process of registering each covariance matrix to a unified registration center using a log-Euclidean registration mechanism to obtain a set of registered covariance matrices includes: The geometric relations characterized by affine invariant Riemannian metrics are used to determine the matrix logarithmic mapping and matrix exponential mapping; Select an identity matrix or any covariance matrix of any source domain as a reference point, and project each covariance matrix of each source domain onto the tangent space of the reference point through matrix logarithmic mapping to obtain a tangent space projection matrix that corresponds one-to-one with each covariance matrix. In the tangent space, the weights of each source domain are determined based on the proportion of battery samples contained in each source domain to the total number of samples, and the weighted Euclidean mean is calculated for all tangent space projection matrices based on the weights of each source domain. The weighted Euclidean mean is transformed back to a symmetric positive definite manifold through matrix exponential mapping to obtain the registration center; Based on the registration center, a registration transformation is performed on each covariance matrix of each source domain to obtain the set of registered covariance matrices.

[0010] Further, training the meta-knowledge distillation model based on the registered covariance matrix set includes: Using the registered set of covariance matrices and the battery health status category labels corresponding to each covariance matrix in the set, the teacher model is trained by minimizing the classification loss. A student model is constructed, using the registered covariance matrix and the corresponding battery health status category label as training samples. The parameters of the student model are optimized based on the joint distillation loss to obtain the initial student model. The registered covariance matrix set is divided into multiple data subsets according to the source domain. For each data subset, it is divided into a first sample group and a second sample group. The gradient of the joint distillation loss is calculated using the first sample group, and a temporary student model is obtained by performing a single gradient descent update on the initial student model based on the gradient and the preset learning rate. The temporary student model is used to perform forward inference on the corresponding second sample group, and the meta-loss is calculated based on the inference results, the corresponding battery health status category labels, and the output of the teacher model. The total loss of the outer meta-optimization is constructed based on the sum of the meta-losses corresponding to all data subsets and the regularization term for the global shared parameters of the initial student model. The global shared parameters of the initial student model are iteratively optimized based on the total loss to obtain the target student model.

[0011] Furthermore, the joint distillation loss is composed of a weighted sum of a KL divergence term and a cross-entropy term. The KL divergence term is used to constrain the consistency between the predicted distribution output by the student model and the predicted distribution output by the teacher model. The cross-entropy term is used to measure the classification error between the probability distribution of the battery health status category output by the student model and the corresponding battery health status category label.

[0012] Furthermore, the mode of predicting the covariance matrix of the test using the target student model includes: In zero-sample mode, the covariance matrix to be tested is input into the target student model for prediction.

[0013] Furthermore, the method of predicting the covariance matrix of the test using the target student model also includes: In the unsupervised adaptation mode, the gradient of the unsupervised adaptation loss function is calculated using multiple covariance matrices corresponding to the battery under test, and the parameters of the target student model are updated step by step based on the gradient. The covariance matrix under test is then input into the updated target student model for prediction. The unsupervised adaptation loss function is a label-free loss function constructed based on the distribution consistency of the multiple covariance matrices.

[0014] Furthermore, the method of predicting the covariance matrix of the test using the target student model also includes: In the few-sample adaptation mode, the gradient of the few-sample adaptation loss function is calculated using a preset number of covariance matrices corresponding to the battery under test, and the parameters of the target student model are updated step by step based on the gradient. The covariance matrix under test is then input into the updated target student model for prediction. The few-sample adaptation loss function is a cross-entropy loss function constructed using the preset number of covariance matrices as input and the corresponding battery health status category label as a supervision signal.

[0015] This invention also provides an electronic device, including: a processor, and a memory storing a program, the program including instructions that, when executed by the processor, cause the processor to perform the method described above.

[0016] Compared with the prior art, the present invention has at least the following beneficial effects:

[0017] (1) This invention performs geometric modeling of the covariance matrix of each source domain battery sample on a symmetric positive definite manifold using affine invariant Riemannian metric, and uses a log-Euclidean registration mechanism to unify them to the same registration center. This eliminates the heterogeneity of covariance structure caused by differences in cell production batches, vehicle uses or regional environments, thereby significantly enhancing the model's adaptability to unseen operating conditions or new batches of batteries and effectively improving the generalization ability of SOH estimation in cross-operating condition and cross-batch scenarios.

[0018] (2) This invention strictly maintains the non-Euclidean geometric structure of the health factor covariance matrix by performing a matrix logarithmic mapping to the tangent space on a symmetric positive definite manifold, calculating the weighted Euclidean mean to obtain the registration center, and then returning to the manifold through a matrix exponential mapping. This avoids feature distortion caused by traditional Euclidean space alignment and effectively improves the SOH estimation accuracy for new operating conditions or new batches of batteries.

[0019] (3) This invention effectively addresses the challenge of lacking historical aging data in the early stages of deployment of new batteries by integrating three prediction modes: zero-shot inference, unsupervised adaptation, and small-shot fine-tuning. When there are no target domain samples, zero-shot prediction can be performed directly using the target student model. When there is only the covariance matrix generated by the battery under test without corresponding battery health status category labels, the model parameters can be updated step by step by constructing an unsupervised adaptation loss function guided by distribution consistency. When a very small number of samples with battery health status category labels are provided, fast supervised fine-tuning is achieved based on cross-entropy loss. This multimodal adaptation mechanism significantly alleviates the cold start problem caused by missing labels or scarce data in the early stages of operation of new battery packs.

[0020] (4) This invention divides the registered covariance matrix set into multiple data subsets according to the source domain, and performs a single gradient update on the initial student model in each subset using the first sample group to obtain a temporary student model. Then, based on the inference results of the temporary model on the second sample group, the true label and the output of the teacher model, the meta-loss is constructed. Finally, the global shared parameters of the initial student model are optimized by the sum of the meta-losses of all source domains. This training mechanism enables the target student model to learn an initial state that is easy to adapt quickly. Thus, when faced with the new covariance matrix of each source domain covered in the training, effective single-step parameter updates can be achieved through a small number of samples of the source domain, which significantly improves the generalization ability and online adaptation performance of the model in the multi-source heterogeneous battery scenario. Attached Figure Description

[0021] Figure 1 This is a flowchart of a battery health state estimation method based on symmetric positive definite manifold and log-Euclidean registration; Figure 2 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0022] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments.

[0023] To overcome the challenges of weak generalization ability, covariance structure mismatch, and cold start in existing SOH estimation methods across different operating conditions and batches, such as... Figure 1 As shown, this invention proposes a battery health state estimation method based on symmetric positive definite manifold and log-Euclidean registration, including: Historical charge and discharge data of battery samples contained in each source domain are obtained from multiple source domains. For each battery sample, multiple health factor vectors are extracted based on its historical data from multiple charge-discharge cycles, and the corresponding covariance matrix is ​​calculated based on these multiple health factor vectors. The covariance matrix is ​​a symmetric positive definite matrix, which forms geometric points on a symmetric positive definite manifold and is used to characterize the statistical characteristics of the health factors of the battery sample. The health factor vector includes at least one of the following: the peak position and amplitude of the voltage response curve, the peak position and peak area of ​​the differential voltage curve, the rate of change of internal resistance with the number of cycles, and the area of ​​the capacity increment curve.

[0024] The covariance matrix is ​​obtained by averaging the deviations between the multiple health factor vectors corresponding to the battery sample and the mean of the multiple health factor vectors.

[0025] In this embodiment, the construction of the covariance matrix on a symmetric positive definite manifold includes: For each source domain ( ), from the source domain contained In the historical charge / discharge data of each battery sample, for each battery sample ( Extract the health factor vectors corresponding to multiple charge-discharge cycles. ( ),in This is the charge / discharge cycle number. Indicates the first In the source domain, the first The total number of charge-discharge cycles corresponding to each battery sample. Indicates the first In the source domain, the first In the battery sample, the first The health factor vector corresponding to each charge-discharge cycle. Represents a d-dimensional real space. This represents a d-dimensional real vector, meaning the health factor vector consists of d real values, used to characterize the health status of the battery in the m-th charge-discharge cycle.

[0026] Based on this battery sample Given a vector of health factors, calculate their covariance matrix: ; ; in: : indicates the first In the source domain, the first The covariance matrix of each battery sample; : Represents the mean vector of the health factor vectors of this battery sample; : Matrix transpose; : Correction of degrees of freedom for unbiased estimation.

[0027] It is a symmetric positive definite matrix ( ), forming a geometric point on a symmetrical positive definite manifold, used to characterize the statistical properties of the health factors of the battery sample. Indicates all The set of 1 / 2 symmetric positive definite matrices, where is the matrix dimension.

[0028] Each battery sample generates a unique covariance matrix. A source domain Include One battery sample, therefore including Each covariance matrix.

[0029] The geometric relationship between the covariance matrices on the symmetric positive definite manifold is characterized by an affine invariant Riemannian metric, and is further characterized by a log-Euclidean registration mechanism (i.e., using a reference point). Using a geometric registration method centered on a matrix, the covariance matrices are registered to a unified registration center by mapping the matrix logarithm to the tangent space, calculating the weighted Euclidean mean, and then exponentially mapping back to the manifold, resulting in a set of registered covariance matrices. The process involves registering each covariance matrix to a unified registration center using a logarithmic Euclidean registration mechanism, resulting in a set of registered covariance matrices, including: The geometric relations characterized by the affine invariant Riemannian metric are used to determine the matrix logarithmic and matrix exponential mappings; their mathematical forms are uniquely determined by the affine invariant Riemannian metric (AIRM). Matrix-logarithmic mapping: Defined as satisfying The principal logarithm; where, It is a symmetric positive definite matrix (here referring to the covariance matrix mentioned above). ), Represents the distance under the affine-invariant Riemannian metric. It is the Frobenius norm. and These represent the matrix logarithmic mapping and the matrix exponential mapping, respectively.

[0030] Matrix exponential mapping: for The inverse mapping satisfies (in (It is a symmetric positive definite matrix).

[0031] Together, they constitute the differential homeomorphism from the SPD manifold to the tangent space, which is the geometric basis for subsequent registration operations.

[0032] Select an identity matrix or any covariance matrix of any source domain as a reference point, and project each covariance matrix of each source domain onto the tangent space of the reference point through matrix logarithmic mapping to obtain a tangent space projection matrix corresponding one-to-one with each covariance matrix; specifically: Selecting a reference point It can be the identity matrix I or any covariance matrix of any source domain (e.g., For each covariance matrix Perform logarithmic mapping: ; in: The projection matrix of the tangent space (is a symmetric matrix) (representing a real matrix with d rows and d columns), located at the reference point. Tangent space at the point middle; Indicates at the reference point Symmetric positive definite matrix manifold The tangent space, which consists of all lines intersecting the tangent space. The relevant symmetric matrices are used to map the nonlinear covariance matrix into a vector form that can be linearly operated on.

[0033] The inverse square root of the reference point; and They correspond one-to-one, forming a set of tangent space projection matrices.

[0034] In the tangent space, the weights of each source domain are determined based on the proportion of battery samples contained in each source domain to the total number of samples. Then, a weighted Euclidean mean is calculated for all tangent space projection matrices based on the weights of each source domain. Specifically: Let the first Each source domain contains There are [number] battery samples, and the total number of samples in all source domains is [number]. Then the first The weights of each source domain are: ; Note: The weights here are allocated based on the number of battery samples in the source domain (not the number of cycles), reflecting the impact of data scale on the registration center.

[0035] The formula for calculating the weighted Euclidean mean is: ; in: : Represents the weighted Euclidean mean in the tangent space; Inner layer For the first Local mean of all samples within a source domain; The outer weighting ensures that the large source domain contributes more to the global center.

[0036] The weighted Euclidean mean is transformed back to a symmetric positive definite manifold through matrix exponential mapping to obtain the registration center; specifically: Will Map back via matrix exponent Manifold: ,in: For registration center; : The square root of the reference point; It lies strictly on a symmetric positive definite manifold and is a first-order approximation of the Fréchet mean of all covariance matrices under the affine invariant Riemannian metric.

[0037] Based on the registration center, a registration transformation is performed on each covariance matrix of each source domain to obtain the set of registered covariance matrices. Specifically: For each original covariance matrix Perform the following registration transformation: ; in: : No. In the source domain, the first Covariance matrix of each battery sample after registration; This transformation will Normalization to registration center The reference representation is such that all registered matrices are centered on the identity matrix I (the identity matrix) on the SPD manifold; All registered covariance matrices form a set with a unified geometric center (identity matrix I), equivalent to having the same geometric center in the original manifold. It serves as a public reference center for subsequent training of meta-knowledge distillation models.

[0038] This invention employs a logarithmic registration mechanism that maps the covariance matrix to the tangent space on a symmetric positive definite manifold, calculates the weighted Euclidean mean to obtain the registration center, and then returns to the manifold via a matrix exponential mapping. This mechanism strictly maintains the non-Euclidean geometric structure of the health factor covariance matrix, avoids feature distortion caused by traditional Euclidean space alignment, and effectively improves the SOH estimation accuracy for new operating conditions or new batches of batteries.

[0039] The target student model is obtained by training a meta-knowledge distillation model based on the registered covariance matrix set. The meta-knowledge distillation model includes a teacher model and a student model. In this embodiment, the teacher model uses a 3-layer LSTM network as its backbone structure, and its data processing and network forward propagation process are as follows: First, perform a vectorization transformation on the registered covariance matrix to flatten it into... A one-dimensional feature vector is fed into a three-layer LSTM backbone network as the input for the time step. Each LSTM layer has 128 hidden units to extract the temporal evolution features of battery health status. The output sequence of the last layer of the LSTM backbone network is then aggregated by a global temporal pooling layer to obtain a fixed-length temporal global feature representation. Finally, this global feature representation is input to a fully connected classification head, which outputs a logits vector of dimension C. This vector is then mapped to a C-dimensional probability distribution using a softmax activation function, where C is the preset number of battery health status categories.

[0040] The student model uses a 2-layer GRU network as a lightweight backbone structure. Its data processing and network forward propagation processes are isomorphic to those of the teacher model. The specific steps are as follows: First, the same vectorization transformation is performed on the registered covariance matrix to flatten it. A one-dimensional feature vector is fed into a two-layer GRU backbone network as the input for the temporal step. Each GRU layer has 64 hidden units to extract temporal evolution features of battery health status with lower computational cost. The output sequences of the final layer of the GRU backbone network are then aggregated through the same global temporal pooling layer to obtain a fixed-length temporal global feature representation. Finally, this global feature representation is input to a fully connected classification head with the same structure as the teacher model, outputting a logits vector of dimension C, which is then mapped to a C-dimensional probability distribution using a softmax activation function. Through this isomorphic output space design, the soft label probability distribution dimensions of the student model and the teacher model are strictly aligned to meet the computational requirements of jointly optimizing KL divergence and cross-entropy loss in knowledge distillation.

[0041] Where d is the dimension of the health factor vector extracted from a single battery sample, and C is the preset total number of battery health state categories, which can be set according to engineering needs (e.g., 2, 3, or 5 categories). The number of battery health state categories C is determined by dividing continuous battery health state values ​​into multiple continuous intervals, each interval corresponding to a discrete health state category; for example, battery health states SOH∈[80%,100%] are classified as healthy, SOH∈[60%,80%) as warning, and SOH<60% as failure, for a total of 3 categories. This discretization and grading transforms the model output from a continuous regression task to a classification task, satisfying the calculation requirements of cross-entropy loss and KL divergence in knowledge distillation. Furthermore, the accuracy requirements of different engineering scenarios can be adapted by adjusting the interval boundaries and the number of categories.

[0042] The training of the meta-knowledge distillation model based on the registered covariance matrix set includes: Using the registered set of covariance matrices and the battery health status category labels corresponding to each covariance matrix in the set, the teacher model is trained by minimizing the classification loss. In this embodiment, minimizing the classification loss specifically refers to using the registered covariance matrix as input and the battery health status category label as the supervision target, employing the cross-entropy loss function to calculate the difference between the teacher model's output probability distribution and the true label, and iteratively updating the model parameters through the backpropagation algorithm until the loss converges.

[0043] A student model is constructed, using the registered covariance matrix and the corresponding battery health status category label as training samples. The parameters of the student model are optimized based on the joint distillation loss to obtain the initial student model. The joint distillation loss is composed of a weighted sum of a KL divergence term and a cross-entropy term. The KL divergence term is used to constrain the consistency between the predicted distribution output by the student model and the predicted distribution output by the teacher model. The cross-entropy term is used to measure the cross-entropy loss between the probability distribution of the battery health status category output by the student model and the corresponding battery health status category label.

[0044] In this embodiment, the specific formula for calculating the combined distillation loss is as follows: ; In the formula, Indicates the loss during combined distillation; This represents the weight of the KL divergence term, used to balance the two losses; Represents the KL divergence function; The teacher model is represented by the registered covariance matrix. When used as input, the output is the probability distribution of the predicted battery health status; The student model is represented by the registered covariance matrix. When used as input, the output is the probability distribution of the predicted battery health status; Represents the cross-entropy function; This represents the battery health status category label. In this formula, the KL divergence term is used to constrain the consistency between the predicted distribution of the student model output and the predicted distribution of the teacher model output, and the cross-entropy term is used to measure the classification error between the probability distribution of the battery health status category output by the student model and the corresponding true label. This joint distillation loss allows the student model to inherit the generalization knowledge of the teacher model during the pre-training stage.

[0045] The registered covariance matrix set is divided into multiple data subsets according to the source domain. For each data subset, it is further divided into a first sample group and a second sample group. The gradient of the joint distillation loss is calculated using the first sample group, and a temporary student model is obtained by performing a single gradient descent update on the initial student model based on the gradient and a preset learning rate. Specifically: The meta-learning training phase is then initiated to enhance the model's rapid adaptation capability: the registered covariance matrix set is divided into multiple data subsets according to the source domain. For each data subset v, it is further divided into a first sample group (support set) and a second sample group (query set). The gradient of the joint distillation loss is calculated using the first sample group, and based on this gradient and a preset learning rate (i.e., the inner loop learning rate)... A single gradient descent update is performed on the globally shared parameters of the initial student model to obtain a temporary student model, whose parameter update formula is: ; In the formula, This represents the temporary student model parameters corresponding to the v-th data subset; Represents the globally shared parameters of the initial student model; Indicates the learning rate of the inner loop; Indicates global shared parameters Find the gradient; Indicates globally shared parameters Student model with parameters First sample group The combined distillation loss calculated above.

[0046] The temporary student model is used to perform forward inference on the corresponding second sample group, and the meta-loss is calculated based on the inference results, the corresponding battery health status category labels, and the output of the teacher model. The total loss of the outer meta-optimization is constructed based on the sum of the meta-losses corresponding to all data subsets and the regularization term for the global shared parameters of the initial student model. The global shared parameters of the initial student model are iteratively optimized based on the total loss to obtain the target student model.

[0047] In this embodiment, the formula for calculating the total loss of outer element optimization is: ; ; In the formula, This represents the total loss of the outermost element optimization; The joint distillation loss (which is the joint distillation loss calculated on all registered samples with the current global parameter θ, and has the same form as defined above) is used as an auxiliary supervision term to participate in the outer layer optimization to stabilize the training process. This represents the sum of the meta-losses of all data subsets; To use temporary student model parameters The student model in the second sample group The elemental loss calculated above, where Indicated by The student model mapping function for network parameters is calculated as follows: the second sample group is input into the temporary student model to obtain the inference result. Combined with the corresponding battery health status category label and the teacher model's output for the same input The loss is calculated according to the aforementioned combined distillation loss formula, namely: ; The weights are the original loss values; For regularization weights; The L2 norm squared is used to prevent overfitting and improve cross-domain generalization performance for globally shared parameters.

[0048] Subsequently, based on the total loss optimized by the outermost layer, the global shared parameter θ is iteratively optimized using the outer loop learning rate η until convergence, yielding the target student model. This design enables the model to achieve high-precision battery health state estimation with only 1-3 gradient updates when facing new domains.

[0049] This invention divides the registered covariance matrix set into multiple data subsets according to the source domain. In each subset, a temporary student model is obtained by performing a single gradient update on the initial student model using the first sample group. Then, a meta-loss is constructed based on the inference results of the temporary model on the second sample group, the true labels, and the output of the teacher model. Finally, the global shared parameters of the initial student model are optimized by the sum of the meta-losses of all source domains. This training mechanism enables the target student model to learn an initial state that is easy to adapt to quickly. Thus, when faced with new covariance matrices of each source domain covered in the training, effective single-step parameter updates can be achieved through a small number of samples from that source domain. This significantly improves the model's generalization ability and online adaptation performance in multi-source heterogeneous battery scenarios.

[0050] Obtain the health factor vector of the battery under test and generate the corresponding covariance matrix. Use the registration center to perform a registration transformation on the covariance matrix to obtain the registered covariance matrix under test. Use the target student model to predict the covariance matrix under test and output the battery health status estimate of the battery under test.

[0051] The mode of predicting the covariance matrix of the test using the target student model includes: In zero-sample mode, the covariance matrix to be tested is input into the target student model for prediction.

[0052] The method of using the target student model to predict the covariance matrix of the test also includes: In the unsupervised adaptation mode, the gradient of the unsupervised adaptation loss function is calculated using multiple covariance matrices corresponding to the battery under test, and the parameters of the target student model are updated step by step based on the gradient. The covariance matrix under test is then input into the updated target student model for prediction. The unsupervised adaptation loss function is a label-free loss function constructed based on the distribution consistency of the multiple covariance matrices.

[0053] In this embodiment, the unsupervised adaptation loss function is used to measure the intrinsic consistency among multiple covariance matrices of the battery under test. For example, by minimizing the difference between these covariance matrices and their mean, the model can maintain stable prediction behavior for different cyclic data of the same battery after the model is updated.

[0054] In this embodiment, the unsupervised adaptation loss function is implemented as follows: based on multiple covariance matrices corresponding to the battery under test, the registration centers obtained during the training phase are used to perform registration transformations on each covariance matrix to obtain multiple registered covariance matrices under test; the mean of the multiple registered covariance matrices under test is calculated, and the sum of the differences between each registered covariance matrix under test and the mean is used as the value of the unsupervised adaptation loss function. This loss function minimizes the dispersion of the registered covariance matrices under test, enabling the model to adapt to the aging structure of the battery under test under unlabeled conditions.

[0055] The method of using the target student model to predict the covariance matrix of the test also includes: In the few-sample adaptation mode, the gradient of the few-sample adaptation loss function is calculated using a preset number of covariance matrices corresponding to the battery under test, and the parameters of the target student model are updated step by step based on the gradient. The covariance matrix under test is then input into the updated target student model for prediction. The few-sample adaptation loss function is a cross-entropy loss function constructed using the preset number of covariance matrices as input and the corresponding battery health status category label as a supervision signal.

[0056] In this embodiment, the prediction modes for estimating battery health status using the target student model to measure the covariance matrix include zero-shot mode, unsupervised adaptation mode, and small-shot adaptation mode.

[0057] In zero-sample mode, when there is no available tag data for the battery to be tested, the covariance matrix to be tested is directly input into the target student model for prediction without any updates to the model parameters.

[0058] In the few-sample adaptation mode, when a preset number of samples (e.g., 5) of the batteries to be tested are obtained (including the covariance matrix and its corresponding battery health status category label), fast adaptation is performed to optimize the model's estimation accuracy on the batteries to be tested. Specifically, the parameters of the target student model are updated step by step using the following formula: ; in, : Represents the target student model parameters after a single-step update, used for subsequent prediction of the covariance matrix to be tested; : These are the globally shared parameters obtained after the meta-training converges, used as the initial parameters for adaptation; : Represents the learning rate during the small sample adaptation stage; in this embodiment, the adaptation learning rate is set to the same value as the meta-training outer loop learning rate (i.e., 0.001), which is used to control the step size of single-step parameter updates.

[0059] : Represents the gradient operator; : Represents the small sample adaptation loss function, which is a cross-entropy loss function constructed with the preset number of covariance matrices as input and the corresponding battery health status category labels as supervision signals; : This refers to the training dataset consisting of the predetermined number of samples (e.g., 5) of the batteries to be tested; : Represents the globally shared parameters obtained after convergence of meta-training. This is the mapping function corresponding to the target student model of the network weights.

[0060] Since the target student model's rapid adaptation capability has been optimized through meta-knowledge distillation during the training phase, the aforementioned single-step gradient update can fully adapt the model to the data distribution characteristics of the battery under test, thereby reducing the mean absolute error of battery health state estimation to below 1%. It should be noted that the preset quantity in this embodiment is preferably the covariance matrix data corresponding to 3 to 5 charge-discharge cycles, which minimizes the dependence on labeled data while ensuring the adaptation effect.

[0061] This invention effectively addresses the challenge of lacking historical aging data in the early stages of new battery deployment by integrating three prediction modes: zero-shot inference, unsupervised adaptation, and few-shot fine-tuning. When there are no target domain samples, zero-shot prediction can be performed directly using the target student model. When only the covariance matrix generated by the battery itself is available without corresponding battery health status category labels, an unsupervised adaptation loss function guided by distribution consistency can be constructed to drive single-step updates of model parameters. When a very small number of samples with battery health status category labels are provided, rapid supervised fine-tuning is achieved based on cross-entropy loss. This multimodal adaptation mechanism significantly alleviates the cold start problem caused by missing labels or scarce data in the early stages of new battery pack operation.

[0062] To verify the effectiveness of the method of the present invention in scenarios with scarce data, a specific example is given below: 1. Data Source Source domain data includes: Source Domain 1: Taxi fleet (500 vehicles / 2 years); Source Domain 2: Bus fleet (200 vehicles / 3 years); Source Domain 3: Energy Storage Power Station (50 clusters / 1 year).

[0063] The battery data to be tested comes from a fleet of private cars (50 vehicles / 1 month of mixed operating conditions), which is the target domain referred to in this embodiment. It is used to simulate the application scenarios of the three prediction modes described in this invention: zero-sample, small-sample, and unsupervised adaptation.

[0064] 2. Model Configuration The teacher model uses a 3-layer LSTM (128 hidden units), and the student model uses a 2-layer GRU (64 hidden units); the weights of the KL divergence term in the joint distillation loss are... The inner loop learning rate is 0.7. The outer loop learning rate η is 0.01, and the meta-loss weights are 0.001. The regularized weight is 0.5. for .

[0065] 3. Verification Results After generating the covariance matrix for each battery under test (target domain) and performing a registration center transformation, prediction is performed using the target student model: Zero-sample mode: Mean absolute error is 1.75%; Small sample fitting mode: Mean absolute error is 0.72%, RMSE is 0.91%; Unsupervised adaptation mode: Mean absolute error is 1.21%.

[0066] The results show that the proposed method can achieve high-precision battery health status estimation under the condition of scarce battery data.

[0067] This invention performs geometric modeling of the covariance matrix of each source domain battery sample on a symmetric positive definite manifold using affine invariant Riemannian metrics, and uses a log-Euclidean registration mechanism to unify them to the same registration center. This eliminates the heterogeneity of covariance structure caused by differences in cell production batches, vehicle applications, or regional environments, thereby significantly enhancing the model's adaptability to unseen operating conditions or new batches of batteries and effectively improving the generalization ability of SOH estimation in cross-operating condition and cross-batch scenarios.

[0068] This invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of this invention.

[0069] The present invention also provides a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of the present invention.

[0070] This invention also provides a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of this invention.

[0071] refer to Figure 2 The present invention will now describe a structural block diagram of an electronic device that can serve as a server or client in embodiments of the present invention, which is an example of a hardware device that can be applied to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0072] like Figure 2As shown, the electronic device includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 402 or a computer program loaded from a storage unit 408 into a random access memory (RAM) 403. The RAM 403 may also store various programs and data required for the operation of the electronic device. The computing unit 401, ROM 402, and RAM 403 are interconnected via a bus 404. An input / output (I / O) interface 405 is also connected to the bus 404.

[0073] Multiple components in the electronic device are connected to I / O interface 405, including: input unit 406, output unit 407, storage unit 408, and communication unit 409. Input unit 406 can be any type of device capable of inputting information into the electronic device. Input unit 406 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 407 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 408 may include, but is not limited to, disks and optical discs. Communication unit 409 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, wireless communication transceivers, and / or chipsets, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0074] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the methods and processes described above. For example, in some embodiments, the method embodiments of the present invention may be implemented as a computer program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program may be loaded and / or installed on an electronic device via ROM 402 and / or communication unit 409. In some embodiments, the computing unit 401 may be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0075] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0076] In the context of embodiments of the present invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0077] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0078] Furthermore, in this invention, descriptions involving terms such as "first," "second," and "a" are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0079] In this invention, unless otherwise explicitly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection or an electrical connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0080] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

Claims

1. A method for estimating battery health status based on symmetric positive definite manifold and logarithmic Euclidean registration, characterized in that, include: Historical charge and discharge data of battery samples contained in each source domain are obtained from multiple source domains. For each battery sample, multiple health factor vectors are extracted based on its historical data from multiple charge-discharge cycles, and the corresponding covariance matrix is ​​calculated based on these multiple health factor vectors. The covariance matrix is ​​a symmetric positive definite matrix, which forms geometric points on a symmetric positive definite manifold and is used to characterize the statistical characteristics of the health factors of the battery sample. The geometric relationship between the covariance matrices on the symmetric positive definite manifold is characterized by the affine invariant Riemannian metric, and the covariance matrices are registered to a unified registration center by the log-Euclidean registration mechanism to obtain the set of registered covariance matrices. The target student model is obtained by training a meta-knowledge distillation model based on the registered covariance matrix set. The meta-knowledge distillation model includes a teacher model and a student model. Obtain the health factor vector of the battery under test and generate the corresponding covariance matrix. Use the registration center to perform a registration transformation on the covariance matrix to obtain the registered covariance matrix under test. Use the target student model to predict the covariance matrix under test and output the battery health status estimate of the battery under test.

2. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The health factor vector includes at least one of the following: the peak position and amplitude of the voltage response curve, the peak position and peak area of ​​the differential voltage curve, the rate of change of internal resistance with the number of cycles, and the area of ​​the capacity increment curve.

3. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The covariance matrix is ​​obtained by averaging the deviations between the multiple health factor vectors corresponding to the battery sample and the mean of the multiple health factor vectors.

4. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The process involves registering each covariance matrix to a unified registration center using a logarithmic Euclidean registration mechanism, resulting in a set of registered covariance matrices, including: The geometric relations characterized by affine invariant Riemannian metrics are used to determine the matrix logarithmic mapping and matrix exponential mapping; Select an identity matrix or any covariance matrix of any source domain as a reference point, and project each covariance matrix of each source domain onto the tangent space of the reference point through matrix logarithmic mapping to obtain a tangent space projection matrix that corresponds one-to-one with each covariance matrix. In the tangent space, the weights of each source domain are determined based on the proportion of battery samples contained in each source domain to the total number of samples, and the weighted Euclidean mean is calculated for all tangent space projection matrices based on the weights of each source domain. The weighted Euclidean mean is transformed back to a symmetric positive definite manifold through matrix exponential mapping to obtain the registration center; Based on the registration center, a registration transformation is performed on each covariance matrix of each source domain to obtain the set of registered covariance matrices.

5. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The training of the meta-knowledge distillation model based on the registered covariance matrix set includes: Using the registered set of covariance matrices and the battery health status category labels corresponding to each covariance matrix in the set, the teacher model is trained by minimizing the classification loss. A student model is constructed, using the registered covariance matrix and the corresponding battery health status category label as training samples. The parameters of the student model are optimized based on the joint distillation loss to obtain the initial student model. The registered covariance matrix set is divided into multiple data subsets according to the source domain. For each data subset, it is divided into a first sample group and a second sample group. The gradient of the joint distillation loss is calculated using the first sample group, and a temporary student model is obtained by performing a single gradient descent update on the initial student model based on the gradient and the preset learning rate. The temporary student model is used to perform forward inference on the corresponding second sample group, and the meta-loss is calculated based on the inference results, the corresponding battery health status category labels, and the output of the teacher model. The total loss of the outer meta-optimization is constructed based on the sum of the meta-losses corresponding to all data subsets and the regularization term for the global shared parameters of the initial student model. The global shared parameters of the initial student model are iteratively optimized based on the total loss to obtain the target student model.

6. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 5, characterized in that, The joint distillation loss is composed of a weighted sum of a KL divergence term and a cross-entropy term. The KL divergence term is used to constrain the consistency between the predicted distribution output by the student model and the predicted distribution output by the teacher model. The cross-entropy term is used to measure the classification error between the probability distribution of the battery health status category output by the student model and the corresponding battery health status category label.

7. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The mode of predicting the covariance matrix of the test using the target student model includes: In zero-sample mode, the covariance matrix to be tested is input into the target student model for prediction.

8. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The method of using the target student model to predict the covariance matrix of the test also includes: In the unsupervised adaptation mode, the gradient of the unsupervised adaptation loss function is calculated using multiple covariance matrices corresponding to the battery under test, and the parameters of the target student model are updated step by step based on the gradient. The covariance matrix under test is then input into the updated target student model for prediction. The unsupervised adaptation loss function is a label-free loss function constructed based on the distribution consistency of the multiple covariance matrices.

9. The battery health state estimation method based on symmetric positive definite manifold and logarithmic Euclidean registration as described in claim 1, characterized in that, The method of using the target student model to predict the covariance matrix of the test also includes: In the few-sample adaptation mode, the gradient of the few-sample adaptation loss function is calculated using a preset number of covariance matrices corresponding to the battery under test, and the parameters of the target student model are updated step by step based on the gradient. The covariance matrix under test is then input into the updated target student model for prediction. The few-sample adaptation loss function is a cross-entropy loss function constructed using the preset number of covariance matrices as input and the corresponding battery health status category label as a supervision signal.

10. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 9.

Images