Lithium-sodium cross-system migration battery performance and life prediction method and system

Abstract

This invention provides a method and system for predicting the performance and lifespan of lithium-sodium cross-system batteries, comprising: constructing an aging feature encoder, a sequence decoder, and a decay kinetics network for a prediction model; pre-training the aging feature encoder, sequence decoder, and decay kinetics network; freezing the backbone network parameters of the aging feature encoder and performing domain-adaptive training on the aging feature encoder; extracting sodium-ion battery data through the domain-adaptive trained aging feature encoder and inserting corresponding adapters into the sequence decoder and decay kinetics network to obtain a prediction model adapted to the sodium-ion battery system; inputting the sodium-ion battery data to be tested into the prediction model adapted to the sodium-ion battery system, and outputting the corresponding capacity decay trajectory or remaining cycle life prediction results. This invention realizes the cross-system mechanistic knowledge transfer from data-rich lithium-ion battery systems to data-scarce sodium-ion battery systems.

Description

Technical Field

[0001] This invention belongs to the field of battery health management and lifespan prediction technology, specifically relating to a method and system for predicting battery performance and lifespan through lithium-sodium cross-system transfer. More specifically, it is a method for predicting battery performance and lifespan based on lithium-sodium cross-system mechanism transfer, that is, it involves a method for transferring mature aging mechanism knowledge and models of lithium-ion batteries to sodium-ion batteries, achieving accurate prediction of sodium-ion battery performance and lifespan under small sample conditions. Background Technology

[0003] Traditional battery performance and life prediction methods based on large-scale test data face challenges such as long development cycles, high costs, and poor applicability when applied to sodium-ion batteries, which severely restricts their R&D efficiency and industrialization process.

[0004] Currently, research on battery performance and lifespan prediction mainly falls into three technical paths, each with certain limitations: First, there are material-specific mechanism modeling methods. These methods construct physical models based on fundamental electrochemical principles to describe processes such as ion diffusion, charge transfer, and interfacial reactions within the battery. While they can reveal local mechanisms, the model parameters need to be recalibrated for different material systems, resulting in poor universality and difficulty in supporting the rapid screening of massive amounts of new formulations. Second, there are pure data-driven prediction methods. These methods use machine learning / deep learning to directly learn the mapping relationship between capacity decay and lifespan from historical data. While more flexible, they heavily rely on large-scale homogeneous data, lacking generalization ability and reliability in the real-world data environment of sodium batteries, which is characterized by "small samples, multiple systems, and non-stationarity," and also lacking physical interpretability. Third, there are transfer learning methods. In recent years, this direction has attempted to transfer knowledge from one battery system to another. However, existing work mostly focuses on transfers within the same system under different operating conditions, and even the few cross-system transfers only remain at the level of shallow feature transfer, failing to deeply explore the commonalities and differences between lithium and sodium systems at the electrochemical mechanism level, resulting in limited transfer effects.

[0005] Patent document CN202411618901.7 discloses a lithium-ion battery lifetime prediction method based on data transfer. This method, which is based on the Newman model, relies on a high-precision electrochemical model and requires that the source battery and the target battery belong to the same material system. It cannot achieve cross-chemical system knowledge transfer, such as from lithium ions to sodium ions. At the same time, the model involves a large number of physical parameters, such as diffusion coefficients and reaction kinetic parameters, which are difficult to calibrate and complex to calculate. Furthermore, experimental calibration of the target battery is still required, resulting in a high implementation threshold and low transfer efficiency.

[0006] Patent document CN202411184567.9 discloses a lithium-ion battery health state assessment modeling method based on transfer learning. This method, based on a stacked autoencoder, divides the health state into finite discrete levels, which is a classification task rather than regression prediction. Therefore, it cannot provide accurate SOH values, which is detrimental to refined battery management. Furthermore, the features used rely on manual design, failing to fully utilize deep learning for end-to-end feature extraction. The transfer scenarios are limited to high-stress to low-stress conditions within the same battery system, neglecting to consider mechanistic transferability.

[0007] Patent document CN201710341928.X discloses a deep learning-based method for predicting the lifespan migration of lithium-ion batteries. This method selects a reference battery based on the average Euclidean distance and matches them solely based on the similarity of their capacity curves, without incorporating aging mechanisms and multi-dimensional operating condition information, which can easily lead to bias in reference battery selection. Furthermore, it requires the target battery and the reference battery to be under the same temperature and rate conditions, making cross-condition migration impossible. In addition, the prediction relies on early degradation data of the target battery, such as when the capacity decays to 90%, resulting in a late prediction start time, and it only uses a single feature of capacity, without integrating multi-source information such as voltage and temperature.

[0008] Patent document CN202111157103.5 discloses a lithium-ion battery lifetime prediction method based on sample migration. This scheme relies on manual feature extraction and inflection point prediction for sample selection. The method depends on manually extracting 17 aging features, which involves complex feature engineering, poor generalization, and is limited by voltage range, making it difficult to apply to batteries with different chemical systems. The inflection point prediction method it uses suffers from error propagation problems, affecting the accuracy of lifetime prediction. The sample selection strategy is based only on curve similarity or inflection point distance, failing to achieve alignment at the distribution level, thus limiting its performance in small sample scenarios.

[0009] Therefore, there is an urgent need to develop a new method that can quickly and accurately predict the performance and lifespan of sodium-ion batteries using limited experimental data.

[0010] This problem urgently needs to be solved. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for predicting the performance and lifespan of lithium-sodium cross-system migration batteries.

[0012] A method for predicting the performance and lifetime of a lithium-sodium cross-system migration battery according to the present invention includes: Model construction steps: Construct the aging feature encoder, sequence decoder, and decay dynamics network of the prediction model; Pre-training steps: The aging feature encoder, sequence decoder, and decay dynamics network are pre-trained using lithium-ion battery data; Migration adaptation steps: Freeze the backbone network parameters of the aging feature encoder and perform domain adaptation training on the aging feature encoder; extract sodium-ion battery data through the domain-adapted aging feature encoder and insert corresponding adapters into the sequence decoder and decay kinetics network to obtain a prediction model adapted to the sodium-ion battery system. Prediction steps: Input the data of the sodium-ion battery to be tested into the prediction model of the adapted sodium-ion battery system, and output the corresponding capacity decay trajectory or remaining cycle life prediction results.

[0013] Preferably, in the pre-training step and the transfer adaptation step, a joint loss function including physical constraints is used to train the prediction model; the expression of the joint loss function is:

[0014] Among them, in the formula middle, Denotes the joint loss function. This represents the data fitting loss. Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

[0015] Preferably, the pre-training step includes: Step A1: Using lithium-ion battery data, train the aging feature encoder separately with the goal of minimizing the capacity estimation error; Step A2: Freeze the parameters of the aging feature encoder after training, and jointly train the sequence decoder and the decay dynamics network based on the lithium-ion battery data and the hidden state sequence extracted by the aging feature encoder through a joint loss function that includes physical constraints. In step A1, the objective of minimizing the capacity estimation error, i.e., the training objective is to minimize the mean squared error of the capacity estimation, is expressed as:

[0016] Among them, in the formula middle, Indicate the training objective; Indicates the number of samples; Indicates the first Prediction capacity of each cycle; Indicates the first The actual capacity of each loop.

[0017] Preferably, in the migration adaptation step, a lightweight adapter module is inserted into the output layer of the aging feature encoder. Furthermore, the adapter module is trained using the maximum mean difference loss as the domain adaptation training objective. The lightweight adapter module The expression is:

[0018] in, Indicates input data; The weight matrix represents the dimensionality reduction, used to reduce the dimensionality of the input data to form a bottleneck structure; Represents the activation function of the rectified linear unit; The weight matrix represents the dimension increase, used to restore the features of the bottleneck layer to the original feature space dimension; (symbol) Indicates product; The maximum mean difference loss is expressed as:

[0019] in, This represents the loss due to the maximum mean difference. The norm of the Reproducing Kernel Hilbert Space (RKHS) measures the distance in the mapped feature space. Represents the data distribution of sodium-ion batteries samples Expected value; This indicates the distribution of data from lithium-ion batteries. samples Expected value; Represents the characteristic mapping function in the reproducing kernel Hilbert space; and These represent the hidden state vectors extracted from sodium-ion battery data and lithium-ion battery data by the aging feature encoder, respectively. In the migration adaptation step, corresponding adapters are inserted for the sequence decoder and the attenuation dynamics network, including: Step B1: Freeze the backbone network parameters of the sequence decoder and the decay dynamics network; Step B2: Insert the lightweight adapter module into the sequence decoder A lightweight adapter module is inserted into the decay dynamics network. ; Step B3: Train the lightweight adapter module using the global loss function With lightweight adapter module ; The lightweight adapter module Its learning rate is higher than that of the lightweight adapter module. The learning rate; The expression for the overall loss function is:

[0020] Among them, in the formula middle, Represents the overall loss function; To account for the prediction error loss based on sodium-ion battery data, Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

[0021] Preferably, the aging feature encoder adopts a hybrid architecture; the hybrid architecture sequentially includes a one-dimensional convolutional layer and a self-attention mechanism layer.

[0022] A lithium-sodium cross-system migration battery performance and lifetime prediction system according to the present invention includes: Model building module: Constructs the aging feature encoder, sequence decoder, and decay dynamics network for the prediction model; Pre-training module: The aging feature encoder, sequence decoder, and decay dynamics network are pre-trained using lithium-ion battery data; Migration adaptation module: Freeze the backbone network parameters of the aging feature encoder and perform domain adaptation training on the aging feature encoder; extract sodium-ion battery data through the domain-adapted aging feature encoder and insert corresponding adapters into the sequence decoder and decay kinetics network to obtain a prediction model adapted to the sodium-ion battery system. Prediction module: Input the sodium-ion battery data to be tested into the prediction model of the adapted sodium-ion battery system, and output the corresponding capacity decay trajectory or remaining cycle life prediction results.

[0023] Preferably, in the pre-training module and the transfer adaptation module, a joint loss function including physical constraints is used to train the prediction model; the expression of the joint loss function is:

[0024] Among them, in the formula middle, Denotes the joint loss function. This represents the data fitting loss. Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

[0025] Preferably, the pre-training module includes: Module A1: Using lithium-ion battery data, train the aging feature encoder separately with the goal of minimizing capacity estimation error; Module A2: Freeze the parameters of the aging feature encoder after training, and jointly train the sequence decoder and the decay dynamics network based on the lithium-ion battery data and the hidden state sequence extracted by the aging feature encoder through a joint loss function that includes physical constraints. In module A1, the training objective is to minimize the capacity estimation error, specifically, to minimize the mean squared error of the capacity estimation. The expression for this objective is:

[0026] Among them, in the formula middle, Indicate the training objective; Indicates the number of samples; Indicates the first Prediction capacity of each cycle; Indicates the first The actual capacity of each loop.

[0027] Preferably, in the migration adaptation module, a lightweight adapter module is inserted into the output layer of the aging feature encoder. Furthermore, the adapter module is trained using the maximum mean difference loss as the domain adaptation training objective. The lightweight adapter module The expression is:

[0028] in, Indicates input data; The weight matrix represents the dimensionality reduction, used to reduce the dimensionality of the input data to form a bottleneck structure; Represents the activation function of the rectified linear unit; The weight matrix represents the dimension increase, used to restore the features of the bottleneck layer to the original feature space dimension; (symbol) Indicates product; The maximum mean difference loss is expressed as:

[0029] in, This represents the loss due to the maximum mean difference. It represents the norm in the reproducing kernel Hilbert space, which measures the distance in the mapped feature space; Represents the data distribution of sodium-ion batteries samples Expected value; This indicates the distribution of data from lithium-ion batteries. samples Expected value; Represents the characteristic mapping function in the reproducing kernel Hilbert space; and These represent the hidden state vectors extracted from sodium-ion battery data and lithium-ion battery data by the aging feature encoder, respectively. In the migration adaptation step, corresponding adapters are inserted for the sequence decoder and the attenuation dynamics network, including: Module B1: Backbone network parameters of the frozen sequence decoder and decay dynamics network; Module B2: Insert a lightweight adapter module for the sequence decoder A lightweight adapter module is inserted into the decay dynamics network. ; Module B3: Training a lightweight adapter module using the global loss function With lightweight adapter module ; The lightweight adapter module Its learning rate is higher than that of the lightweight adapter module. The learning rate; The expression for the overall loss function is:

[0030] Among them, in the formula middle, Represents the overall loss function; To account for the prediction error loss based on sodium-ion battery data, Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

[0031] Preferably, the aging feature encoder adopts a hybrid architecture; the hybrid architecture sequentially includes a one-dimensional convolutional layer and a self-attention mechanism layer.

[0032] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention overcomes the limitation of existing methods that can only transfer within the same chemical system, and realizes cross-system mechanism knowledge transfer from the data-rich lithium-ion battery system to the data-scarce sodium-ion battery system. Specifically, the aging feature encoder provided by this invention adopts a hybrid architecture that combines multi-scale convolution and self-attention mechanism, which can automatically extract deep features that are highly correlated with the battery health status from the original charge and discharge time sequence signal. There is no need for manual feature engineering design, which avoids the limitations and subjectivity of traditional methods that rely on domain expert knowledge for feature selection.

[0033] 2. This invention introduces a degradation kinetics network to implicitly learn the physical laws of battery capacity degradation. Through PDE residual constraints and monotonically decreasing constraints, it ensures that the prediction results conform to the basic physical laws of battery aging, balancing prediction accuracy and physical interpretability.

[0034] 3. This invention introduces physical information neural network constraints, which implicitly learn the overall battery degradation dynamics through neural networks. Without explicitly decoupling each aging mechanism, it achieves capacity degradation trajectory prediction guided by physical constraints, balancing physical interpretability and prediction accuracy.

[0035] 4. This invention adopts a parameter-efficient fine-tuning strategy. By training only the lightweight adapter module on the basis of freezing the backbone network, it effectively avoids the risk of overfitting in small sample scenarios, while retaining the general aging feature representation capability obtained by pre-training in the source domain. Attached Figure Description

[0036] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the overall framework of the cross-system mechanism transfer-driven small sample prediction method provided by the present invention. Detailed Implementation

[0037] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0038] Existing purely data-driven methods heavily rely on large-scale homogeneous data for model training. In the context of scarce data in sodium-ion battery data environments, this can easily lead to overfitting, resulting in insufficient generalization ability and predictive reliability. See also Figure 1This invention forms the overall framework for predicting the lifespan of lithium-ion batteries. By learning general characteristics and aging kinetics in the source domain and utilizing adapter and domain adversarial techniques for cross-domain transfer, this invention achieves high-precision capacity trajectory prediction in scenarios where target domain data is scarce. It overcomes the limitation of existing technologies that can only transfer knowledge within the same chemical system, establishing a cross-system knowledge transfer framework from lithium-ion batteries to sodium-ion batteries. This effectively transfers the massive amounts of aging data and modeling experience accumulated in the lithium-ion battery field to the sodium-ion battery system, significantly reducing the time and experimental costs of developing new battery systems.

[0039] This invention proposes a method for predicting the performance and lifetime of sodium-ion batteries using cross-system mechanism transfer. This method constructs a deep learning architecture with three cascaded modules: an encoder, a decoder, and a dynamic network, and uses a four-stage cross-domain adaptation training strategy as its core to achieve cross-system knowledge transfer from lithium-ion batteries to sodium-ion batteries.

[0040] The technical solution of this invention will be described in detail below from three aspects: model architecture design, physical constraint implementation mechanism, and cross-system transfer training strategy.

[0041] At the model architecture level, the entire system consists of an aging feature encoder module, denoted as Encoder. abbreviation The sequence decoder module, denoted as Decoder abbreviation And the decay dynamics network, denoted as It consists of three modules connected in series.

[0042] Aging feature encoder module, i.e., Encoder As the signal sensing layer of the system, it is used to extract deep feature representations related to the battery health state from the raw time-series measurement data of a single charge-discharge cycle. And achieve the estimated output of the current training capacity. The encoder receives voltage signals acquired during the charge-discharge cycle. Current signal Temperature signal .in, Represents the real number field, footnote This represents the deep hidden state representation related to battery health extracted by the encoder, indicated by the superscript. Represents the dimension of the feature space; Indicates time, indicating data at different points in time during the charging and discharging process. This represents the capacity estimate.

[0043] To eliminate the domain shift problem caused by differences in numerical dimensions among different battery systems, this invention employs a unified normalization strategy for all input data, mapping the original data to... Interval:

[0044] in, This represents the normalization result. This represents the input data for the prediction model, and the normalization parameters. and The statistical properties of the lithium-ion battery training set are determined and remain unchanged in the subsequent sodium-ion battery migration stage. This ensures that the lithium-ion battery and sodium-ion battery data always use a consistent normalization standard, fundamentally avoiding feature space drift caused by differences in numerical ranges.

[0045] The encoder employs a hybrid architecture combining convolutional neural networks and an attention mechanism, specifically optimized for local feature patterns and long-range temporal dependencies in charge-discharge time-series data. First, multi-scale one-dimensional convolutional layers are used. That is, a 1D-CNN scans the local features of the input time-series signal, achieving simultaneous capture of features at different time scales by configuring multiple sets of convolutional kernels of different sizes in parallel.

[0046] in, Indicates local features; This represents a one-dimensional convolution operator; This represents the normalized input tensor, which consists of multi-channel time-series data of voltage, current, and temperature. These represent the kernel size parameters for small, medium, and large scales, respectively. This is the activation function for the rectifier linear unit. This multi-scale design can effectively extract various local feature information that is highly correlated with the battery aging state, such as the slope change trend of the charge-discharge curve, the peak position and shape in the incremental capacity characteristics.

[0047] Secondly, based on the local feature sequences output by the convolutional layers, a self-attention mechanism is introduced to model global associations:

[0048] in, This represents a self-attention computation function based on query, key, and value, used to model the global correlation between features at different time steps in a sequence; The superscript represents the normalized exponential function used to map similarity scores to attention weights in the form of a probability distribution. This represents the matrix transpose operation, querying the matrix. Key matrix Value matrix All are composed of local features Obtained through different linear transformations; This is a scaling factor used to prevent the gradient vanishing problem caused by excessively large dot product values. This mechanism models the long-range dependencies of features at different time points during the charging and discharging process, automatically identifies the key moments most discriminative for judging aging states, and outputs a global feature representation. .

[0049] Finally, a fully connected projection layer is used to compress the global features into a fixed-dimensional hidden state vector:

[0050] in, Represents a hidden state vector of fixed dimensions. and These are the weight matrix and bias vector of the projection layer, respectively; The hyperbolic tangent activation function restricts the output to... Within the interval; the output hidden state vector , The hidden state dimension fully encodes the battery health status information reflected in the current charge-discharge cycle, providing input conditions for subsequent degradation trajectory prediction. To provide effective supervision signals for the encoder during the pre-training phase, this invention connects two sensing layers after the hidden state vector to form a capacity estimation regression head:

[0051] in, For the weights of the fully connected layer, As a bias term, the capacity estimation task ensures that the hidden state vector effectively encodes the core feature information related to capacity, thus establishing a reliable feature foundation for subsequent cross-system migration.

[0052] Sequence decoder module, i.e., Decoder As the prediction layer of the system, it is used to predict future capacity decay trajectories based on the aging feature sequences of historical cycles. The decoder's input is a sequence of hidden states from several early cycles. Each hidden state is generated by the aging feature encoder module Encoder. Extract and characterize the battery aging features corresponding to the cycle. This invention employs a Transformer-Decoder architecture to implement the sequence decoder, and uses a lower triangular mask matrix to achieve causal self-attention, ensuring that the model can only access information from historical moments during prediction, thus enabling the network to correctly learn the temporal causal laws of battery capacity decay. Since the Transformer architecture itself is insensitive to sequence order, this invention uses sine-cosine positional encoding to inject temporal positional information into the input sequence:

[0053] in, This represents positional encoding, used to assign different positions in a sequence. Inject explicit timing location information; It is a position index. It is a dimensional index. This refers to the hidden state dimension. The decoder consists of multiple stacked networks, each layer containing a masked multi-head self-attention sublayer, a feedforward neural network sublayer, residual connections, and a layer normalization module. The decoder's output is mapped to future capacity trajectory predictions through a linear output layer.

[0054] in, and For output layer parameters, the total number of predicted iterations. Able to be much larger than the input This allows for early data-driven, full-lifecycle life prediction. The sequence decoder module is used to generate a predictive feature representation of future capacity evolution; Indicates the encoder at the 1st The hidden state vector corresponding to each charge-discharge cycle characterizes the health and aging features of the battery under that cycle. Indicates the first Predicted capacity value at each charge-discharge cycle.

[0055] Attenuation dynamics network This is used to model the dynamics of battery capacity degradation, providing physical constraints for prediction. The battery capacity degradation process essentially follows a nonlinear differential equation:

[0056] Among them, in the formula middle, Indicates battery capacity, Indicates the number of loops. To encode the hidden state vector that contains the current aging characteristics of the battery, These are the model parameters. This represents the decay kinetic function.

[0057] Since the battery aging mechanism involves multiple complex physicochemical processes, such as SEI film growth, electrolyte consumption, and electrode material structure evolution, and these processes are highly coupled, it is necessary to analytically derive an accurate decay kinetic function. This is difficult to implement in engineering. This invention uses a neural network. The decay kinetics function is implicitly learned, eliminating the need for explicit decoupling modeling of each aging mechanism. This design avoids the difficulty of accurately calibrating a large number of physical parameters in traditional physical modeling methods, while also avoiding model biases that may be introduced by strong simplification assumptions.

[0058] Attenuation dynamics network The network architecture employs a multilayer perceptron, whose input includes a normalized number of iterations. Hidden state vector and current predicted capacity value The network structure consists of an input layer, multiple fully connected hidden layers (each with an activation function), and an output layer; the network output is the predicted theoretical decay rate. Due to the decoder Directly output discrete capacity sequences The rate of change of capacity is calculated by the difference between adjacent cycles:

[0059] This discretization process simplifies the computational implementation and matches the discrete sampling characteristics of battery cycle data. Among other things, Indicates the first The and the first The capacity change between charge-discharge cycles is used to approximate the capacity decay within that cycle interval. This represents a partial derivative operator used to describe capacity. With the number of loops The rate of change of the change.

[0060] Regarding the physical constraint mechanism, this invention defines the PDE residual term. This is used to measure the deviation between the actual attenuation rate output by the decoder and the theoretical attenuation rate predicted by the dynamics network, where... Indicates the predicted first The theoretical decay rate per cycle. When This indicates that the predicted decay trajectory satisfies the constraints of the kinetic equation, that is, it follows the physical laws of battery aging.

[0061] Based on this, the present invention provides a joint loss function to simultaneously optimize data fitting accuracy and physical constraint satisfaction. This loss function consists of three parts: data fitting loss, PDE dynamics loss, and capacity monotonically decreasing constraint loss.

[0062] in, Denotes the joint loss function. This represents the data fitting loss. Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

[0063] Specifically, the data fitting loss uses the mean squared error to measure the deviation between the predicted capacity value and the actual measured value, expressed as:

[0064] in, This represents the data fitting loss. For the sample size, For the first The model predicts capacity using a cyclical approach. This corresponds to the true capacity. The PDE dynamics loss forces the decoder to minimize the sum of squares of the PDE residuals. Actual attenuation rate of output With decay dynamics network Predicted theoretical decay rate Maintain consistency:

[0065] This constraint ensures that the degradation pattern learned by the model follows implicit physical dynamics equations, rather than simply fitting a pre-defined degradation formula. The monotonically decreasing constraint loss is designed based on the fundamental physical law that battery capacity decreases monotonically under normal aging conditions.

[0066] in, This represents the loss due to the monotonically decreasing capacity constraint. The function is only for The model penalizes non-physical capacity increases but not normal capacity decreases, effectively preventing aberrant predictions that violate physical laws. Weighting coefficients and To balance the relative importance of data fitting and physical constraints, adjustments are made during training based on actual needs.

[0067] This invention provides a four-stage cross-domain adaptation training strategy, enabling the model to effectively transfer capacity estimation and degradation prediction knowledge learned from large-scale lithium-ion battery data to sodium-ion battery systems. The first two stages perform pre-training on lithium-ion battery data, i.e., the source domain, while the latter two stages perform adaptation fine-tuning on sodium-ion battery data, i.e., the target domain. This design fully utilizes the abundant data resources of lithium-ion batteries, overcoming the application bottleneck of scarce sodium-ion battery data.

[0068] Phase 1: Pre-training of aging feature encoder based on lithium-ion battery data.

[0069] The training objective of this phase is to train a feature encoder using large-scale lithium-ion battery cycle aging data, enabling it to extract deep feature representations highly correlated with battery health. Lithium-ion battery technology is mature, and abundant cycle data resources covering different models, specifications, operating conditions, and aging stages have been accumulated. The raw time-series data of each charge-discharge cycle is input into the feature encoder. Output a hidden state vector of fixed dimensions. Then, the capacity estimation regression head outputs the capacity prediction value for the current loop. The training objective is to minimize the mean squared error of the capacity estimation.

[0070] Where, formula middle, This indicates the training objective for the first stage.

[0071] After sufficient training, the encoder can extract stable aging features from the noisy raw charge and discharge signals, and the hidden state vector is highly correlated with the battery health status.

[0072] Phase 2: Joint pre-training of sequence decoder and decay dynamics network based on lithium-ion battery data.

[0073] The training objective of this stage is to train the sequence decoder. and decay dynamics network This aims to enable the battery to predict the general aging trajectory and degradation dynamics of lithium-ion batteries. It utilizes full lifecycle data of lithium-ion batteries, encompassing all cycles from battery activation to capacity degradation reaching the failure threshold. At the start of this phase, the encoder weights pre-trained in the first phase are loaded, and all encoder parameters are frozen and remain unchanged during this training phase to preserve the encoder's learned feature extraction capabilities and prevent feature representation drift. A full lifecycle observation sequence of lithium-ion batteries is then provided. Indicates the first The raw observation data from each cycle is used to perform feature mapping on each cycle to obtain the hidden state sequence. ,in, , This indicates the aging characteristic encoder module; Indicates the relationship with the first The hidden state sequence corresponding to the original observation data of each cycle.

[0074] Serial decoder Training is performed using a forced teaching method, namely Teacher Forcing, which divides the battery life sequence into segments of length [length missing]. The time window is constructed using a sliding method to create training samples.

[0075] For any window The decoder input is The target of the prediction is .

[0076] In the formula Subscript Indicates the starting loop index of the sliding time window, i.e., the... The initial charge / discharge cycle number corresponding to each training sample. Indicates the first The actual capacity corresponding to each cycle.

[0077] At each time step The encoder, under the constraint of causal masking, focuses only on historical hidden state information and learns the conditional probability distribution. In the formula middle, Indicates a sequence decoder; Indicates the deadline step Historical capacity sequence; Indicates the deadline step The historical hidden state sequence. The aging hidden state sequence. As conditional information, it is injected into the decoding process through a cross-attention mechanism to guide the capacity decay trajectory prediction to remain consistent with the battery aging state. This stage employs a composite loss function that includes data fitting, dynamic constraints, and monotonicity constraints.

[0078] Where, formula middle, This indicates the training objective for the second phase.

[0079] After sufficient training, the sequence decoder can accurately predict the capacity decay trajectory of lithium-ion batteries, the decay kinetics network learns the general kinetic laws of battery aging, the PDE residual can be reduced to a low level, and the predicted trajectory meets physical constraints.

[0080] Phase 3: Encoder Domain Adaptation Migration for Sodium-Ion Batteries The training objective in this phase is to enable the pre-trained encoder to effectively adapt to the data distribution characteristics of sodium-ion batteries using only a limited amount of sodium-ion battery data, while preserving to the greatest extent possible its ability to extract general aging features learned from lithium-ion battery data. Currently, the available labeled training data for sodium-ion batteries is limited in scale. Full parameter fine-tuning of the encoder could easily lead to overfitting in small-sample scenarios and may disrupt the stable feature representations formed by the encoder on lithium-ion battery data, triggering catastrophic forgetting.

[0081] This invention employs Parameter Efficient Fine Tuning (PEFT) for domain adaptation. Specifically, the scheme involves: freezing all parameters of the encoder backbone network to maintain their stability during training; and inserting a lightweight adapter module after the encoder output layer. , recorded as Only the parameters of this adapter are trained. The adapter uses a bottleneck structure, expressed as:

[0082] in, Used to transform the hidden state vector from dimension Compress to the bottleneck dimension Specifically, , Then it is mapped back to the original dimension. Since the adapter parameter size is much smaller than the encoder body, this structure can effectively learn data distribution corrections for sodium-ion batteries while significantly reducing the risk of overfitting. To promote cross-system knowledge transfer and enhance feature space consistency, this invention introduces the Maximum Mean Discrepancy (MMD) loss as a domain adaptation regularization term, with the expression:

[0083] in, Describes the feature mapping function in the reproducing kernel Hilbert space. and These are the hidden state vectors extracted from sodium-ion battery data and lithium-ion battery data by the encoder, respectively. By minimizing this loss term, the distribution difference between sodium-ion and lithium-ion batteries in the hidden state space can be reduced, making the encoder output tend towards a domain-invariant representation. This ensures that subsequent modules pre-trained on lithium-ion batteries still have good generalization ability in sodium-ion battery scenarios. The overall training objective function for this stage is:

[0084] in, This represents the overall training objective function for the third stage. To account for the prediction error loss based on sodium-ion battery data, The weighting coefficients are used to determine the fitness strength of the control domain.

[0085] Phase 4: Joint migration fine-tuning of sequence decoder and decay kinetics network for sodium-ion batteries.

[0086] The training objective of this stage is to enable the sequence decoder to... With decay dynamics network This method can accurately predict the capacity decay trajectory of sodium-ion batteries using limited full-lifecycle data and adapt to the unique decay kinetics of sodium-ion batteries. Due to differences in material composition, interfacial reactions, and structural evolution between lithium-ion and sodium-ion batteries, the two battery systems exhibit different characteristics in terms of capacity decay rate, decay curve morphology, and accelerated decay stages. Encoder domain adaptation alone cannot fully cover these differences; further adaptation of the sequence modeling module and the decay kinetics modeling module is required.

[0087] This stage also employs adapter technology, specifically: frozen sequence decoder. With decay dynamics network The backbone network parameters are respectively introduced into lightweight adapter modules. and Only the adapter parameters are trained. The temporal morphological characteristics used to adapt to the capacity decay trajectory of sodium-ion batteries enable the sequence decoder to learn the decay mode unique to sodium-ion batteries. This is used to adapt the degradation kinetics of sodium-ion batteries, enabling the kinetic network to characterize the capacity change rate of sodium-ion batteries at different aging stages.

[0088] Specifically, the and In terms of structural form, it is similar to the aforementioned To maintain consistency, both adopt a lightweight residual adapter based on a bottleneck structure, the specific mathematical form of which is the same as... They are the same, only differing in their embedding location and function module.

[0089] Different learning rate settings are used for the two adapters to achieve fine-tuning. Sequence decoder adapter A relatively high learning rate is used to accelerate the model's adaptation to the capacity decay trajectory distribution of sodium-ion batteries; decay kinetics network adapter. A lower learning rate is used to adapt to the dynamic characteristics of sodium-ion batteries while avoiding violation of the general decay dynamics constraints obtained in the pre-training phase. The overall loss function form in this phase remains consistent with that in the second phase.

[0090] in, This represents the overall loss function for the fourth stage.

[0091] The loss terms are used to constrain capacity prediction accuracy, decay kinetic consistency, and capacity evolution monotonicity, respectively. After this training phase, the model can stably output capacity decay prediction results that conform to physical constraints under limited sodium-ion battery data, completing the cross-system knowledge transfer from lithium-ion batteries to sodium-ion batteries.

[0092] According to the present invention, a method for predicting the performance and lifespan of sodium-ion batteries with small sample data driven by cross-system mechanism transfer is provided. By constructing a deep learning architecture with three cascaded modules of "encoder-decoder-dynamic network" and combining it with a four-stage cross-domain adaptation training strategy, the method realizes cross-system knowledge transfer from lithium-ion batteries to sodium-ion batteries and completes accurate prediction of capacity decay trajectory and cycle life under the condition of a small amount of sodium-ion battery data.

[0093] Specifically, the aging feature encoder adopts a hybrid architecture that combines multi-scale one-dimensional convolution with a self-attention mechanism. The multi-scale one-dimensional convolutional layer achieves synchronous capture of features at different time scales by configuring convolutional kernels of different sizes, and the self-attention mechanism achieves long-distance dependency modeling of features at different time points during the charging and discharging process, and finally outputs a hidden state vector containing battery health status information.

[0094] Specifically, the implicit physical modeling method applicable to systems with incomplete understanding of mechanisms learns the dynamics of battery capacity decay through neural networks. It eliminates the need for explicit modeling and parameter calibration of specific aging mechanisms such as SEI film growth, electrolyte consumption, and electrode material structure evolution. It is particularly suitable for emerging systems such as sodium-ion batteries, where mechanism research is not yet in-depth and failure modes are complex.

[0095] Specifically, the physical constraint mechanism is implemented through a joint loss function, which consists of three parts: data fitting loss, PDE dynamics loss, and capacity monotonically decreasing constraint loss. The PDE dynamics loss forces the predicted trajectory to follow physical dynamics by minimizing the residual between the actual decay rate output by the decoder and the theoretical decay rate predicted by the dynamics network.

[0096] The four-stage cross-domain adaptation training strategy includes: First, an aging feature encoder based on lithium-ion battery data is pre-trained, with the training objective being capacity estimation. Secondly, a sequence decoder based on lithium-ion battery data and a decay dynamics network are jointly pre-trained, and a physically constrained composite loss function is adopted. Third, for encoder domain adaptation migration for sodium-ion batteries, a parameter-efficient fine-tuning strategy and maximum mean difference loss are adopted. Fourth, a sequence decoder and decay kinetics network are jointly used for migration fine-tuning in sodium-ion batteries.

[0097] The efficient parameter fine-tuning strategy is implemented by freezing the backbone network parameters and inserting a lightweight bottleneck structure adapter module. The adapter adopts a bottleneck residual structure and only the adapter parameters are trained.

[0098] In the first embodiment, the data preparation involves collecting complete lifecycle cycle data from multiple sets of lithium-ion batteries. This data must cover at least different battery models and specifications, different cathode material systems, different operating conditions, and different ambient temperature conditions. During each charge-discharge cycle, voltage, current, temperature signals, and corresponding cycle capacity data are simultaneously collected and recorded. The lithium-ion battery dataset should contain complete lifecycle data from at least 100 batteries to ensure sufficient statistical representativeness of the model in the source domain.

[0099] Furthermore, a small amount of cycling data from sodium-ion batteries was collected, with the data format, sampling frequency, and signal type consistent with those of lithium-ion batteries. The sodium-ion battery data consisted of complete lifecycle data from 5 to 20 batteries, which was used for subsequent small-sample transfer training and validation.

[0100] The raw data described above undergoes unified data preprocessing, which includes outlier removal, time axis alignment, and normalization. The normalization parameters are determined based on the statistical characteristics of the lithium-ion battery training set and remain constant during the sodium-ion battery transfer learning phase to avoid introducing additional distribution shifts.

[0101] Secondly, model construction. The model and system consist of three modules: an aging feature encoder, a sequence decoder, and a decay dynamics network.

[0102] The aging feature encoder employs a hybrid structure combining multi-scale one-dimensional convolutions and a self-attention mechanism to extract latent features related to aging states from the raw battery operating signals. The multi-scale one-dimensional convolutional layers use kernel sizes of 3, 7, and 15, with 32 channels at each scale, to capture battery degradation features at different time scales. A self-attention mechanism is then introduced to enhance the global correlation modeling ability between features; this mechanism is configured as a 4-head attention structure, with each head having a feature dimension of 24. The encoder ultimately outputs a 128-dimensional latent state vector. A capacity estimation regression head is connected to the encoder output. This regression head consists of two fully connected layers with 64 and 1 neurons, respectively, used to perform regression estimation of single-cycle capacity.

[0103] The sequence decoder employs a 4-layer Transformer-Decoder architecture, with each layer containing an 8-head attention mechanism and a corresponding feedforward network. Each attention head has a feature dimension of 16, and the feedforward network's hidden layer has a dimension of 512. The decoder's input is the hidden state sequence corresponding to 100 consecutive cycles, and its output is the capacity prediction sequence for the next 100 cycles, enabling serialized prediction of battery capacity decay trajectories.

[0104] The decay dynamics network adopts a fully connected neural network structure with 4 hidden layers. The number of neurons in the hidden layers are 128, 64, 32 and 16 respectively. The final output of the network is the theoretical decay rate of battery capacity decay, which is used to introduce aging dynamics constraints.

[0105] Third, lithium-ion battery source domain pre-training.

[0106] First, the aging feature encoder and its capacity estimation regression head were pre-trained using cyclic data from the entire lifecycle of lithium-ion battery samples. During training, the batch size was set to 128, the Adam optimizer was used, and the initial learning rate was [missing information]. It employs a cosine annealing learning rate scheduling strategy, with the minimum learning rate set to... The total number of training rounds is 500.

[0107] After encoder pre-training, the trained encoder is used to extract features from the full lifecycle data of lithium-ion battery samples, obtaining the corresponding hidden state sequences, and the encoder parameters are frozen. Subsequently, based on these hidden state sequences, the sequence decoder and decay dynamics network are jointly pre-trained. Training samples are constructed using a sliding window method, with a window length of 100 cycles and a sliding step size of 6 cycles. During joint training, the batch size is set to 64, and the learning rate is... The training rounds consist of 300 rounds.

[0108] Fourth, target domain migration adaptation for sodium-ion batteries.

[0109] Domain adaptation transfer training of a sodium-ion battery encoder was performed using a small amount of sodium-ion battery sample data. An adapter module with a bottleneck dimension of 16 was inserted after the output layer of the pre-trained encoder. The number of parameters in the adapter module was much smaller than the number of parameters in the encoder backbone, and the encoder backbone network parameters were frozen. The batch size was set to 32, and the learning rate was... Training for 100 epochs. The MMD loss uses a Gaussian RBF kernel with a bandwidth of 1.0 and weight coefficients... It is 0.2.

[0110] The latent state sequence of sodium-ion battery data was extracted using an encoder that had completed domain adaptation training, and the encoder and its adapter parameters were frozen. Subsequently, adapter modules were inserted into the sequence decoder and decay kinetics network, respectively, and the corresponding backbone network parameters were frozen; training was performed only on the adapter parameters. The batch size for the transfer fine-tuning phase was set to 16, and differentiated learning rates were set for the sequence decoder adapter and the decay kinetics network adapter, respectively. and , train 100 rounds.

[0111] Specifically, in terms of data preprocessing, Min-Max normalization can be replaced with Z-score standardization, and feature scaling can be performed using the mean and standard deviation.

[0112] Specifically, in the aging feature encoder architecture, the hybrid architecture of multi-scale one-dimensional convolution and self-attention can be replaced with a pure Transformer encoder architecture, or the multi-scale convolutional layers can be replaced with a temporal convolutional network.

[0113] Specifically, the parameter-efficient fine-tuning adapter strategy can be replaced with a low-rank adaptive method, which achieves efficient parameter updates by superimposing a low-rank decomposition matrix on the original weight matrix, and may obtain more flexible adaptation capabilities under the same parameter budget.

[0114] The present invention also provides a lithium-sodium cross-system migration battery performance and life prediction system. The lithium-sodium cross-system migration battery performance and life prediction system can be implemented by executing the process steps of the lithium-sodium cross-system migration battery performance and life prediction method. That is, those skilled in the art can understand the lithium-sodium cross-system migration battery performance and life prediction method as a preferred embodiment of the lithium-sodium cross-system migration battery performance and life prediction system.

[0115] A lithium-sodium cross-system migration battery performance and lifetime prediction system according to the present invention includes: Model building module: Constructs the aging feature encoder, sequence decoder, and decay dynamics network for the prediction model; Pre-training module: The aging feature encoder, sequence decoder, and decay dynamics network are pre-trained using lithium-ion battery data; Migration adaptation module: Freeze the backbone network parameters of the aging feature encoder and perform domain adaptation training on the aging feature encoder; extract sodium-ion battery data through the domain-adapted aging feature encoder and insert corresponding adapters into the sequence decoder and decay kinetics network to obtain a prediction model adapted to the sodium-ion battery system. Prediction module: Input the sodium-ion battery data to be tested into the prediction model of the adapted sodium-ion battery system, and output the corresponding capacity decay trajectory or remaining cycle life prediction results.

[0116] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0117] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for predicting the performance and lifetime of lithium-sodium cross-system migration batteries, characterized in that, include: Model construction steps: Construct the aging feature encoder, sequence decoder, and decay dynamics network of the prediction model; Pre-training steps: The aging feature encoder, sequence decoder, and decay dynamics network are pre-trained using lithium-ion battery data; Migration adaptation steps: Freeze the backbone network parameters of the aging feature encoder and perform domain adaptation training on the aging feature encoder; Sodium-ion battery data is extracted by an aging feature encoder trained by domain adaptation, and corresponding adapters are inserted into the sequence decoder and decay kinetics network to obtain a prediction model adapted to the sodium-ion battery system. Prediction steps: Input the data of the sodium-ion battery to be tested into the prediction model of the adapted sodium-ion battery system, and output the corresponding capacity decay trajectory or remaining cycle life prediction results.

2. The method for predicting the performance and lifetime of lithium-sodium cross-system migration batteries according to claim 1, characterized in that, In the pre-training and transfer adaptation steps, a joint loss function incorporating physical constraints is used to train the prediction model; the expression for the joint loss function is: Among them, in the formula middle, Denotes the joint loss function. This represents the data fitting loss. Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

3. The method for predicting the performance and lifetime of lithium-sodium cross-system migration batteries according to claim 2, characterized in that, The pre-training step includes: Step A1: Using lithium-ion battery data, train the aging feature encoder separately with the goal of minimizing the capacity estimation error; Step A2: Freeze the parameters of the aging feature encoder after training, and jointly train the sequence decoder and the decay dynamics network based on the lithium-ion battery data and the hidden state sequence extracted by the aging feature encoder through a joint loss function that includes physical constraints. In step A1, the objective of minimizing the capacity estimation error, i.e., the training objective is to minimize the mean squared error of the capacity estimation, is expressed as: Among them, in the formula middle, Indicate the training objective; Indicates the number of samples; Indicates the first Prediction capacity of each cycle; Indicates the first The actual capacity of each loop.

4. The method for predicting the performance and lifetime of lithium-sodium cross-system migration batteries according to claim 1, characterized in that, In the migration adaptation step, a lightweight adapter module is inserted into the output layer of the aging feature encoder. Furthermore, the adapter module is trained using the maximum mean difference loss as the domain adaptation training objective. The lightweight adapter module The expression is: in, Indicates input data; Represents the weight matrix for dimensionality reduction; Represents the activation function of the rectified linear unit; Represents the weight matrix for increasing dimensionality; symbol Indicates product; The maximum mean difference loss is expressed as: in, This represents the loss due to the maximum mean difference. Denotes the norm in the reproducing kernel Hilbert space; Represents the data distribution of sodium-ion batteries samples Expected value; This indicates the distribution of data from lithium-ion batteries. samples Expected value; Represents the characteristic mapping function in the reproducing kernel Hilbert space; and These represent the hidden state vectors extracted from sodium-ion battery data and lithium-ion battery data by the aging feature encoder, respectively. In the migration adaptation step, corresponding adapters are inserted for the sequence decoder and the attenuation dynamics network, including: Step B1: Freeze the backbone network parameters of the sequence decoder and the decay dynamics network; Step B2: Insert the lightweight adapter module into the sequence decoder A lightweight adapter module is inserted into the decay dynamics network. ; Step B3: Train the lightweight adapter module using the global loss function With lightweight adapter module ; The lightweight adapter module Its learning rate is higher than that of the lightweight adapter module. The learning rate; The expression for the overall loss function is: Among them, in the formula middle, Represents the overall loss function; To account for the prediction error loss based on sodium-ion battery data, Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

5. The method for predicting the performance and lifetime of lithium-sodium cross-system migration batteries according to claim 1, characterized in that, The aging feature encoder adopts a hybrid architecture; the hybrid architecture includes a one-dimensional convolutional layer and a self-attention mechanism layer in sequence.

6. A lithium-sodium cross-system migration battery performance and lifetime prediction system, characterized in that, include: Model building module: Constructs the aging feature encoder, sequence decoder, and decay dynamics network for the prediction model; Pre-training module: The aging feature encoder, sequence decoder, and decay dynamics network are pre-trained using lithium-ion battery data; Transfer adaptation module: Freezes the backbone network parameters of the aging feature encoder and performs domain adaptation training on the aging feature encoder; Sodium-ion battery data is extracted by an aging feature encoder trained by domain adaptation, and corresponding adapters are inserted into the sequence decoder and decay kinetics network to obtain a prediction model adapted to the sodium-ion battery system. Prediction module: Input the sodium-ion battery data to be tested into the prediction model of the adapted sodium-ion battery system, and output the corresponding capacity decay trajectory or remaining cycle life prediction results.

7. The lithium-sodium cross-system migration battery performance and lifetime prediction system according to claim 6, characterized in that, In the pre-training module and the transfer adaptation module, a joint loss function including physical constraints is used to train the prediction model; the expression of the joint loss function is: Among them, in the formula middle, Denotes the joint loss function. This represents the data fitting loss. Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

8. The lithium-sodium cross-system migration battery performance and lifetime prediction system according to claim 7, characterized in that, The pre-training module includes: Module A1: Using lithium-ion battery data, train the aging feature encoder separately with the goal of minimizing capacity estimation error; Module A2: Freeze the parameters of the aging feature encoder after training, and jointly train the sequence decoder and the decay dynamics network based on the lithium-ion battery data and the hidden state sequence extracted by the aging feature encoder through a joint loss function that includes physical constraints. In module A1, the training objective is to minimize the capacity estimation error, specifically, to minimize the mean squared error of the capacity estimation. The expression for this objective is: Among them, in the formula middle, Indicate the training objective; Indicates the number of samples; Indicates the first Prediction capacity of each cycle; Indicates the first The actual capacity of each loop.

9. The lithium-sodium cross-system migration battery performance and lifetime prediction system according to claim 6, characterized in that, In the migration adaptation module, a lightweight adapter module is inserted into the output layer of the aging feature encoder. Furthermore, the adapter module is trained using the maximum mean difference loss as the domain adaptation training objective. The lightweight adapter module The expression is: in, Indicates input data; Represents the weight matrix for dimensionality reduction; Represents the activation function of the rectified linear unit; Represents the weight matrix for increasing dimensionality; symbol Indicates product; The maximum mean difference loss is expressed as: in, This represents the loss due to the maximum mean difference. Denotes the norm in the reproducing kernel Hilbert space; Represents the data distribution of sodium-ion batteries samples Expected value; This indicates the distribution of data from lithium-ion batteries. samples Expected value; Represents the characteristic mapping function in the reproducing kernel Hilbert space; and These represent the hidden state vectors extracted from sodium-ion battery data and lithium-ion battery data by the aging feature encoder, respectively. In the migration adaptation step, corresponding adapters are inserted for the sequence decoder and the attenuation dynamics network, including: Module B1: Backbone network parameters of the frozen sequence decoder and decay dynamics network; Module B2: Insert a lightweight adapter module for the sequence decoder A lightweight adapter module is inserted into the decay dynamics network. ; Module B3: Training a lightweight adapter module using the global loss function With lightweight adapter module ; The lightweight adapter module Its learning rate is higher than that of the lightweight adapter module. The learning rate; The expression for the overall loss function is: Among them, in the formula middle, Represents the overall loss function; To account for the prediction error loss based on sodium-ion battery data, Indicates PDE kinetic loss, This represents the loss due to the monotonically decreasing capacity constraint. and This represents the weighting coefficient.

10. The lithium-sodium cross-system migration battery performance and lifetime prediction system according to claim 6, characterized in that, The aging feature encoder adopts a hybrid architecture; the hybrid architecture includes a one-dimensional convolutional layer and a self-attention mechanism layer in sequence.

Images