Adaptive Gated Method and System for Estimating the Health Status of Lithium-ion Batteries Across Operating Conditions

Abstract

This invention discloses an adaptive gating method and system for estimating the state of health (SOH) of lithium-ion batteries across operating conditions, relating to the fields of battery management and artificial intelligence. The invention includes: preprocessing operating data of tagged batteries in the source domain and untagged batteries in the target domain; evaluating task heterogeneity using surrogate A distance (PAD); adaptively selecting a lightweight temporal branch or an enhanced multi-scale branch based on the PAD results; performing feature alignment using MK-MMD under highly heterogeneous tasks; dynamically adjusting the domain alignment loss weights using a progressive preheating strategy; and finally outputting the SOH estimation results for the target domain battery. This invention can adaptively adjust the model structure and alignment strength for different cross-operating condition migration tasks, reducing the risk of negative migration and improving the accuracy, stability, and adaptability of lithium-ion battery SOH estimation under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the fields of battery management and artificial intelligence technology, specifically to an adaptive gating method and system for estimating the health status of lithium-ion batteries across operating conditions. Background Technology

[0002] Lithium-ion batteries, due to their high energy density, long cycle life, and good power performance, are widely used in electric vehicles and energy storage systems. With increasing service life, batteries undergo irreversible degradation during repeated charge-discharge cycles, manifesting as capacity decay, increased internal resistance, and heightened safety risks. Therefore, accurate estimation of the battery's State of Health (SOH) is crucial to support operational monitoring, maintenance decisions, and lifespan management within battery management systems. While existing SOH estimation methods can learn battery degradation patterns from historical operational data, most assume that training and test data follow the same distribution. When charging rates, fast-charging protocols, environmental conditions, or degradation trajectories change, significant distribution shifts can easily occur between the source and target domains, leading to a decrease in the model's generalization ability across operating conditions. Existing unsupervised domain adaptation methods typically employ fixed network structures and fixed alignment strategies. In low-domain-shift scenarios, this can introduce unnecessary forced alignment, while in high-domain-shift scenarios, insufficient representation capabilities may result in inadequate alignment, making it difficult to meet the adaptation requirements of different transfer tasks. In addition, existing methods often perform degradation representation learning and cross-domain distribution alignment simultaneously throughout the training process. However, in the early stages of training, when stable degradation features have not yet been formed, introducing strong alignment constraints too early can easily interfere with the learning of key degradation patterns and reduce training stability.

[0003] To address this, the present invention proposes an adaptive gating method and system for estimating the health status of lithium-ion batteries across operating conditions. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive gating method and system for estimating the state of health (SOH) of lithium-ion batteries across operating conditions, so as to solve the problems of insufficient adaptability, easy generation of negative transfer and poor training stability in the existing technology for cross-operating condition SOH estimation.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions, comprising the following steps: Receive source domain tagged battery operation data and target domain untagged battery operation data, and preprocess the data to obtain source domain input time series samples and target domain input time series samples; The predefined proxy A distance PAD is calculated based on the source domain input time series samples and the target domain input time series samples. The proxy A distance PAD is used to characterize the degree of inter-domain heterogeneity between the source domain and the target domain. The distance of agent A to PAD is compared with a preset heterogeneity threshold, and a gating signal is generated based on the comparison result; A health status estimation model is constructed, which includes a feature extraction network, a selective domain alignment module, and a health status prediction layer. The feature extraction network includes a lightweight temporal feature extraction branch for extracting temporal dynamic features and an enhanced multi-scale temporal feature extraction branch for extracting multi-scale degradation features. Based on the gated signal and the source domain input time series samples and the target domain input time series samples, the feature extraction branch is adaptively selected, and the source domain features and target domain features are output. The adaptive selection strategy is as follows: When the distance of agent A to PAD is less than the heterogeneity threshold, a lightweight temporal feature extraction branch is used, and domain alignment is weakened or turned off; when the distance of agent A to PAD is greater than or equal to the heterogeneity threshold, an enhanced multi-scale temporal feature extraction branch is used, and domain alignment is enabled. When the gating signal indicates the start of domain alignment, the multi-core maximum mean difference (MK-MMD) is introduced to perform distribution alignment between the source domain features and the target domain features, and the aligned source domain features and target domain features are output. A progressive warm-up strategy is adopted to dynamically adjust the domain alignment loss weights, and the source domain supervised regression loss and domain alignment loss are jointly optimized to train the health status estimation model. Input the time-series samples of the target domain to be detected into the trained health state estimation model, and output the battery health state (SOH) estimation results of the target domain.

[0006] Furthermore, the system receives operating data from tagged batteries in the source domain and operating data from untagged batteries in the target domain, and preprocesses the data to obtain source domain input time-series samples and target domain input time-series samples, as follows: Voltage, current and capacity time-series data during the constant current charging phase of the battery are selected as input features, and each cycle sample is uniformly resampled and normalized to obtain a fixed-length time-series input sample. The State of Health (SOH) of a lithium-ion battery is defined as the ratio of the current capacity to the initial capacity or rated capacity. in, Let be the capacity corresponding to the t-th iteration. This refers to the battery's initial capacity or rated capacity.

[0007] Furthermore, the predefined agent A distance PAD is calculated based on the source domain input time series samples and the target domain input time series samples, as follows: Source domain sample matrix Target domain sample matrix Flattened to form eigenvectors; Random resampling is performed on the source domain samples and the target domain samples to balance the number of samples in the two domains; Principal component analysis (PCA) was used to reduce the dimensionality of the combined features after equilibrium. A linear support vector machine domain discriminant is constructed based on the dimensionality reduction results, and the average generalization error is obtained through cross-validation. ; According to the average generalization error Calculate the distance of agent A to PAD: In the formula, D t Indicates the target domain. D s Represents the source domain. PAD represents the average generalization error of the domain discriminator in a linear support vector machine. The larger the PAD value, the greater the macroscopic distribution difference between the source domain and the target domain.

[0008] Furthermore, the distance of agent A to PAD is compared with a preset heterogeneity threshold, and a gating signal is generated based on the comparison result, as follows: in, To preset the heterogeneity threshold; when At that time, a lightweight temporal feature extraction branch is used, and distribution alignment is weakened or turned off; when At that time, an enhanced multi-scale temporal feature extraction branch is adopted, and cross-domain distribution alignment is enabled.

[0009] Furthermore, the enhanced multi-scale temporal feature extraction branch includes one-dimensional convolutional feature extraction, temporal coding, and self-attention aggregation, as detailed below: The self-attention aggregation process is represented as: Finally, the sequence encoder outputs the features at the last moment. With attention aggregation features By concatenating the data, we obtain a robust feature representation: in, H cnn For local features extracted by one-dimensional convolution, H seq is the output feature of time-series coding. Let Q be the key vector dimension, and let K and V be the query vector matrix, key vector matrix, and value vector matrix in the self-attention mechanism, respectively.h last The output features of the sequence encoder at the last moment. h att For attention-aggregated features, Concat represents the feature concatenation operation.

[0010] Furthermore, the lightweight temporal feature extraction branch includes a cyclic temporal coding unit, which is used to directly extract long-short-term dependency features from the preprocessed input temporal samples.

[0011] Furthermore, when the gating signal indicates the start of domain alignment, multi-core maximum mean difference (MK-MMD) is introduced to align the source domain features and the target domain features, and the aligned source domain features and target domain features are output, as follows: Multi-kernel maximum mean difference (MK-MMD) employs a family of multi-bandwidth Gaussian kernel functions, whose kernel function is expressed as: in, This represents the output value of the multi-kernel function, where x and y represent the two feature vectors input to the kernel function. This represents the square of the Euclidean distance between eigenvectors x and y. The number of kernel functions. For the first The bandwidth parameters of a Gaussian kernel; The corresponding distribution alignment loss is expressed as: in, For source domain features, For target domain features, , Let i and j represent the features of the i-th and j-th samples in the source domain in the current batch, respectively. , Let i and j represent the features of the i-th and j-th samples in the target domain in the current batch, respectively. and These represent the number of source and target domain samples in the current batch, respectively.

[0012] Furthermore, a progressive warm-up strategy is adopted to dynamically adjust the domain alignment loss weights, and the source domain supervised regression loss and domain alignment loss are jointly optimized to train the health state estimation model, as detailed below: in, For the current training round, Based on alignment strength, This refers to the length of the preheating stage; The joint optimization objective function is expressed as: in, For source domain supervision regression loss, This represents the cross-domain distribution alignment loss.

[0013] Furthermore, during model training and model selection, the target domain real labels are not used for parameter updates or early stopping. Training is performed using only source domain supervision information and unlabeled target domain distribution information to avoid target domain label leakage.

[0014] According to a second aspect of the present invention, the present invention provides a lithium-ion battery cross-operating condition health state estimation system for implementing the adaptive gating lithium-ion battery cross-operating condition health state estimation method described in the first aspect, comprising: The data receiving and preprocessing module is used to receive source domain tagged battery operation data and target domain untagged battery operation data, and preprocess the data to obtain source domain input time series samples and target domain input time series samples. The heterogeneity assessment module is used to calculate a predefined proxy A distance PAD based on the source domain input time series samples and the target domain input time series samples. The proxy A distance PAD is used to characterize the degree of inter-domain heterogeneity between the source domain and the target domain. The gating decision module is used to compare the distance of agent A to PAD with a preset heterogeneity threshold and generate a gating signal based on the comparison result. The model building module is used to build a health status estimation model. The health status estimation model includes a feature extraction network, a selective domain alignment module, and a health status prediction layer. The feature extraction network includes a lightweight temporal feature extraction branch for extracting temporal dynamic features and an enhanced multi-scale temporal feature extraction branch for extracting multi-scale degradation features. The adaptive feature extraction module is used to adaptively select a feature extraction branch based on the gated signal and the source domain input time series samples and the target domain input time series samples, and output source domain features and target domain features. The adaptive selection strategy is as follows: When the distance of agent A to PAD is less than the heterogeneity threshold, a lightweight temporal feature extraction branch is used, and domain alignment is weakened or turned off; when the distance of agent A to PAD is greater than or equal to the heterogeneity threshold, an enhanced multi-scale temporal feature extraction branch is used, and domain alignment is enabled. The selective distribution alignment module is used to introduce multi-core maximum mean difference (MK-MMD) to perform distribution alignment of the source domain features and the target domain features when the gating signal indicates the start domain alignment, and outputs the aligned source domain features and target domain features. The model training module is used to dynamically adjust the domain alignment loss weights using a progressive warm-up strategy, and jointly optimize the source domain supervised regression loss and the domain alignment loss to train the health status estimation model. The SOH output module is used to input the time-series samples of the target domain to be detected into the trained health state estimation model and output the SOH estimation results of the battery health state in the target domain.

[0015] This invention has at least the following beneficial effects: 1. This invention uses the proxy A distance as a quantitative indicator to assess the degree of distribution offset between the source and target domains before model training. Based on this indicator, a gating signal is generated, and the feature extraction branch and domain alignment strategy are adaptively selected. This enables the invention to flexibly handle various cross-scenario migration tasks, from slight fluctuations in operating conditions to extreme changes in operating conditions, and to maintain high estimation accuracy and stability without human intervention.

[0016] 2. In the process of jointly optimizing the source domain supervised regression loss and the domain alignment loss, this invention introduces a progressive warm-up strategy to dynamically adjust the alignment loss weight, which overcomes the mutual interference problem caused by optimizing two objectives at the same time in the early stage of training in traditional methods. This makes the model convergence process smoother and more stable, and ultimately obtains better cross-condition estimation performance.

[0017] 3. The enhanced multi-scale temporal feature extraction branch designed in this invention integrates one-dimensional convolutional local feature extraction, temporal coding long-term dependency modeling, and self-attention global aggregation, enabling it to capture subtle change patterns in the battery degradation process at multiple time scales. Combined with multi-kernel maximum mean difference alignment using multi-bandwidth Gaussian kernel functions, it can measure and reduce the difference in feature distribution between the source and target domains at multiple scales. This allows the model to extract robust, domain-invariant health state features even under highly complex operating conditions where charging rate, ambient temperature, and aging path change simultaneously, thus obtaining stable and reliable health state estimation results.

[0018] 4. Throughout the entire model training and model selection process, this invention strictly avoids using the actual health status labels of the target domain for parameter updates or early stopping judgments. Instead, it relies solely on source domain supervision information and unlabeled target domain distribution information for optimization. This fully aligns with the practical constraints of obtaining target domain labels in real-world cross-condition application scenarios, enabling the technical solution of this invention to have real deployability while ensuring the fairness and objectivity of comparisons between different methods.

[0019] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the estimation method described in this invention; Figure 2 This is a diagram of the data-aware adaptive gating network structure described in this invention. Part A is a schematic diagram of multi-condition data preprocessing, and Part B is a schematic diagram of the adaptive feature extraction and multi-kernel maximum mean difference alignment structure. Figure 3 This is a schematic diagram illustrating how the alignment loss weight changes with training rounds in the progressive warm-up mechanism described in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] Example 1: Please see Figure 1 This invention provides a technical solution: an adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions, comprising the following steps: S1. Receive the operating data of the tagged battery in the source domain and the operating data of the untagged battery in the target domain, and preprocess the data to obtain the input time series samples of the source domain and the input time series samples of the target domain; Acquire tagged battery cycle data from the source domain and untagged battery cycle data from the target domain. Preferably, the voltage, current, and capacity sequences during the constant current charging phase are selected as input features, such as... Figure 2 As shown in Part A, this embodiment performs unified preprocessing on battery operating data under different operating conditions. Using battery cycle data from different sources as input, capacity-related information and voltage, current, and capacity time-series features during the constant-current charging phase are extracted from each cycle sample. These time-series features are then resampled and normalized to form a fixed-length input sample that can be used for subsequent heterogeneity assessment and network modeling. Figure 2 Part A of the document primarily describes the data organization and preprocessing methods under multiple operating conditions. The target domain, State of Health (SOH), is defined as the ratio of the current capacity to the initial capacity or rated capacity. in, Let be the capacity corresponding to the t-th iteration. The initial or rated capacity of the battery S2. Calculate the predefined surrogate A distance PAD based on the source domain input time series samples and the target domain input time series samples. The surrogate A distance PAD is used to characterize the degree of inter-domain heterogeneity between the source domain and the target domain. (S21) The source domain sample matrix Target domain sample matrix Flattened to form eigenvectors; (S22) Perform random resampling on the source domain samples and the target domain samples to balance the number of samples in the two domains; (S23) Perform principal component analysis (PCA) to reduce the dimensionality of the joint features after equilibrium. (S24) Construct a linear support vector machine domain discriminant based on the dimensionality reduction results, and obtain the average generalization error through cross-validation. ; (S25) Based on the average generalization error Calculate the distance of agent A to PAD: In the formula, D t Indicates the target domain. D s Represents the source domain. PAD represents the average generalization error of the domain discriminator in a linear support vector machine. The larger the PAD value, the greater the macroscopic distribution difference between the source domain and the target domain. S3. Compare the distance of agent A to PAD with a preset heterogeneity threshold, and generate a gating signal based on the comparison result, as follows: in, To preset the heterogeneity threshold; when At that time, a lightweight temporal feature extraction branch is used, and distribution alignment is weakened or turned off; when At that time, an enhanced multi-scale temporal feature extraction branch is adopted, and cross-domain distribution alignment is enabled; S4. Construct a health status estimation model, which includes a feature extraction network, a selective domain alignment module, and a health status prediction layer. The feature extraction network includes a lightweight temporal feature extraction branch for extracting temporal dynamic features and an enhanced multi-scale temporal feature extraction branch for extracting multi-scale degradation features. S5. Based on the gating signal and the source domain input time series samples and the target domain input time series samples, adaptively select the feature extraction branch and output the source domain features and target domain features. The adaptive selection strategy is as follows: When the distance of agent A to PAD is less than the heterogeneity threshold, a lightweight temporal feature extraction branch is used, and domain alignment is weakened or turned off; when the distance of agent A to PAD is greater than or equal to the heterogeneity threshold, an enhanced multi-scale temporal feature extraction branch is used, and domain alignment is enabled. like Figure 2 As shown in Part B, the adaptive feature extraction module of this embodiment receives source domain input samples and target domain input samples, and selects different feature extraction paths under the control of gating signals. For low heterogeneity tasks, a lightweight temporal feature extraction branch is adopted to extract battery degradation dynamic features with low model complexity. For high heterogeneity tasks, an enhanced multi-scale feature extraction branch is adopted to jointly model local features, temporal features and global features. The lightweight branch only requires a small number of convolutional layers or recurrent units to complete feature extraction in low heterogeneity tasks. Its forward inference time is short and its memory usage is small, making it very suitable for deployment on vehicle battery management systems or edge computing terminals. The enhanced branch is only activated when necessary, avoiding the waste of resources caused by using a heavy model uniformly in all migration tasks, and realizing the intelligent design of "allocating computing power on demand". At the same time, using a fixed strong alignment strategy or complex feature extractor may destroy the original good feature structure due to over-alignment in low heterogeneity scenarios, which will reduce the accuracy of target domain estimation (i.e. "negative transfer"). This embodiment guides different tasks to the most suitable feature extraction path through a gating mechanism, so that the model always achieves an appropriate balance between "fidelity" and "alignment", which significantly improves the adaptability to diverse cross-condition tasks. Furthermore, the lightweight temporal feature extraction branch includes a cyclic temporal coding unit, which is used to directly extract long-short-term dependency features from the preprocessed input temporal samples. Furthermore, the enhanced multi-scale temporal feature extraction branch includes one-dimensional convolutional feature extraction, temporal coding, and self-attention aggregation, as detailed below: The self-attention aggregation process is represented as: Finally, the sequence encoder outputs the features at the last moment. With attention aggregation features By concatenating the data, we obtain a robust feature representation: in, H cnn For local features extracted by one-dimensional convolution, H seq is the output feature of time-series coding. Let Q be the key vector dimension, and let K and V be the query vector matrix, key vector matrix, and value vector matrix in the self-attention mechanism, respectively. h last The output features of the sequence encoder at the last moment. h att For attention-aggregated features, Concat represents the feature concatenation operation; Specifically, such as Figure 2 As shown in Part B, the enhanced multi-scale feature extraction branch includes a local feature extraction module, a temporal feature extraction module, and a global feature extraction module. The local feature extraction module extracts local degradation morphological information in the input sequence through one-dimensional convolution, normalization, activation, and pooling operations. The temporal feature extraction module extracts long-term dependencies in the degradation process through recurrent temporal coding units. The global feature extraction module constructs a self-attention aggregation process through query vector Q, key vector K, and value vector V, and concatenates and linearly maps the attention aggregation result with the temporal coding output to obtain a robust feature representation for target domain SOH estimation. S6. When the gating signal indicates that domain alignment is initiated, the multi-core maximum mean difference (MK-MMD) is introduced to align the source domain features and target domain features, and the aligned source domain features and target domain features are output as follows: Multi-kernel maximum mean difference (MK-MMD) employs a family of multi-bandwidth Gaussian kernel functions, whose kernel function is expressed as: in, This represents the output value of the multi-kernel function, where x and y represent the two feature vectors input to the kernel function. This represents the square of the Euclidean distance between eigenvectors x and y. The number of kernel functions. For the first The bandwidth parameters of a Gaussian kernel; The corresponding distribution alignment loss is expressed as: in, For source domain features, For target domain features, , Let i and j represent the features of the i-th and j-th samples in the source domain in the current batch, respectively. , Let i and j represent the features of the i-th and j-th samples in the target domain in the current batch, respectively. and These represent the number of source and target domain samples in the current batch, respectively. Specifically, such as Figure 2As shown in the lower half of section B, in highly heterogeneous tasks, this embodiment further performs multi-kernel maximum mean difference (MK-MMD) distribution alignment on the source domain features and the target domain features. The distribution alignment module maps the source domain features and the target domain features to the regenerating kernel Hilbert space, and uses multiple Gaussian kernel functions with different bandwidths to jointly characterize the distribution difference of the two domain features, thereby improving the cross-domain feature consistency under complex working conditions. Figure 2 The multi-kernel maximum mean difference alignment structure illustrates the distribution relationship between source domain features and target domain features in the original feature space and kernel mapping space. This shows that the present invention enhances the representation ability of multi-scale domain shift through multi-kernel mapping. For low heterogeneity tasks, since the difference between source domain and target domain features is relatively small, the distribution alignment constraint is weakened or turned off to avoid unnecessary interference to the original degenerate structure. S7. A progressive warm-up strategy is adopted to dynamically adjust the domain alignment loss weights, and the source domain supervised regression loss and domain alignment loss are jointly optimized to train the health status estimation model, as follows: like Figure 3 As shown, a progressive warm-up mechanism is used to dynamically adjust the cross-domain distribution alignment loss weights. In the early stages of training, the alignment loss weights gradually increase with each training epoch to avoid the model being prematurely affected by strong distribution alignment constraints before stable degradation features are formed. After the training epochs reach the preset warm-up stage length, the alignment loss weights remain constant, thus continuously performing stable cross-domain alignment in subsequent training processes. The formula for dynamically adjusting the alignment loss weights is as follows: in, For the current training round, Based on alignment strength, This refers to the length of the preheating stage; Figure 3 The left-hand region in the diagram represents the progressive warm-up phase, during which the alignment loss weight gradually increases from zero. Figure 3 The right-hand region in the diagram represents the constant alignment phase, during which the alignment loss weights remain stable. Through this mechanism, this embodiment can progressively enhance cross-domain distribution constraints while ensuring the quality of the degraded representation, thereby improving model training stability and transfer robustness. The joint optimization objective function is expressed as: in, For source domain supervision regression loss, Cross-domain distribution alignment loss; It should be further explained that during model training and model selection, the true labels of the target domain are not used for parameter updates or early stopping. Training is carried out using only source domain supervision information and unlabeled target domain distribution information to avoid target domain label leakage. S8. Input the time-series samples of the target domain to be detected into the trained health state estimation model, and output the battery health state (SOH) estimation results for the target domain, as follows: After training, the target domain input samples are fed into the health status estimation model, and the target domain lithium-ion battery SOH estimation results are output. The prediction process can be expressed as follows: in, The first output of the model represents the... SOH estimates for each target domain sample; This indicates that after training and optimization in step S6, the network parameters are optimal. The mapping function (i.e., the health status estimation model after training); Indicates the input number of the first... Features of the target domain samples to be tested; This represents the total number of samples to be evaluated in the target domain.

[0024] Example 2: This embodiment provides a lithium-ion battery cross-operating condition health state estimation system to implement the adaptive gating lithium-ion battery cross-operating condition health state estimation method described in Embodiment 1, including: The data receiving and preprocessing module is used to receive the operating data of the source domain tagged battery and the target domain untagged battery, and to preprocess the data to obtain the source domain input time series samples and the target domain input time series samples. The heterogeneity assessment module is used to calculate a predefined proxy A-distance (PAD) based on the source domain input time series samples and the target domain input time series samples. The proxy A-distance (PAD) is used to characterize the degree of inter-domain heterogeneity between the source domain and the target domain. The gating decision module is used to compare the distance of agent A to PAD with a preset heterogeneity threshold and generate a gating signal based on the comparison result. The model building module is used to build a health status estimation model. The health status estimation model includes a feature extraction network, a selective domain alignment module, and a health status prediction layer. The feature extraction network includes a lightweight temporal feature extraction branch for extracting temporal dynamic features and an enhanced multi-scale temporal feature extraction branch for extracting multi-scale degradation features. The adaptive feature extraction module is used to adaptively select the feature extraction branch based on the gating signal and the source domain input time series samples and the target domain input time series samples, and output source domain features and target domain features. The adaptive selection strategy is as follows: When the distance of agent A to PAD is less than the heterogeneity threshold, a lightweight temporal feature extraction branch is used, and domain alignment is weakened or turned off; when the distance of agent A to PAD is greater than or equal to the heterogeneity threshold, an enhanced multi-scale temporal feature extraction branch is used, and domain alignment is enabled. The selective distribution alignment module is used to introduce multi-core maximum mean difference (MK-MMD) to perform distribution alignment of source domain features and target domain features when the gating signal indicates the start domain alignment, and outputs the aligned source domain features and target domain features. The model training module is used to dynamically adjust the domain alignment loss weights using a progressive warm-up strategy, and jointly optimize the source domain supervised regression loss and the domain alignment loss to train the health status estimation model. The SOH output module is used to input the time-series samples of the target domain to be detected into the trained health state estimation model and output the SOH estimation results of the battery health state in the target domain.

[0025] Example 3: The present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. The memory stores the computer program capable of running on the processor. When the processor loads and executes the computer program, it adopts the adaptive gating lithium-ion battery cross-operating condition health state estimation method described in Embodiment 1.

[0026] It should be noted that the terminal device can be a computer device such as a desktop computer, a laptop computer, or a cloud server, and the terminal device includes, but is not limited to, a processor and a memory. For example, the terminal device may also include input / output devices, network access devices, and buses.

[0027] Furthermore, the processor can be a central processing unit (CPU). Of course, depending on the actual use, other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be used. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it in this regard.

[0028] Example 4: The present invention provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform the adaptive gating method for estimating the cross-condition health status of a lithium-ion battery described in Embodiment 1.

[0029] The computer program can be stored in a computer-readable medium. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or certain middleware. The computer-readable medium includes any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the computer-readable medium includes, but is not limited to, the above-mentioned components.

[0030] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of this application. Those skilled in the art should understand that this application is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of this application. Various changes and modifications can be made to this application without departing from the spirit and scope thereof, and all such changes and modifications fall within the scope of the claims of this application.

Claims

1. An adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions, characterized in that, Includes the following steps: Receive source domain tagged battery operation data and target domain untagged battery operation data, and preprocess the data to obtain source domain input time series samples and target domain input time series samples; The predefined proxy A distance PAD is calculated based on the source domain input time series samples and the target domain input time series samples. The proxy A distance PAD is used to characterize the degree of inter-domain heterogeneity between the source domain and the target domain. The distance of agent A to PAD is compared with a preset heterogeneity threshold, and a gating signal is generated based on the comparison result; A health status estimation model is constructed, which includes a feature extraction network, a selective domain alignment module, and a health status prediction layer. The feature extraction network includes a lightweight temporal feature extraction branch for extracting temporal dynamic features and an enhanced multi-scale temporal feature extraction branch for extracting multi-scale degradation features. Based on the gated signal and the source domain input time series samples and the target domain input time series samples, the feature extraction branch is adaptively selected, and the source domain features and target domain features are output. The adaptive selection strategy is as follows: When the distance of agent A to PAD is less than the heterogeneity threshold, a lightweight temporal feature extraction branch is used, and domain alignment is weakened or turned off; when the distance of agent A to PAD is greater than or equal to the heterogeneity threshold, an enhanced multi-scale temporal feature extraction branch is used, and domain alignment is enabled. When the gating signal indicates the start of domain alignment, the multi-core maximum mean difference (MK-MMD) is introduced to perform distribution alignment between the source domain features and the target domain features, and the aligned source domain features and target domain features are output. A progressive warm-up strategy is adopted to dynamically adjust the domain alignment loss weights, and the source domain supervised regression loss and domain alignment loss are jointly optimized to train the health status estimation model. Input the time-series samples of the target domain to be detected into the trained health state estimation model, and output the battery health state (SOH) estimation results of the target domain.

2. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 1, characterized in that: The system receives operating data from tagged batteries in the source domain and operating data from untagged batteries in the target domain, and preprocesses the data to obtain source domain input time series samples and target domain input time series samples, as follows: Voltage, current and capacity time-series data during the constant current charging phase of the battery are selected as input features, and each cycle sample is uniformly resampled and normalized to obtain a fixed-length time-series input sample. The State of Health (SOH) of a lithium-ion battery is defined as the ratio of the current capacity to the initial capacity or rated capacity. in, Let be the capacity corresponding to the t-th iteration. This refers to the battery's initial capacity or rated capacity.

3. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 1, characterized in that: The predefined agent A distance PAD is calculated based on the source domain input time series samples and the target domain input time series samples, as follows: Source domain sample matrix Target domain sample matrix Flattened out, they become eigenvectors; Random resampling is performed on the source domain samples and the target domain samples to balance the number of samples in the two domains; Principal component analysis (PCA) was used to reduce the dimensionality of the combined features after equilibrium. A linear support vector machine domain discriminant is constructed based on the dimensionality reduction results, and the average generalization error is obtained through cross-validation. ; According to the average generalization error Calculate the distance of agent A to PAD: In the formula, D t Indicates the target domain. D s Represents the source domain. PAD represents the average generalization error of the domain discriminator in a linear support vector machine. The larger the PAD value, the greater the macroscopic distribution difference between the source domain and the target domain.

4. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 1, characterized in that: The distance of agent A to PAD is compared with a preset heterogeneity threshold, and a gating signal is generated based on the comparison result, as follows: in, To preset the heterogeneity threshold; when At that time, a lightweight temporal feature extraction branch is used, and distribution alignment is weakened or turned off; when At that time, an enhanced multi-scale temporal feature extraction branch is adopted, and cross-domain distribution alignment is enabled.

5. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 4, characterized in that: The enhanced multi-scale temporal feature extraction branch includes one-dimensional convolutional feature extraction, temporal coding, and self-attention aggregation, as detailed below: The self-attention aggregation process is represented as: Finally, the sequence encoder outputs the features at the last moment. With attention aggregation features By concatenating the data, robust feature representations are obtained. F out : in, H cnn For local features extracted by one-dimensional convolution, H seq is the output feature of time-series coding. Let Q be the key vector dimension, and let K and V be the query vector matrix, key vector matrix, and value vector matrix in the self-attention mechanism, respectively. h last The output features of the sequence encoder at the last moment. h att For attention-aggregated features, Concat represents the feature concatenation operation.

6. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 1, characterized in that: The lightweight temporal feature extraction branch includes a cyclic temporal coding unit, which is used to directly extract long-short-term dependency features from the preprocessed input temporal samples.

7. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 5, characterized in that: When the gating signal indicates the start of domain alignment, multi-core maximum mean difference (MK-MMD) is introduced to align the source domain features and the target domain features, and the aligned source domain features and target domain features are output as follows: Multi-kernel maximum mean difference (MK-MMD) employs a family of multi-bandwidth Gaussian kernel functions, whose kernel function is expressed as: in, This represents the output value of the multi-kernel function, where x and y represent the two feature vectors input to the kernel function. This represents the square of the Euclidean distance between eigenvectors x and y. The number of kernel functions. For the first The bandwidth parameters of a Gaussian kernel; The corresponding distribution alignment loss is expressed as: in, For source domain features, For target domain features, , Let i and j represent the features of the i-th and j-th samples in the source domain in the current batch, respectively. , Let i and j represent the features of the i-th and j-th samples in the target domain in the current batch, respectively. and These represent the number of source and target domain samples in the current batch, respectively.

8. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 1, characterized in that: A progressive warm-up strategy is adopted to dynamically adjust the domain alignment loss weights, and the source domain supervised regression loss and domain alignment loss are jointly optimized to train the health status estimation model, as follows: in, For the current training round, Based on alignment strength, This refers to the length of the preheating stage; The joint optimization objective function is expressed as: in, For source domain supervision regression loss, This represents the cross-domain distribution alignment loss.

9. The adaptive gating method for estimating the health status of lithium-ion batteries across operating conditions according to claim 1, characterized in that: During model training and model selection, the target domain real labels are not used for parameter updates or early stopping. Training is performed using only source domain supervision information and unlabeled target domain distribution information to avoid target domain label leakage.

10. A lithium-ion battery cross-operating-condition health state estimation system, used to implement the adaptive gating lithium-ion battery cross-operating-condition health state estimation method according to any one of claims 1 to 9, characterized in that, include: The data receiving and preprocessing module is used to receive source domain tagged battery operation data and target domain untagged battery operation data, and preprocess the data to obtain source domain input time series samples and target domain input time series samples. The heterogeneity assessment module is used to calculate a predefined proxy A distance PAD based on the source domain input time series samples and the target domain input time series samples. The proxy A distance PAD is used to characterize the degree of inter-domain heterogeneity between the source domain and the target domain. The gating decision module is used to compare the distance of agent A to PAD with a preset heterogeneity threshold and generate a gating signal based on the comparison result. The model building module is used to build a health status estimation model. The health status estimation model includes a feature extraction network, a selective domain alignment module, and a health status prediction layer. The feature extraction network includes a lightweight temporal feature extraction branch for extracting temporal dynamic features and an enhanced multi-scale temporal feature extraction branch for extracting multi-scale degradation features. The adaptive feature extraction module is used to adaptively select a feature extraction branch based on the gated signal and the source domain input time series samples and the target domain input time series samples, and output source domain features and target domain features. The adaptive selection strategy is as follows: When the distance of agent A to PAD is less than the heterogeneity threshold, a lightweight temporal feature extraction branch is used, and domain alignment is weakened or turned off; when the distance of agent A to PAD is greater than or equal to the heterogeneity threshold, an enhanced multi-scale temporal feature extraction branch is used, and domain alignment is enabled. The selective distribution alignment module is used to introduce multi-core maximum mean difference (MK-MMD) to perform distribution alignment of the source domain features and the target domain features when the gating signal indicates the start domain alignment, and outputs the aligned source domain features and target domain features. The model training module is used to dynamically adjust the domain alignment loss weights using a progressive warm-up strategy, and jointly optimize the source domain supervised regression loss and the domain alignment loss to train the health status estimation model. The SOH output module is used to input the time-series samples of the target domain to be detected into the trained health state estimation model and output the SOH estimation results of the battery health state in the target domain.

Images