A lithium battery state of health estimation method and system fusing physical information diffusion causality transformer

Abstract

The present application belongs to the technical field of battery state of health estimation, and specifically discloses a lithium battery state of health estimation method and system fusing physical information diffusion causality Transformer, which comprises: obtaining lithium battery charging and discharging data and preprocessing; constructing a PI-DART network model comprising patch embedding, causality Transformer encoder, diffusion module, physical information module and estimator module; training the model using a hybrid loss function combining prediction error and physical constraint loss; and obtaining SOH estimation value through the trained model. The present application prevents information leakage through the causality attention mechanism, enhances the robustness to noise and uncertainty through the diffusion process, and improves the generalization ability and interpretability of the model by introducing the physical laws of electrolyte diffusion and SEI growth; solves the problems of poor reliability and insufficient physical consistency of existing methods, and is suitable for high-precision state of health estimation of lithium batteries.

Description

Technical Field

[0001] This invention belongs to the field of battery health state estimation technology, and more specifically, relates to a lithium battery health state estimation method and system that integrates physical information diffusion causality Transformer. Background Technology

[0002] Lithium-ion batteries have become the core power source for electric vehicles and energy storage systems due to their high energy density and long cycle life. However, irreversible performance degradation occurs during battery cycling, and accurately estimating their State of Health (SOH) is crucial for ensuring system safety and preventing failures.

[0003] Existing SOH estimation methods mainly fall into three categories: model-based methods, data-driven methods, and hybrid methods. Model-based methods (such as equivalent circuit models and pseudo-two-dimensional P2D electrochemical models) rely on precise physicochemical equations, making parameter identification difficult and computationally complex, thus hindering their application in real-time battery management systems (BMS). Data-driven methods (such as CNNs, LSTMs, and Transformers) can learn degradation patterns from historical data, but typically require large amounts of high-quality data and lack physical consistency, exhibiting poor generalization ability under noise interference and unknown operating conditions. Furthermore, the traditional Transformer architecture may suffer from future information leakage when processing time series data, leading to unreliable predictions.

[0004] Hybrid methods that have emerged in recent years attempt to combine the advantages of data-driven and physical models, but still have several limitations: (1) lack of strict constraints on the causal relationship of time series, affecting the reliability of prediction; (2) insufficient ability to model measurement noise and random uncertainty in the degradation process.

[0005] Therefore, how to provide a SOH estimation method that can guarantee prediction accuracy and causality, integrate the physical mechanism of battery degradation, and have strong anti-interference ability has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a lithium battery state of health estimation method and system that integrates physical information diffusion causality Transformer. The aim is to achieve a high-precision, robust, and physically interpretable estimation of the state of health (SOH) of lithium batteries.

[0007] To achieve the above objectives, according to a first aspect of the present invention, a method for estimating the health status of lithium batteries by fusing physical information diffusion causality Transformer is proposed, comprising the following steps: The PI-DART network model is trained using a training set, and the health status of lithium batteries is estimated using the trained PI-DART network model. The training set includes time series tensors and corresponding lithium battery SOH data, wherein the time series tensors are preprocessed lithium battery charge-discharge cycle data. The PI-DART network model includes a patch embedding layer, a causal Transformer encoder, a diffusion module, and an estimator module. The patch embedding layer divides the time-series tensor into non-overlapping patches and adds learnable positional codes to each patch to obtain a patch embedding vector. The causal Transformer encoder, based on the patch embedding vector, uses a masked self-attention mechanism to obtain a first feature. The diffusion module performs forward noise addition and backward noise removal on the patch embedding vector to obtain a noise vector. The first feature and the noise vector are combined to form the final feature, and the estimator module predicts the state of harmonics (SOH) of the lithium battery based on the final feature. The loss function used during training is a weighted average of the model data prediction loss and the physical information loss.

[0008] As a further preferred option, the loss function for:

[0009] in, This represents the mean squared error loss in the model's predictions. For physical information loss, , Electrolyte diffusion loss, For SEI growth loss; This represents the weighting coefficient for physical information loss.

[0010] As a further preferred embodiment, the electrolyte diffusion loss The finite difference approximation based on Fick's second law is constructed as follows:

[0011] in, Electrolyte concentration, The learnable electrolyte diffusion coefficient; For time, It represents a one-dimensional spatial location, describing the spatial distribution of concentration within the battery; i For spatial indexing, k For time indexing, N The total number of discrete points in one-dimensional space. K The number of steps in the time direction.

[0012] As a further preferred embodiment, the SEI growth loss The capacity loss model, based on the number of cycles and average current, is constructed and expressed as follows:

[0013] in, The learnable SEI growth rate constant. The average absolute current, This refers to the battery's rated capacity. m This serves as an index for the number of charge / discharge cycles. M This represents the total number of charge-discharge cycles. The model predicts the first m SOH after one charge-discharge cycle.

[0014] As a further preferred embodiment, the diffusion module performs forward noise addition and reverse noise removal on the patch embedding vector to obtain a noise vector, specifically: The patch embedding vector is subjected to forward noise addition for a random number of steps. Then, a multilayer perceptron (MLP) is used as the denoising network to predict the added noise and obtain the noise vector.

[0015] As a further preferred embodiment, the lithium battery charge-discharge cycle data includes current, voltage, temperature, and time data of the lithium battery charge-discharge cycle.

[0016] As a further preferred option, the lithium battery charge-discharge cycle data is preprocessed, including: resampling the data for each charging cycle, unifying the sequence length, and performing maximum and minimum normalization on the current, voltage, temperature, and time data to convert them to the range of [-1, 1], and then combining them to obtain a time series tensor.

[0017] According to a second aspect of the present invention, a lithium battery health state estimation system incorporating a physical information diffusion causal Transformer is provided, comprising a processor for executing the above-described lithium battery health state estimation method incorporating a physical information diffusion causal Transformer.

[0018] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described lithium battery health state estimation method using a fused physical information diffusion causal Transformer.

[0019] According to a fourth aspect of the present invention, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the above-described lithium battery health state estimation method using a fused physical information diffusion causal Transformer.

[0020] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages: 1. This invention ensures strict causality in the SOH estimation process through the masked attention mechanism in the causal Transformer, effectively preventing the leakage of future information and improving the reliability of prediction; 2. This invention introduces a diffusion module, which enhances the model's robustness to sensor measurement noise and random uncertainties caused by battery degradation by simulating the noise addition and removal process; 3. This invention introduces physical information loss into the loss function. Specifically, the simplified physical laws of the P2D model (electrolyte diffusion and SEI growth) are embedded into the neural network in the form of learnable parameters and loss functions, so that data-driven prediction is consistent with electrochemical principles, which significantly improves the generalization ability and interpretability of the model. 4. The method of this invention achieves leading estimation accuracy on public datasets, solves the problems of poor reliability and insufficient physical consistency of existing methods, and demonstrates excellent performance and engineering application potential. It is suitable for high-precision health state estimation of lithium batteries. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the lithium battery health state estimation method based on the physical information diffusion causal Transformer according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of a diffusion autoregressive network driven by physical information fusion according to an embodiment of the present invention; Figure 3 This is the SOH estimation result of lithium batteries in the MIT battery dataset of this invention embodiment; Figure 4 This is the SOH estimation result of lithium batteries in the XJTU battery dataset of this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0023] This invention provides a lithium battery health state estimation method that integrates physical information diffusion causality Transformer, such as... Figure 1 As shown, it includes the following steps: S1. Training set construction: Obtain charge and discharge cycle data of N lithium batteries and preprocess them to obtain normalized multi-channel time series data. Then, construct the multi-channel time series data as the model input tensor and combine it with the corresponding lithium battery SOH data to construct the training set.

[0024] Specifically, lithium battery charge-discharge cycle data includes current, voltage, temperature, and time data. Each lithium battery undergoes multiple charge-discharge cycles, and the data from each charging cycle is resampled, with the sequence length standardized. (Preferred) The current, voltage, temperature, and time data are then subjected to maximum and minimum normalization to convert them to the range of [-1, 1].

[0025] The preprocessed data is then constructed into a time series tensor. Where N is the batch size, Number of channels (current, voltage, temperature, time). The sequence length is provided; at the same time, a loop count tensor is provided as the time context.

[0026] S2. Establish the PI-DART network model: Use a patch embedding layer, a causal Transformer encoder, a diffusion module, an estimator module, and a physical information module to build a diffusion autoregressive (PI-DART) network model driven by physical information fusion for lithium battery SOH estimation.

[0027] like Figure 2 As shown, the patch embedding layer divides the time series tensor into non-overlapping patches and adds learnable positional encodings to the patches to obtain patch embedding vectors. The causal Transformer encoder, based on these patch embedding vectors, employs a masked self-attention mechanism, ensuring that the information at each position depends only on the current time step and previous sequence information. This guarantees that the model focuses only on current and historical information, preventing future information leakage, and yields the first feature. The diffusion module enhances the features through forward denoising and backward denoising processes. Specifically, it performs a random number of forward denoising steps on the patch embedding vectors and predicts the added noise through a denoising network. The predicted noise is then added to the first feature to form the enhanced feature representation, resulting in the final feature. The estimator module, a multilayer perceptron (MLP), outputs the final SOH estimate based on the final feature. The physical information module, based on a simplified P2D electrochemical model, constructs electrolyte diffusion loss and SEI growth loss.

[0028] Specifically, the model design method is as follows: (1) Design a Patch embedding module to capture richer local information and fully represent the degradation process of lithium batteries.

[0029] The raw battery data only contains four dimensions: current, voltage, and time. Using this low-dimensional data directly for the Transformer's attention mechanism makes it difficult for the mechanism to capture complex multi-head dependencies; therefore, patch embedding is necessary. Before partitioning the data into patches, a Conv1d is used to embed the raw data... The dimension is transformed from 4 to The new input data obtained Represented as:

[0030] In the formula, This represents the first Conv1d, where the kernel is set to 1, and the padding and stride are set to 0 and 1 respectively.

[0031] Furthermore, to prevent information leakage, the length is... Data Divided into growth stages Multiple non-overlapping patches, where the number of patches is .

[0032] To achieve the above patch division, both the convolutional kernel and the stride are set to the patch length. Conv1d is adopted. This process can be represented as:

[0033] In the formula, This indicates the second Conv1d, where padding is set to 0.

[0034] After completing the patch embedding, the causal Transformer encoder module is used to capture temporal dependencies and enforce causality.

[0035] First, the traditional Transformer encoder is initialized as a representation network. During training, a learnable start-of-sequence (SFOS) flag is embedded before the first patch, while the last patch is removed to obtain a new patch sequence.

[0036] To add positional information to the sequence after embedding, each patch is given a learnable positional code:

[0037] In the formula, This indicates the result of patch embedding. The embeddings represent the learning positions. Initialize the sample from a standard normal distribution (mean 0, variance 1). Its mathematical expression can be represented as:

[0038] In the formula, , representing a multidimensional standard normal distribution.

[0039] Subsequently, a position-encoded embedding sequence was added and processed by the Transformer encoder, and a causal mask was used to ensure that each patch only focuses on itself and the patches preceding it.

[0040] In the formula, The output of the Transformer encoder, It is a triangular matrix to ensure autoregressive dependency.

[0041] (2) Design a patch-level diffusion module to mitigate the impact of uncertainty in lithium battery monitoring data and enhance the robustness and expressive power of the time-series features extracted by the causal Transformer.

[0042] Patch-level diffusion operations include a forward diffusion step, which progressively adds noise to the patch embedding, and a backward denoising step, which recovers a clean representation by predicting the added noise.

[0043] 1) Forward process.

[0044] For each patch embedding layer output The forward propagation process of the diffusion module gradually increases the noise. , represented as:

[0045] in, Indicates a Gaussian distribution. This indicates the added noise. For the first Step noise dispatcher, , A linear interpolation noise scheduling from arrive Cross One diffusion step. The cumulative product can be expressed as:

[0046] Therefore, the clean patch directly represents the noisy patch at time step. Distribution at various locations.

[0047] Noise is added to the forward process of the diffusion model given an initial patch. The patch obtained after this can be represented as:

[0048] in, For the first The original embedding vectors of each patch, For the first The embedding vector after adding noise at step s of the patch. This is a noise scheduling function.

[0049] Noise is added independently to each patch embedding, ensuring the model can learn to handle different noise levels in the sequence. This design avoids oversimplification of the denoising task and encourages the model to capture subtle local patterns. Therefore, for the ... The noisy patch embedding of each patch sample is represented as:

[0050] in It is the time step for random sampling for each patch.

[0051] 2) Reverse process.

[0052] The reverse process aims to reconstruct a clean patch embedding by predicting the noise components added during the forward process. A multilayer perceptron (MLP) is used as the denoising network to receive the noisy patch embedding. As input, it predicts noise. The MLP consists of two linear layers, with a ReLU activation function used in between, from the embedding dimension. Mapped to Back to This process can be represented as:

[0053] Finally, the predicted noise Combined with the output of the causal Transformer encoder, the resulting... The final feature representation of each patch:

[0054] (3) Design an estimator module that integrates two-stage data to output the SOH estimate of lithium-ion battery.

[0055] The input to the estimator is a shape of The tensor, in which Indicates batch size, This indicates the number of patches derived from the input sequence. For embedded dimensions To ensure compatibility with fully connected layers, the feature tensor is first reshaped into a shape of... The flattened form.

[0056] Subsequently, the flattened features are passed through a Multilayer Perceptron (MLP) regressor, which is defined as a sequential neural network consisting of two linear transformations. This structure allows the model to capture the complex interactions within the flattened features while projecting them onto a one-dimensional regression objective. This mitigates the risk of vanishing gradients and enhances the model's ability to represent nonlinearities by inserting a Rectified Linear Unit (ReLU) activation function between the two linear transformations. The MLP architecture is formalized as follows:

[0057]

[0058] in It is the input feature tensor (final feature) from the previous module. It is a flattened vector. and These are the weights and biases of the first linear layer (where...) This indicates the hidden intermediate dimension, which is set to 128 in this invention. Represents the ReLU activation function. and It outputs the weights and biases of the linear layer. The predicted SOH value corresponds to each sample in the batch.

[0059] S3. Model Training: The PI-DART network model is trained using the training set.

[0060] The training objective is to minimize the mixture loss function. ,in This represents the mean squared error loss between the predicted and actual values. This is a loss of physical information, specifically electrolyte diffusion loss. and SEI growth loss sum; The physical loss weighting coefficient is determined through grid search.

[0061] Furthermore, electrolyte diffusion loss Based on Fick's diffusion law:

[0062] in, Electrolyte concentration, The learnable electrolyte diffusion coefficient; For time, It represents a one-dimensional spatial location, describing the spatial distribution of concentration within the battery. i For spatial indexing, identify spatial grid points i Location, k For time indexing, N This represents the total number of discrete points in a one-dimensional spatial direction (typically along the x-axis of the battery thickness). K The number of steps in the time direction.

[0063] SEI growth loss Based on the capacity loss model:

[0064] in, The learnable SEI growth rate constant. The average absolute current, This refers to the battery's rated capacity. m This serves as an index for the number of charge / discharge cycles. M This represents the total number of charge-discharge cycles. The model predicts the first m SOH after one charge-discharge cycle.

[0065] Specifically, this invention improves the generalization of downstream SOH prediction tasks by designing physical losses based on the P2D model to constrain the model. The invention focuses on two key electrochemical processes: electrolyte diffusion and SEI growth. The constructed physical losses include electrolyte diffusion losses and solid electrolyte interphase (SEI) growth losses.

[0066] 1) Electrolyte diffusion process To model ion transport in the electrolyte, a simplified one-dimensional Fick diffusion equation was used to model transient electrolyte concentrations. The model can be approximated as follows:

[0067] in It is the electrolyte concentration. It is the electrolyte diffusion coefficient (a learnable parameter with an initial value of...). ), These are spatial coordinates.

[0068] Furthermore, the spatial second derivative is approximated using a finite difference scheme:

[0069] in It is a fixed spatial step size. Therefore, the residual of the diffusion equation can be expressed as:

[0070] 2) SEI growth process The SEI growth model quantifies the capacity loss due to the formation of the SEI layer, which is driven by current and accumulates with cycling. The capacity loss rate is modeled as follows:

[0071] in It is capacity loss. It is the number of loops. It is the SEI growth rate constant (a learnable parameter, initialized to). , It is the absolute value of the current amplitude.

[0072] The average absolute current of the lithium-ion battery in the input model is expressed as:

[0073] As the number of charge-discharge cycles increases, the cumulative capacity loss is expressed as:

[0074] The physical SOH of a lithium-ion battery can be calculated using the above formula:

[0075] in This refers to the maximum capacity of the lithium battery, i.e., the rated capacity.

[0076] S4. Model Application: Input the lithium battery test data to be estimated into the trained PI-DART network model to obtain its SOH estimate.

[0077] The following are specific examples: Based on the MIT dataset and the XJTU battery dataset, this paper compares the results of comparative experiments using methods such as CNN, LSTM, AttMoE, ELSTM, MMMe, Pathformer, and SAMformer, demonstrating the effectiveness and advancement of the invention.

[0078] like Figure 3 As shown in Table 1, the results of each metric are presented based on the experimental results verified by the MIT Battery Dataset.

[0079] Table 1. Performance test results of SOH estimation in the MIT dataset experiment.

[0080] Compared to the models based on the aforementioned infrastructure, models based on the Transformer architecture achieve better results. These Transformer-based models achieve an MSE of no more than 0.109 across all lithium-ion batteries in the MIT Battery Dataset, with R... 2 All values ​​were above 0.900. Furthermore, among all comparative methods, SAMformer achieved the second-best results, with the average MSE and average R0 across all cells in the 9 batches. 2 The values ​​are 0.099 and 0.918, respectively. Although the baseline methods based on the Transformer architecture have strong long-range dependency modeling capabilities, their performance still has significant room for improvement due to the lack of consideration for uncertainties in the monitoring data and the absence of constraints on lithium battery degradation mechanisms.

[0081] Compared to baseline methods based on the Transformer architecture, the proposed PI-DART achieves dominant best performance, ranking first across all nine batches and their averages. Specifically, compared to its strongest competitor (SAMformer), PI-DART significantly reduces the average MSE ratio by 84.849%, and the average R... 2 This represents a significant improvement of 7.407%. More importantly, the model of this invention exhibits excellent robustness, with minimal performance fluctuations across batches, ranging from 0.006 to 0.039, which is a key advantage for battery management system applications.

[0082] like Figure 4 As shown in Table 2, the results of each indicator are presented based on the experimental results verified using the XJTU battery dataset.

[0083] It can be seen that the PI-DART proposed in this invention still performs optimally. In terms of average MSE, the proposed PI-DART significantly reduces MSE by 14.783% compared to LSTM. In terms of average R... 2 In terms of performance, the proposed PI-DART achieves a 1.059% improvement over SAMformer. More importantly, among all baseline methods, R... 2 The number of batches exceeding 0.950 does not exceed 3, and only the PI-DART proposed in this invention implements R in all batches except Batch 6. 2 The value exceeds 0.950. This result indicates that the method of the present invention exhibits minimal performance fluctuations across batches, demonstrating great potential for practical battery management system applications.

[0084] Table 2. Performance test results of SOH estimation in the XJTU dataset experiment.

[0085] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for estimating the health status of lithium batteries by integrating physical information diffusion causality Transformer, characterized in that, Includes the following steps: The PI-DART network model is trained using a training set, and the health status of lithium batteries is estimated using the trained PI-DART network model. The training set includes time series tensors and corresponding lithium battery SOH data, wherein the time series tensors are preprocessed lithium battery charge-discharge cycle data. The PI-DART network model includes a patch embedding layer, a causal Transformer encoder, a diffusion module, and an estimator module. The patch embedding layer divides the time-series tensor into non-overlapping patches and adds learnable positional codes to each patch to obtain a patch embedding vector. The causal Transformer encoder, based on the patch embedding vector, uses a masked self-attention mechanism to obtain a first feature. The diffusion module performs forward noise addition and backward noise removal on the patch embedding vector to obtain a noise vector. The first feature and the noise vector are combined to form the final feature, and the estimator module predicts the state of harmonics (SOH) of the lithium battery based on the final feature. The loss function used during training is a weighted average of the model data prediction loss and the physical information loss.

2. The lithium battery health state estimation method based on the fusion of physical information diffusion causality Transformer as described in claim 1, characterized in that, The loss function for: in, This represents the mean squared error loss in the model's predictions. For physical information loss, , Electrolyte diffusion loss, For SEI growth loss; This represents the weighting coefficient for physical information loss.

3. The lithium battery health state estimation method based on the fusion of physical information diffusion causality Transformer as described in claim 2, characterized in that, Electrolyte diffusion loss The finite difference approximation based on Fick's second law is constructed as follows: in, Electrolyte concentration, The learnable electrolyte diffusion coefficient; For time, It represents a one-dimensional spatial location, describing the spatial distribution of concentration within the battery; i For spatial indexing, k For time indexing, N The total number of discrete points in one-dimensional space. K The number of steps in the time direction.

4. The lithium battery health state estimation method based on the fusion of physical information diffusion causality Transformer as described in claim 2, characterized in that, SEI growth loss The capacity loss model, based on the number of cycles and average current, is constructed and expressed as follows: in, The learnable SEI growth rate constant. The average absolute current, This refers to the battery's rated capacity. m This serves as an index for the number of charge / discharge cycles. M This represents the total number of charge-discharge cycles. The model predicts the first m SOH after one charge-discharge cycle.

5. The lithium battery health state estimation method based on the fusion of physical information diffusion causality Transformer as described in claim 1, characterized in that, The diffusion module performs forward denoising and backward denoising on the patch embedding vector to obtain a noise vector, specifically: The patch embedding vector is subjected to forward noise addition for a random number of steps. Then, a multilayer perceptron (MLP) is used as the denoising network to predict the added noise and obtain the noise vector.

6. The lithium battery health state estimation method based on the fusion of physical information diffusion causality Transformer as described in any one of claims 1-5, characterized in that, The lithium battery charge-discharge cycle data includes current, voltage, temperature, and time data for lithium battery charge-discharge.

7. The lithium battery health state estimation method based on the fusion of physical information diffusion causality Transformer as described in claim 6, characterized in that, The lithium battery charge-discharge cycle data is preprocessed, including: resampling the data for each charging cycle, unifying the sequence length, and performing max-min normalization on the current, voltage, temperature and time data to convert them to the range of [-1, 1], and then combining them to obtain a time series tensor.

8. A lithium battery health state estimation system integrating physical information diffusion causality Transformer, characterized in that, Includes a processor for executing the lithium battery health state estimation method according to any one of claims 1-7, which incorporates a physical information diffusion causal Transformer.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the lithium battery health state estimation method according to any one of claims 1-7, which incorporates physical information diffusion causal Transformer.

10. A computer program product, characterized in that, The system includes a computer program that, when executed by a processor, implements the lithium battery health state estimation method according to any one of claims 1-7, which incorporates a physical information diffusion causal Transformer.

Images