Lithium battery state of health estimation method, system, storage medium and electronic device

Abstract

The application provides a lithium battery health state estimation method, system, storage medium and electronic equipment, and relates to the technical field of lithium battery state estimation. The application realizes high-dimensional feature nonlinear dimension reduction through a dynamic activation mechanism based on B-splines of a KAN network, extracts deep features in combination with a multi-scale one-dimensional CNN, models a time sequence dependency relationship through an LSTM network, introduces an attention mechanism on the output side of the LSTM to adaptively weight key recession period information, and finally outputs an SOH prediction value by using a double-layer KAN to fit the nonlinear attenuation law of the battery. The method effectively improves the accuracy of lithium battery health state prediction, enhances the adaptation capability to the nonlinear change of the whole life cycle of the battery, can more accurately capture the capacity attenuation law, optimizes the operation efficiency of the model, reduces parameter redundancy, improves prediction stability and reliability, adapts to the deployment requirements of BMS edge devices, and finally realizes accurate estimation of the SOH of the lithium battery.

Description

Technical Field

[0001] This invention relates to the field of lithium battery state estimation technology, specifically to a lithium battery health state estimation method, system, storage medium, and electronic device. Background Technology

[0002] As a core energy storage component in the new energy field, the accurate estimation of the State of Health (SOH) of lithium batteries is crucial for ensuring safe and stable operation and extending their service life, and is widely used in new energy vehicles, energy storage systems, and other scenarios. Current lithium battery SOH prediction methods largely rely on traditional modeling approaches, which are difficult to adapt to the complex operating conditions throughout the battery's entire lifecycle. Therefore, there is an urgent need for an efficient and accurate prediction solution to address the technical challenges of nonlinear fitting and multi-feature fusion.

[0003] Among related technologies, the existing lithium battery SOH prediction includes invention application CN121432214A, which achieves SOH estimation through data acquisition and preprocessing, multi-level feature extraction, and fusion prediction model. Specifically, it first acquires multi-cycle test data of lithium batteries, performs autoencoder preprocessing on the data to obtain low-dimensional feature representations, and then constructs a fusion prediction model through a one-dimensional convolutional neural network (CNN), a long short-term memory network (LSTM), and a Transformer encoder to extract short-term and long-term time-dependent features respectively, and finally outputs the SOH estimate value. It also includes hyperparameter optimization and anomaly attribution diagnosis functions.

[0004] However, the above solutions have significant shortcomings and fail to address the core pain points in practical applications: First, using a single-scale convolutional kernel for feature extraction results in a fixed receptive field, which cannot fully capture the dynamic response features at different time scales during the charging and discharging process of lithium batteries, leading to insufficient feature extraction. Second, using a traditional Transformer encoder for time series modeling results in computational complexity that is quadratic with the sequence length, a large number of parameters, and high memory consumption, making it difficult to deploy on BMS edge devices with limited computing power. Third, using a traditional fully connected layer (MLP) for SOH (State of Health) regression prediction results in a fixed activation function that is difficult to fit the nonlinear accelerated degradation trajectory at the end of the lithium battery's life cycle, leading to insufficient SOH prediction accuracy. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method, system, storage medium, and electronic device for estimating the state of health (SOH) of lithium batteries, solving the technical problems that traditional models struggle to accurately fit the nonlinear variation law of SOH in lithium batteries and cannot effectively handle the fusion of multiple types of features and time-series dependencies.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A method for predicting the state of health (SOH) of a lithium battery includes: Collect multi-cycle test data of lithium batteries and construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation. The discharge timing curve feature sequence and the incremental capacity IC curve feature are concatenated along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN). The deep curve features of multi-scale local polarization and capacity decay are extracted through multi-receptive field parallel convolution branches. After flattening the deep curve features, they are connected to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism, and then reshaped through temporal dimension reshaping to obtain the reshaped features. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edge. The reshaped features and single-valued scalar features are fused and concatenated, and then input into a Long Short-Term Memory (LSTM) network to model temporal dependencies, thereby obtaining the full temporal hidden state sequence and the final hidden state. An attention mechanism is introduced on the output side of the LSTM. The final hidden state is used as the query vector to perform adaptive weighting on the full-time hidden state sequence to obtain a context vector that enhances key decay cycle information. The context vector and the final hidden state are concatenated and fused according to their feature dimensions. A two-layer KAN is used to fit the nonlinear decay law of the battery to obtain the predicted SOH value of the lithium battery.

[0007] Preferably, the CNN includes three parallel convolutional branches with different receptive fields, each branch corresponding to a convolutional kernel of a different size; Among them, the smallest size convolution kernel is used to capture instantaneous voltage polarization abrupt changes; Intermediate-sized convolutional kernels are used to capture the temporal changes of steady states during charging and discharging. The largest size convolution kernel is used to capture a gradual capacity decay trend.

[0008] Preferably, the adaptive weighting process uses Softmax normalization to allocate weights.

[0009] Preferably, the double-layer KAN comprises two linear KAN layers; The first KAN linear layer uses a dynamic activation function based on B-spline curves to perform high-order nonlinear mapping and compression on the spliced ​​and fused features, and outputs low-dimensional hidden regression features. The second KAN linear layer receives low-dimensional hidden regression features and uses the approximation capability of high-order functions to fit the highly nonlinear capacity decay trajectory at the end of the lithium battery's life cycle and output the SOH prediction value.

[0010] Preferably, the single-valued scalar feature sequence includes at least one of the following: number of cycles, temperature statistics, internal resistance, and charging time.

[0011] Preferably, the discharge timing curve is characterized as a voltage or capacity timing sequence with fixed sampling points.

[0012] Preferably, the incremental capacity IC curve characteristics are obtained by differentiating the charge-discharge curve.

[0013] A lithium battery state of health (SOH) prediction system, comprising: The heterogeneous feature construction module is used to collect multi-cycle test data of lithium batteries and construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation. The multi-scale convolutional feature extraction module is used to concatenate the discharge timing curve feature sequence and the incremental capacity IC curve feature along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN) and extracted deep curve features of multi-scale local polarization and capacity decay through multi-receptive field parallel convolutional branches. The KAN nonlinear dimensionality reduction module is used to flatten deep curve features and then connect them to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism, and the reshaped features are obtained after temporal dimension reshaping. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edges. The cross-domain feature fusion and temporal modeling module is used to fuse and concatenate reshaped features with single-valued scalar feature sequences, input the Long Short-Term Memory (LSTM) network to model temporal dependencies, and obtain the full temporal hidden state sequence and the final hidden state. The dynamic attention weighting module is used to introduce an attention mechanism on the output side of the LSTM. Using the final hidden state as the query vector, it adaptively weights the full-time hidden state sequence to obtain a context vector that enhances key decay cycle information. The SOH prediction output module is used to concatenate and fuse the context vector and the final hidden state by feature dimension, and use a two-layer KAN to fit the nonlinear decay law of the battery to obtain the SOH prediction value of the lithium battery.

[0014] A storage medium storing a computer program, wherein the computer program causes a computer to execute the lithium battery state of health (SOH) prediction method as described above.

[0015] An electronic device, the electronic device comprising: Processor and memory; The memory stores program instructions; The processor is configured to run the program instructions to execute the lithium battery state of health (SOH) prediction method as described above.

[0016] (III) Beneficial Effects This invention provides a method, system, storage medium, and electronic device for estimating the state of health of a lithium battery. Compared with the prior art, it has the following advantages: This invention constructs a dual-channel feature input, combines it with a multi-scale one-dimensional CNN to extract deep features, and then uses a KAN network based on a B-spline dynamic activation mechanism to achieve nonlinear dimensionality reduction of high-dimensional features. Next, an LSTM network is used to model temporal dependencies, and an attention mechanism is introduced on the LSTM output to adaptively weight key degradation cycle information. Finally, a two-layer KAN is used to fit the nonlinear degradation law of the battery and output the SOH prediction value. This method effectively improves the accuracy of lithium battery health state prediction, enhances its adaptability to nonlinear changes throughout the battery's life cycle, more accurately captures capacity degradation patterns, optimizes model computational efficiency, reduces parameter redundancy, and improves prediction stability and reliability to meet the deployment requirements of BMS edge devices, ultimately achieving accurate prediction of lithium battery SOH. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a lithium battery state of health (SOH) prediction method provided in an embodiment of the present invention; Figure 2 A SOH prediction curve of a traditional LSTM model provided in an embodiment of the present invention; Figure 3 A SOH prediction curve of an MS-CLA-KAN model provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0019] Component labeling explanation: 100 - electronic device, 101 - memory, 102 - processor, 103 - display. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solutions in this application are intended to solve the aforementioned technical problems. The overall concept and complete inventive objective are as follows: 1) Employing multi-scale CNNs to achieve comprehensive perception of multi-dimensional degradation features. This invention, by setting convolutional kernels of different sizes in parallel, can simultaneously extract short-term local polarization abrupt changes and medium- to long-term gradual capacity decay trends. This multi-scale architecture aligns with the electrochemical reaction mechanisms of lithium batteries under different scenarios, enhances the model's ability to capture initial minor aging precursors, and overcomes the problem of incomplete feature extraction caused by the limited receptive field of a single convolutional kernel.

[0022] 2) Design a task-oriented attention mechanism to replace the highly redundant Transformer encoder. After modeling temporal dependencies using LSTM and outputting the full-time hidden state sequence and the final hidden state, this embodiment of the invention employs a task-oriented attention mechanism to replace the computationally complex Transformer encoder. This attention mechanism uses the final hidden state as the query vector and adaptively weights the full-time hidden state sequence, focusing on the key decay cycle information that has the greatest impact on the current SOH decay. This reduces computational complexity, adapts to the real-time deployment requirements of BMS edge devices, and improves the model's interpretability, addressing the shortcomings of excessive computational complexity and parameter redundancy in the Transformer.

[0023] 3) Introducing KAN to overcome the bottleneck of nonlinear fitting and improve the prediction accuracy throughout the entire life cycle. This invention introduces a Kolmogorov-Arnold Network (KAN) to replace the traditional fully connected network (MLP). KAN uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edges, which has a stronger ability to approximate nonlinear functions. It can perform nonlinear dimensionality reduction of high-dimensional features while retaining nonlinear degradation information, effectively improving the problem of insufficient prediction accuracy of traditional MLP in the stage of accelerated capacity decay at the end of the battery life cycle. It accurately fits the nonlinear degradation trajectory of lithium battery SOH, especially improving the prediction accuracy of the inflection point of accelerated degradation at the end of the life cycle.

[0024] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0025] Example 1: like Figure 1 As shown, this embodiment of the invention provides a method for predicting the state of health (SOH) of a lithium battery, including: S1. Collect multi-cycle test data of lithium battery and construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation. S2. The discharge timing curve feature sequence and the incremental capacity IC curve feature are spliced ​​along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN). The deep curve features of multi-scale local polarization and capacity decay are extracted through multi-receptive field parallel convolution branches. S3. After flattening the deep curve features, they are connected to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism, and the reshaped features are obtained after temporal dimension reshaping. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edge. S4. Fuse and concatenate the reshaped features with the single-valued scalar feature sequence, input it into the Long Short-Term Memory (LSTM) network to model the temporal dependency relationship, and obtain the full temporal hidden state sequence and the final hidden state. S5. An attention mechanism is introduced on the output side of the LSTM. The final hidden state is used as the query vector to perform adaptive weighting on the full-time hidden state sequence to obtain the context vector that enhances the key decay cycle information. S6. The context vector and the final hidden state are concatenated and fused according to their feature dimensions. A two-layer KAN is used to fit the nonlinear decay law of the battery to obtain the predicted SOH value of the lithium battery.

[0026] The embodiments of the present invention effectively improve the accuracy of lithium battery health state prediction, enhance the adaptability to nonlinear changes throughout the battery's life cycle, more accurately capture capacity decay patterns, optimize model computation efficiency, reduce parameter redundancy, and improve prediction stability and reliability to adapt to the deployment requirements of BMS edge devices, ultimately achieving accurate prediction of lithium battery SOH.

[0027] Specifically, in step S1, multi-cycle test data of lithium batteries are collected to construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation.

[0028] This step involves acquiring and constructing feature data for multi-source heterogeneous batteries.

[0029] By acquiring test data of lithium batteries at various cycle periods, three types of heterogeneous input features are constructed, as detailed in an optional implementation: Single-valued scalar feature sequence (Scalar Features) include cycle count, temperature statistics, internal resistance, and charging time; Discharge timing curve characteristic sequence (Curve Features) are voltage or capacity timing sequences with a fixed number of sampling points (e.g., 100 data points); Incremental capacity IC curve characteristic sequence (IC Features) By differentiating the charge-discharge curves, the incremental capacity (IC) curve, which reflects the state of electrochemical reaction inside the battery, is extracted.

[0030] In step S2, the discharge timing curve feature sequence and the incremental capacity IC curve feature are spliced ​​along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN). The deep curve features of multi-scale local polarization and capacity decay are extracted through multi-receptive field parallel convolution branches.

[0031] This step performs curve feature extraction based on multi-scale one-dimensional convolution (Multi-Scale CNN). By setting multiple convolution kernels of different sizes in parallel, it can simultaneously extract short-term local polarization abrupt features and medium- to long-term gradual capacity decay trends.

[0032] In an optional implementation, the CNN specifically includes three parallel convolutional branches with different receptive fields, each branch corresponding to a convolutional kernel of a different size; Among them, the smallest size convolution kernel is used to capture instantaneous voltage polarization abrupt changes; Intermediate-sized convolutional kernels are used to capture the temporal changes of steady states during charging and discharging. The largest size convolution kernel is used to capture a gradual capacity decay trend.

[0033] Specifically, this step involves extracting the characteristic sequence of the discharge timing curves mentioned above. Characteristic sequence of incremental capacity IC curve The sequences are concatenated along the channel dimension to construct a dual-channel curve feature sequence. The sequence is then input into a multi-scale CNN module for parallel extraction of local polarization features.

[0034] For example, for parallel multi-scale feature extraction, the model sets up three parallel one-dimensional convolutional branches with different receptive fields: the first branch uses a small convolutional kernel with size 3 and padding of 1 to capture instantaneous voltage polarization features; the second branch uses a medium-sized convolutional kernel with size 5 and padding of 2; and the third branch uses a large convolutional kernel with size 11 and padding of 5 to capture the gradual capacity decay trend. For feature concatenation and pooling, the features extracted by the three branches are concatenated along the channel dimension, and then sequentially passed through a batch normalization layer (BatchNorm1d), a ReLU activation function, and a max pooling layer (MaxPool1d) for feature compression and nonlinear mapping. This process is repeated twice (Layer 1 and Layer 2) to output deep curve features. .

[0035] In step S3, the deep curve features are flattened and then connected to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism and then reshaped through temporal dimension reshaping to obtain the reshaped features. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edge.

[0036] This step performs feature dimensionality reduction based on the KAN (Kolmogorov-Arnold Network). Specifically, the traditional multilayer perceptron (MLP) at the end of the model is replaced with the KAN network. Because the KAN network abandons fixed node activation functions and instead uses dynamic basis functions based on B-spline curves to perform nonlinear mapping on the network edges (weights), it has a very strong ability to approximate nonlinear functions.

[0037] Specifically, this step flattens the deep multi-scale curve features output from step two, transforming them into a one-dimensional high-dimensional feature vector. The specific formula is as follows: (1) In the formula, Indicates the feature flattening operation; The deep curve features output from step two; This is a reshaped one-dimensional high-dimensional feature vector suitable for network input; This is the total feature dimension of the one-dimensional high-dimensional feature vector. Subsequently, this sequence is input into the first-layer KANLinear network module.

[0038] Traditional models typically use Multilayer Perceptrons (MLPs) for dimensionality reduction, but here we use KAN layers. Unlike traditional MLPs that place fixed activation functions on neuron nodes, KAN layers dynamically learn activation functions at the network edges (weights) using B-splines. The input neuron to the first The connection edges of the output neurons and their dynamic nonlinear activation functions Represented as: (2) In the formula, x For a single feature scalar of input (i.e. The (each component); Based on the mapping weights; The basic nonlinear activation function; For the first One B-spline basis function; These are control point coefficients that can be dynamically learned; and These represent the configured grid size and spline order, respectively.

[0039] Based on the aforementioned edge activation mechanism, the KAN layer... The eigenvalues ​​of each output node It is a linear sum of all input features after being processed by the corresponding edge activation functions: (3) In the formula, For KAN layer number The feature values ​​of each output node; The total dimension of the input features; The traversal index for the input feature nodes; This represents the v-th input component in the flattened feature vector; This is the corresponding dynamic edge activation function.

[0040] Through this aggregation operation, the KAN layer can perform nonlinear mapping on extremely high-dimensional flattened features and compress them into a preset low-dimensional hidden layer space (such as reducing the dimension to 64 dimensions).

[0041] Subsequently, to adapt to the data format of downstream time-series networks, the dimensionality-reduced features are reshaped back into a time-series structure containing batches and sliding windows (i.e., feature reshaping): (4) In the formula, () represents the dimension reshaping operation, where Batch_Size and Window_Size are the batch size and sliding window length, respectively. is the dimension of the hidden layer features after dimensionality reduction, and is the final output dimensionality-reduced KAN feature sequence.

[0042] Understandably, the above processing significantly reduces the number of feature parameters and computational complexity while preserving the nonlinear degradation information of the battery physical curve characteristics to the greatest extent.

[0043] In step S4, the reshaped features and single-valued scalar feature sequences are fused and concatenated, and the temporal dependencies are modeled by the Long Short-Term Memory (LSTM) network to obtain the full temporal hidden state sequence and the final hidden state.

[0044] This step performs feature fusion and LSTM-based temporal dependency modeling. The specific implementation details are as follows: For feature concatenation, the single-valued scalar feature sequence obtained in step S1 is... Reshaping features output from step S3 The features are concatenated along the feature dimension to form a complete fused feature sequence. The specific formula is as follows: (5) In the formula, Indicates the current time step within the sliding window; Indicates time step Physical scalar characteristics (such as resistance, charging time, average temperature, etc.). Indicates time step Microscopic curve characteristics after dimensionality reduction using a KAN network; This represents a concatenation operation along the feature dimension. For time step The fusion feature sequence.

[0045] For time series modeling, feature sequences will be fused. By time step (set the sliding window size to...) The input is fed into the Long Short-Term Memory (LSTM) network.

[0046] LSTM utilizes its gating mechanism (input gate, forget gate, output gate) to model the sequential dependencies of sequences over time.

[0047] For any time step in the sequence The nonlinear update process of its internal cell state and hidden state can be expressed by the mathematical formula as follows: (6) (7) (8) (9) (10) In the formula, , , These represent the activation state vectors of the forget gate, input gate, and output gate, respectively. and These are the learnable weight matrix and bias term for the corresponding gating unit, respectively; It is a sigmoid non-linear activation function used to map the gate signal between 0 and 1 to control the throughput of the information flow; It is the hyperbolic tangent activation function; This represents the Hadamard product of matrix elements (i.e., element-wise multiplication). and Each is the previous time step ( The output cell state and the retained hidden state; Cell state preserved for the current time step (long-term memory); The hidden state (short-term memory) output at the current time step.

[0048] After all After recursive computation at each time step, the network outputs a full-time hidden state sequence. And the final hidden state of the last time step. : (11) (12) In the formula, Indicates time step The hidden layer states (where ,Right now Corresponding to the 1st to the 2nd (Hidden state output at each time step); It contains all historical aging evolution information within the entire sliding window period and will serve as the global input matrix for subsequent attention mechanisms; This condenses the battery life state most recent to the current moment, which will be used as the query vector of the attention mechanism and the core feature of subsequent bypass propagation to participate in the calculation of the end-of-life inflection point.

[0049] In step S5, an attention mechanism is introduced on the output side of the LSTM. The final hidden state is used as the query vector to perform adaptive weighting on the full-time hidden state sequence to obtain a context vector that enhances the key decay cycle information.

[0050] This step performs key information enhancement based on the attention mechanism. In an optional implementation, the adaptive weighting process uses Softmax normalization to allocate weights, thereby achieving accurate enhancement of key decay cycle information and improving the reliability of the context vector.

[0051] Specifically, the full-time hidden state sequence output by the LSTM in step S4 And the final hidden state of the last time step. The input is fed into the Attention Layer for cross-cycle feature weight allocation.

[0052] The attention layer will eventually hide the state. As a query vector, the full-time hidden state sequence As a key-value vector (Key and Value). Specific implementation details are as follows: First, each historical time step is calculated through the internal fully connected layer. The correlation score between the hidden state and the final state ( The calculation formula is: (13) In the formula, For the first The hidden state at each time step is... ; and These are the learnable weight matrices for the corresponding fully connected layers, used to map the two sets of features to the same dimensional space; It is the hyperbolic tangent activation function; The score-mapped weight matrix is ​​used to transform non-linear features into a single relevance scalar score. .

[0053] Subsequently, the scores at all time steps are globally normalized using the Softmax function, transforming them into attention weights representing the degree of importance. : (14) In the formula, This represents the total time step length of the sliding window; This is the index variable for summing the time steps within the entire sliding window; It is an exponential function with the natural constant e as its base; The value of is between 0 and 1, and the sum of the weights of all time steps is 1.

[0054] Next, the attention weights ( ) and corresponding number The hidden layer state at each time step ( The weighted summation is performed to output a context vector that adaptively focuses on key historical recession cycles. : (15) In the formula, This represents the total time step length of the sliding window; Index for time step traversal; For the first Attention weights for each time step; For the first The hidden layer state at each time step.

[0055] Understandably, through the above mechanism, the model breaks through the information forgetting bottleneck in long sequence modeling, enabling the subsequent network regression layer to prioritize capturing the core historical degradation cycle that contributes the most to the current battery aging (SOH).

[0056] In step S6, the context vector and the final hidden state are concatenated and fused according to their feature dimensions. The nonlinear decay law of the battery is fitted using a two-layer KAN to obtain the predicted SOH value of the lithium battery.

[0057] This step performs a health status (SOH) prediction based on a two-layer KAN regressor.

[0058] In an optional implementation, a double-layer KAN is designed, comprising two linear KAN layers. Through successive nonlinear transformations, the fitting accuracy of the nonlinear decay law is improved, wherein: The first KAN linear layer uses a dynamic activation function based on B-spline curves to perform high-order nonlinear mapping and compression on the spliced ​​and fused features, and outputs low-dimensional hidden regression features. The second KAN linear layer receives low-dimensional hidden regression features and uses the approximation capability of high-order functions to fit the highly nonlinear capacity decay trajectory at the end of the lithium battery's life cycle and output the SOH prediction value.

[0059] Specifically, this step will use the context vector output in step S5. With the final hidden state in LSTM Bypass concatenation is performed to generate regressor input features that include historical attributions for decline and the current instantaneous state. The specific splicing and fusion formula is as follows: (16) In the formula, This represents a tensor concatenation operation along the feature dimension; It encompasses key information on global long-term historical decline, weighted by an attention mechanism; It provides the most recent instantaneous state features.

[0060] Understandably, the bypass design described above can effectively prevent the critical state at the latest moment from being overly diluted in the global attention weighting.

[0061] Then, the regressor generated by splicing is input into the features. The input is fed into a two-layer KAN network that serves as the end predictor.

[0062] The first KAN linear layer utilizes a dynamic activation function based on B-spline curves to input features into the regressor. Perform high-order nonlinear mapping and compression to output low-dimensional hidden regression features. , its first The formula for calculating the feature components of each output node is: (17) In the formula, This is the total dimension of the concatenated features (i.e., twice the dimension of the LSTM hidden layer). For the first layer of the KAN network, connect the th The input feature node and the first The dynamic edge activation function of each output node (composed of a basic linear mapping and B-spline basis functions); These are the hidden regression features after nonlinear dimensionality reduction.

[0063] The second-layer KANLinear receives this low-dimensional hidden regression feature and, utilizing its excellent high-order function approximation capability, further fits the highly nonlinear capacity decay trajectory of the lithium battery at the end of its life cycle. Its final output formula is: (18) In the formula, The dimension of the hidden features after dimensionality reduction of the first layer KAN; This is the dynamic edge activation function corresponding to the second-layer KAN network; The network's final output is a single continuous scalar, which is the current model's extremely accurate estimate of the lithium battery's state of health (SOH).

[0064] To help understand the embodiments of the present invention, the following explanation uses SOH estimation based on the MIT-Stanford-Toyota battery dataset as an example: Taking the MIT lithium battery dataset as an example, this paper details the specific operation process and parameter passing specifications based on the architecture of multi-scale convolution, long short-term memory network, attention mechanism and Kolmogorov-Arnold network (MS-CLA-KAN).

[0065] A. Implementation Environment and Basic Parameter Settings Hardware and software operating environment: The model building and training in this embodiment are based on the deep learning framework PyTorch, and the computing device adopts a GPU environment that supports CUDA acceleration to meet the requirements of high-concurrency tensor computation. To ensure the absolute reproducibility of experimental results, the system's global random seed is set to 0.

[0066] The predicted target_col is the battery capacity, which is converted to SOH during the evaluation phase by dividing it by the initial capacity (the maximum value of the first 5 cycles). The failure threshold of SOH is set to 0.8 (i.e. 80%).

[0067] Rated capacity: set to 1.1Ah.

[0068] The sliding window size, window_size, is set to 32, meaning that the state of the next cycle is predicted using data from the past 32 cycles.

[0069] Batch size: set to 128.

[0070] LSTM network parameters: Set the number of layers to single layer (lstm_layers=1) and the hidden state dimension to 64 (lstm_hidden=64).

[0071] Kolmogorov-Arnold Network (KAN) parameters: Set the grid size of the B-spline curve to 3 to balance the accuracy of nonlinear fitting with the computational cost of edge deployment.

[0072] Regularization parameter: Set the Dropout rate to 0.2 to prevent the network from overfitting on limited battery cycle data.

[0073] B. Data Acquisition and Preprocessing Stage The system extracts battery charge and discharge data from the test equipment and constructs three sets of synchronized time-series features. During this process, a specific window of moving average smoothing is performed to eliminate measurement noise. (i) Physical scalar characteristics Scalar data was extracted from three dimensions: internal resistance (resistance), charging time (charge_time), and average temperature (temp_avg). After Log transformation, a sliding smoothing was performed using a window size of 50.

[0074] (ii) Characteristics of voltage curves Extract 100 voltage sampling points (v_0 to v_99) for each cycle and perform sliding smoothing with a window size of 5.

[0075] (iii) Characteristics of incremental capacity curve Extract 100 IC sampling points (ic_0 to ic_99) for each cycle, and use a window of size 5 for smoothing.

[0076] The above features are each min-max normalized using FeatureScaler and then sliced ​​according to a window size of 32. During training, a single batch of data input to the model contains: The dimensions are (128, 32, 3). The dimensions are (128, 32, 100). The dimensions are (128, 32, 100).

[0077] C. Model forward propagation and feature evolution process Step c-1: Local feature extraction of the multi-scale CNN backbone network MS-CNN Input feature channel splicing: combining voltage curve features and incremental capacity curve characteristics After reshaping the sequence length dimension, it is spliced ​​in the channel dimension to construct a dual-channel curve feature sequence. The dimensions are transformed into (4096, 2, 100), where 4096 is the product of the batch size of 128 and the window size of 32.

[0078] Layer 1 (Shallow Extraction): Data enters the first multi-scale convolutional layer, simultaneously passing through three parallel 1D convolutional branches (kernel sizes of 3, 5, and 11, with corresponding padding of 1, 2, and 5), each branch outputting 16 channels. After concatenation, it undergoes batch normalization (BatchNorm1d, 48 channels), ReLU activation, and a max-pooling layer with a kernel size of 2. At this point, the tensor dimension is compressed to (4096, 48, 50).

[0079] Layer 2 (Deep Extraction): After the above operations, the data enters the second layer of multi-scale convolutional branches. Each branch outputs 32 channels. After feature concatenation and pooling, the deep multi-scale curve features are output. Its tensor dimension is (4096, 96, 25).

[0080] Step c-2: Nonlinear dimensionality reduction based on KANLinear The deep multi-scale curve features output by the multi-scale CNN backbone network ( The sequence is flattened along the sequence dimension, and the feature dimension becomes 96×25=2400.

[0081] The flattened tensor (4096, 2400) is input into the first layer of a KANLinear network with a grid_size of 3. Using a B-spline-based edge activation function, it is non-linearly reduced to 64 dimensions, then reshaped back to the batch sequence shape (128, 32, 64), finally outputting the dimensionality-reduced KAN feature sequence. .

[0082] Step c-3: Long and short-term time series modeling and dynamic attention weight allocation Multi-source feature fusion: The reshaped dimensionality-reduced KAN feature sequence With 3D single-valued scalar features The features are concatenated along the feature dimension to construct a complete fused feature sequence. The dimensions are (128, 32, 67).

[0083] LSTM temporal extraction: fusing feature sequences The time-step input is fed into a single-layer (lstm_layers=1) LSTM network with 64 hidden nodes (lstm_hidden=64). Utilizing its internal gating mechanism, it outputs a temporal hidden layer state sequence containing all historical time steps. (dimensions 128, 32, 64), and the final hidden state representing the final memory of the entire sequence. (Dimensions are 128, 64).

[0084] Attention weighting: and All these elements enter the Attention Layer. The fully connected layer calculates the relevance scores of the 32 time steps in the sequence to the final state. These scores are then spatially normalized using a Softmax function and transformed into attention weights. The weighted summation outputs a context vector focused on key historical decay cycles. The dimensions are (128, 64).

[0085] Step c-4: Regression prediction based on two-layer KAN Bypassing and concatenation: passing the context vector With final hidden state Bypass concatenation is performed to generate regressor input features that include historical attributions for decline and the current instantaneous state. The dimension is (128, 128). It is then fed into a regressor constructed from two layers of KANLinear (grid_size=3).

[0086] Two-layer regression fitting: The first layer KANLinear will Map and compress from 128 dimensions to output low-dimensional hidden regression features. (64-dimensional). The second layer KANLinear receives this low-dimensional feature and further approximates the highly nonlinear capacity decay trajectory at the end of the lifetime, finally mapping it to output dimension 1 (output_size=1), outputting the predicted capacity value corresponding to the current sliding window (the final shape of the tensor is (128,1)), and converting it into an SOH estimate by dividing it by the initial capacity during evaluation.

[0087] D. Model Training and Evaluation Strategies Loss function and optimizer: Mean squared error (MSE) is used as the loss function. The optimizer selected is AdamW, with an initial learning rate of 0.0001 and a weight decay parameter of 0.0001, and is used in conjunction with a Cosine Annealing (LR) learning rate scheduler.

[0088] Training cycle and early stopping mechanism: The maximum number of training rounds is set to 100. The system monitors the loss performance on the validation set. When the decrease in the validation set loss is less than 1e-6, the patience value is accumulated. If no effective improvement is seen for 20 consecutive rounds, the early stopping mechanism is triggered to terminate training and prevent overfitting.

[0089] Cross-validation: A 5-fold cross-validation strategy was used to evaluate the model's generalization ability. Finally, the root mean square error (RMSE) and mean absolute percentage error (MAPE) of the five experiments were used as the performance evaluation indicators of the system.

[0090] E. Experimental Results and Verification of Beneficial Effects To verify the effectiveness and advancement of the proposed Multi-Scale CNN-LSTM-Attention with Kolmogorov-Arnold Network (MS-CLA-KAN) model in this embodiment of the invention, rigorous comparative and ablation experiments were conducted based on the MIT-Stanford-Toyota lithium battery public dataset. Five-fold cross-validation was used to ensure the statistical reliability of the results.

[0091] (1) Explanation of evaluation indicators The industry-standard root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as evaluation metrics for system performance. The smaller the RMSE and MAPE values, the higher the overall prediction accuracy of the model.

[0092] (2) Baseline model comparison experiment and analysis of beneficial effects The model of this invention is compared with existing mainstream battery health state prediction models (including multilayer perceptron MLP, standard LSTM, standard Transformer, and conventional CNN-LSTM-Transformer cascade architecture). The average error test results of each model under 5-fold cross-validation are shown in Table 1: Table 1. Average Error Test Results (3) Analysis of the ability to track the nonlinear "accelerated decline inflection point" at the end of life. Traditional deep learning models perform reasonably well in the early stages of battery cycling, but often exhibit "linear prediction inertia" at the end of battery life (after SOH drops below 80%), causing the prediction curve to deviate significantly from the actual capacity degradation inflection point.

[0093] Figure 2 The SOH prediction curve of the traditional LSTM model is disclosed, as well as Figure 3 The SOH prediction curve of the MS-CLA-KAN model has been published, such as... Figure 2 and Figure 3 As can be seen from the intuitive comparison, the embodiments of the present invention have an accuracy advantage in predicting the fit of the end of life that is unmatched by traditional models.

[0094] Example 2: This invention provides a lithium battery state of health (SOH) prediction system, comprising: The heterogeneous feature construction module is used to collect multi-cycle test data of lithium batteries and construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation. The multi-scale convolutional feature extraction module is used to concatenate the discharge timing curve feature sequence and the incremental capacity IC curve feature along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN) and extracted deep curve features of multi-scale local polarization and capacity decay through multi-receptive field parallel convolutional branches. The KAN nonlinear dimensionality reduction module is used to flatten deep curve features and then connect them to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism, and the reshaped features are obtained after temporal dimension reshaping. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edges. The cross-domain feature fusion and temporal modeling module is used to fuse and concatenate reshaped features with single-valued scalar feature sequences, input the Long Short-Term Memory (LSTM) network to model temporal dependencies, and obtain the full temporal hidden state sequence and the final hidden state. The dynamic attention weighting module is used to introduce an attention mechanism on the output side of the LSTM. Using the final hidden state as the query vector, it adaptively weights the full-time hidden state sequence to obtain a context vector that enhances key decay cycle information. The SOH prediction output module is used to concatenate and fuse the context vector and the final hidden state by feature dimension, and use a two-layer KAN to fit the nonlinear decay law of the battery to obtain the SOH prediction value of the lithium battery.

[0095] The embodiments of the present invention effectively improve the accuracy of lithium battery health state prediction, enhance the adaptability to nonlinear changes throughout the battery's life cycle, more accurately capture capacity decay patterns, optimize model computation efficiency, reduce parameter redundancy, and improve prediction stability and reliability to adapt to the deployment requirements of BMS edge devices, ultimately achieving accurate prediction of lithium battery SOH.

[0096] In an optional implementation, the CNN includes three parallel convolutional branches with different receptive fields, each branch corresponding to a convolutional kernel of a different size; Among them, the smallest size convolution kernel is used to capture instantaneous voltage polarization abrupt changes; Intermediate-sized convolutional kernels are used to capture the temporal changes of steady states during charging and discharging. The largest size convolution kernel is used to capture a gradual capacity decay trend.

[0097] In an optional implementation, the adaptive weighting process uses Softmax normalization to allocate weights.

[0098] In an optional implementation, the double-layer KAN comprises two linear KAN layers; The first KAN linear layer uses a dynamic activation function based on B-spline curves to perform high-order nonlinear mapping and compression on the spliced ​​and fused features, and outputs low-dimensional hidden regression features. The second KAN linear layer receives low-dimensional hidden regression features and uses the approximation capability of high-order functions to fit the highly nonlinear capacity decay trajectory at the end of the lithium battery's life cycle and output the SOH prediction value.

[0099] In an optional implementation, the single-valued scalar feature sequence includes at least one of the following: number of cycles, temperature statistics, internal resistance, and charging time.

[0100] In an optional implementation, the discharge timing curve is characterized as a voltage or capacity timing sequence with fixed sampling points.

[0101] In an alternative implementation, the incremental capacity IC curve characteristics are obtained by differentiating the charge-discharge curve.

[0102] Example 3: This invention provides a storage medium storing a computer program, wherein the computer program causes a computer to execute a lithium battery state of health (SOH) prediction method as provided in any embodiment of this invention.

[0103] In embodiments of the present invention, any combination of one or more storage media may be used. The storage medium may be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable computer disk, a hard disk, RAM, ROM, erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in connection with an instruction execution system, apparatus, or device.

[0104] Example 4: This invention provides an electronic device. Figure 4 The diagram shown illustrates the structure of an electronic device 100 provided in an embodiment of the present invention. In some embodiments, the electronic device may be a mobile phone, tablet computer, wearable device, in-vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), or other terminal device. Furthermore, the lithium battery state of health (SOH) prediction method provided in this embodiment can also be applied to databases, servers, and service response systems based on terminal artificial intelligence. This embodiment does not limit the specific application scenarios of the lithium battery SOH prediction method.

[0105] like Figure 4 As shown, the electronic device 100 provided in this embodiment of the invention includes a memory 101 and a processor 102.

[0106] The memory 101 is used to store computer programs; preferably, the memory 101 includes various media that can store program code, such as ROM, RAM, magnetic disk, USB flash drive, memory card or optical disk.

[0107] Specifically, memory 101 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. Electronic device 100 may further include other removable / non-removable, volatile / non-volatile computer system storage media. Memory 101 may include at least one program product having a set (e.g., at least one) of program modules configured to perform the functions of the embodiments of the present invention.

[0108] The processor 102 is connected to the memory 101 and is used to execute the computer program stored in the memory 101 so that the electronic device 100 executes the lithium battery state of health (SOH) prediction method provided in any embodiment of the present invention.

[0109] In an optional implementation, the processor 102 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it may also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, evaluation index gate or transistor logic devices, or evaluation index hardware components.

[0110] In an optional embodiment, the electronic device 100 of this invention may further include a display 103. The display 103 is communicatively connected to the memory 101 and the processor 102, and is used to display the relevant GUI interactive interface of the lithium battery state of health (SOH) prediction method.

[0111] It is understood that the lithium battery state of health (SOH) prediction system, storage medium, and electronic device provided in the embodiments of the present invention correspond to the lithium battery state of health (SOH) prediction method provided in the embodiments of the present invention. The explanations, examples, and beneficial effects of the relevant contents can be referred to the corresponding parts of the method, and will not be repeated here.

[0112] In summary, compared with existing technologies, it has the following beneficial effects: 1. Overcoming the bottleneck of nonlinear lifetime degradation prediction and significantly improving the accuracy of full life cycle prediction. Related technologies typically use traditional fully connected layers (MLPs) with fixed activation functions (such as ReLU) at the end of the model, which struggle to accurately fit the highly nonlinear capacity "knee point" phenomenon at the end of the battery's life cycle. This invention innovatively introduces a Kolmogorov-Arnold network (KAN), utilizing B-splines to dynamically learn nonlinear activation functions on the network's edge weights. This high-order function approximation capability makes the model more sensitive to tracking the inflection point of accelerated capacity degradation, significantly reducing the estimation error of the mid-to-late stage state of health (SOH).

[0113] 2. Achieving comprehensive perception of multi-scale micro-polarization features and single-valued scalar features, enhancing model robustness. Related techniques only use convolutional kernels of a single size (e.g., size 3) and typically only process single charge-discharge sequences. This invention constructs a dual-channel micro-input of voltage curves (Curve) and incremental capacity curves (IC), and sets parallel multi-scale one-dimensional convolutional branches with sizes of 3, 5, and 11. This structure acts like a multi-resolution "microscope," capable of simultaneously capturing instantaneous ohmic polarization abrupt changes (small convolutional kernel) and gradual decay trends spanning longer periods (large convolutional kernel). Combined with single-valued scalar features such as internal resistance and temperature, it completely eliminates the feature omission problem caused by a single-scale receptive field.

[0114] 3. Reduced model computational complexity and parameter redundancy, facilitating BMS edge hardware deployment. The relevant technology employs a standard Transformer encoder for long-term time series modeling; its global self-attention mechanism has a computational complexity quadratic with the sequence length (O(N)). 2 The rapid growth of data leads to significant computational consumption. This invention utilizes a Long Short-Term Memory (LSTM) network to extract temporal dependencies and then designs a lightweight dynamic temporal attention layer. Weights are allocated by calculating the correlation between the hidden state and the final state at each historical time step. This improvement significantly reduces the model's computational load while retaining the ability to "focus" on key historical degradation cycles, meeting the real-time performance and low memory requirements of the underlying Battery Management System (BMS).

[0115] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0116] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the state of health (SOH) of a lithium battery, characterized in that, include: Collect multi-cycle test data of lithium batteries and construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation. The discharge timing curve feature sequence and the incremental capacity IC curve feature are concatenated along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN). The deep curve features of multi-scale local polarization and capacity decay are extracted through multi-receptive field parallel convolution branches. After flattening the deep curve features, they are connected to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism, and then reshaped through temporal dimension reshaping to obtain the reshaped features. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edge. The reshaped features and single-valued scalar features are fused and concatenated, and then input into a Long Short-Term Memory (LSTM) network to model temporal dependencies, thereby obtaining the full temporal hidden state sequence and the final hidden state. An attention mechanism is introduced on the output side of the LSTM. The final hidden state is used as the query vector to perform adaptive weighting on the full-time hidden state sequence to obtain a context vector that enhances key decay cycle information. The context vector and the final hidden state are concatenated and fused according to their feature dimensions. A two-layer KAN is used to fit the nonlinear decay law of the battery to obtain the predicted SOH value of the lithium battery.

2. The lithium battery state of health (SOH) prediction method as described in claim 1, characterized in that, The CNN includes three parallel convolutional branches with different receptive fields, each branch corresponding to a convolutional kernel of a different size; Among them, the smallest size convolution kernel is used to capture instantaneous voltage polarization abrupt changes; Intermediate-sized convolutional kernels are used to capture the temporal changes of steady states during charging and discharging. The largest size convolution kernel is used to capture a gradual capacity decay trend.

3. The lithium battery state of health (SOH) prediction method as described in claim 1, characterized in that, The adaptive weighting process uses Softmax normalization to allocate weights.

4. The lithium battery state of health (SOH) prediction method as described in claim 1, characterized in that, The double-layer KAN comprises two linear KAN layers; The first KAN linear layer uses a dynamic activation function based on B-spline curves to perform high-order nonlinear mapping and compression on the spliced ​​and fused features, and outputs low-dimensional hidden regression features. The second KAN linear layer receives low-dimensional hidden regression features and uses the approximation capability of high-order functions to fit the highly nonlinear capacity decay trajectory at the end of the lithium battery's life cycle and output the SOH prediction value.

5. The lithium battery state of health (SOH) prediction method as described in claim 1, characterized in that, The single-valued scalar feature sequence includes at least one of the following: number of cycles, temperature statistics, internal resistance, and charging time.

6. The lithium battery state of health (SOH) prediction method as described in claim 1, characterized in that, The discharge timing curve is characterized by a voltage or capacity timing sequence at fixed sampling points.

7. The lithium battery state of health (SOH) prediction method as described in claim 1, characterized in that, The incremental capacity IC curve characteristics are obtained by differentiating the charge and discharge curves.

8. A lithium battery state of health (SOH) prediction system, characterized in that, include: The heterogeneous feature construction module is used to collect multi-cycle test data of lithium batteries and construct three types of heterogeneous input features: single-value scalar feature sequence, discharge time sequence curve feature sequence, and incremental capacity IC curve feature sequence obtained by differentiation. The multi-scale convolutional feature extraction module is used to concatenate the discharge timing curve feature sequence and the incremental capacity IC curve feature along the channel dimension to form a dual-channel sequence, which is then input into a multi-scale one-dimensional convolutional neural network (CNN) and extracted deep curve features of multi-scale local polarization and capacity decay through multi-receptive field parallel convolutional branches. The KAN nonlinear dimensionality reduction module is used to flatten deep curve features and then connect them to the Kolmogorov-Arnold network (KAN). The high-dimensional features are nonlinearly reduced through a dynamic edge activation mechanism, and the reshaped features are obtained after temporal dimension reshaping. The dynamic edge activation mechanism uses dynamic basis functions based on B-spline curves to perform nonlinear mapping at the network edges. The cross-domain feature fusion and temporal modeling module is used to fuse and concatenate reshaped features with single-valued scalar feature sequences, input the Long Short-Term Memory (LSTM) network to model temporal dependencies, and obtain the full temporal hidden state sequence and the final hidden state. The dynamic attention weighting module is used to introduce an attention mechanism on the output side of the LSTM. Using the final hidden state as the query vector, it adaptively weights the full-time hidden state sequence to obtain a context vector that enhances key decay cycle information. The SOH prediction output module is used to concatenate and fuse the context vector and the final hidden state by feature dimension, and use a two-layer KAN to fit the nonlinear decay law of the battery to obtain the SOH prediction value of the lithium battery.

9. A storage medium, characterized in that, It stores a computer program, wherein the computer program causes the computer to execute the lithium battery state of health (SOH) prediction method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes: Processor and memory; The memory stores program instructions; The processor is configured to run the program instructions to execute the lithium battery state of health (SOH) prediction method as described in any one of claims 1 to 7.

Images