A method and system for estimating the state of oxygen (SOH) of batteries across chemical systems
By employing a single-stage end-to-end technology loop of parallel multi-scale CNN encoding, homologous complementary dual embedding fusion, and sparse MoE dynamic routing, the high maintenance cost and prediction reliability issues of SOH estimation for cross-chemical systems batteries are resolved, achieving high-precision, low-cost, and reliable battery health status assessment.
Patent Information
- Application Number
- CN202611618628.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-20
- Publication Date
- 2026-11-17
AI Technical Summary
Existing battery SOH estimation techniques cannot be adapted to multiple chemical systems within the same framework, resulting in high maintenance costs, insensitivity of feature representation to chemical systems, insufficient multi-scale time-series feature extraction, lack of quantitative evaluation of prediction reliability, and complex MoE training process, making it difficult to deploy in embedded BMS.
A single-stage end-to-end technical loop is adopted, which includes parallel multi-scale CNN encoding, homologous complementary dual embedding fusion, sparse MoE dynamic routing, and gradient isolation confidence calibration. Through hierarchical data preprocessing, multi-scale feature extraction, chemical system embedding, and confidence evaluation, unified high-precision estimation across chemical systems is achieved.
It achieves high-precision estimation of SOH in batteries across chemical systems, reduces maintenance costs, improves the reliability of predictions and safety decision-making capabilities, simplifies the engineering implementation of MoE, and is suitable for embedded BMS deployment.
Smart Images

Figure CN123112751A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of new energy power system technology, and in particular to a method and system for estimating the state of charge (SOH) of batteries across chemical systems. Background Technology
[0002] This invention addresses the battery state of health (SOH) estimation scenario, specifically involving online assessment of the health status of lithium-ion batteries (including various chemical systems such as NCM ternary lithium batteries and LFP lithium iron phosphate batteries) in applications such as electric vehicles, energy storage systems, and power tools. Batteries with different chemical systems exhibit significant differences in charge-discharge curve morphology, internal resistance characteristics, and capacity decay modes. Traditional methods typically require building separate models for each chemical system, resulting in high maintenance costs and the inability to achieve unified platform management.
[0003] Existing battery SOH estimation techniques face the following critical issues that urgently need to be addressed: 1. Difficulty in unified modeling across chemical systems. NCM and LFP have fundamental differences in charge-discharge plateau voltage, internal resistance growth mode, and capacity decay trajectory. Existing models cannot be adapted to multiple chemical systems within the same framework. It is necessary to model, collect data, and train models separately for each chemical system, resulting in high maintenance costs.
[0004] 2. Lack of explicit injection of chemical system differentiation information. Existing deep models do not explicitly encode battery type and chemical system identifiers as structured information into feature representations, resulting in weak model differentiation ability among different chemical systems and insensitivity of feature representations to chemical systems.
[0005] 3. Insufficient extraction of multi-scale temporal features. Existing convolutional or recurrent encoders typically use a single time scale for modeling, making it difficult to capture the short, medium, and long-term dynamic features contained in the charge-discharge curve in parallel.
[0006] 4. Lack of quantitative assessment of prediction reliability. Existing methods only output SOH estimates, which cannot identify high-uncertainty samples when the input data contains noise, distribution bias, or the model encounters cognitive blind spots. This makes it impossible to provide a reliable basis for battery management systems and poses a risk to safety decisions.
[0007] 5. The training process for MoE (Expert Model) is extremely complex and difficult to engineer. In the battery field, MoE solutions require multi-stage gradient isolation training, multi-checkpoint version management, and independent hyperparameter tuning at each stage, resulting in high engineering complexity and hindering deployment in industrial scenarios such as embedded BMS.
[0008] Therefore, a new method and system for estimating the state of oxygen (SOH) of batteries across chemical systems is needed to achieve unified and high-precision estimation across chemical systems in a single model and a single forward computation path, and to simultaneously output the confidence level associated with the prediction error to support safety decisions. Summary of the Invention
[0009] This invention addresses the shortcomings of existing technologies by providing a method and system for estimating the state of oxygen (SOH) of batteries across chemical systems.
[0010] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for estimating the state of harm (SOH) of a battery across chemical systems, comprising the following steps: Acquire time-series aging data of batteries containing at least two different chemical systems, hierarchically aggregate fine-grained model identifiers under each chemical system to generate chemical system level identifiers, and perform outlier filtering and sequence length unification on SOH tags to construct a normalized time-series input tensor paired with SOH tags dataset; Multi-scale temporal features of the temporal input tensor are extracted to generate a global temporal feature vector. The chemical system level identifier is mapped to a chemical system embedding vector through an independently parameterized embedding layer. The global temporal feature vector and the chemical system embedding vector are fused to generate a fused feature vector. The fused feature vector is input into the hybrid expert model. The routing probability of each expert is calculated by the router. The K experts with the highest routing probability are selected for forward calculation. The SOH prediction values output by each activated expert are weighted and fused to obtain the final SOH prediction value. After performing gradient truncation on the fused feature vector, an independent confidence estimation head is input, and a confidence score characterizing the reliability of the final SOH prediction value is output. The gradient truncation operation is used to truncate the backpropagation gradient, so that the training error of the confidence estimation head only updates the parameters of the confidence head itself. Obtain the chemical system type of the target battery corresponding to the time-series input tensor, determine the corresponding rejection threshold based on the chemical system type, and compare the confidence score with the rejection threshold; if the confidence score is lower than the rejection threshold, refuse to output the final SOH prediction value, otherwise output the final SOH prediction value.
[0011] Furthermore, the hierarchical aggregation specifically involves mapping the fine-grained identifiers of different cell models in the original data to coarse-grained identifiers at the chemical system level, wherein the chemical system includes at least the lithium iron phosphate (LFP) system and the nickel-cobalt-manganese ternary NCM system. The outlier filtering specifically involves: grouping by chemical system and removing outliers that satisfy the condition |SOH- group mean|>3. Outlier labels for group standard deviations and their corresponding time series data.
[0012] Furthermore, extracting multi-scale temporal features specifically includes: The temporal input tensor is convolutionally encoded using at least two parallel 1D-CNN branches, with each branch configured with a convolution kernel of a different size, to capture the charging and discharging dynamic features at different time scales. After each branch undergoes a stepwise dimensionality reduction process including residual connections, the outputs of each branch are concatenated and fused along the feature dimension, and then global pooling is used to obtain the global temporal feature vector.
[0013] Furthermore, the parallel 1D-CNN branches are specifically three branches with convolutional kernel sizes of 3, 5, and 7, respectively; The dimensionality reduction path for each branch is 256→128→32→16. Each dimensionality reduction block contains two fully connected layers and residual connections. The residual connections directly superimpose the input of the current dimensionality reduction block onto the output of the dimensionality reduction block.
[0014] Furthermore, the chemical system embedding vector is generated as follows: The battery type identifier and the chemical system level identifier are mapped to a low-dimensional type embedding vector and a low-dimensional chemical system embedding vector through two independent parameterized embedding layers, respectively. The global temporal feature vector, the low-dimensional type embedding vector, and the low-dimensional chemical system embedding vector are concatenated and then subjected to a projection layer nonlinear transformation to generate the fused feature vector. The low-dimensional type embedding vector and the low-dimensional chemical system embedding vector form a complementary verification in the projection fusion layer.
[0015] Furthermore, the router is a data-driven multilayer perceptron, whose output is scaled by temperature parameters and then used to generate the routing probability distribution of each expert through the Softmax function; When selecting the K experts with the highest routing probability, K=2; Each of the activated experts is a lightweight multilayer perceptron with a consistent structure but independent parameters. The weighted fusion is achieved by summing the SOH predictions of the K experts according to their respective routing probabilities.
[0016] Furthermore, the confidence estimation head is trained using a binary cross-entropy calibration loss, the soft label of which is generated based on a smoothing indicator of the SOH prediction error relative to a calibration error threshold. When the prediction error is less than the calibration error threshold, the soft tag tends to 1; when the prediction error is greater than the calibration error threshold, the soft tag tends to 0; when the prediction error is equal to the calibration error threshold, the soft tag is accurate to 0.5. Furthermore, in the first N rounds of model training, the backpropagation of the calibration loss is masked, and only the SOH prediction main task loss is optimized.
[0017] Furthermore, the method for determining the rejection threshold is as follows: On the validation set, candidate confidence thresholds are traversed independently for each chemical system type. With maximizing the F1 score as the optimization objective, the candidate confidence threshold corresponding to maximizing the F1 score is selected as the exclusive rejection threshold for that chemical system type. Specifically, samples with SOH prediction errors higher than a preset error threshold are considered positive, and rejection decisions with confidence scores lower than the current candidate threshold are considered positive detections.
[0018] In a second aspect, the present invention also provides a cross-chemical system battery SOH estimation system for implementing the cross-chemical system battery SOH estimation method of the first aspect, comprising: The data preprocessing module is used to acquire time-series aging data of batteries containing at least two different chemical systems, hierarchically aggregate the fine-grained model identifiers under each chemical system to generate chemical system-level identifiers, filter outliers and unify the sequence length of SOH tags, and construct a time-series input tensor. The feature encoding and fusion module is used to extract multi-scale temporal features from the temporal input tensor to generate a global temporal feature vector, and to map the chemical system level identifier to a chemical system embedding vector through an independently parameterized embedding layer. The global temporal feature vector and the chemical system embedding vector are then fused to generate a fused feature vector. The SOH prediction module includes a hybrid expert model, which is used to input the fused feature vector into the router to calculate the routing probability of each expert, select the K experts with the highest routing probability for forward calculation, and weight and fuse the SOH prediction values output by each activated expert to obtain the final SOH prediction value. The confidence assessment module includes an independent confidence estimation head and a gradient truncation unit. It is used to input the fused feature vector into the confidence estimation head after performing a gradient truncation operation on the gradient truncation unit, and output a confidence score that characterizes the reliability of the final SOH prediction value. The gradient truncation operation is used to truncate the backpropagation gradient, so that the training error of the confidence assessment module only updates the parameters of the confidence head itself. The rejection output module is used to obtain the chemical system type of the target battery corresponding to the time-series input tensor, determine the corresponding rejection threshold based on the chemical system type, and compare the confidence score with the rejection threshold. If the confidence score is lower than the rejection threshold, the final SOH prediction value is rejected; otherwise, the final SOH prediction value is output. Furthermore, in the rejection output module, the rejection threshold is determined as follows: On the validation set, candidate confidence thresholds are traversed independently for each chemical system type. With maximizing the F1 score as the optimization objective, the candidate confidence threshold corresponding to maximizing the F1 score is selected as the exclusive rejection threshold for that chemical system type. Specifically, samples with SOH prediction errors higher than a preset error threshold are considered positive, and rejection decisions with confidence scores lower than the current candidate threshold are considered positive detections.
[0019] In summary, compared with the prior art, the beneficial effects of the above technical solution are: This invention constructs a single-stage end-to-end technical closed loop consisting of "parallel multi-scale CNN encoding - homologous complementary dual embedding fusion - sparse MoE dynamic routing - gradient isolation confidence calibration - chemical system adaptive rejection." Within a unified model framework, it simultaneously adapts to both LFP and NCM chemical systems, eliminating the high maintenance costs of traditional methods that require separate modeling, data collection, and model training for each system. The parallel multi-scale CNN and dual embedding fusion enable the model to explicitly perceive differences in chemical systems at the feature level and simultaneously capture short-term fluctuations, medium-term slopes, and long-term decay trends. Sparse MoE activates the most proficient expert subset in a data-driven manner to achieve dynamic division of labor, ensuring accuracy. Simultaneously, it reduces inference computation; the gradient isolation confidence calibration mechanism outputs a confidence score highly correlated with the true prediction error without interfering with the feature representation of the main prediction network. Combined with an adaptive rejection threshold based on the chemical system type, the BMS can actively reject low-reliability samples, significantly improving the SOH estimation accuracy and safety decision-making capability of effective samples. At the same time, the single-stage end-to-end joint training outputs a single model checkpoint, avoiding the engineering complexity of traditional MoE multi-stage gradient isolation training. This makes the architecture feasible for practical deployment on an embedded BMS platform, truly achieving a balance of high accuracy, high reliability, and low maintenance cost in cross-chemical system battery SOH estimation. Attached Figure Description
[0020] Figure 1 This is a flowchart of the method according to an embodiment of the present invention; Figure 2 The SOH curve of LFP-HUST; Figure 3 The SOH curve of NCM-Tshinghua; Figure 4 An estimated SOH diagram for LFP-HUST battery cells; Figure 5 A diagram showing the error distribution for SOH estimation in LFP-HUST cells; Figure 6A diagram showing the estimated SOH of NCM-Tshinghua battery cells; Figure 7 Error distribution diagram for SOH estimation of NCM-Tshinghua battery cells. Detailed Implementation
[0021] The principles and features of the present invention are described below with reference to all the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.
[0022] This invention discloses a method and system for estimating the state of oxygen (SOH) of batteries across chemical systems.
[0023] Reference Figure 1 to Figure 7 A method for estimating the state of matter (SOH) of a battery across chemical systems includes the following steps: Acquire time-series aging data of batteries containing at least two different chemical systems, hierarchically aggregate fine-grained model identifiers under each chemical system to generate chemical system level identifiers, and perform outlier filtering and sequence length unification on SOH tags to construct a normalized time-series input tensor paired with SOH tags dataset; Multi-scale temporal features of the temporal input tensor are extracted to generate a global temporal feature vector. The chemical system level identifier is mapped to a chemical system embedding vector through an independently parameterized embedding layer. The global temporal feature vector and the chemical system embedding vector are fused to generate a fused feature vector. The fused feature vector is input into the hybrid expert model. The routing probability of each expert is calculated by the router. The K experts with the highest routing probability are selected for forward calculation. The SOH prediction values output by each activated expert are weighted and fused to obtain the final SOH prediction value. After performing gradient truncation on the fused feature vector, an independent confidence estimation head is input, and a confidence score characterizing the reliability of the final SOH prediction value is output. The gradient truncation operation is used to truncate the backpropagation gradient, so that the training error of the confidence estimation head only updates the parameters of the confidence head itself. Obtain the chemical system type of the target battery corresponding to the time-series input tensor, determine the corresponding rejection threshold based on the chemical system type, and compare the confidence score with the rejection threshold; if the confidence score is lower than the rejection threshold, refuse to output the final SOH prediction value, otherwise output the final SOH prediction value.
[0024] The following is a detailed explanation of each of the above contents: S1: Hierarchical data preprocessing for multi-chemical systems.
[0025] Acquire cycle aging data of batteries containing at least two different chemical systems, unify the fine-grained multi-model identifiers under each chemical system into coarse-grained identifiers at the chemical system level, perform outlier filtering on SOH tags grouped by chemical system, truncate or pad the time series data of each cycle to a uniform length, and construct a paired dataset of normalized time series input tensors and SOH tags.
[0026] Hierarchical aggregation specifically includes: The fine-grained identifiers of different cell models in the original data are uniformly mapped to coarse-grained identifiers at the chemical system level. The chemical system includes at least the lithium iron phosphate (LFP) system and the nickel-cobalt-manganese (NCM) ternary system. The outlier filtering specifically involves: grouping by chemical system and removing outliers that satisfy the condition |SOH- group mean|>3. Outlier labels for group standard deviations and their corresponding time series data.
[0027] Specifically, hierarchical chemical system identifier mapping: The original data contains a large number of fine-grained battery model identifiers. For example, the LFP and NCM systems each contain multiple cell models (with different cell manufacturers and cell capacities), which can be regarded as fine-grained model identifiers.
[0028] This step unifies the fine-grained multi-model identifiers into coarse-grained identifiers at the chemical system level, mapping them to two chemical system level identifiers (LFP=0, NCM=1).
[0029] The core design of this hierarchical mapping strategy is to compress the granularity of embedding learning from the fine-grained model that is prone to overfitting to the chemical system level with a sufficient sample size and the ability to generalize across models. This allows the embedding layer to focus on the essential differences at the chemical system level, while retaining the battery type embedding as a complementary signal.
[0030] 3-sigma outlier filtering: Outliers in the SOH labels are filtered using 3-sigma (three standard deviations) based on chemical system grouping, eliminating those with |SOH-group mean| > 3. Sigma's anomaly labels and their corresponding time-series data.
[0031] Unified sequence length and pairing construction: The time series data of each cycle is truncated or zero-padded to a fixed length T to construct a pair between the input sample X and the SOH label.
[0032] Dataset partitioning and Min-Max normalization: The dataset is partitioned into training, validation, and test sets according to a preset ratio. Training set statistics are used to perform dimension-wise normalization of the time-series features. Min-Max Normalized to the [0,1] interval, as shown in the following formula: ; in, XIt is the original temporal characteristic. The features are normalized. and These are the statistical minimum and maximum values of each dimension of features on the training set.
[0033] S2: Parallel multi-scale CNN temporal coding.
[0034] The normalized temporal input is fed into at least two parallel 1D-CNN branches. Each branch is configured with convolutional kernels of different sizes to capture charging and discharging temporal features at different time scales. After the branches are progressively reduced in dimensionality with residual connections, they are spliced and fused in the feature dimension, and then global pooling is used to obtain the global temporal feature vector.
[0035] In this embodiment, the parallel 1D-CNN branches are specifically three branches with convolution kernel sizes of 3, 5, and 7, respectively, to capture short-term voltage fluctuations, medium-term slope changes, and long-term decay trends; each branch contains multiple consecutive dimensionality reduction blocks, and each dimensionality reduction block contains two fully connected layers and introduces residual connections.
[0036] For the k-th branch, perform temporal convolution first: ; in, The output of the k-th branch convolution; Conv1D The filling method is same Padding (padding=(k-1) / 2) ensures that the temporal length T remains unchanged after convolution; X The time-series input tensor output in step S1; BatchNorm , ReLU and Dropout These are batch normalization, activation function, and random deactivation operation, respectively.
[0037] Each branch contains three consecutive dimensionality reduction blocks (DimReductionBlock), and the dimensionality reduction path of each branch is 256→128→32→16. Each dimensionality reduction block contains two fully connected layers. Linear + LayerNorm + ReLU + Dropout It also introduces residual connections, which directly superimpose the input of the current dimensionality reduction block onto the output of the dimensionality reduction block.
[0038] For the first dimensionality reduction block, the operation is as follows: ; The residual connections ensure stable propagation of deep gradients. After three dimensionality reduction blocks, each branch outputs... k=1,2,3. l Number the dimensionality reduction blocks; For the first k The first branch l Input features of each dimension reduction block; This refers to the output features of the dimensionality reduction block; Linear , LayerNorm , ReLU and Dropout These are, respectively, fully connected layer, layer normalization, activation function, and random deactivation operation.
[0039] The outputs of the three branches are concatenated along the feature dimension, fused by linear projection, and then global mean pooling is performed along the temporal dimension.
[0040] Cross-branch fusion: The outputs of the three branches are concatenated along the feature dimension, as follows: ; In the above formula, , , These are the output features of the three branches; This is the fused feature obtained by splicing the three elements along the feature dimension; "||" indicates the splicing operation along the feature dimension.
[0041] Then, the mixture is fused via linear projection. ; In the above formula, the fully connected layer The spliced features are projected into 32 dimensions, and then randomly deactivated using Dropout. It is the temporal feature sequence after linear projection.
[0042] Global mean pooling: Taking the mean along the time series dimension to obtain a global time series representation. .
[0043] In the above formula, This represents the global temporal feature vector after taking the mean along the temporal dimension; T is the temporal length, and t is the temporal step number. hseq [:, t ,:] represents the time-series feature sequence The eigenvector corresponding to the t-th time step.
[0044] S3: Homologous complementary dual embedding feature fusion.
[0045] The battery type identifier and chemical system identifier are mapped to low-dimensional embedding vectors through independent parameterized embedding layers. The low-dimensional embedding vectors are then concatenated with the global time-series feature vectors output in step S2 and nonlinearly fused through a projection layer to form a fused feature vector.
[0046] Specifically, although the battery type identifier and the chemical system identifier are from the same source, they adopt an embedding layer with independent parameters. The type embedding is learned from the perspective of battery classification, while the chemical system embedding is learned from the perspective of chemical system aggregation. The two form complementary information in the projection fusion layer, which verifies each other and enhances the robustness of the representation of differences in chemical systems. The fused feature vector is simultaneously sent to the router and expert in step S4 and the confidence head in step S5 to ensure that all prediction paths have the ability to perceive chemical systems.
[0047] It should be noted that: For type embedding: Identify the battery type. Mapped to a low-dimensional vector through the embedding layer: ; In the above formula, Battery type identifier; For type embedding layer, The size of the value range for the battery type identifier; Let be the dimension of the low-dimensional vector obtained by the mapping; It is a low-dimensional type of embedding vector.
[0048] For chemical system embedding (Model Embedding): the hierarchically mapped chemical system identifiers. (LFP=0, NCM=1) Mapped through independent embedding layers: ; In the above formula, This serves as the level identifier for the chemical system after hierarchical mapping. As an embedding layer for chemical systems; The size of the value space for identifying the level of a chemical system; For low-dimensional chemical system embedding vectors.
[0049] That is, although type embedding and chemical system embedding share the same origin (both characterizing NCM / LFP chemical systems), they employ embedding layers with independent parameters, enabling the model to learn chemical system representations from two complementary perspectives. Type embedding learns from the perspective of battery classification, while chemical system embedding learns from the perspective of chemical system aggregation. The values are 0 (LFP) and 1 (NCM). The two form complementary information in the fusion layer, which verifies each other and enhances the robustness of characterizing differences in chemical systems.
[0050] Feature concatenation and projection fusion: The 32-dimensional pooled features output by the CNN encoder are concatenated with two 8-dimensional embedding vectors, and then normalized by layers. ReLU Activation and Dropout The fully connected layer is projected back to 32 dimensions.
[0051] The splicing characteristics are: ; In the above formula, This is the 32-dimensional global temporal feature vector output by the CNN encoder. and These are respectively an 8-dimensional low-dimensional type embedding vector and a low-dimensional chemical system embedding vector. The 48-dimensional splicing feature is obtained by splicing the three elements.
[0052] The characteristics after fusion are: .
[0053] In the above formula, It features 48-dimensional splicing characteristics, through a fully connected layer. The projection is 32-dimensional, and then successively normalized by layers. , Activation and Random inactivation, resulting It is a 32-dimensional fused feature vector.
[0054] S4: Data-driven sparse MoE routing and multi-expert collaborative prediction.
[0055] The lightweight router driven by the fused feature vector input data calculates the routing probability of each expert. The Top-K sparse selection mechanism activates only the K experts with the highest routing probabilities. Each activated expert independently outputs the SOH prediction value, and the final SOH prediction value is obtained by weighting and fusing them according to the routing probability.
[0056] Specifically, the router is a purely data-driven multilayer perceptron. The output is scaled by an adjustable temperature parameter and then passed through Softmax to obtain the routing probability distribution of each expert. The routing decision is entirely driven by data learning and does not rely on manually hard-coded rules. Top-K sparse selection only activates the K experts with the highest weights, and the other experts do not participate in the forward computation.
[0057] Data-driven routing computation: fusing features The input is a lightweight router called SimpleRouter. This router is a purely data-driven, Layer 2 router. MLP : ; In the above formula, This refers to the two-layer, multi-layer sensor corresponding to the router. The original routing score vector output by the router has a dimension equal to the number of experts, E.
[0058] The specific structure is as follows: Linear( 32→16 )→ReLU→Dropout→Linear (16→E), output with adjustable temperature parametertau (Default 1.0) Perform temperature scaling: ; In the above formula, max(tau, 0.1 ) Indicates taking tau The larger of 0.1 and 0.1 is used to avoid the temperature parameter being too small and causing numerical instability. This is the temperature-scaled routing score vector.
[0059] Then through Softmax The routing probabilities of each expert are obtained: ; Where E represents the number of experts. Temperature parameter. tau The hardware and software capabilities of the router can be dynamically adjusted. tau The higher the value, the more uniform the probability distribution (soft router); tau The lower the probability, the more concentrated it is on a single expert (hard router). p This represents the routing probability distribution vector for each expert.
[0060] Top-K sparse expert selection: Select the K experts with the highest routing probability from E experts (K=2, and K<=E): ; in, It is the route weight of the selected expert. Select the expert index. p To route probabilities for each expert, only these K experts are activated for forward computation, while the remaining experts do not participate, thus reducing the computational load of inference.
[0061] Lightweight expert independent prediction: Each expert is an independent three-layer lightweight MLP structure, with consistent structure but no shared parameters. j Independent SOH predictions by individual experts: ; In the above formula, j Number the experts; This is the SOH prediction value output by the j-th expert. This expert was successively... Linear (32→32) ReLU activation, Linear (32→16) ReLU Activation and Linear (16→1) Layer-by-layer mapping transforms the 32-dimensional fusion features into single-dimensional SOH prediction values. Each expert processes the fusion features of the same sample in parallel and outputs SOH prediction values independently.
[0062] Weighted fusion: The predictions of the K selected Top-K experts are weighted and fused according to their corresponding route probabilities. ; in, K For the Top-K sparse selection number, For the first i The route weight of each selected expert Predict its SOH; This represents the final SOH prediction value obtained by weighted fusion based on routing probability.
[0063] S5: Gradient isolation learning confidence calibration.
[0064] The fused feature vector is fed into an independent confidence estimation head after gradient truncation, and the confidence score associated with the SOH prediction error is output. The gradient truncation operation ensures that the backpropagation gradient of the confidence calibration only updates the parameters of the confidence head itself and does not propagate back to the upstream shared encoder and fusion layer.
[0065] Specifically, the confidence estimation head is an independent lightweight multilayer perceptron, and its output is mapped to the [0,1] interval via a sigmoid function. It is trained using a binary cross-entropy calibration loss, where the soft label is a smooth indicator of the SOH prediction error relative to a calibration error threshold. The soft label tends to 1 when the prediction error is less than the calibration error threshold, and tends to 0 when the prediction error is greater than the calibration error threshold. The soft label is precisely 0.5 when the prediction error equals the calibration error threshold.
[0066] Gradient cutoff (Detach): This method fuses features. After being processed by PyTorch's detach() operation, the confidence header is fed into the confidence header. This operation truncates the backpropagation gradient, ensuring that the training gradient of the confidence header only updates the parameters of the confidence header itself; error signals during the confidence learning process are never propagated back to the upstream shared CNN encoder, dual embedding fusion layer, and router.
[0067] The learning objective of the confidence head is not to improve its confidence score by changing the upstream feature representation, but to learn to read and judge the reliability of the feature representation already produced by the upstream. Without this gradient isolation design, the master prediction error is large in the early training stage, and the BCE calibration loss will drive the confidence score to 0. The backpropagation of the calibration gradient will destroy the prediction feature representation already learned by the shared encoder, causing the master regression task and the confidence task to interfere with each other, and ultimately the performance of both will deteriorate simultaneously.
[0068] Confidence estimation: Independent lightweight MLP confidence header structure is : Linear (32→32)→ ReLU → Dropout → Linear (32→16)→ ReLU → Linear (16→1), output viaSigmoid Mapping to the [0,1] interval: ; in, c The confidence score is... sigma for Sigmoid Activation function This represents the fused features after gradient truncation.
[0069] BCE Calibration Loss Training: The training objective of the confidence head is achieved through BCE (Binary Cross-Entropy) loss. `soft_label` is defined as a smoothing indicator of the prediction error magnitude. ; in, To calibrate the error threshold, y This represents the true value of SOH. This is the predicted SOH value. 's' is a soft label; its physical meaning is: when | -y|< When the error is less than 2%, s→1, that is, s tends to 1, and should be given a high confidence level; when | -y|> When the error is greater than 2%, s→0, that is, s tends to 0, and a low confidence level should be assigned. when | -y|= When s=0.5, it is the smooth transition point.
[0070] The calibration loss is: .
[0071] In the above formula, B Batch size; b For sample numbers within a batch; Let be the confidence score of the b-th sample; For the first b Soft labels for each sample, The confidence level is calibrated for the loss.
[0072] S6: Single-stage end-to-end joint training.
[0073] Construct a total loss function that includes the main regression loss, expert diversity loss, and confidence calibration loss. All module parameters are jointly updated in a single optimization loop, and a single model checkpoint is output.
[0074] Specifically, the total loss function is as follows: ; In the formula, The mean squared error is the loss from the main regression. For expert diversity loss based on hinge loss, During training, a penalty is applied when the mean standard deviation of each expert's prediction falls below a preset margin to prevent expert collapse. For confidence level calibration loss, and Automatically masking is performed in the first N training epochs to calibrate the loss. Warmup Mechanism. The definitions of each sub-loss are as follows: Mean squared error main regression loss : ; In the above formula, For the first b The true SOH value of each sample. For the first b SOH prediction values for each sample.
[0075] Loss of expert diversity (hinge loss) : ; In the formula, This represents the independent prediction matrix for each expert across all samples. The loss is calculated when the mean standard deviation of each expert's predictions for each sample is lower than the preset threshold. The system generates penalties that force experts to produce differentiated predictions, thus preventing the common expert collapse problem in MoE.
[0076] Calibration loss Warmup Mechanism: The confidence calibration loss is automatically masked for the first N epochs of training, optimizing only the main regression loss and expert diversity loss. This mechanism ensures that the SOH main prediction path has reached a relatively stable initial convergence state before entering the confidence calibration phase.
[0077] In this embodiment, all module parameters are jointly updated and a single model checkpoint is output in a single optimization loop. The AdamW optimizer is used, with an initial learning rate of 5e-4, weight decay of 1e-5, batch size of 32, maximum number of training epochs of 20, diversity loss weight of 0.01, calibration loss weight of 0.03, calibration error threshold of 0.02, and calibration... Warmup Round number 5, gradient clipping max_norm=1.0, early stop patience value 10.
[0078] S7: Multi-stage evaluation post-processing and low-confidence rejection.
[0079] The confidence temperature parameter was optimized on the validation set, and the optimal rejection threshold was searched for each chemical system type. During testing, the SOH prediction value was rejected for samples with a confidence level lower than the rejection threshold of their respective chemical system type.
[0080] Specifically, temperature scaling optimization: on the validation set, the confidence scores are... logit Learnable temperature parameters Re-process after scaling Sigmoid : ; In the above formula, This represents the log-odds ratio corresponding to the confidence score. This is the calibration confidence score obtained after temperature scaling and remapping via Sigmoid. Optimized using the LBFGS optimizer. To minimize BCE loss and reduce ECE (expected calibration error).
[0081] Per-type adaptive rejection threshold search: On the validation set, candidate confidence thresholds are traversed for each chemical system type (the search range in the example is [0.2, 0.8], with a step size of 0.02), and the F1 score is calculated (high SOH error samples are regarded as positive classes, and low confidence rejection is regarded as positive detection). The threshold corresponding to the maximization of F1 is selected as the exclusive rejection threshold for this type.
[0082] During test inference, if the confidence level of a sample is lower than the optimal rejection threshold for its chemical system type, the predicted SOH value of that sample will not be output, and only the SOH estimation results of high-confidence samples will be reported.
[0083] To further verify the technical effectiveness of the cross-chemical system battery SOH estimation method provided by this invention, specific examples are given below: Example 1: Unified estimation of SOH across chemical systems based on 3-Expert MoE Application scenario: The battery management system (BMS) platform needs to perform online health status assessments on NCM ternary lithium batteries and LFP lithium iron phosphate batteries in operation. It requires the same model framework to cover both chemical systems simultaneously and to have predictive reliability warning capabilities under high uncertainty scenarios.
[0084] Dataset: LFP-HUST public dataset: lithium iron phosphate battery, containing 29,214 cycle samples of multiple cells under operating conditions; The NCM-Tshinghua-48ah public dataset contains 6884 cycle samples of ternary lithium batteries under different temperature and rate conditions.
[0085] The time-series data for each cycle contains 6-dimensional features [t, I, V, dV, Q, dq]. The SOH label is obtained from the ratio of discharge capacity to rated capacity through capacity increment analysis.
[0086] The implementation steps are as follows: Step S1 (Data Preprocessing): Group the SOH labels by chemical system and perform 3... sigma Outlier filtering: After removing outliers, the time series length was uniformly filled to 256 sampling points. The original 13 fine-grained battery model identifiers were aggregated into 2 chemical system level identifiers (LFP=0, NCM=1). The data was divided into training / validation / test sets at 60% / 20% / 20%, and training set statistics were used to analyze the 6-dimensional features dimension by dimension. Min-Max Normalize to [0,1].
[0087] Step S2 (Multi-scale encoding): The temporal input is processed by three parallel 1D-CNN branches (k=3 / 5 / 7). Each branch is concatenated into 48 dimensions after three levels of dimensionality reduction (256→128→32→16). The 48 dimensions are then linearly projected to 32 dimensions and mean pooling is used to obtain 32-dimensional global features.
[0088] Step S3 (Dual Embedding Fusion): Type embedding (vocab=3, dim=8) and chemical system embedding (vocab=2, dim=8) are concatenated with 32-dimensional pooling features to form 48-dimensional features, and then fused back to 32-dimensional features through a projection layer.
[0089] Step S4 (MoE Routing and Prediction): Lightweight Router ( Linear 32→16→E=3) Output probability, Top-2 select two experts to activate, 3 layers of lightweight MLP experts make independent predictions and then fuse them according to probability weight.
[0090] Step S5 (Confidence Calibration): The fused features are detached and then sent to the independent ConfidenceHead ( Linear 32→32→16→1), via Sigmoid Output confidence levels of [0,1]. BCE calibration loss is... tau =0.02 is used as the threshold to generate soft_label.
[0091] Step S6 (Single-stage end-to-end training): Total loss = MSE + 0.01 Diversity loss + (when epoch ≥ 5) 0.03 Calibration loss. AdamW optimizer, lr=5e-4, Cosine annealing scheduling, gradient clipping max_norm=1.0, early stopping patience=10, batch size 32.
[0092] Step S7 (Post-evaluation processing): LBFGS temperature scaling optimization on the validation set + per-type threshold search (range 0.2~0.8, step size 0.02, F1 maximization).
[0093] The core parameters of Example 1 are summarized in Table 1: Table 1 Summary of Core Parameters
[0094] The technical effects are shown in Table 2: Table 2 Technical Effects
[0095] Experimental results show that the present invention achieves high-precision estimation of overall MAE less than 1% on mixed data across chemical systems, with small difference in accuracy between NCM and LFP, verifying the effectiveness of unified modeling.
[0096] Example 2: Expert Ablation and Interpretability Analysis To verify the independent contributions of each expert in the 3-Expert MoE architecture, a per-expert disabling ablation experiment was conducted: the routing probability of each individual expert was sequentially forced to zero and then renormalized, and the change in the SOH-estimated MAE (delta_mae) was evaluated on the test set. As shown in Table 3: Table 3 Ablation Experiment
[0097] This ablation experiment validates that each expert assumes different subspace prediction responsibilities, while providing diagnostic-level interpretability that traditional single models lack.
[0098] Example 3: Confidence calibration and low-confidence rejection effect The test set performance after performing temperature scaling optimization (LBFGS) and per-type threshold search on the validation set is shown in Table 4: Table 4 Test Set Performance
[0099] Under the low-confidence rejection mechanism, approximately 8%–12% of low-reliability samples are correctly rejected, and the estimated mean efficiency (MAE) of the state of harm (SOH) for valid samples is significantly reduced. Samples that are correctly rejected mainly include: samples with abnormal segments in their charge-discharge curves, samples located in ambiguous regions at the boundaries of chemical systems, and extremely aged samples with very low SOH (<0.7).
[0100] Based on the same inventive concept, embodiments of the present invention also provide a second aspect, which discloses a cross-chemical system battery SOH estimation system, comprising: The data preprocessing module is used to acquire time-series aging data of batteries containing at least two different chemical systems, hierarchically aggregate fine-grained model identifiers under each chemical system to generate chemical system-level identifiers, filter outliers and unify sequence lengths of SOH tags, and construct a normalized time-series input tensor paired with SOH tags dataset. The feature encoding and fusion module is used to extract multi-scale temporal features from the temporal input tensor to generate a global temporal feature vector, and to map the chemical system level identifier to a chemical system embedding vector through an independently parameterized embedding layer. The global temporal feature vector and the chemical system embedding vector are then fused to generate a fused feature vector. The SOH prediction module includes a hybrid expert model, which is used to input the fused feature vector into the router to calculate the routing probability of each expert, select the K experts with the highest routing probability for forward calculation, and weight and fuse the SOH prediction values output by each activated expert to obtain the final SOH prediction value. The confidence evaluation module includes an independent confidence estimation head and a gradient truncation unit. It is used to input the fused feature vector into the confidence estimation head after performing a gradient truncation operation through the gradient truncation unit, and output a confidence score that characterizes the reliability of the final SOH prediction value. The gradient truncation operation is used to truncate the backpropagation gradient, so that the training error of the confidence evaluation module only updates the parameters of the confidence head itself. The rejection output module is used to obtain the chemical system type of the target battery corresponding to the time-series input tensor, determine the corresponding rejection threshold based on the chemical system type, and compare the confidence score with the rejection threshold. If the confidence score is lower than the rejection threshold, the final SOH prediction value is rejected; otherwise, the final SOH prediction value is output.
[0101] In the rejection output module, the rejection threshold is determined as follows: On the validation set, candidate confidence thresholds are traversed independently for each chemical system type. With maximizing the F1 score as the optimization objective, the candidate confidence threshold corresponding to maximizing the F1 score is selected as the exclusive rejection threshold for that chemical system type. Specifically, samples with SOH prediction errors higher than a preset error threshold are considered positive, and rejection decisions with confidence scores lower than the current candidate threshold are considered positive detections.
[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for estimating the state of matter (SOH) of a battery across chemical systems, characterized in that, Includes the following steps: Acquire time-series aging data of batteries containing at least two different chemical systems, hierarchically aggregate fine-grained model identifiers under each chemical system to generate chemical system level identifiers, and perform outlier filtering and sequence length unification on SOH tags to construct a normalized time-series input tensor paired with SOH tags dataset; Multi-scale temporal features of the temporal input tensor are extracted to generate a global temporal feature vector. The chemical system level identifier is mapped to a chemical system embedding vector through an independently parameterized embedding layer. The global temporal feature vector and the chemical system embedding vector are fused to generate a fused feature vector. The fused feature vector is input into the hybrid expert model. The routing probability of each expert is calculated by the router. The K experts with the highest routing probability are selected for forward calculation. The SOH prediction values output by each activated expert are weighted and fused to obtain the final SOH prediction value. After performing gradient truncation on the fused feature vector, an independent confidence estimation header is input, and a confidence score characterizing the reliability of the final SOH prediction value is output. The gradient truncation operation is used to truncate the backpropagation gradient, so that the training error of the confidence estimation head only updates the parameters of the confidence head itself. Obtain the chemical system type of the target battery corresponding to the time-series input tensor, determine the corresponding rejection threshold based on the chemical system type, and compare the confidence score with the rejection threshold. If the confidence score is lower than the rejection threshold, the final SOH prediction value is rejected; otherwise, the final SOH prediction value is output.
2. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 1, characterized in that: The hierarchical aggregation specifically involves mapping the fine-grained identifiers of different cell models in the original data to coarse-grained identifiers at the chemical system level. The chemical system includes at least the lithium iron phosphate (LFP) system and the nickel-cobalt-manganese ternary NCM system. The outlier filtering specifically involves: grouping by chemical system and removing outliers that satisfy the condition |SOH- group mean|>3. Outlier labels for group standard deviations and their corresponding time series data.
3. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 1, characterized in that: Extracting multi-scale temporal features, specifically including: The temporal input tensor is convolutionally encoded using at least two parallel 1D-CNN branches, with each branch configured with a convolution kernel of a different size, to capture the charging and discharging dynamic features at different time scales. After each branch undergoes a stepwise dimensionality reduction process including residual connections, the outputs of each branch are concatenated and fused along the feature dimension, and then global pooling is used to obtain the global temporal feature vector.
4. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 3, characterized in that: The parallel 1D-CNN branches are specifically three branches with kernel sizes of 3, 5, and 7 respectively; The dimensionality reduction path for each branch is 256→128→32→16. Each dimensionality reduction block contains two fully connected layers and residual connections. The residual connections directly superimpose the input of the current dimensionality reduction block onto the output of the dimensionality reduction block.
5. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 1, characterized in that: The method for generating the embedding vector of the chemical system is as follows: The battery type identifier and the chemical system level identifier are mapped to a low-dimensional type embedding vector and a low-dimensional chemical system embedding vector through two independent parameterized embedding layers, respectively. The global temporal feature vector, the low-dimensional type embedding vector, and the low-dimensional chemical system embedding vector are concatenated and then subjected to a projection layer nonlinear transformation to generate the fused feature vector. The low-dimensional type embedding vector and the low-dimensional chemical system embedding vector form a complementary verification in the projection fusion layer.
6. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 1, characterized in that: The router is a data-driven multilayer perceptron, whose output is scaled by temperature parameters and then used to generate the routing probability distribution of each expert through the Softmax function. When selecting the K experts with the highest routing probability, K=2; Each of the activated experts is a lightweight multilayer perceptron with a consistent structure but independent parameters. The weighted fusion is achieved by summing the SOH predictions of the K experts according to their respective routing probabilities.
7. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 1, characterized in that: The confidence estimation head is trained using a binary cross-entropy calibration loss, the soft label of which is generated based on a smoothing indicator of the SOH prediction error relative to a calibration error threshold. When the prediction error is less than the calibration error threshold, the soft tag tends to 1; when the prediction error is greater than the calibration error threshold, the soft tag tends to 0; when the prediction error is equal to the calibration error threshold, the soft tag is accurate to 0.
5. Furthermore, in the first N rounds of model training, the backpropagation of the calibration loss is masked, and only the SOH prediction main task loss is optimized.
8. The method for estimating the state of charge (SOH) of a battery across chemical systems according to claim 1, characterized in that: The rejection threshold is determined as follows: On the validation set, candidate confidence thresholds are traversed independently for each chemical system type. With maximizing the F1 score as the optimization objective, the candidate confidence threshold corresponding to maximizing the F1 score is selected as the exclusive rejection threshold for that chemical system type. Specifically, samples with SOH prediction errors higher than a preset error threshold are considered positive, and rejection decisions with confidence scores lower than the current candidate threshold are considered positive detections.
9. A cross-chemical system battery SOH estimation system, characterized in that, A method for estimating the state of charge (SOH) of a battery across chemical systems as described in any one of claims 1-8 includes: The data preprocessing module is used to acquire time-series aging data of batteries containing at least two different chemical systems, hierarchically aggregate fine-grained model identifiers under each chemical system to generate chemical system-level identifiers, filter outliers and unify sequence lengths of SOH tags, and construct a normalized time-series input tensor paired with SOH tags dataset. The feature encoding and fusion module is used to extract multi-scale temporal features from the temporal input tensor to generate a global temporal feature vector, and to map the chemical system level identifier to a chemical system embedding vector through an independently parameterized embedding layer. The global temporal feature vector and the chemical system embedding vector are then fused to generate a fused feature vector. The SOH prediction module includes a hybrid expert model, which is used to input the fused feature vector into the router to calculate the routing probability of each expert, select the K experts with the highest routing probability for forward calculation, and weight and fuse the SOH prediction values output by each activated expert to obtain the final SOH prediction value. The confidence assessment module includes an independent confidence estimation head and a gradient truncation unit. It is used to input the fused feature vector into the confidence estimation head after performing a gradient truncation operation on the gradient truncation unit, and output a confidence score that characterizes the reliability of the final SOH prediction value. The gradient truncation operation is used to truncate the backpropagation gradient, so that the training error of the confidence assessment module only updates the parameters of the confidence head itself. The rejection output module is used to obtain the chemical system type of the target battery corresponding to the time-series input tensor, determine the corresponding rejection threshold based on the chemical system type, and compare the confidence score with the rejection threshold. If the confidence score is lower than the rejection threshold, the final SOH prediction value is rejected; otherwise, the final SOH prediction value is output.
10. A cross-chemical system battery SOH estimation system according to claim 9, characterized in that: In the rejection output module, the rejection threshold is determined as follows: On the validation set, candidate confidence thresholds are traversed independently for each chemical system type. With maximizing the F1 score as the optimization objective, the candidate confidence threshold corresponding to maximizing the F1 score is selected as the exclusive rejection threshold for that chemical system type. Specifically, samples with SOH prediction errors higher than a preset error threshold are considered positive, and rejection decisions with confidence scores lower than the current candidate threshold are considered positive detections.
