Lithium battery life prediction method and system based on dynamic graph convolution and multi-modal fusion

Abstract

The invention discloses a lithium battery life prediction method and system based on dynamic graph convolution and multi-modal fusion. The method comprises a battery data acquisition and preprocessing module, a dynamic graph construction module, a dynamic graph convolution calculation module, a multi-modal fusion module, a life prediction module and an optimization and verification module. Preprocessing the multi-modal data of the lithium battery, and constructing a heterogeneous graph data structure; performing time sequence feature extraction on the battery data by adopting a dynamic graph convolutional neural network, and capturing long-term and short-term dynamic features; deep fusion of multi-source data is realized in combination with time sequence features, a self-attention mechanism and a feature weighting strategy; and predicting the remaining service life of the lithium battery by using the deep neural network model, and performing visual display. The method can accurately describe the nonlinear attenuation trend of the battery life, improves the prediction accuracy, and is suitable for the fields of new energy vehicles, power grid energy storage and consumer electronics.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery state of health evaluation and life prediction, and particularly relates to a lithium battery life prediction method and system based on dynamic graph convolution and multi-modal fusion. BACKGROUND

[0002] As one of the most important energy storage devices, lithium ion batteries are widely used in new energy vehicles, grid energy storage systems and other fields. However, the state of health (SOH) and the remaining useful life (RUL) of lithium batteries are affected by environmental temperature, charging and discharging rate, material degradation and other factors during long-term operation, resulting in battery capacity attenuation, electrochemical performance degradation and even safety accidents. Therefore, accurately predicting the state of health of lithium batteries and optimizing battery management strategies have become a key research topic in the academic and industrial fields.

[0003] At present, lithium battery life prediction methods mainly include physical models, statistical analysis models, machine learning and deep learning methods. Physical models simulate the chemical reaction and charge transport process inside the battery to describe the aging mechanism of the battery, which is greatly affected by the environment and usage conditions, and has poor universality. Statistical analysis models use regression analysis, Markov chains, Kalman filtering and other methods to establish a mathematical relationship between battery capacity degradation and usage time, which is difficult to adapt to complex and variable working conditions. Machine learning methods often require a large amount of feature engineering, and feature selection and parameter adjustment can only rely on expert experience.

[0004] With the development of deep learning, graph neural networks have gradually become an important research direction in time series analysis. In the field of battery life prediction, a dynamic graph model is constructed to map the battery operating state to the nodes in the dynamic graph, and graph convolution (GCN) or graph attention mechanism (GAT) is used for feature extraction to model the complex correlation in the battery aging process. However, most of the current graph neural network methods still face the problems of high computational complexity and low model interpretability when processing time series data. How to further optimize the dynamic graph structure and improve the computational efficiency is the focus of current research. SUMMARY

[0005] Therefore, the present application provides a lithium battery life prediction method and system based on dynamic graph convolution and multi-modal fusion, which optimizes the graph structure by dynamically adjusting the node relationship through the construction of a dynamic graph adjacency matrix, and improves the extraction ability of time series features. The self-attention mechanism is used to weight and fuse multi-modal data, improving the utilization rate of different sensor data. The problems of high computational complexity and low prediction accuracy of existing graph convolution are solved.

[0006] For this purpose, in a first aspect, the present application proposes a lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion, comprising the following steps:

[0007] (1) Collecting battery state data of different modalities during the operation of lithium batteries, including voltage, current, temperature, internal resistance, and charge-discharge cycle number, and preprocessing the battery state data;

[0008] (2) Constructing a heterogeneous dynamic graph structure based on battery operation characteristics, wherein the nodes represent the battery state at different time steps, the edges represent the time evolution relationship, and the weights are assigned to obtain a heterogeneous graph feature vector;

[0009] (3) Using a dynamic graph convolutional neural network to extract features from the heterogeneous graph data, capturing the temporal dynamic changes of the lithium battery state, and obtaining a time feature vector;

[0010] (4) Combining attention mechanism and feature weighting strategy to process the time feature vector, and fusing data of different modalities to obtain multi-modal time series features;

[0011] (5) Based on the extracted multi-modal time series features, using a TCN-Attention model to predict the remaining useful life of the lithium battery, and obtaining the prediction result.

[0012] Further, the dynamic graph modeling adopts a strategy based on a dynamic adjacency matrix, and the adjacency matrix changes with time steps to adapt to the battery aging process, comprising the following steps:

[0013] 1) According to the historical operation data of the lithium battery, an initial adjacency matrix is constructed to describe the correlation between the battery states at different time steps;

[0014] 2) The cosine similarity between nodes is used to represent the correlation of the state;

[0015] 3) An adaptive graph updating mechanism based on reinforcement learning is adopted, and the graph updating process is regarded as a decision problem:

[0016] Define State: current battery state feature vector: voltage, current, temperature and SOH estimated value; Recent prediction error statistics: MAE and RMSE; Sparsity index of the current adjacency matrix;

[0017] Define Action: adjustable parameters: adjustment amount of decay factor, re-distribution coefficient of different time step weights in sliding window and connection threshold adjustment;

[0018] Define Reward: based on the improvement of prediction accuracy at subsequent time steps, add a regularization term;

[0019] 4) Based on the error of the current prediction and the effect of historical updates, use the reinforcement learning algorithm DQN to actively learn and adjust the adjacency matrix, including decay strength, weight distribution within the window, and node connection relationship. Trigger the update decision after each prediction time step, or when the prediction error exceeds the threshold, to minimize future prediction errors and dynamically learn the optimal strategy.

[0020] 5) Use a sliding window strategy to retain only the information of the last N time steps, reducing computational complexity while enhancing attention to the latest battery state.

[0021] 6) Dynamically adjust the size of the time window in combination with the operating cycle of the battery.

[0022] Further, the dynamic graph convolution calculation uses an adaptive graph convolution mechanism to dynamically adjust the graph structure, including:

[0023] Using GCN-based convolution operations, aggregate node information at adjacent time steps, calculate the weighted sum of neighbor node information, use attention mechanisms to assign different weights to different neighbor nodes, and use a two-path dynamic graph convolution to learn the battery state changes at different time scales.

[0024] Further, the multi-modal fusion module includes a self-attention mechanism that can assign weights based on the importance of the data modalities, including:

[0025] Using an adaptive weighted fusion method, dynamically adjust the contribution of the modalities based on the current battery state, use a multi-head attention mechanism to weight the sum of features from different modalities, calculate the correlation between modalities through the Query, Key, and Value mechanisms, and use a gating mechanism to control the flow of features from different modalities.

[0026] Further, the life prediction module uses a time convolution model based on a self-attention mechanism (TCN-Attention model) to learn the decay trend of lithium batteries, including:

[0027] Using a causal convolution structure, the current prediction only depends on past information, using a dilated convolution mechanism to expand the time window, combining multi-head attention to learn features at different time scales, and using a sliding window smoothing process to smooth the prediction results.

[0028] Further, the model optimization module uses a multi-objective optimization strategy, including minimizing prediction error and quantifying uncertainty, including:

[0029] Adopt MSE, MSLE, Huber loss function, combine Adam, RMSProp optimization algorithm, combine Monte Carlo Dropout for multiple forward propagation, adopt time series disturbance, local noise data enhancement method, adopt cross validation and independent test set evaluation.

[0030] In a second aspect, the application further provides a lithium battery life prediction system based on dynamic graph convolution and multi-modal fusion, which comprises:

[0031] A data acquisition module is configured to acquire battery state data of different modalities during lithium battery operation, including voltage, current, temperature, internal resistance, and charge / discharge cycle number, and to preprocess the battery state data

[0032] A dynamic graph construction module is configured to construct a heterogeneous dynamic graph structure based on battery operation characteristics, wherein nodes represent battery states at different time steps, edges represent time evolution relationships, and weights are assigned to obtain a heterogeneous graph feature vector.

[0033] A dynamic graph convolution calculation module is configured to use a dynamic graph convolution neural network to extract features from the heterogeneous graph data, capture the temporal dynamic changes of the lithium battery state, and obtain a time feature vector.

[0034] A multi-modal fusion module is configured to process the time feature vector using an attention mechanism and a feature weighting strategy, fuse data of different modalities, and obtain multi-modal time series features.

[0035] A life prediction module is configured to predict the remaining useful life of the lithium battery based on the extracted multi-modal time series features using a TCN-Attention model, and obtain a prediction result.

[0036] In a third aspect, the application further provides a lithium battery life prediction device based on dynamic graph convolution and multi-modal fusion, comprising a memory and one or more processors, wherein the memory stores executable code, and the processor executes the executable code to implement the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion.

[0037] In a fourth aspect, the application further provides a computer-readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion.

[0038] In a fifth aspect, the application further provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion.

[0039] The application has the following advantages:

[0040] 1. The present application combines dynamic graph modeling, graph neural network, self-attention mechanism and time convolution network for lithium battery life prediction. By constructing a dynamic graph adjacency matrix, dynamically adjusting the node relationship, optimizing the graph structure and improving the extraction ability of time series features. The self-attention mechanism is used to weight and fuse multi-modal data, improve the utilization rate of different sensor data, and use the TCN-Attention model for time series prediction to enhance the modeling ability of long-range dependencies. A multi-objective optimization strategy is adopted, including loss functions such as mean square error and mean square logarithmic error, and Monte Carlo Dropout for uncertainty quantification to improve the stability and prediction accuracy of the model.

[0041] 2. Compared with the traditional method, the present application adopts dynamic graph convolutional neural network, which can dynamically model the time evolution of battery state and improve the feature extraction ability; combined with the self-attention mechanism, the multi-modal data is weighted and fused to improve the utilization rate of different modal data; the TCN-Attention model is used to combine causal convolution and dilated convolution to improve the modeling ability of time series; the multi-objective optimization strategy is used to improve the robustness of the model and improve the reliability of the prediction results.

[0042] 3. The present application can be deployed in embedded devices, cloud platforms or edge computing terminals to realize intelligent battery management. It can be widely used in new energy vehicles, power grid energy storage, intelligent battery management systems and other fields, providing more accurate and efficient solutions for lithium battery health management and life prediction. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The flowchart of the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion provided by the embodiments of the present application.

[0044] Figure 2 The structure diagram of the lithium battery life prediction system based on dynamic graph convolution and multi-modal fusion provided by the embodiments of the present application.

[0045] Figure 3 The structure diagram of the lithium battery life prediction device based on dynamic graph convolution and multi-modal fusion provided by the embodiments of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme of the embodiments of the present application will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0047] As Figure 1 shown, the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion provided by the embodiment of the application comprises the following steps: first, the multi-modal data (including voltage, current, temperature, internal resistance, cycle number, etc.) of the lithium battery is preprocessed, and a heterogeneous graph data structure is constructed. Second, a dynamic graph convolutional neural network is used to extract time series features of the battery data to capture long-term and short-term dynamic characteristics. Subsequently, a multi-modal fusion module is designed to combine time series features, self-attention mechanisms and feature weighting strategies to achieve deep fusion of multi-source data. Finally, a prediction model is used to predict the remaining useful life of the lithium battery, and the effectiveness of the method is verified through experiments.

[0048] As Figure 2 shown, the lithium battery life prediction system structure diagram based on dynamic graph convolution and multi-modal fusion provided by the embodiment of the application comprises the following steps:

[0049] Through the multi-modal data acquisition module, data of different modalities during the operation of the lithium battery are acquired, including but not limited to voltage, current, temperature, internal resistance, charge and discharge cycle number, SOC and SOH, etc. Then, the acquired data is cleaned, normalized and time-aligned, and high-quality time series data is constructed, including:

[0050] X={V,I,T,R int ,C,SOC,SOH}

[0051] Wherein, V is voltage, I is current, T is temperature, R int is internal resistance, C is charge and discharge cycle number, SOC is state of charge, and SOH is state of health.

[0052] Data preprocessing includes:

[0053] 1) Normalization:

[0054]

[0055] 2) Outlier detection: Z-score method is used to remove outliers:

[0056]

[0057] Wherein, i is the time step, μ is the mean, σ is the standard deviation, and if |Z i > 3 |, it is considered as an outlier.

[0058] 3) Time alignment: interpolation method is used to align the time series of different sensors.

[0059] A heterogeneous dynamic graph structure is constructed, where nodes represent battery states at different time steps, and edges represent time evolution relationships. A dynamic adjacency matrix strategy is adopted, which makes the adjacency matrix dynamically change with time steps to adapt to the aging process and nonlinear decay trend of the battery.

[0060] A sliding window strategy is adopted, only retaining information of the last N time steps, while dynamically adjusting the size of the time window in combination with the running period of the battery, to reduce computational complexity and improve the degree of attention to the latest battery state information.

[0061] At each time step t, a dynamic graph G t =(V t , E t ) is constructed, where:

[0062] Node V t represents the state of the lithium battery at time t; edge E t represents the time evolution relationship; let the sliding window size be N, then the dynamic graph node set is:

[0063] V t ={v t-N , v t-N+1 ,..., v t}

[0064] where v t represents the battery state vector at time t.

[0065] The initial edge weight is calculated based on the cosine similarity:

[0066]

[0067] where is the basic adjacency matrix at time t, v i , v j are the battery state vectors at time steps i and j.

[0068] Next, the reinforcement learning state representation is defined.

[0069] First, the reinforcement learning state vector is defined:

[0070]

[0071] where v t is the current battery state vector, is the average of the prediction error in the last K steps, and φ(·) is the adjacency matrix feature extraction function (returns the statistics of sparsity / maximal weight).

[0072] Second, the reinforcement learning action output is calculated, which adopts the policy network π θ outputs the update parameters:

[0073]

[0074] where λ t ∈(0, 1) is a time-varying decay factor, γ t > 0 is a global weight scaling factor, τ t ≥ 0 is a dynamic sparsification threshold.

[0075] After outputting the action at the current time step, the adjacency matrix needs to be dynamically updated:

[0076]

[0077] where d(i, j) is the time distance metric, is the time decay function, γ t is the global weight regulator, τ t is the sparsification threshold.

[0078] The reward function is generated by combining the accuracy improvement term and the sparsity penalty term:

[0079]

[0080] where η, β are hyperparameters, ||A(t)||0 is the number of non-zero elements of the adjacency matrix, e t-1 is the historical average prediction error, y t is the true value at time step t, is the predicted value made by the model at time step t.

[0081] Finally, the network parameters are updated by the policy gradient method:

[0082]

[0083] where Q(·) is the state-action value function, is the gradient of the objective function J(θ) with respect to the policy network parameters θ.

[0084] The dynamic graph is constructed and the dynamic graph convolutional neural network is used for time series feature extraction. The node information of adjacent time steps is aggregated using GCN-based convolution operation, and its calculation process can be represented as:

[0085]

[0086] where H (l) represents the features of the l-th layer, A is the adjacency matrix, D is the degree matrix, W (l) is the weight matrix, and σ is the nonlinear activation function. The weighted sum of neighbor node information is calculated, and different weights are given to different neighbor nodes through the attention mechanism.

[0087] The dynamic graph convolution structure of the dual-path GCN is adopted to capture the short-term H local and long-term H global battery state change characteristics:

[0088] H global = GCN global (A, X), H local = GCN local (A, X)

[0089] Final feature fusion:

[0090] H = aH global + (1-a)H local

[0091] Wherein, a is a learnable weight, A is an adjacency matrix.

[0092] The fusion of different modal data features is realized by introducing a self-attention mechanism.

[0093] The multi-modal feature set is: F m = {F V , F I , F R , F SOC , F SOH}

[0094] Wherein, F V is the voltage feature, F I is the current feature, F R is the internal resistance feature, F SOC is the state of charge feature, and F SOH is the health state feature.

[0095] The flow of different modal features is controlled by using a gating mechanism, and the gating value is:

[0096] g m = sigmoid(W g ·F m +b g )

[0097] Wherein, W g is the gating weight matrix, and b g is the gating bias term.

[0098] The fused multi-modal feature is:

[0099] F m ′=∑g m ·F m

[0100] Further, the relevance between modalities is calculated through the Query-Key-Value mechanism. An adaptive weighted fusion method is adopted to dynamically adjust the contribution of each modality according to the current state of the battery, and a multi-head attention mechanism is used to weight and sum the features of different modalities to obtain the inter-modal attention weight matrix A:

[0101]

[0102] H fusion =AV

[0103] where Q, K, V are query, key, value matrices, and F m ′ is the assignment; d is the feature dimension, H fusion is the fused multi-modal feature representation.

[0104] The TCN-Attention model is used for lithium battery life prediction. The model uses a causal convolution structure to ensure that the prediction at the current time only depends on past information, and the output features of a specific layer are obtained:

[0105]

[0106] where * represents convolution operation, l is the network layer index (l = 0, 1,..., L-1), H (l) is the input feature of the l-th layer, W (l) is the causal convolution kernel, α(·) is the activation function, is the output feature of the l-th layer.

[0107] The multi-head attention mechanism is used to learn features of different time scales:

[0108] H final =MultiHead(H TCN , H fusion )

[0109] where H TCN is the time series feature matrix extracted by the TCN module, and H fusion is the multi-modal fusion matrix.

[0110] Finally, the prediction results are smoothed through a sliding window to achieve life prediction:

[0111] RUL_pred(t)=TCN_Attention(H final (t))

[0112] where H final (t) is the feature vector at time t, TCN_Attention(·) is the prediction network, and RUL_pred(t) is the remaining useful life prediction value at time t.

[0113] Multi-objective optimization strategy is used for training and optimization of the model. Various loss function combinations including MSE, MSLE and Huber loss are used:

[0114]

[0115] Wherein λ1, λ2, λ3∈[0, 1] are loss weight coefficients (satisfying ∑λ i =1).

[0116] Adam is used to accelerate convergence in the early stage, and RMSProp is switched to improve accuracy in the later stage:

[0117]

[0118] Wherein η is the learning rate, is the current gradient, θ is all weight parameters of the model, and θ' is the updated weight parameter.

[0119] The Monte Carlo Dropout method is used for uncertainty quantification, and the mean and variance are taken T times of forward propagation:

[0120]

[0121]

[0122] Wherein RULpred t represents the tth time of forward propagation sampling of Monte Carlo dropout.

[0123] Data augmentation (time series disturbance, local noise addition), cross-validation, independent test set verification and other methods are used to comprehensively evaluate the prediction accuracy and generalization ability of the model.

[0124] Corresponding to the foregoing embodiment of the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion, the present application also provides an embodiment of a lithium battery life prediction device based on dynamic graph convolution and multi-modal fusion.

[0125] Referring to Figure 3 , the lithium battery life prediction device based on dynamic graph convolution and multi-modal fusion provided by the embodiment of the present application comprises a memory and one or more processors, the memory stores executable code, and the processor executes the executable code to implement the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion in the foregoing embodiment.

[0126] The embodiment of the lithium battery life prediction device based on dynamic graph convolution and multi-modal fusion provided by the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be realized by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory for running by the processor of the device with data processing capability. From the hardware level, as shown in Figure 3 The embodiment of the lithium battery life prediction device based on dynamic graph convolution and multi-modal fusion provided by the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be realized by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory for running by the processor of the device with data processing capability. From the hardware level, as shown in Figure 3 In addition to the processor, the memory, the network interface, and the non-volatile memory shown in the figure, the device with data processing capability in the embodiment can also include other hardware according to the actual function of the device with data processing capability, and details are not described here.

[0127] The implementation process of the functions and roles of each unit in the above device is specifically described in the implementation process of the corresponding steps in the above method, and details are not described here.

[0128] For the device embodiment, since it basically corresponds to the method embodiment, the related part is described in the part of the method embodiment. The device embodiment described above is only schematic, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, that is, they can be located in one place, or distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purpose of the application scheme. Those skilled in the art can understand and implement without creative labor.

[0129] The embodiment of the application further provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the lithium battery life prediction method based on dynamic graph convolution and multi-modal fusion in the above embodiment.

[0130] The computer readable storage medium can be an internal storage unit of any of the aforementioned devices with data processing capability, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device of any of the aforementioned devices with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both an internal storage unit and an external storage device of any of the aforementioned devices with data processing capability. The computer readable storage medium is used to store the computer program and other programs and data required by the aforementioned devices with data processing capability, and can also be used to temporarily store data that has been output or is to be output.

[0131] The application further provides a computer program product comprising a computer program, which, when executed by a processor, implements the method for predicting lithium battery life based on dynamic graph convolution and multi-modal fusion.

[0132] Although the embodiments of the present application are described in conjunction with the drawings, various modifications and changes can be made by those skilled in the art without departing from the spirit and scope of the present application, and such modifications and changes are intended to fall within the scope of the appended claims.

Claims

1. A lithium battery life prediction method based on dynamic graph convolution and multimodal fusion, characterized in that: The following steps are involved: (1) Collect battery status data of different modes during the operation of lithium batteries, including voltage, current, temperature, internal resistance, and number of charge and discharge cycles, and pre-process the battery status data; (2) Construct a heterogeneous dynamic graph structure based on battery operation characteristics, where nodes represent battery states at different time steps, edges represent time evolution relationships, and weights are assigned to obtain heterogeneous graph feature vectors; (3) Using dynamic graph convolutional neural networks to extract features from heterogeneous graph data, capture the temporal dynamic changes of lithium battery status, and obtain temporal feature vectors; (4) Combining the attention mechanism and feature weighting strategy to process the temporal feature vector, the data of different modalities are fused to obtain multimodal temporal features; (5) Based on the extracted multimodal time series features, the TCN-Attention model is used to predict the remaining service life of the lithium battery and obtain the prediction results.

2. The lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to claim 1 is characterized in that: The dynamic graph modeling adopts a strategy based on a dynamic adjacency matrix, which changes with time steps to adapt to the battery aging process, including the following steps: 1) Based on the historical operating data of the lithium battery, an initial adjacency matrix is ​​constructed to describe the correlation between the battery states at different time steps; 2) Cosine similarity is used between nodes to represent the correlation of states; 3) Adopting an adaptive graph update mechanism based on reinforcement learning, the graph update process is regarded as a decision problem: Define the state (State): the current battery state feature vector: voltage, current, temperature and SOH estimation value; recent prediction error statistics: MAE and RMSE; the sparsity index of the current adjacency matrix; Define the action: Adjustable parameters: the amount of adjustment of the attenuation factor, the redistribution coefficient of the weights of different time steps in the sliding window, and the connection threshold adjustment; Define reward: Add regularization based on the improvement of prediction accuracy in subsequent time steps; 4) Based on the current prediction error and historical update results, the reinforcement learning algorithm DQN is used to actively learn and adjust the adjacency matrix, including attenuation strength, weight distribution within the window, and node connection relationships. After each prediction time step, or when the prediction error exceeds a threshold, an update decision is triggered to minimize future prediction errors and dynamically learn the optimal strategy; 5) Using a sliding window strategy, only the information of the latest N time steps is retained, reducing computational complexity while increasing attention to the latest battery status; 6) Dynamically adjust the size of the time window based on the battery's operating cycle.

3. The lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to claim 1 is characterized in that: The dynamic graph convolution calculation adopts an adaptive graph convolution mechanism to dynamically adjust the graph structure, including: A GCN-based convolution operation is used to aggregate node information of adjacent time steps, calculate the weighted sum of neighbor node information, use the attention mechanism to assign different weights to different neighbor nodes, and adopt dual-channel dynamic graph convolution to learn battery state changes at different time scales.

4. The lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to claim 1, characterized in that: The multimodal fusion module includes a self-attention mechanism that can assign weights based on the importance of the data modality, including: An adaptive weighted fusion method is adopted to dynamically adjust the contribution of the modality according to the current battery status. A multi-head attention mechanism is used to perform weighted summation of the features of different modalities. The correlation between modalities is calculated through the query, key, and value mechanism, and a gating mechanism is used to control the flow of different modal features.

5. The lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to claim 1 is characterized in that: The life prediction module uses a temporal convolution model based on the self-attention mechanism (TCN-Attention model) to learn the lithium battery degradation trend, including: A causal convolution structure is adopted to make the current prediction depend only on past information. A void convolution mechanism is adopted to expand the time window, combined with multi-head attention, to learn features of different time scales, and a sliding window is used to smooth the prediction results.

6. The lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to claim 1, characterized in that: The model optimization module adopts a multi-objective optimization strategy, including prediction error minimization and uncertainty quantification, including: The MSE, MSLE, and Huber loss functions are used, combined with the Adam and RMSProp optimization algorithms, and Monte Carlo Dropout is used for multiple forward propagations. Time series perturbation and local noise are used for data enhancement. Cross-validation and independent test set evaluation are used.

7. A lithium battery life prediction system based on dynamic graph convolution and multimodal fusion that implements the method according to any one of claims 1 to 6, characterized in that: The system includes: Data acquisition module: used to collect battery status data of different modes during the operation of lithium batteries, including voltage, current, temperature, internal resistance, and number of charge and discharge cycles, and pre-process the battery status data Dynamic graph construction module: used to construct a heterogeneous dynamic graph structure based on battery operation characteristics, where nodes represent battery states at different time steps, edges represent time evolution relationships, and weights are assigned to obtain heterogeneous graph feature vectors; Dynamic graph convolution calculation module: used to extract features from heterogeneous graph data using a dynamic graph convolutional neural network, capture the temporal dynamic changes of lithium battery status, and obtain temporal feature vectors; Multimodal fusion module: It is used to process the temporal feature vector by combining the attention mechanism and feature weighting strategy, fuse the data of different modalities, and obtain multimodal temporal features; Life prediction module: It is used to predict the remaining service life of lithium batteries based on the extracted multimodal time series features and use the TCN-Attention model to obtain prediction results.

8. A lithium battery life prediction device based on dynamic graph convolution and multimodal fusion, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the processor executes the executable code, it implements a lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to any one of claims 1 to 6.

9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, a lithium battery life prediction method based on dynamic graph convolution and multimodal fusion according to any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements a lithium battery life prediction method based on dynamic graph convolution and multimodal fusion as described in any one of claims 1 to 6.

Images