Lithium battery residual life prediction method and device based on cross-temperature-domain transfer learning

Abstract

The invention discloses a lithium battery residual life prediction method based on cross-temperature-domain transfer learning. The method comprises the following steps: collecting charge-discharge cycle data and EIS characteristics under multiple temperature domains; obtaining a degradation curve based on the charge-discharge cycle data and carrying out nonlinear time alignment; constructing a multi-level temperature sensing vector; carrying out sub-band feature extraction and multi-modal fusion on the EIS features to obtain EIS fusion features; constructing a Hybrid Net model and carrying out pre-training; taking the aligned degradation curve and EIS fusion features as input, and respectively carrying out feature extraction and fusion; generating a temperature gating weight based on a multi-level temperature sensing vector, and carrying out the temperature self-adaptive weighting of the fusion features; outputting a predicted value of the remaining life of the lithium battery based on the fusion features after temperature gating; and to-be-predicted charge and discharge cycle data and EIS features are collected and input into the trained Hybrid Net model, and a life prediction value is output. According to the method, errors during temperature-crossing migration can be remarkably reduced, and prediction stability and interpretability are improved.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery health management technology, specifically to a method and apparatus for predicting the remaining life of lithium-ion batteries based on cross-temperature domain transfer learning. Background Technology

[0002] Under various ambient temperature conditions, the degradation kinetics of lithium-ion batteries exhibit significant differences, manifested in changes in the time scale and morphology of capacity decay curves, as well as the drift of health characteristics across temperature domains. Traditional single-temperature-domain modeling methods are prone to temporal mismatches and reduced generalization ability when migrating across domains; while pure physical priors can characterize the effect of temperature on reaction rates, they are difficult to achieve precise characterization under complex degradation pathways. EIS can reflect internal mechanisms such as ohmic impedance, charge transfer impedance, and diffusion processes, but there is still a lack of unified engineering solutions for effectively integrating it with degradation time sequences such as capacity and internal resistance. For example, Chinese patent CN120217090A discloses a method, system, device, and medium for classifying the health of power batteries based on deep learning algorithms and EIS technology. This method acquires EIS data of the target power battery, constructs an equivalent circuit model including components such as solution resistance, SEI layer resistance, and charge transfer impedance, extracts feature parameter factors, and uses recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) to build a prediction model for training. Furthermore, Chinese patent CN117031306A discloses a method for predicting the remaining lifetime (RUL) of a battery using EIS based on a multi-model combination, including the following steps: S1 testing and collecting battery electrochemical impedance spectroscopy data and RUL data; S2 using a PCA model to extract principal components from the EIS input data; S3 using a TCN model to extract the spatial features of EIS; S4 using a BiLSTM model to extract the temporal series variation patterns of EIS; S5 using an Attention model to extract important parts of the spatiotemporal features of EIS; S6 predicting the remaining lifetime (RUL) of the battery: after the fully connected layer is the model output layer, and the predicted value is the remaining lifetime of the battery.

[0003] Therefore, there is an urgent need for a RUL prediction scheme that can achieve coordinated optimization of time alignment, temperature adaptation, mechanism enhancement and domain distribution alignment in cross-temperature domain scenarios. Summary of the Invention

[0004] This invention proposes a method and apparatus for predicting the remaining life (RUL) of lithium-ion batteries based on cross-temperature domain transfer learning. This method and apparatus can significantly reduce the error during cross-temperature transfer and improve the prediction stability and interpretability.

[0005] To achieve the above objectives, the present invention includes the following technical steps: A method for predicting the remaining life of a lithium battery using cross-temperature domain transfer learning, the method comprising the following steps: S1. Collect lithium battery charge-discharge cycle data under multiple temperature domains and construct a time-series feature matrix, and simultaneously collect EIS features; S2. Based on the charge-discharge cycle data collected in S1, the degradation curves at different temperatures are obtained, and nonlinear time alignment is performed based on the time-series feature matrix to obtain the aligned degradation curves. S3. Construct a multi-level temperature sensing vector based on temperature and EIS features; S4. Perform frequency band feature extraction and multimodal fusion on the EIS features to obtain the EIS fused features; S5. Construct the HybridNet model and pre-train it, including: S5-1. Using the degenerate curve aligned by S2 and the EIS fusion feature of S4 as input, and performing feature extraction separately, the fusion feature is obtained by fusing them together. S5-2. Based on the multi-level temperature sensing vector constructed in S3, temperature-gated weights are generated, and the fused features are subjected to temperature-adaptive weighting to obtain the temperature-gated fused features. S5-4. Based on the temperature-gated fusion features, output the predicted value of the remaining life of the lithium battery. S6. Collect the charge-discharge cycle data and EIS features of the lithium battery to be predicted and input them into the trained HybridNet model to output the lifetime prediction value.

[0006] In S1, the lithium battery charge-discharge cycle data under the multi-temperature domain includes voltage V(t), current I(t), capacity C(t), and temperature T.

[0007] In S2, based on the time-series feature matrix constructed in S1, a dynamic time warping algorithm with adaptive window width is used to perform nonlinear time alignment on the battery degradation curves in different temperature domains, resulting in aligned degradation curves.

[0008] In S2, feature alignment and knowledge transfer across temperature domains are achieved through a joint loss function in nonlinear time alignment.

[0009] In S3, the method for constructing a multi-level temperature sensing vector based on temperature and EIS features is as follows: S3-1. Construct the temperature-EIS feature mapping matrix; S3-2. Extract the temperature-sensitive principal components from the matrix constructed in S3-1 through principal component analysis to obtain the temperature-sensitive subspace; S3-3. Project the EIS features onto the temperature-sensitive subspace to obtain the temperature-sensitive subspace mapping based on EIS; S3-4. Construct a hybrid temperature vector as a multi-level temperature sensing vector based on temperature and EIS-based temperature-sensitive subspace mapping.

[0010] In S4, the methods for frequency band feature extraction and multimodal fusion of EIS features include: S4-1. The spectrum is divided into three sub-bands: low frequency band (0.01-1Hz), mid frequency band (1-100Hz), and high frequency band (100Hz-10kHz) using a dynamic segmentation method based on information gain. S4-2. The equivalent circuit parameters, including ohmic resistance, charge transfer resistance, double-layer capacitance and diffusion impedance, are extracted using an identification method based on equivalent circuits. S4-3. The cross-attention mechanism is used to perform multimodal fusion of the segmented spectrum of S4-1 and the equivalent circuit parameters of S4-2 to obtain the EIS fusion characteristics.

[0011] In S5, the HybridNet model includes: A multimodal input layer is used to receive the aligned degradation curve of S2 and the EIS fusion feature of S4 as input, and the actual lifetime prediction value is used as the supervision label. A time-series encoder is used to perform time-series modeling and degradation feature extraction on the aligned degradation curve to obtain a time-series feature representation that characterizes the battery degradation trend. Impedance encoders are used to encode the frequency domain mechanism features of EIS fusion features to obtain impedance feature representations characterizing ohmic impedance, charge-transfer impedance and diffusion process; The feature fusion layer is used to perform cross-modal deep fusion of the temporal feature representation and impedance feature representation through a cross-attention mechanism to obtain a unified fused feature representation; The temperature gating layer is used to generate temperature gating weights from the multi-level temperature sensing vector constructed based on S3, and to perform temperature-adaptive weighting on the fused feature representation in order to achieve dynamic modulation and consistency of features across temperature domains. The decoder is used to perform residual lifetime RUL decoding regression on the temperature-gated fused features and can simultaneously perform uncertainty estimation. The output layer is used to output the predicted remaining life of the lithium battery and its uncertainty parameters.

[0012] The HybridNet model includes an improved Arrhenius compensation module, which performs temperature compensation on the temperature gating layer based on a learnable Arrhenius function. The physical constraints in the learnable Arrhenius function include: Monotonicity constraint: Ensures that the reaction rate constant increases monotonically within the operating temperature range; Continuity constraint: The compensation function maintains first-order continuity at the temperature boundary; Boundary condition constraint: At standard temperature, the compensation function value is 1; Physical rationality constraints: Activation energy must be kept within a reasonable range. Internal, pre-exponential factor .

[0013] To improve the physical consistency of the model, this invention designs a learnable Arrhenius temperature compensation equation, which enhances cross-temperature generalization ability while maintaining the monotonicity and continuity of the reaction rate. Finally, by combining multi-domain alignment and transfer learning strategies, a joint loss function is used to achieve feature distribution consistency and knowledge transfer, and the remaining lifetime prediction and its uncertainty are output based on the HybridNet model.

[0014] The present invention also provides a lithium battery remaining life prediction device based on cross-temperature domain transfer learning. The device includes a memory and one or more processors. The memory stores an executable program. When the one or more processors execute the program, they can implement the lithium battery remaining life prediction method based on cross-temperature domain transfer learning.

[0015] The present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the lithium battery remaining life prediction method based on cross-temperature domain transfer learning.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The method provided by this invention achieves robust lifetime prediction and few-sample transfer adaptation under different temperature conditions through the collaborative design of time alignment, temperature adaptive gating, EIS multimodal fusion, temperature compensation for physical constraints, and domain alignment training strategies, and solves the problems of time misalignment of degradation curves across temperature regions, feature drift, and insufficient generalization ability. The method provided by this invention takes into account both data-driven and physical priors; it can significantly reduce errors during cross-temperature migration and improve prediction stability and interpretability, and experiments have shown that it has significant improvements over several typical baseline models in various cross-temperature migration scenarios. This invention can effectively solve the problems of misalignment of battery degradation curves, feature drift and poor generalization at different temperatures. Compared with traditional models, it improves the prediction of cross-temperature domains by an average of about 30%-50%, and has high engineering application value and universality. Attached Figure Description

[0017] Figure 1 A flowchart of a lithium battery remaining life prediction method using cross-temperature domain transfer learning provided as an example; Figure 2 This is the system architecture of the HybridNet model in the embodiment; Figure 3 This is a schematic diagram of cross-temperature domain (DTW) alignment in the embodiment; Figure 4 The embodiment shows a temperature-adaptive gating structure; Figure 5 This is the EIS feature extraction and fusion process in the example; Figure 6 This is an example of an improved Arrhenius temperature compensation relationship. Figure 7 This is the transfer learning training and domain alignment process in the embodiment; Figure 8 This is a schematic diagram of relative enhancement and ablation in the embodiments; Figure 9 This example compares the lifetime predictions of different prediction models. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but the present invention is not limited to the following embodiments.

[0019] like Figure 1 As shown in the figure, this embodiment provides a method for predicting the remaining life of a lithium battery through cross-temperature domain transfer learning. The specific method includes the following steps.

[0020] S1. Collect lithium battery charge-discharge cycle data under multiple temperature domains and construct a time-series feature matrix, and simultaneously collect EIS features; This embodiment performs unified preprocessing and feature construction on battery time-series data acquired under multiple temperature domains. The raw data includes voltage data acquired under different ambient temperatures and operating conditions. ),temperature and optional internal resistance V(t): Voltage at time t. The current at time t. : Capacity at time t, both are used equally in this invention, T(t): Temperature at time t, R(t): Internal resistance at time t (as provided in the dataset).

[0021] S1-1. Data Cleaning: Outlier detection and missing value handling are performed on the original time series. Outliers can be detected and removed or replaced using the 3σ criterion or median filtering based on a sliding window; missing values ​​can be filled using linear interpolation or spline interpolation to ensure sequence continuity.

[0022] S1-2. Sampling Unification: For cases where sampling intervals are inconsistent under different operating conditions, all sequences are unified to a preset time step through resampling or interpolation. The above yields an equally spaced sampling sequence.

[0023] S1-3. Normalization and Standardization: For Features are normalized or standardized respectively, for example, using Standardization or Normalization maps features of different dimensions to a similar numerical range, which facilitates subsequent model training.

[0024] S1-4. Sliding Window Construction: On each processed sequence, a sliding slice is made with a window length n and a step size s. The feature matrix is ​​constructed for the k-th window. ; in, for The feature matrix, where d is the feature dimension. This is the starting time of the loop. For the first The starting time index of each sliding window, n is the number of time steps within a single sliding window, t is the time series sampling time index, and k is the sample or window index; The set of temporal features is obtained by overlaying all windows: N is the total number of windows (sample size); Through the above preprocessing, this embodiment ensures that the original data under different temperature zones and operating conditions meet the same standard in terms of time sampling, scale range and feature form, providing a foundation for cross-temperature migration.

[0025] S2. Based on the charge-discharge cycle data collected in S1, the degradation curves at different temperatures are obtained, and nonlinear time alignment is performed based on the time-series feature matrix to obtain the aligned degradation curves. In this embodiment, an improved DTW dynamic time warping alignment step is adopted. An adaptive window width dynamic time warping algorithm is used to perform nonlinear time alignment of battery degradation curves in different temperature domains. During the alignment process, a joint loss function is used to achieve feature alignment and knowledge transfer across temperature domains. Wherein: A temperature-differential adaptive Sakoe-Chiba constraint is employed, with a constraint bandwidth of [missing information]. Where α is the temperature sensitivity coefficient and ΔT is the temperature difference. For reference temperature; The cost function uses the weighted Mahalanobis distance: , where Σ is the covariance matrix, and i and j are the time indices of the two sequences when aligned by DTW; The step constraint uses a symmetrical P-type constraint: This ensures the monotonicity and continuity of time alignment.

[0026] The improved adaptive window width strategy in the DTW algorithm is as follows: Initial window width Determined based on sequence length and temperature difference: , where k is the temperature sensitivity coefficient; Dynamic adjustment strategy: During the DTW path search process, the window width is dynamically adjusted according to the cumulative cost gradient. When the gradient change rate exceeds the threshold θ, the window width is expanded to allow for greater time deformation. Path smoothing constraint: Path smoothing regularization term is used. Ensure the alignment path is smooth.

[0027] In this embodiment, as Figure 3 As shown, this embodiment employs an improved Dynamic Time Warping (DTW) algorithm to achieve nonlinear time alignment for degradation curves at different temperatures; S2 specifically includes the following steps: S2-1. Based on the time-series feature matrix X obtained in step S1, at the reference temperature... Calculate the temperature difference at the current temperature T. ,use The DTW search deformation bandwidth w(ΔT) is determined (adaptively adjusted with temperature difference) to limit excessive deformation; where, T is the reference temperature, L is the current sample temperature, k and α are hyperparameters for controlling bandwidth.

[0028] S2-2. During the path search process, the DTW cumulative cost matrix D(i,j) adopts the following recursive form. ,in Adaptively select from Euclidean distance, Mahalanobis distance, or cosine distance using the aforementioned method. For example, if weighted Mahalanobis distance is used: .

[0029] S2-3, DTW Alignment Path Satisfy step constraints: This ensures the monotonicity and continuity of the path. Among them, The length of the DTW alignment path P.

[0030] like Figure 7 As shown, this embodiment uses a joint loss function to align the feature distributions of the source and target domains. The specific method is as follows: S2-4. The specific calculation method for the multi-domain alignment loss function is as follows: The total loss is: ; MMD loss: These are the mean embeddings of the source and target domains in the reproducing kernel Hilbert space, respectively. CORAL loss: Let be the covariance matrix of the source and target domains, and d be the feature dimension; Combat losses: Where D is the domain discriminator and G is the feature extractor; Adaptive weight adjustment: This achieves dynamic balance during the training process.

[0031] S2-5. Source and Target Domain Division: Based on temperature or operating conditions, different temperature zones or datasets are divided into a source domain and a target domain, denoted as X^(s) and X^(t) respectively, with corresponding label sets denoted as Y^(s) and Y^(t). X, X^(s), and X^(t) serve as the unified input data format for subsequent steps. X^(s): Source domain feature set; X^(t): Target domain feature set.

[0032] In the early stages of training, this embodiment increases To enhance alignment, the number of parameters is gradually reduced during the later stages of training to ensure stable convergence of the prediction head. By pre-training on multi-temperature region data and fine-tuning with a small number of samples at the target temperature, rapid adaptation to new temperature regions can be achieved under limited labeling conditions.

[0033] In this embodiment, the meanings of each parameter are as follows: : RUL prediction loss (e.g., mean squared error); Maximum mean difference (MMD) loss; Covariance alignment (CORAL) loss or domain correlation alignment loss; : Adversarial domain alignment loss; Physical constraint loss (including physical constraints on k(T), etc.); , , , : Weights of each loss; Total loss function; : Feature covariance matrix of source and target domains; : Feature vectors of the source and target domains; , : Mean embedding of the source and target domains in the reproducing kernel Hilbert space; D(·): Domain discriminator; G(·): Feature extractor.

[0034] After the above alignment, this embodiment obtains the aligned features. This serves as the input for step S5.

[0035] S3. Construct a multi-level temperature sensing vector based on temperature and EIS features; In this embodiment, S3 specifically includes the following steps: S3-1. Constructing the Temperature-EIS Feature Mapping Matrix ,in For frequency points, This represents the number of temperature sampling points. S3-2. Extract the temperature-sensitive principal components from the matrix constructed in S3-1 using principal component analysis to obtain the temperature-sensitive subspace: Extract the temperature-sensitive principal components using principal component analysis (PCA): The first k principal components are retained as temperature-sensitive features. S3-3, Subspace Projection: Projecting EIS features onto a temperature-sensitive subspace yields an EIS-based temperature-sensitive subspace mapping. Specifically, this involves projecting the input EIS features onto the temperature-sensitive subspace to obtain a temperature-related EIS representation. ,in It is the projection matrix formed by the first k principal components; S3-4. Construct a hybrid temperature vector as a multi-level temperature sensing vector based on temperature and EIS-based temperature-sensitive subspace mapping: Temperature encoding adopts a hybrid strategy. ,in One-hot encoding for discrete temperatures. These are continuous temperature normalized values. For temperature-sensitive subspace mapping based on EIS; In this embodiment, EIS feature matrix, dimension 1 ,in For frequency points, This represents the number of temperature samples. :right The first k principal component matrices obtained by performing principal component analysis (PCA); : Feature vector of a single EIS measurement; Features projected onto the temperature-sensitive subspace; Discrete temperature region one-hot encoded vector; : Normalized continuous temperature scalar; The constructed multi-level temperature sensing vector is applied to step S5 to adaptively weight the fused features.

[0036] S4. Perform frequency band feature extraction and multimodal fusion on the EIS features to obtain the EIS fused features; In the case of available EIS data, this embodiment performs adaptive segmentation and multimodal fusion on the EIS data in the 0.01Hz to 10kHz frequency band. S4 specifically includes the following steps: S4-1. Divide the spectrum on the logarithmic frequency axis according to information gain or spectral characteristics. ,For example: For low frequency, For intermediate frequency, The frequency band is high. In this embodiment, a dynamic segmentation method based on information gain is used to divide the spectrum into three sub-bands: low frequency band (0.01-1Hz), mid frequency band (1-100Hz), and high frequency band (100Hz-10kHz).

[0037] S4-2. Extract equivalent circuit parameters using an equivalent circuit-based identification method. Including ohmic resistors Charge transfer resistance Double-layer capacitors Warburg diffusion impedance (Optional); specifically includes: S4-2-1 Circuit Topology Identification: Automatically identify Randle circuits, dual time constant circuits, or transmission line models based on impedance spectrum shape characteristics; S4-2-2, Initial Parameter Estimation: Estimate the initial values ​​of circuit parameters using frequency domain feature points (high-frequency intercept, low-frequency intercept, characteristic frequency); S4-2-3, Nonlinear Fitting Optimization: The Levenberg-Marquardt algorithm is used for accurate parameter fitting. For each frequency point f, the equivalent circuit model Z_model(f;θ) is used to fit the measured impedance. The optimization objective is: , where θ is the circuit parameter vector; where, The complex impedance measured at frequency f. The complex impedance given by the equivalent circuit model under parameter θ; S4-2-4. Fit Quality Assessment: The fit quality is assessed through chi-square test and residual analysis, and the optimal circuit model is automatically selected.

[0038] S4-3. A cross-attention mechanism is used to perform multimodal fusion of the segmented spectrum of S4-1 and the equivalent circuit parameters of S4-2 to obtain the EIS fusion features: feature vectors are extracted for each frequency band. And obtain weights through attention mechanism The final fusion features are: If a dataset does not provide EIS data, this embodiment will automatically... If set to a zero vector or an empty vector, this invention degenerates into an implementation that relies solely on timing branches. : Feature vector extracted from the b-th frequency band Attention weights for the b-th frequency band. : The fused EIS feature vector; In this embodiment, the EIS fusion feature obtained in step S4 is used as the input for S5.

[0039] S5. Construct the HybridNet model and pre-train it, including: S5-1, a multimodal input layer, is used to receive the aligned degradation curve of S2 and the EIS fusion feature of S4 as input, and the actual lifetime prediction value as the supervision label. S5-2, a time encoder, is used to perform time-series modeling and degradation feature extraction on aligned degradation curves / time-series features to obtain a time-series feature representation characterizing the battery degradation trend; S5-3, Impedance encoder, used to encode the frequency domain mechanism features of EIS fusion features to obtain impedance feature representations characterizing ohmic impedance, charge-transfer impedance and diffusion process; S5-4, Feature Fusion Layer, used to perform cross-modal deep fusion of the time-series feature representation and impedance feature representation through a cross-attention mechanism to obtain a unified fused feature representation; S5-5, Temperature Gating Layer, is used to generate temperature gating weights from the multi-level temperature sensing vector constructed based on S3, and to perform temperature-adaptive weighting on the fused feature representation in order to achieve dynamic modulation and consistency of features across temperature domains. S5-6. Improved Arrhenius compensation module: temperature compensation is achieved for the temperature gating layer based on a learnable Arrhenius function. S5-7, the decoder, is used to perform residual lifetime (RUL) decoding regression on the fused features after adaptive weighting by the temperature-gated layer and compensation by the improved Arrhenius compensation module, and can simultaneously perform uncertainty estimation. S5-8, Output Layer, is used to output the predicted remaining lifespan of the lithium battery and its uncertainty parameters.

[0040] In the multimodal input layer of S5-1: for the lifetime prediction task, based on the cycle number or remaining capacity corresponding to the end time of each window, a remaining lifetime (RUL) label for that window is constructed, forming... Sample pairs.

[0041] In S5-4, multimodal fusion: ; in, , and For attention mechanism parameters; In the temperature gating layer of S5-5, such as Figure 4 As shown, this embodiment constructs a hybrid temperature vector. A gating network is introduced to adaptively weight the feature contributions of different temperature zones. The gating network calculates the temperature gating weights. For the encoded features With residual characteristics Perform weighted fusion: By incorporating smoothing and entropy regularization terms into the loss, this embodiment ensures that the constraint gating weights are not overly concentrated in a single temperature range, thereby improving the cross-temperature generalization capability; wherein. : Weights and biases of the temperature-gated network, g: temperature-gated weight vector, H: principal feature representation of the encoder output. Residual bypass characteristics Characteristics after temperature gating The function "⊙" represents element-wise multiplication.

[0042] In the improved Arrhenius compensation module of S5-6, a learnable temperature compensation equation with physical constraints is used:

[0043] in The gas constant is As the reference activation energy, For piecewise linear basis functions, The parameters are learnable and satisfy physical constraints: exist Within the specified range, ensure the physical rationality of temperature compensation. Improvements to the physical constraints in the Arrhenius equation include: Monotonicity constraint: Ensure that the reaction rate constant k(T) increases monotonically within the operating temperature range; Continuity constraint: The compensation function maintains C1 continuity at the temperature boundary; Boundary condition constraints: at standard temperature At that point, the compensation function value is 1, that is... ; Physical rationality constraint: activation energy Keep in a reasonable Internal, pre-exponential factor .

[0044] This embodiment defines a temperature compensation function: ,in, As a preset temperature basis function, These are the weights to be learned. Through monotonicity, continuity, and standardization constraints, we ensure... Consistent with actual physical laws of degradation. Applying a temperature compensation factor to the gated feature: Simultaneously construct physical constraint loss Penalties are imposed for violations of the above properties, thereby improving the reliability of cross-temperature migration. Where: k(T): temperature compensation factor, : No. A temperature basis function, : with the The learnable weights corresponding to the basis functions, M: the number of temperature basis functions. Standard temperature, at which temperature there is =1 : Upper and lower limits of the operating temperature range; and satisfy: (1) In Monotonous inside: ; (2) First-order continuity at each segment boundary; (3) At standard temperature Satisfying ; In this embodiment, the compensated features are: .

[0045] like Figure 2 As shown, this embodiment presents the specific structure of the HybridNet model. The model includes a temporal encoder, an EIS encoder, a temperature gating layer, an improved Arrhenius temperature compensation layer, a domain alignment module, and a RUL prediction and uncertainty estimation head. This embodiment fuses the compensated features H^ with the EIS features. Input encoder: RUL Predictor Head Output Mean Uncertainty estimation head output .in: A coding network used to process temporal features and EIS features. The fused feature representation output by the encoder. RUL prediction head network, parameters are: , Uncertainty estimation head network, with parameters as follows: , Predict the mean. Predict the standard deviation or its parameterized form. :final Prediction results.

[0046] In this embodiment, RUL is modeled as a Gaussian distribution. The output interface can simultaneously provide point estimates and confidence intervals for engineering decision-making. Specifically, it adopts a Transformer encoder-decoder architecture combined with a pre-training-fine-tuning strategy. In the pre-training stage, it uses multi-temperature domain data for joint training, and in the fine-tuning stage, it performs few-sample adaptation for the target temperature domain. The Transformer encoder consists of 4 layers, each containing a multi-head self-attention network (8 heads) and a feedforward network (2048 hidden dimensions).

[0047] S6. Collect the charge-discharge cycle data and EIS features of the lithium battery to be predicted and input them into the trained HybridNet model to output the lifetime prediction value.

[0048] Experimental verification and application expansion On publicly available multi-temperature battery datasets (such as the NASA dataset), a source-target domain transfer task was constructed with various temperature region combinations. The results were compared with baseline models such as LSTM, CNN-LSTM, Support Vector Regression, Random Forest, and Gradient Boosting Tree, as well as ablation configurations including removal of DTW, gating, EIS, and domain alignment. Experimental results are as follows: Figure 8 and Figure 9 As shown, the method of the present invention achieves significant error reduction in various cross-temperature migration scenarios, verifying the effectiveness and versatility of the method.

[0049] This invention's method can be applied to multi-temperature operating scenarios such as new energy vehicles and grid energy storage. Through online fine-tuning and small-step updates, model parameters can be dynamically adjusted according to changes in ambient temperature, improving the intelligence and robustness of the battery health management system during long-term operation.

[0050] To facilitate engineering replication, this embodiment further explains the data acquisition scheme, data processing flow, and quality control.

[0051] 1. Data Acquisition Implementation Plan The sampling rate, recording format, metadata structure, and naming conventions of the battery data all conform to the existing conventions of public datasets or engineering systems. This invention only performs unified preprocessing, windowing, and characterization steps after data acquisition, without adding any additional hardware or experimental equipment. Input data includes charge-discharge time-series data under multiple temperature domains and optional EIS data; the output is the RUL prediction result and its uncertainty index.

[0052] 2. Overall Data Processing Flow In the project implementation, the overall process corresponds to: Raw data → Quality inspection → Standardization and noise reduction → Sliding window construction → Preprocessing yields → Cross-temperature DTW alignment obtained →Temperature adaptive gating is obtained → Improved Arrhenius temperature compensation yielded → Transfer learning and domain alignment (based on) → HybridNet inference output .

[0053] 3. Pretreatment and Quality Standards This embodiment employs a robust denoising and normalization strategy, with the sliding window length and step size adaptively optimized based on the signal-to-noise ratio and degradation rate. Deformation bands and step constraints are applied during the alignment stage, and the distance metric is automatically selected from Euclidean, Markovian, and cosine based on the validation set performance. EIS segments are dynamically divided based on information gain or spectral peaks, and quality checks are primarily based on anomaly rate and reproducibility.

[0054] Based on the same inventive concept, the embodiment also provides a lithium battery remaining life prediction device based on cross-temperature domain transfer learning, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned lithium battery remaining life prediction method based on cross-temperature domain transfer learning.

[0055] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for lithium battery remaining life prediction by cross-temperature domain transfer learning, characterized in that, The method comprises the following steps: S1, collecting lithium battery charge-discharge cycle data under multiple temperature domains and constructing a time sequence feature matrix, and synchronously collecting EIS features; S2, obtaining degradation curves at different temperatures based on the charge-discharge cycle data collected in S1, and performing nonlinear time alignment based on the time sequence feature matrix to obtain aligned degradation curves; S3, constructing a multi-level temperature perception vector according to temperature and EIS features; S4, performing frequency band feature extraction and multi-modal fusion on the EIS features to obtain EIS fusion features; S5, constructing a HybridNet model and pre-training, including: S5-1, taking the aligned degradation curves in S2 and the EIS fusion features in S4 as inputs, and performing feature extraction and fusion to obtain fusion features; S5-2, generating temperature gating weights based on the multi-level temperature perception vector constructed in S3, and performing temperature self-adaptive weighting on the fusion features to obtain temperature-gated fusion features; S5-4, outputting a lithium battery remaining life prediction value based on the temperature-gated fusion features; S6, collecting lithium battery charge-discharge cycle data and EIS features to be predicted and inputting them into the trained HybridNet model to output a life prediction value.

2. The method of claim 1, wherein, In S1, the lithium battery charge-discharge cycle data under multiple temperature domains includes voltage V(t), current I(t), capacity C(t), and temperature T.

3. The method of claim 1, wherein, In S2, based on the time sequence feature matrix constructed in S1, a dynamic time warping algorithm with adaptive window width is used to perform nonlinear time alignment on the battery degradation curves in different temperature domains to obtain aligned degradation curves.

4. The method of claim 3, wherein, In S2, in the nonlinear time alignment, cross-temperature domain feature alignment and knowledge transfer are realized through a joint loss function.

5. The method of claim 1, wherein, In S3, the method for constructing a multi-level temperature perception vector according to temperature and EIS features is: S3-1, constructing a temperature-EIS feature mapping matrix; S3-2, extracting temperature-sensitive principal components in the matrix constructed in S3-1 through principal component analysis to obtain a temperature-sensitive subspace; S3-3, projecting the EIS features into the temperature-sensitive subspace to obtain EIS-based temperature-sensitive subspace mapping; S3-4, constructing a hybrid temperature vector as a multi-level temperature perception vector according to temperature and EIS-based temperature-sensitive subspace mapping.

6. The method of claim 1, wherein, In S4, the method for performing frequency band feature extraction and multi-modal fusion on the EIS features includes: S4-1, using a dynamic segmentation method based on information gain to divide the spectrum into three sub-bands: low frequency band 0.01-1Hz, medium frequency band 1-100Hz, and high frequency band 100Hz-10kHz; S4-2, using an equivalent circuit-based identification method to extract equivalent circuit parameters, including ohmic resistance, charge transfer resistance, double-layer capacitance, and diffusion impedance; S4-3, using a cross-attention mechanism to perform multi-modal fusion on the segmented spectrum in S4-1 and the equivalent circuit parameters in S4-2 to obtain EIS fusion features.

7. The method of claim 1, wherein, In S5, the HybridNet model includes: a multi-modal input layer for receiving the aligned degradation curve of S2 and the EIS fusion feature of S4 as input, and taking the actual life prediction value as a supervised label; a time-series encoder for time-series modeling and degradation feature extraction on the aligned degradation curve, obtaining a time-series feature representation representing the battery degradation trend; an impedance encoder for frequency-domain mechanism feature encoding on the EIS fusion feature, obtaining an impedance feature representation representing ohmic impedance, charge transfer impedance and diffusion process; a feature fusion layer for cross-modal deep fusion of the time-series feature representation and the impedance feature representation through cross-attention mechanism, obtaining a unified fusion feature representation; a temperature gating layer for generating temperature gating weights based on the multi-level temperature perception vector constructed by S3, and performing temperature adaptive weighting on the fusion feature representation to realize dynamic modulation and unification of cross-temperature domain features; a decoder for RUL decoding regression on the fusion feature after temperature gating, and can simultaneously perform uncertainty estimation; an output layer for outputting the lithium battery remaining useful life prediction value and its uncertainty parameters.

8. The method of claim 7, wherein, The HybridNet model includes an improved Arrhenius compensation module, which realizes temperature compensation for the temperature gating layer based on a learnable Arrhenius type function, and the physical constraints in the learnable Arrhenius type function include: monotonicity constraint: to ensure that the reaction rate constant is monotonically increasing in the working temperature range; continuity constraint: the compensation function maintains first-order continuity at the temperature boundary; boundary condition constraint: the compensation function value is 1 at the standard temperature; Physical plausibility constraints: activation energies remain reasonable within, the pre-factor . 9.A device for lithium battery remaining life prediction across temperature domains transfer learning, characterized in that, The device includes a memory and one or more processors, the memory has a executable program stored therein, and the one or more processors can implement the cross-temperature domain transfer learning lithium battery remaining useful life prediction method of any one of claims 1-8 when executing the program.

10. A computer readable storage medium having a program stored thereon, when the program is executed by a processor, for implementing the cross-temperature domain transfer learning lithium battery remaining useful life prediction method of any one of claims 1-8.

Images