A method and system for estimating the state of health of lithium batteries based on perturbation recovery enhancement and cross-domain adaptation mechanisms.

Abstract

This invention discloses a method and system for estimating the state of health (SOH) of lithium batteries based on perturbation recovery enhancement and cross-domain adaptation mechanisms. Relating to the field of lithium battery technology, the method simulates degradation and regeneration of raw data tagged with SOH, generating high-fidelity extended samples and constructing an enhanced dataset to address the scarcity of original labeled data. Hierarchical analysis is performed on the enhanced data to uncover the long-term and short-term dynamic patterns of lithium battery operation data, outputting initial SOH prediction results. Utilizing a cross-domain knowledge adaptation optimization strategy, the method quantifies the difference in feature distribution between the source and target domains through a distribution similarity measure mechanism, dynamically adjusting parameters to achieve adaptive alignment of data distribution in the new environment, correcting the prediction results, and outputting high-precision SOH values. This invention achieves high-precision estimation of the SOH of distributed energy storage lithium batteries, is suitable for scenarios where the amount of raw SOH-tagged data is scarce, and improves the applicability and accuracy of the prediction model.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and more specifically to a method and system for estimating the state of health of lithium batteries based on perturbation recovery enhancement and cross-domain adaptation mechanisms. Background Technology

[0002] With the rapid development of new energy technologies, lithium batteries are increasingly being used in distributed energy storage systems. However, due to complex operating environments and incomplete data collection, existing SOH estimation models often face problems such as scarce raw data, incomplete labels, and distribution bias, resulting in poor model generalization ability and large prediction errors in real-world scenarios.

[0003] Traditional lithium battery state of health (SOH) estimation methods mostly rely on large amounts of labeled historical operational data for training. In practical applications, especially in newly deployed sites or scenarios with changing operating conditions, sufficient labeled data is often unavailable. Existing data augmentation techniques fail to adequately consider the degradation mechanisms during data evolution; traditional transfer learning methods also struggle to effectively align significant distributional differences between the source and target domains.

[0004] Therefore, how to provide a lithium battery health state estimation method and system based on perturbation recovery enhancement and cross-domain adaptation mechanism, combined with generative enhancement and efficient migration technology, to achieve high-precision estimation of the SOH of distributed energy storage lithium batteries is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a lithium battery health state estimation method and system based on perturbation recovery enhancement and cross-domain adaptation mechanisms. By combining data perturbation enhancement strategies, deep sequence feature extraction techniques, and adaptive feature distribution alignment methods, it overcomes the problems of existing SOH estimation methods, such as data scarcity, insufficient model generalization ability, and poor cross-scenario adaptability. Through the data perturbation enhancement model, the present invention can generate diverse training samples from a small amount of raw lithium battery data labeled with SOH. Based on the sequence feature extraction module, the system can effectively capture the dynamic features in lithium battery operating data. Combined with a cross-domain knowledge adaptation optimization strategy, the present invention achieves adaptive processing of cross-domain data by measuring and minimizing the feature differences between different data distributions. Overall, the present invention provides an accurate and highly generalizable SOH estimation method and system that can obtain stable health state prediction results even with scarce raw data.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism, comprising:

[0007] Raw data of lithium batteries with health status tags are collected, and redistributed data is generated through a data perturbation enhancement model.

[0008] Sequence features are extracted and health status is estimated from the redistributed data, and preliminary prediction results of the health status of lithium batteries are output. The health status estimation bias is also calculated.

[0009] Feature distribution alignment is performed based on the method of measuring the difference in distribution mean to obtain the feature distribution difference;

[0010] A joint loss function is constructed based on the health status estimation bias and the feature distribution difference. By iteratively updating the parameters, the cross-domain generalization ability of the model is improved, and the final lithium battery health status prediction result is obtained.

[0011] Preferably, the data perturbation enhancement model includes a perturbation generation unit and a dynamic feature identification unit, used to introduce perturbations into the original lithium battery data, generate perturbation data samples, and then restore the perturbation data samples to obtain the restored redistributed data.

[0012] Preferably, the disturbance generation unit is used to gradually add Gaussian noise to the original lithium battery data based on a preset disturbance scheduling strategy until a preset disturbance level is reached, thereby generating multi-stage degradation samples.

[0013] The dynamic feature identification unit is used to fuse time step information, distinguish the differences between generated samples and real data at different stages during training, and improve the reverse recovery capability of the perturbation generation unit through an adversarial optimization mechanism.

[0014] The perturbation generation unit simulates data degradation by gradually introducing Gaussian noise, and then the reverse recovery module iteratively restores the perturbed samples to recover a data representation close to the original distribution; the dynamic feature identification unit is used to identify the difference features between the restored samples and the real data, and optimizes the accuracy and stability of the restoration process through an adversarial training mechanism.

[0015] Preferably, after the perturbation generation unit is trained, it can generate initial pseudo-samples from random Gaussian noise and reconstruct their feature structure through a reverse recovery mechanism.

[0016] Preferably, the reverse recovery mechanism is used to gradually remove noise and reconstruct the original data features until the noise-free state is restored, thus obtaining pseudo samples;

[0017] A neighborhood feature matching strategy is adopted to calculate the feature similarity between pseudo-samples and original samples, determine the corresponding health status labels, and merge pseudo-samples with the original dataset to construct an extended dataset for subsequent training.

[0018] The initial pseudo-samples are restored to the feature scale of the original data after inverse normalization, and the health status labels are back-mapped based on the neighborhood similarity with the original samples in the feature space using the nearest neighbor distribution strategy.

[0019] The pseudo-samples labeled with health status are merged with the original samples to construct a sample dataset for subsequent modeling training.

[0020] Based on the time-step encoding strategy, time information is embedded into data features, and a probability assessment of the data authenticity is output.

[0021] Preferably, the redistributed data undergoes sequence feature extraction and health status estimation, including:

[0022] A selective state-space sequence model incorporating multi-layer nonlinear transformations and state control mechanisms is used to mine temporal features of redistributed data.

[0023] Preferably, the selective state-space sequence model includes a deep dynamic filtering unit, an information filtering unit, a multi-scale dependency capture unit, and a health state estimation unit connected in sequence.

[0024] Preferably, the deep dynamic filtering unit processes the input sequence data based on an adaptive filtering strategy;

[0025] The information filtering unit filters the processed data through a gated linear structure;

[0026] The multi-scale dependency capture unit extracts local and global temporal dependency features of the time series in parallel across multiple feature spaces;

[0027] The health status estimation unit performs aggregate analysis on the extracted time-dependent features and outputs preliminary prediction results of the lithium battery health status.

[0028] Preferably, feature distribution alignment is performed based on the mean difference measurement method to reduce the impact of different data distribution differences on model performance; obtaining feature distribution differences includes: calculating the feature distribution differences between the source domain and the target domain samples in the high-dimensional feature space, and dynamically adjusting the parameters in the feature extraction process based on the measurement results using an adaptive matching strategy to minimize the feature distribution differences;

[0029] Among them, during the feature distribution alignment process, the health status estimation bias is dynamically monitored;

[0030] When the estimation error exceeds the preset threshold, the weight of the distribution mean difference measure is automatically adjusted, the feature distribution difference is recalculated and the parameters are updated until the prediction error converges to the preset range.

[0031] Preferably, a lithium battery health state estimation system based on perturbation recovery enhancement and cross-domain adaptation mechanism includes:

[0032] The data redistribution module is used to collect raw data of lithium batteries with health status tags and generate redistributed data through a data perturbation enhancement model.

[0033] The feature extraction and estimation module is used to extract sequence features and estimate the health status of the redistributed data, output the preliminary prediction results of the health status of the lithium battery, and calculate the health status estimation bias.

[0034] The cross-domain knowledge adaptation and optimization module is used to align feature distributions based on the distribution mean difference measurement method, thereby obtaining feature distribution differences and reducing the impact of cross-domain bias on prediction accuracy.

[0035] The results output module is used to construct a joint loss function based on the health status estimation bias and the feature distribution difference, and improve the cross-domain generalization ability of the model by iteratively updating the parameters to obtain the final lithium battery health status prediction result.

[0036] As can be seen from the above technical solution, compared with the prior art, this invention discloses a method and system for estimating the state of health (SOH) of lithium batteries based on perturbation recovery enhancement and cross-domain adaptation mechanisms. The method includes: collecting raw lithium battery operating data tagged with SOH; expanding the raw lithium battery operating data using a data perturbation enhancement model to obtain an expanded dataset; extracting features from the expanded dataset using a selective state-space sequence model to learn the potential dynamic patterns in the lithium battery time series and obtain a preliminary prediction result of the lithium battery SOH; and using a strategy to minimize feature distribution differences for cross-domain knowledge adaptation optimization to correct the preliminary prediction result and output the final SOH prediction value. This invention achieves high-precision estimation of the SOH of distributed energy storage lithium batteries, is suitable for scenarios where the amount of raw SOH-tagged data is scarce, and can improve the applicability and accuracy of the prediction model by generating effective expanded training samples and optimizing the data distribution in new environments through cross-domain adaptation. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0038] Figure 1 This is a schematic flowchart of a lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism, provided in an embodiment of the present invention.

[0039] Figure 2 The image shows the SOH prediction results provided in this embodiment of the invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] This invention discloses a method for estimating the state of health of lithium batteries based on perturbation recovery enhancement and cross-domain adaptation mechanisms, such as... Figure 1 As shown, the entire process from data acquisition, data generation, feature learning and prediction to cross-domain knowledge adaptation and optimization, and result output is clearly presented. The specific steps are as follows:

[0042] S100: Collect raw lithium battery data with health status tags, and generate redistributed data through a data perturbation enhancement model to obtain an expanded dataset;

[0043] S200. Perform sequence feature extraction and health status estimation on the redistributed data, output the preliminary prediction results of the lithium battery health status, and calculate the health status estimation bias; among them, a selective state-space sequence model is used to model and extract features from the expanded dataset, learn the potential dynamic laws in the lithium battery time series, and obtain the preliminary prediction results of the lithium battery health status.

[0044] S300. Feature distribution alignment is performed based on the distribution mean difference measurement method to obtain the feature distribution difference; a joint loss function is constructed based on the health status estimation bias and the feature distribution difference, and the parameters are iteratively updated to improve the cross-domain generalization ability of the model, thus obtaining the final lithium battery health status prediction result. This embodiment of the invention performs cross-domain feature adaptation optimization to correct the preliminary prediction result of the lithium battery health status and outputs the final SOH prediction value.

[0045] This invention addresses a small amount of raw lithium battery data with health status tags by gradually introducing perturbations to generate diverse data samples. The process involves gradually applying Gaussian noise to the raw data according to a set perturbation scheduling strategy, forming multiple noisy data versions until a preset perturbation level is reached.

[0046] Based on the data samples with added perturbations, noise interference is gradually removed through a reverse recovery process to obtain the restored redistributed data; this achieves effective restoration of the original data structure and maintains the consistency between the health status labels and the restored samples.

[0047] Based on the restored data, a structure including multi-layer nonlinear transformation and state control mechanism is adopted to explore the potential time-series evolution patterns in lithium battery operation data and output preliminary health status prediction results.

[0048] For lithium battery samples from different data distributions, the distribution differences between their features are calculated, and the differences are minimized through a cross-domain knowledge adaptation optimization strategy, so that the distribution of source domain and target domain samples in the high-dimensional feature space tends to be consistent.

[0049] A joint loss function is constructed, which incorporates the bias in health status estimation and the differences in feature distribution into the optimization objective. By iteratively updating the parameters to minimize the joint loss, a stable and accurate health status estimation is achieved.

[0050] Specifically, the data perturbation enhancement model includes a perturbation generation unit and a dynamic feature identification unit, which are used to introduce perturbations into the original lithium battery data, generate perturbation data samples, and then restore the perturbation data samples to obtain the restored redistributed data.

[0051] Specifically, the disturbance generation unit is used to gradually add Gaussian noise to the original lithium battery data based on a preset disturbance scheduling strategy until a preset disturbance level is reached, thereby generating multi-stage degradation samples.

[0052] The dynamic feature identification unit is used to fuse time step information, distinguish the differences between generated samples and real data at different stages during training, and improve the reverse recovery capability of the perturbation generation unit through an adversarial optimization mechanism.

[0053] The data perturbation enhancement model integrates a noise perturbation mechanism and dynamic feature identification. The noise perturbation mechanism gradually introduces noise into the original lithium battery data through a forward process to simulate the degradation process of the running data. Then, it combines a reverse recovery strategy to iteratively restore the perturbed samples and recover their effective representation that is close to the original distribution.

[0054] The dynamic feature identification is used to identify the distribution differences between the restored sample and the real data, and improves the accuracy of the identification by introducing time step embedding information.

[0055] Specifically, after the data perturbation enhancement model is trained, pseudo-samples are generated from Gaussian noise, and the SOH labels are determined by inverse normalization and spatial proximity analysis. These pseudo-samples are then merged with the original samples to form an expanded dataset.

[0056] Specifically, the data perturbation enhancement model integrates a noise perturbation mechanism and a dynamic feature identification unit, including:

[0057] During forward perturbations, scheduling is performed according to a pre-defined variance. Gradually moving towards the original lithium battery operating data Adding Gaussian noise results in data contaminated with noise. ;

[0058] During the reverse recovery process, it starts from the area heavily polluted by noise. The original data was gradually recovered. ;

[0059] The input to the dynamic feature identification unit includes data from the forward perturbation process. The time step t is encoded using a multilayer perceptron and converted into a time feature vector; the output of the dynamic feature identification unit is the probability of data authenticity.

[0060] Specifically, pseudo-samples are generated from Gaussian noise, and their SOH labels are determined using inverse normalization and spatial proximity analysis. These pseudo-samples are then merged with the original samples to form an expanded dataset, including:

[0061] Initial noise samples are randomly sampled from a standard normal distribution;

[0062] Next, using the trained reverse diffusion process model, reverse diffusion is performed iteratively from t=T to t=1 to obtain pseudo samples; T is the total number of time steps in the forward perturbation process.

[0063] Based on the location of the sample in the original feature space, the spatial proximity analysis method is used to find the K nearest neighbors of the generated sample in the original sample set, and a weighted average method is used to generate SOH labels to form pseudo sample pairs;

[0064] The generated set of pseudo-samples is merged with the original set of real samples to form the expanded dataset.

[0065] In one specific embodiment of the present invention, to overcome the problem of limited quantity of original SOH data, the original collected lithium battery operating data with SOH tags is first enhanced. This embodiment employs a data perturbation enhancement model for data expansion. The original data includes input feature data and corresponding SOH tags. Due to the limited number of samples, directly training the model easily leads to overfitting and makes it difficult to generalize to new scenarios.

[0066] The data perturbation enhancement model innovatively integrates noise perturbation mechanism and dynamic feature identification unit. Its theoretical basis is based on the reversible mapping strategy of gradually perturbing and restoring the distribution of time series data. The core includes two stages: forward perturbation process and reverse recovery process.

[0067] The model takes raw lithium-ion battery operating data samples with a time-series structure as input. Each sample consists of input features and a corresponding SOH label. All feature data undergoes normalization and time alignment before input to ensure the effectiveness and consistency of subsequent noise perturbation and sequence modeling.

[0068] S101, Forward Disturbance Process

[0069] In this process, scheduling is based on a pre-set variance. Gradually moving towards the initial data In this embodiment of the invention, Gaussian noise is added. This represents the original lithium battery characteristic data, specifically a vector composed of collected characteristic values ​​such as voltage, current, and temperature. As the time step t progresses from 1 to T, the data... Gradually dominated by noise, the characteristics of the original data are gradually masked, eventually approaching a pure noise distribution. T is the total number of time steps in the forward perturbation process, which determines the number of steps required for the data to go from its original state to being completely noise-enhanced.

[0070] To ensure the stability and effectiveness of the noise addition process, variance scheduling... Using a fixed linear schedule, its expression is as follows:

[0071] (1)

[0072] Where T represents the total number of diffusion steps, and These represent the minimum and maximum noise intensities, respectively. (Take...) , This ensures that the noise gradually increases without becoming excessive.

[0073] The entire forward perturbation process can be precisely described by the following formula:

[0074] (2)

[0075] (3)

[0076] in, Given the initial data In this case, all data samples from time step 1 to T The joint probability distribution, Indicates that at the known previous time step Data samples In the case of the current time step Data samples The conditional probability distribution, Let I represent the Gaussian distribution, and let I denote the identity matrix. Choosing the identity matrix I as the covariance matrix of the Gaussian noise implies that each feature is independent when noise is added, and that the noise of each feature has the same variance. This method of gradually adding noise simulates the degradation process of data under complex environments, providing a foundation for subsequent reverse recovery of the original data features.

[0077] S102, Reverse Recovery Process

[0078] In contrast to the forward perturbation process, the backward recovery process aims to recover from areas heavily contaminated by noise. The original data was gradually recovered. x 0. The process can be represented as:

[0079] (4)

[0080] (5)

[0081] in, Indicates in model parameters Under the influence of this, all data samples from time step 0 to T The joint probability distribution. This represents a data sample at a known current time step t. In the case of the data sample from the previous time step t-1 The conditional probability distribution is determined by the model parameters. Decision. In this process, the model uses the learned parameters... Continuously estimate the mean at each time step Covariance The process of gradually removing noise and restoring the original features of the data is essentially about optimizing the variational lower bound so that the model can accurately learn the distribution characteristics of the data, thereby achieving effective recovery from noise to the original data.

[0082] S103, Introducing a dynamic feature identification unit

[0083] Data perturbation enhancement models utilize progressive noise perturbation and reverse recovery mechanisms to synthesize diverse data samples to enhance the training dataset. However, traditional high-order progressive recovery processes often rely on long noise restoration paths, which not only increase computational overhead but may also affect the quality of synthesized samples due to error accumulation. Therefore, this invention introduces a dynamic feature discrimination unit, which improves the efficiency of the noise restoration stage and reduces the structural degradation of generated samples through an adversarial training mechanism, making the reconstructed lithium battery operating data more consistent with the real distribution.

[0084] The main function of the dynamic feature discrimination unit is to distinguish between data generated at different time steps t and real battery data, and to guide the optimization of the reverse recovery strategy through adversarial training, so that the generated data more closely matches the distribution of real SOH data. Its application process mainly includes the following steps:

[0085] Dynamic feature identification unit The input includes data from the diffusion process. (Including real data sampled from the battery dataset and data generated by the data augmentation module) and time step t (representing the current stage of the diffusion process, taking discrete integer values ​​t=1, 2, ..., T). To better utilize the time step information, the time step t is encoded using a multilayer perceptron (MLP), i.e. First, t is normalized, and then converted into a time feature vector through MLP.

[0086] Discriminator Taking the encoding of the data and time step t as input, output the probability that its source is real data or generated data:

[0087] (6)

[0088] in, The weight matrix represents the data features. The weight matrix represents the features at each time step. For bias terms, This is the Sigmoid function, used to output the probability of data being true (values ​​range from 0 to 1).

[0089] The core objective of augmentation models is to learn a reverse recovery process from random noise. Recover the real data Its training process is mainly based on mean squared error (MSE) loss, and the optimization objective is as follows:

[0090] (7)

[0091] in, This is data from the forward perturbation process; It's Gaussian noise. It's the noise in the model's predictions; the goal is to reduce the noise in the model's predictions. Approximates real noise This ensures the stability of the reverse recovery process. A cumulative scaling factor, used to gradually adjust the mixing ratio of data and noise during forward perturbation, is defined as:

[0092] (8)

[0093] Where t is the time step, The variance scheduling parameters for the s-th time step.

[0094] In the data perturbation enhancement model, a dynamic feature identification unit is added to further improve the authenticity and quality of the generated data. Therefore, combating losses It consists of two parts:

[0095] (9)

[0096] The first term indicates that the discriminator expects to analyze the real data. The second term indicates the probability the discriminator expects for the generated data to be true; It gives a lower probability (i.e., the prediction is false).

[0097] Finally, during the training of the data perturbation enhancement model, the adversarial training loss of the dynamic feature discrimination unit is combined with the back-recovery loss, and the optimization objective is:

[0098] (10)

[0099] Through this adversarial training, the discriminator continuously optimizes its discrimination capabilities, and the data augmentation module generates data that more closely approximates real SOH data.

[0100] S104, Enhanced Sample Generation and Application

[0101] After the data perturbation enhancement model is trained, it can be used to generate fake samples with real feature distributions from Gaussian noise to enhance the data scale and diversity of the SOH prediction task.

[0102] Specifically, the first step is to perform Gaussian initialization, which involves randomly sampling initial noise samples from a standard normal distribution. ( (This represents a completely degraded Gaussian noise sample, with the same dimension as the original input features). Then, using the trained diffusion inverse process model, the formula is iteratively executed from t=T to t=1:

[0103] (11)

[0104] in, , To enhance the neural network trained by the model, used to predict current noise, , These are the preset noise dispatcher parameters used during the disturbance process. (where is the variance adjustment coefficient for the sampling process), after T-step reverse diffusion processing, the final pseudo-sample is obtained:

[0105] (12)

[0106] Since the input features have been uniformly normalized to the interval [0, 1] or [-1, 1] before training, it is necessary to perform an inverse normalization operation on the generated results, i.e. Subsequently, based on the sample's location in the original feature space, spatial proximity analysis is used to find the K nearest neighbors to the generated sample's features in the original sample set, and a weighted average is used to generate the SOH label. This ultimately leads to the formation of pseudo-sample pairs. Finally, the generated pseudo-sample set... With the original real sample set The datasets are merged to form the enhanced, complete dataset. This augmented dataset has a richer sample distribution and coverage area, which helps to improve the training stability, robustness and generalization ability of subsequent SOH prediction models (such as selective state space sequence models), especially when the scale of the original labeled data is limited.

[0107] Specifically, sequence feature extraction and health status estimation are performed on the redistributed data, including:

[0108] A selective state-space sequence model incorporating multi-layer nonlinear transformations and state control mechanisms is used to mine temporal features of redistributed data.

[0109] Specifically, the selective state-space sequence model includes a deep dynamic filtering unit, an information filtering unit, a multi-scale dependency capture unit, and a health state estimation unit connected in sequence.

[0110] Specifically, the deep dynamic filtering unit processes the input sequence data based on an adaptive filtering strategy;

[0111] The information filtering unit filters the processed data through a gated linear structure;

[0112] The multi-scale dependency capture unit extracts local and global temporal dependency features of the time series in parallel across multiple feature spaces;

[0113] The health status estimation unit performs aggregate analysis on the extracted time-dependent features and outputs preliminary prediction results of the lithium battery health status.

[0114] The deep dynamic filtering unit performs dynamic convolution operations on the input sequence using a learnable state-space filter to obtain filtered feature data.

[0115] The information filtering unit filters the filtered feature data.

[0116] The multi-scale dependency capture unit captures the local and global dependencies of the lithium battery at different time scales in parallel across multiple subspaces.

[0117] The health status estimation unit integrates the processed features and outputs the predicted state of health (SOH) of the lithium battery through the fully connected layer.

[0118] In one specific embodiment of the present invention, the dataset after being enhanced by the data perturbation enhancement model... Efficient modeling and feature extraction are performed to learn the underlying dynamic patterns in the time series, and accurate regression prediction is made for the SOH state corresponding to each input sequence. A selective state-space sequence model is used to process the enhanced data. The selective state-space sequence model is an efficient sequence model architecture that can effectively capture the complex dependencies in lithium battery time series data. The entire processing can be divided into the following stages:

[0119] S201, Deep Dynamic Filtering Unit: The Selective State-Space Sequence Model innovatively introduces a learnable state-space filter specifically designed to capture long-term dependency features in lithium battery time-series data. The Selective State-Space Sequence Model performs a "dynamic convolution" operation on the input sequence, achieving cross-step information propagation at each time step t through state transitions. Each input sequence is represented by a parameterized state-space model:

[0120] (13)

[0121] in, It is the feature representation of the input sequence at time step t. This is the updated hidden state, which incorporates the hidden state from the previous moment. With current input This mechanism has lower complexity (approximately linear) compared to standard convolution / self-attention, while possessing the ability to remember long-term dependencies, making it suitable for handling long-sequence problems such as battery life.

[0122] S202, Information Filtering Unit: After filtering, features undergo information selection control through the information filtering unit.

[0123] (14)

[0124] Among them, A and B are the linear transformation results extracted from the previous layer (i.e., the deep dynamic filtering unit); This is element-wise multiplication, used to multiply A by a product of elements activated by the Sigmoid function. The processed B is multiplied element by element; The range of B is mapped to the interval (0, 1), thereby enabling the filtering of information.

[0125] GLU can selectively retain important information and mask noise features, making it particularly suitable for handling potentially redundant data. In lithium battery data processing, it can identify key features that truly reflect the changing trends of battery health, eliminate interference information generated during the data generation process, enhance the robustness of the model, and provide higher-quality data features for subsequent processing.

[0126] S203, Multi-scale Dependency Capture Unit: Data enters the multi-scale attention mechanism. The multi-scale attention mechanism in selective state-space sequence models is used to capture the local and global dependencies of the input sequence at different time scales, constructing rich time series representations. First, data perturbation enhances the time series features generated by the model. Through linear transformation, these are mapped to query, key, and value vectors respectively:

[0127] (15)

[0128] in , , It is a learnable weight matrix. , It refers to the number of attention subchannels.

[0129] First, a scaled dot product attention score is computed independently for each attention subchannel:

[0130] (16)

[0131] Then, the above calculation process is performed in parallel on multiple attention subchannels to form multiple subspace attention outputs:

[0132] (17)

[0133] Finally, the outputs of all attention sub-channels are concatenated and then processed by a linear transformation matrix. The fusion process yields the final multi-scale attention output:

[0134] (18)

[0135] The multi-scale mechanism models time dependencies in parallel across multiple subspaces, accurately capturing the characteristic differences of lithium batteries at different aging stages. Compared to traditional sequence models, it has advantages such as modeling long-term dependencies, maintaining temporal resolution, and adapting to multiple operating conditions, significantly improving the robustness of SOH prediction.

[0136] S204, Health Status Estimation Unit: After processing by the deep dynamic filtering unit, information filtering unit, and multi-scale attention mechanism, the selective state-space sequence model integrates the learned feature information and outputs the estimated SOH of the lithium battery through fully connected layers and other methods. During the training phase, the model uses minimizing the mean squared error between the predicted value and the true SOH as the optimization objective, and the loss function is expressed as:

[0137] (19)

[0138] in, The number of training samples. For the first The true SOH value of each sample This is the corresponding predicted value.

[0139] The model uses the backpropagation algorithm to automatically adjust various parameters in the model by propagating the error calculated by the loss function back along the network structure.

[0140] Specifically, feature distribution alignment is performed based on the mean difference measurement method to reduce the impact of different data distribution differences on model performance; obtaining feature distribution differences includes: calculating the feature distribution differences between source domain and target domain samples in high-dimensional feature space, and dynamically adjusting the parameters in the feature extraction process based on the measurement results using an adaptive matching strategy to minimize feature distribution differences;

[0141] Among them, during the feature distribution alignment process, the health status estimation bias is dynamically monitored;

[0142] When the estimation error exceeds the preset threshold, the weight of the distribution mean difference measure is automatically adjusted, the feature distribution difference is recalculated and the parameters are updated until the prediction error converges to the preset range.

[0143] Specifically, a distributed mean-difference algorithm is used for cross-domain knowledge adaptation optimization to correct the lithium battery SOH prediction results output by the selective state-space sequence model, and the final SOH prediction value is output, including:

[0144] Distributed mean difference is used as a metric, and based on the Gaussian kernel function, the source domain feature vector set and the target domain feature vector set are mapped to a high-dimensional kernel Hilbert space respectively.

[0145] Next, calculate the feature distribution alignment loss between the source domain feature vector set and the target domain feature vector set in the high-dimensional kernel Hilbert space;

[0146] The selective state-space sequence model is trained based on the feature distribution alignment loss to obtain a new comprehensive loss function;

[0147] The gradient descent method is used to optimize the weights of the selective state-space sequence model by iteratively updating the weights. This improves the model's feature extraction capabilities.

[0148] Specifically, during the cross-domain knowledge adaptation and optimization process, the prediction error of the selective state-space sequence model is continuously monitored;

[0149] If the prediction error exceeds the set threshold, the weight parameters of the feature distribution alignment loss are dynamically adjusted, the feature distribution alignment loss is recalculated and the model weights are optimized and trained again. This process is repeated until the prediction error of the selective state space sequence model on the target domain data converges to a certain range, and the final SOH prediction value is output.

[0150] In one specific embodiment of the present invention, in order to effectively reduce prediction errors caused by different data distributions, enable the model to quickly and accurately adapt to new operating environments, and enhance the model's generalization ability and prediction accuracy in different distributed energy storage scenarios, a distributed mean difference algorithm is adopted. By measuring and aligning the feature distributions of the source and target domain data, an efficient domain adaptation process is achieved, allowing the model to perform well on new datasets.

[0151] Source domain feature vector set ( )Depend on The intermediate sequence features extracted after selective state-space sequence modeling. These features reflect typical characteristics of battery aging behavior in the training dataset. Target domain feature vector set ( The sequence feature representation is generated from battery operation data from real deployment environments or target test scenarios, after being processed by the same selective state-space sequence model structure.

[0152] To achieve consistent alignment of feature distributions between the source and target domains, distributed mean difference is used as a metric, and based on the Gaussian kernel function, the mean difference is... and The features are mapped to a high-dimensional kernel Hilbert space (RKHS), and their mean distance in that space is calculated. The core idea is to minimize the mean difference between the two distributions in the RKHS, making the distributions of the two feature sets as similar as possible in the latent space. This mean difference is the feature distribution alignment loss, and its mathematical expression is:

[0153] (20)

[0154] in, This represents the result after mapping to a high-dimensional feature space using a kernel function; , These represent samples in the source and target domains, respectively. and This represents the number of samples in the source and target domains.

[0155] Using a Gaussian kernel function, the feature distribution alignment loss can be further expressed as:

[0156] (twenty one)

[0157] in, The kernel function is preferably a Gaussian kernel function, which is defined as follows:

[0158] (twenty two)

[0159] Wherein, σ is the kernel bandwidth parameter, which controls the sensitivity of the mapping space to local differences. The specific value can be determined by cross-validation or automatically optimized as an adjustable hyperparameter.

[0160] The calculated feature distribution alignment loss is incorporated into the model's training objective, forming a new comprehensive loss function together with the original prediction loss:

[0161] (twenty three)

[0162] Among them, the original predicted loss The deviation of the predicted SOH value from the actual SOH value is measured by mean square error (MSE) and is expressed as shown in formula (19). λ This is an adjustable hyperparameter used to balance the weights of prediction loss and feature distribution alignment loss. Subsequently, gradient descent is used to optimize the weights of the selective state-space sequence model network. By iteratively updating the weights, the comprehensive loss function gradually decreases, allowing the model to better adapt to the distribution characteristics of the target domain data while learning the relationship between source domain data features and SOH.

[0163] During the cross-domain knowledge adaptation optimization process, the model's prediction error is continuously monitored. If the prediction error exceeds a set threshold, it indicates that the model's adaptation effect to the target domain is poor. At this point, the weights λ are dynamically adjusted, and the feature distribution alignment loss is recalculated and the model weights are retrained for optimization. This process is repeated until it stabilizes, the model's prediction error on the target domain data converges to an acceptable range, and finally, accurate and reliable SOH prediction values ​​are output.

[0164] In one specific embodiment of the present invention, a lithium battery health state estimation system based on perturbation recovery enhancement and cross-domain adaptation mechanism includes:

[0165] The data redistribution module is used to collect raw data of lithium batteries with health status tags and generate redistributed data through a data perturbation enhancement model.

[0166] The feature extraction and estimation module is used to extract sequence features and estimate the health status of the redistributed data, output the preliminary prediction results of the health status of the lithium battery, and calculate the health status estimation bias.

[0167] The cross-domain knowledge adaptation and optimization module is used to align feature distributions based on the distribution mean difference measurement method to obtain feature distribution differences.

[0168] The result output module is used to construct a joint loss function based on the health status estimation bias and the feature distribution difference, and obtain the final lithium battery health status prediction result by iteratively updating the parameters.

[0169] The present invention proposes a lithium battery state of health estimation method and system based on perturbation recovery enhancement and cross-domain adaptation mechanism, which breaks through the problem of strong dependence of traditional SOH estimation methods on a large amount of labeled data, as well as the dilemma of data scarcity and weak generalization ability that are common in data-driven methods.

[0170] The data perturbation enhancement model efficiently generates highly realistic pseudo-samples through a unique forward diffusion and backward recovery process combined with the adversarial training mechanism of the dynamic feature discrimination unit. This module accurately learns the distribution of a small amount of real data and generates samples with realistic feature distributions, greatly expanding the capacity and diversity of the dataset. This provides a rich and high-quality data foundation for subsequent model training and effectively alleviates the problem of insufficient original labeled data.

[0171] Selective state-space sequence models capture long-term data dependency features with the help of deep dynamic filtering units, filter key information with information filtering units, and mine dependencies at different time scales through multi-scale attention mechanisms. This results in a deep understanding and processing capability of lithium battery operation data time series, significantly improving the accuracy of SOH prediction.

[0172] A cross-domain knowledge adaptation optimization strategy based on feature distribution alignment accurately measures and aligns the feature distributions of data in the source and target domains, dynamically adjusting the optimization process to address data distribution offset issues. By calculating the feature distribution alignment loss and incorporating it into the training objective, and combining it with gradient descent to optimize model weights, the prediction accuracy of the model in the new scenario (target domain) is ensured, effectively solving the key problem of poor model adaptability.

[0173] Specifically, such as Figure 2 As shown, in terms of prediction accuracy, most datasets are concentrated in the range of approximately 0 to 0.0025, with a frequency of over 800 in this range. This indicates that in most cases, the model's predicted values ​​are very close to the true values ​​with minimal deviation, meaning that the model exhibits high accuracy in the SOH estimation task and can accurately predict the true health status values.

[0174] From the perspective of the concentration of data distribution, the difference distribution shows a clear concentration trend, with a large amount of data concentrated in a few intervals. This reflects that the model's prediction results are relatively stable, and the difference between the predicted value and the actual value fluctuates little at different data points, demonstrating good stability and consistency.

[0175] This invention employs a data perturbation enhancement model to simulate the degradation and regeneration of original data labeled with State of Health (SOH). It integrates a time-series discrimination mechanism to generate high-fidelity extended samples, constructing an enhanced dataset to address the scarcity of original labeled data. A selective state-space sequence model is used to perform hierarchical analysis on the enhanced data, employing dynamic filtering, gating, and multi-scale dependency capture to uncover the long- and short-term dynamic patterns of lithium battery operation data, outputting initial SOH prediction results. A cross-domain knowledge adaptation optimization strategy is leveraged, quantifying the differences in feature distribution between the source and target domains through a distribution similarity measure mechanism. Dynamic parameter tuning achieves adaptive alignment of data distribution in new environments, correcting the prediction results and outputting high-precision SOH values. This invention achieves high-precision estimation of SOH for distributed energy storage lithium batteries, suitable for scenarios where the amount of original SOH-labeled data is scarce. It can improve the applicability and accuracy of the prediction model by generating effective extended training samples and adapting to new environmental data distributions across domains.

[0176] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0177] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for estimating the state of health of a lithium battery based on perturbation recovery enhancement and cross-domain adaptation mechanisms, characterized in that, include: Raw data of lithium batteries with health status tags are collected, and redistributed data is generated through a data perturbation enhancement model. Sequence features are extracted and health status is estimated from the redistributed data, and preliminary prediction results of the health status of lithium batteries are output. The health status estimation bias is also calculated. Feature distribution alignment is performed based on the method of measuring the difference in distribution mean to obtain the feature distribution difference; A joint loss function is constructed based on the health status estimation bias and the feature distribution difference. The parameters are updated iteratively to obtain the final lithium battery health status prediction result. The data perturbation enhancement model includes a perturbation generation unit and a dynamic feature identification unit, which are used to introduce perturbations into the original lithium battery data, generate perturbation data samples, and then restore the perturbation data samples to obtain the restored redistribution data. The forward perturbation process is described by the following formula: Where, f(x) 1:T |x0) is the sum of all data samples x from time step 1 to T, given the initial data x0. 1:T The joint probability distribution, f(x) t |x t-1 () represents the data sample x at the previous time step t-1, where the data is known. t-1 In the case of the current time step t, the data sample x t The conditional probability distribution, Let I represent the Gaussian distribution, and let α represent the identity matrix. t Indicates variance scheduling; The reverse recovery process starts from x, which is heavily contaminated by noise. T The original data x0 is gradually recovered; the process is represented as follows: in, Indicates in model parameters Under the influence of this, all data samples x from time step 0 to T 0:T The joint probability distribution of ; This represents the data sample x at the known current time step t. t In the case of the data sample x from the previous time step t-1 t-1 The conditional probability distribution is determined by the model parameters. Decide; Dynamic feature discrimination unit C(x) t The input to ,t) includes data x from the diffusion process. t And time step t; the probability of the output data being true; C(x t ,t)=σ(W d ·x t +W t ·τ t +b d ); Among them, W d W is the weight matrix of the data features. t Let b be the weight matrix of the time step features. d Here, σ(·) is the bias term, and σ(·) is the Sigmoid function; The adversarial loss V(C,θ) consists of two parts: The first term represents the discriminator's desired discrimination of the real data x ~ f(x). 1:T |x0) gives a higher probability; the second term indicates that the discriminator wants the generated data to be of higher probability. Give a lower probability; From random noise x T The real data x0 is recovered, and its training process is based on mean squared error loss. The optimization objective is as follows: in, This is data from the forward perturbation process; It's Gaussian noise. It's the noise in the model's predictions; the goal is to reduce the noise in the model's predictions. Approximates real noise ∩. This is a cumulative scaling factor; During training, the data perturbation enhancement model combines the adversarial training loss of the dynamic feature discrimination unit with the back-recovery loss, with the optimization objective being: Using distributed mean difference as a metric, and based on the Gaussian kernel function, X... s With X t Each feature distribution is mapped to a high-dimensional kernel Hilbert space, and then the mean distance in that space is calculated to obtain the feature distribution alignment loss. The calculated feature distribution alignment loss is incorporated into the model's training objective, forming a new comprehensive loss function together with the original prediction loss: THE total =L SOH +λL DSD ; Among them, L SOH L represents the original prediction loss; λ is an adjustable hyperparameter. DSD Indicates feature alignment loss; Sequence feature extraction and health status estimation of redistributed data, including: A selective state-space sequence model incorporating multi-layer nonlinear transformations and state control mechanisms is used to mine temporal features of redistributed data. The selective state-space sequence model includes a deep dynamic filtering unit, an information filtering unit, a multi-scale dependency capture unit, and a health state estimation unit connected in sequence. The deep dynamic filtering unit processes the input sequence data based on an adaptive filtering strategy; The information filtering unit filters the processed data through a gated linear structure; The multi-scale dependency capture unit extracts local and global temporal dependency features of the time series in parallel across multiple feature spaces; The health status estimation unit performs aggregate analysis on the extracted time-dependent features and outputs preliminary prediction results of the lithium battery health status.

2. The lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism according to claim 1, characterized in that, The disturbance generation unit is used to add Gaussian noise to the original lithium battery data based on a preset disturbance scheduling strategy until a preset disturbance level is reached, thereby generating multi-stage degradation samples. The dynamic feature identification unit is used to fuse time step information, distinguish the differences between generated samples and real data at different stages during training, and improve the reverse recovery capability of the perturbation generation unit through an adversarial optimization mechanism.

3. The lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism according to claim 1, characterized in that, After the perturbation generation unit is trained, it generates initial pseudo-samples from random Gaussian noise and reconstructs their feature structure through a reverse recovery mechanism.

4. The lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism according to claim 3, characterized in that, The reverse recovery mechanism is used to gradually remove noise and reconstruct the original data features until the noise-free state is restored, thus obtaining pseudo samples; A neighborhood feature matching strategy is adopted to calculate the feature similarity between pseudo-samples and original samples, determine the corresponding health status labels, and merge pseudo-samples with the original dataset to construct an extended dataset for subsequent training.

5. The lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism according to claim 1, characterized in that, Feature distribution alignment is performed based on the mean difference measurement method to obtain the feature distribution difference, including: calculating the feature distribution difference between the source domain and the target domain samples in the high-dimensional feature space, and dynamically adjusting the parameters in the feature extraction process based on the measurement result using an adaptive matching strategy to minimize the feature distribution difference; Among them, during the feature distribution alignment process, the health status estimation bias is dynamically monitored; When the estimation error exceeds the preset threshold, the weight of the distribution mean difference measure is automatically adjusted, the feature distribution difference is recalculated and the parameters are updated until the prediction error converges to the preset range.

6. A lithium battery health state estimation system based on perturbation recovery enhancement and cross-domain adaptation mechanism, employing the lithium battery health state estimation method based on perturbation recovery enhancement and cross-domain adaptation mechanism as described in any one of claims 1-5, characterized in that, include: The data redistribution module is used to collect raw data of lithium batteries with health status tags and generate redistributed data through a data perturbation enhancement model. The feature extraction and estimation module is used to extract sequence features and estimate the health status of the redistributed data, output the preliminary prediction results of the health status of the lithium battery, and calculate the health status estimation bias. The cross-domain knowledge adaptation and optimization module is used to align feature distributions based on the distribution mean difference measurement method to obtain feature distribution differences. The result output module is used to construct a joint loss function based on the health status estimation bias and the feature distribution difference, and obtain the final lithium battery health status prediction result by iteratively updating the parameters.

Images