Battery lithium precipitation risk prediction method and system, and storage medium

Abstract

The application relates to a battery lithium precipitation risk prediction method, a system and a storage medium, the method comprising: obtaining multiple types of operation data generated during battery operation and integrating the operation data to obtain battery multi-modal data; taking the battery multi-modal data as nodes, constructing a causal driving coupling dynamic graph for representing the interaction relationship of multiple physical field risks during battery operation from three modalities of electrochemical, thermal and mechanical modalities; based on a preset cross-modal feature extraction network, jointly modeling the causal driving coupling dynamic graph to obtain cross-modal causal correlation information; based on a preset risk prediction network, performing spatio-temporal joint analysis on the causal driving coupling dynamic graph and the cross-modal causal correlation information to obtain a risk prediction result. The application structurally associates multi-modal battery operation data with a lithium precipitation risk evolution process, realizes the depiction of a lithium precipitation risk formation path and a propagation process, and improves the accuracy of battery lithium precipitation risk prediction.

Description

Technical Field

[0001] This invention relates to the field of electrochemical energy storage safety technology, and in particular to a method, system and storage medium for predicting lithium plating risk in batteries. Background Technology

[0002] In the field of battery technology, lithium plating has become a key research focus due to its significant impact on battery performance and safety. Lithium plating not only leads to rapid capacity decay and a substantial reduction in battery lifespan, but it can also trigger internal short circuits, potentially causing serious safety issues such as thermal runaway. This severely limits the application of batteries under complex conditions such as high power and low temperatures. Therefore, accurately and efficiently predicting battery lithium plating is crucial for improving battery performance and ensuring safe battery operation.

[0003] Currently, lithium plating risk prediction technologies mainly cover three categories: single-mode detection, multi-mode fusion, and data-driven approaches.

[0004] Among single-mode detection methods, electrochemical impedance spectroscopy (EIS) can identify lithium plating by deeply analyzing changes in the real part of impedance in the high-frequency region. However, this method only reflects the electrochemical characteristics of the battery interface and cannot establish an effective correlation with the risk of battery thermal runaway, thus having significant shortcomings in comprehensively assessing battery safety. Infrared thermal imaging technology, with its keen ability to capture local temperature rises in the battery, can monitor possible signs of lithium plating to some extent. However, due to its limited detection accuracy, it is difficult to accurately locate the electrode micro-regions where lithium plating begins, making accurate detection of early-stage lithium plating challenging. Acoustic emission detection relies on the mechanical stress signal generated inside the battery due to lithium plating for judgment; however, this method is highly susceptible to interference from environmental noise, leading to a significant reduction in the accuracy of the detection results and poor performance in real-world complex environments.

[0005] Early multimodal fusion methods primarily focused on static feature-level fusion. A typical approach involved directly concatenating multiple parameters such as voltage, temperature, and impedance, and then inputting them into models like Support Vector Machines (SVM) or Random Forests for analysis. While this method improved the accuracy of lithium plating detection to some extent, it lacked a deep understanding of the cross-modal physical mechanisms involved. Consequently, it was poorly adaptable to complex operating conditions, such as lithium plating caused by electro-thermal coupling within the battery during low-temperature fast charging, and could not accurately and effectively predict and analyze these phenomena. Regarding physical field coupling modeling, some studies attempted to construct multiphysics models by simultaneously solving electrochemical and thermal conduction equations, and then used preset thresholds to determine lithium plating status. However, these models have significant drawbacks: their parameters are fixed, making it impossible to update changes in the battery electrode structure during use, such as the decrease in porosity due to SEI film thickening. Furthermore, they did not fully integrate real-time monitoring data such as infrared thermal imaging and acoustic emission, resulting in an incomplete and inaccurate reflection of the battery's actual state.

[0006] In data-driven approaches, deep learning models (such as LSTM and CNN) are widely used to mine features from battery time-series data to detect lithium plating. However, this method relies solely on electrical signals, completely ignoring the spatial inconsistencies between battery cells, thus limiting the applicability of the detection results at the battery pack level. While graph neural networks (GNNs) can model the spatial relationships within battery packs, current applications are mostly limited to cell state estimation and have not yet been optimized for the complex dynamic coupling of multiple physics fields during lithium plating, making it difficult to accurately describe and predict the occurrence and development of lithium plating in battery packs.

[0007] In summary, existing lithium plating risk prediction technologies have certain limitations in terms of accuracy, comprehensiveness, and adaptability to complex operating conditions. There is an urgent need to develop a more advanced, comprehensive, and adaptable battery lithium plating risk prediction technology to meet the growing demands for battery performance and safety. Summary of the Invention

[0008] This invention aims to address the problems in existing battery lithium plating risk prediction methods, such as insufficient utilization of multimodal information, inadequate characterization of coupling relationships between physical fields, and difficulty in depicting the risk evolution process.

[0009] To address the aforementioned technical problems, in a first aspect, the present invention provides a method for predicting the risk of lithium plating in batteries, comprising the following steps:

[0010] S101. Acquire various types of operating data generated during battery operation, and standardize and integrate the different operating data to obtain battery multimodal data; S102. Using the battery multimodal data as nodes, and based on the physical field risk transmission mechanism between electrochemical, thermal, and mechanical modes, a causal-driven coupled dynamic graph is constructed to characterize the directed causal transmission relationship of multiphysical field risks during battery operation. The edges in the causal-driven coupled dynamic graph are directed edges, and the direction of the directed edges is determined based on the physical causal transmission direction between electrochemical, thermal, and mechanical modes. The weight of the directed edges is the risk transmission weight. Specifically, the risk transmission weight between nodes under different modes is calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weight. S103. Through a pre-defined cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, the causal-driven coupling dynamic graph is jointly modeled to extract cross-modal causal association information used to characterize the lithium plating risk triggering and amplification relationship between different physical field states. S104. Using a pre-defined risk prediction network constructed based on a spatiotemporal graph convolutional network, a spatiotemporal joint analysis is performed on the causal driving coupled dynamic graph and the cross-modal causal correlation information to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

[0011] Furthermore, step S101, the step of standardizing and integrating the different operational data, includes: Use spatiotemporal alignment technology to unify the data time base for different operational data; At least one of sliding window, wavelet denoising, and ROI segmentation is used to extract key features and denoise the running data. The Z-score normalization method was used to scale the different running data.

[0012] Furthermore, in step S102, the step of adaptively and dynamically adjusting the topology of the causal-driven coupled dynamic graph according to the risk transmission weight specifically includes: During battery operation, at each preset dynamic update time, the topology between different nodes in the causal-driven coupled dynamic graph is updated based on the risk transmission weight to reflect the evolution and propagation process of lithium plating risk in the battery across multiple physical fields.

[0013] Furthermore, in step S103, the preset cross-modal feature extraction network is implemented based on a cross-modal graph attention network, wherein: From the same modal perspective, the cross-modal graph attention network extracts the time-dependent features of the electrochemical mode through the sequence feature modeling unit, the spatial distribution features of the thermal mode through the spatial feature modeling unit, and the structural features of the mechanical mode through the frequency domain feature modeling unit. From a cross-modal perspective, the cross-modal graph attention network, based on the neighborhood node structure in the causal-driven coupled dynamic graph, performs weighted fusion of different modal features to obtain cross-modal causal association information for characterizing the triggering and amplification relationship of lithium plating risk.

[0014] Furthermore, in step S104, the preset risk prediction network is implemented based on a spatiotemporal graph convolutional network. The preset risk prediction network is used to jointly model the cumulative changes of battery lithium plating risk in the time dimension and the propagation process in the multi-physics topology, and output the battery lithium plating risk prediction result.

[0015] Furthermore, step S104 also includes: Based on the battery risk prediction results, the battery operation process is slowed down or terminated according to a preset risk probability threshold.

[0016] Secondly, the present invention also provides a battery lithium plating risk prediction system, comprising: The data acquisition module is used to acquire various types of operational data generated during battery operation, and to standardize and integrate the different operational data to obtain battery multimodal data. The modal characterization module is used to construct a causal-driven coupled dynamic graph that characterizes the directed causal transmission relationship of multi-physics risks during battery operation, based on the physical field risk transmission mechanism between electrochemical, thermal, and mechanical modes, using the battery multimodal data as nodes. The edges in the causal-driven coupled dynamic graph are directed edges, the direction of which is determined based on the physical causal transmission direction between the electrochemical, thermal, and mechanical modes, and the weight of the directed edge is the risk transmission weight. Specifically, the risk transmission weights between nodes under different modes are calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weights. The feature extraction module is used to jointly model the causal-driven coupled dynamic graph through a preset cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, and extract cross-modal causal association information to characterize the lithium plating risk triggering and amplification relationship between different physical field states; The risk prediction module is used to perform spatiotemporal joint analysis on the causal driving coupled dynamic graph and the cross-modal causal correlation information through a preset risk prediction network constructed based on a spatiotemporal graph convolutional network, so as to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

[0017] Thirdly, the present invention also provides a computer device, including: a memory, a processor, and a battery lithium plating risk prediction program stored in the memory and executable on the processor, wherein when the processor executes the battery lithium plating risk prediction program, it implements the steps in the battery lithium plating risk prediction method as described in any of the above embodiments.

[0018] Fourthly, the present invention also provides a storage medium storing a battery lithium plating risk prediction program, wherein when the battery lithium plating risk prediction program is executed by a processor, it implements the steps in the battery lithium plating risk prediction method as described in any of the above embodiments.

[0019] The beneficial effects achieved by this invention lie in proposing a battery lithium plating risk prediction method based on causal driving and multi-physics field coupled dynamic graphs. The beneficial effects achieved by this method include: I. Overcoming the limitations of a single data source and the lag of static models: By constructing and dynamically adjusting the dynamic graph of the coupling of the electro-thermal-mechanical three fields, the complex multi-physics interactions inside the battery and the evolution of lithium plating risk are structurally correlated in time and space, enabling the model to capture the dynamic changes of the internal structure and physical field relationships of the battery in real time, and significantly improving the predictive adaptability under complex working conditions. Second, it has achieved a qualitative leap from result prediction to path characterization: by mining the causal triggering and amplification relationship between physical fields through cross-modal feature extraction networks, and combining it with spatiotemporal graph convolutional networks for joint analysis, the output is not only a single risk probability, but also the evolution trend and propagation path of lithium plating risk, providing battery management systems with better decision information depth. Third, it generates synergistic effects: dynamic graph construction, cross-modal causal reasoning, and spatiotemporal joint analysis are deeply coupled. The dynamic graph provides a dynamic and realistic physical topological foundation for cross-modal reasoning, the causal reasoning results provide physically meaningful features for spatiotemporal convolution, and the spatiotemporal analysis, in turn, guides the update direction of the dynamic graph. This structured collaborative working mechanism effectively improves the accuracy and robustness of the entire prediction system. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the steps of the battery lithium plating risk prediction method provided in this embodiment of the invention. Figure 2This is an example diagram of the causal-driven coupling dynamic diagram in the battery lithium plating risk prediction method provided in this embodiment of the invention; Figure 3 This is a schematic diagram comparing the battery lithium plating risk prediction method proposed in this embodiment of the invention with existing methods through ablation experiments; Figure 4 This is a schematic diagram of the battery lithium plating risk prediction system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of the battery lithium plating risk prediction method provided in this embodiment of the invention. The battery lithium plating risk prediction method includes the following steps: S101. Acquire various types of operating data generated during battery operation, and standardize and integrate the different operating data to obtain battery multimodal data.

[0023] Since lithium plating in lithium-ion batteries is affected by a variety of factors, a single data source cannot fully present its complex process. Therefore, in this embodiment of the invention, multi-modal acquisition methods such as electrochemical impedance spectroscopy, infrared thermal imaging, acoustic emission signals, and historical cycle data are integrated to comprehensively acquire data on the battery's operating status in multiple dimensions, including electrochemical reactions, thermal changes, and mechanical changes. This breaks through the limitations of traditional single data sources and enables comprehensive capture of information related to lithium plating risk.

[0024] For example, for operational data related to electrochemical reactions during battery operation, the acquisition module collects data on the individual cell voltage at a frequency of 10Hz. Real-time data acquisition is employed to sensitively capture subtle voltage fluctuations; using an electrochemical workstation, offline electrochemical impedance spectroscopy (EIS) is performed weekly, with the frequency range limited to 10 mHz-10 kHz, to accurately obtain charge transfer resistance. Key electrochemical parameters, etc.

[0025] For operational data related to thermal changes, an industrial-grade infrared thermal imager was used to capture images of the cell surface temperature distribution at a frame rate of 5Hz. Where x represents the horizontal pixel coordinate in the image, corresponding to a position on the surface of the battery cell, y represents the vertical pixel coordinate in the image, and t represents time; simultaneously, thermocouples are strategically placed at the battery cell tabs and center position to collect the single-point temperature of the j-th thermocouple at time t in real time. Used to calculate local temperature gradient ,in, The temperature is calculated based on the average temperature of all thermocouples in the battery pack, thus accurately locating potential temperature anomalies inside the battery cells.

[0026] For operational data related to mechanical changes, stress release signals induced by dendrite growth are continuously monitored using acoustic emission sensors operating at frequencies between 100kHz and 1MHz. Electrode thickness was obtained through ultrasonic scanning. With interface roughness These parameters are crucial for assessing electrode structural integrity and predicting lithium plating risk. Furthermore, the capacity of each cycle is continuously recorded from the time the battery is put into use. With internal resistance The rate of increase in internal resistance is calculated using the following formula: ; This provides a crucial basis for assessing battery aging and lithium plating risks.

[0027] Specifically, the collected operational data needs to be preprocessed. Step S101, which involves standardizing and integrating the different operational data, includes: Use spatiotemporal alignment technology to unify the data time base for different operational data; At least one of sliding window, wavelet denoising, and ROI segmentation is used to extract key features and denoise the running data. The Z-score normalization method was used to scale the different running data.

[0028] For example, based on spatiotemporal alignment technology, a hardware clock synchronization device is used to ensure that all sensors that are collecting operational data operate on the same time reference. At the same time, a cubic spline interpolation algorithm is used to synchronize the time of data with different sampling frequencies, and the timestamp error of each modality data is strictly controlled within 50ms.

[0029] For time-series data such as voltage and current, a sliding window of 30 seconds is used for segmentation to generate time-series samples. Where 300 represents the number of sampling points taken at a sampling frequency of 10Hz within 30 seconds, and N represents the number of samples. The standard deviation of voltage within the window is calculated as follows: ; in The average voltage within this window is recorded to reflect voltage fluctuations; the peak current is also recorded. This highlights abnormal changes in the current.

[0030] For the infrared thermal image, an image segmentation algorithm is used to delineate the battery cell area as the region of interest (ROI). By setting an appropriate temperature threshold, the area ratio of the high-temperature anomaly region is calculated using a formula: ; This allows for precise location of potential lithium plating areas.

[0031] For acoustic emission signals, wavelet denoising algorithms are used to filter out environmental noise interference, and the number of valid events in the acoustic emission signal per unit time is counted. The signal's main frequency is obtained through Fourier transform. .

[0032] Finally, all preprocessed running data are normalized using the Z-score normalization method: ; in and These are the mean and standard deviation of the feature within the training set, respectively. Using the voltage standard deviation as an example... For example, it is necessary to calculate all the training sets. mean of values with standard deviation And thus the current The value is converted to: .

[0033] S102. Using the battery multimodal data as nodes, and based on the physical field risk transmission mechanism between electrochemical, thermal, and mechanical modes, a causal-driven coupled dynamic graph is constructed to characterize the directed causal transmission relationship of multiphysical field risks during battery operation. The edges in the causal-driven coupled dynamic graph are directed edges, and the direction of the directed edges is determined based on the physical causal transmission direction between electrochemical, thermal, and mechanical modes. The weight of the directed edges is the risk transmission weight. Specifically, the risk transmission weight between nodes under different modes is calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weight.

[0034] By constructing a coupled dynamic graph model of electro-thermal-mechanical systems, and using nodes to represent the multi-physics state of battery cells, the limitations of traditional static models in tracking the internal physical field relationships and structural changes of batteries in real time can be overcome, providing dynamic and realistic model support for risk prediction. In this embodiment of the invention, the coupled dynamic graph is a graphical model constructed based on the coupling characteristics of electro-thermal-mechanical multi-physics fields. Nodes represent key state parameters of the battery cell, encompassing multi-dimensional information such as electrochemical (e.g., negative electrode overpotential, charge transfer resistance), thermal (e.g., temperature, temperature gradient), and mechanical (e.g., acoustic emission energy, electrode roughness), comprehensively reflecting the internal physical field state of the battery. Edges are used to represent the directed physical transfer relationships between different nodes (i.e., different physical field parameters or different battery cells). For example, if the overpotential change of a battery cell is related to the temperature rise of adjacent battery cells, this interaction will be represented by edges.

[0035] It should be noted that the physical prior basis of the "causal drive" in the embodiments of the present invention is as follows: the abnormal increase in the negative electrode overpotential directly drives the reduction and deposition of lithium ions (electrochemical cause); the lithium dendrites generated by the deposition pierce the SEI film, triggering a micro-short circuit and releasing Joule heat (electrothermal effect); and the dendrite growth generates compressive stress on the electrode particles and releases it through acoustic emission signals (electromechanical effect). The above physical causal chain constitutes the prior basis for the setting of the directed edge direction, making the coupled dynamic graph different from undirected graphs or static graphs constructed solely based on statistical correlation of data.

[0036] Step S102, the step of adaptively and dynamically adjusting the topology of the causal-driven coupling dynamic graph according to the risk transmission weight, specifically includes: During battery operation, at each preset dynamic update time, the topology between different nodes in the causal-driven coupled dynamic graph is updated based on the risk transmission weight to reflect the evolution and propagation process of lithium plating risk in the battery across multiple physical fields.

[0037] like Figure 2The causal-driven coupling dynamic graph example shown in this embodiment of the invention uses key state parameters from electrochemical, thermal, and mechanical modes as nodes. The directed edges between nodes and their weights (i.e., risk transmission weights) characterize the risk transmission relationships between different physical fields. The risk transmission weights are dynamically calculated based on three factors: electrochemical correlations (e.g., the effect of charge transfer on voltage), thermal correlations (e.g., the effect of temperature gradients on heat conduction), and mechanical correlations (e.g., the effect of stress distribution on structural changes). The causal-driven coupling dynamic graph can adaptively adjust itself dynamically as the battery's operating state changes. For example, when the overpotential of a cell increases sharply, the edge weight between its temperature node and that of adjacent cells increases, reflecting the significant impact of the overpotential increase on local temperature rise; conversely, if the correlation is weak, the edge weight decreases. By dynamically adjusting the risk transmission weights, the model can accurately quantify the interaction strength between different physical fields and different cells, providing a quantitative basis for subsequent cross-modal feature extraction and causal inference, making the fusion of multimodal data more physically meaningful. By setting a preset dynamic update time, the structure of the coupled dynamic graph is updated in real time to reflect the dynamic changes in processes such as electrochemical reaction, heat conduction, and mechanical stress propagation, so that the model always fits the actual state of the battery.

[0038] Specifically, in this embodiment of the invention, the constructed electro-thermal-mechanical coupling dynamic diagram is defined as follows: Each node Eigenvectors related to time t Integrating key parameters of multiphysics and coupling dynamic graphs Specifically, it is expressed as follows: ; in, Represents a node The electrical conductivity or related electrical efficiency index at the point. Represents a node Charge transfer resistance at the point, Represents a node The temperature at that location Represents a node Temperature gradient at that location, Represents a node Acoustic emission energy at the location. Represents a node Contact resistance at the point, Represents a node The battery capacity at the nth cycle. Represents a node The change in resistance during the nth cycle.

[0039] The negative electrode overpotential is calculated in the following way: ; This represents the equilibrium potential for lithium deposition.

[0040] Risk transmission weight Dynamic calculations are performed from three modal perspectives. Regarding electrochemical correlation, the electrochemical risk transfer weight between node i and node j at time t is calculated. The calculation method is as follows: ; Where, η i η j Let represent the electrical state indices of nodes i and j at time t, respectively. It is a constant; Regarding thermal correlation, at time t, the thermal risk transmission weight between node i and node j The calculation method is as follows: ; in, This represents the thermodynamic distance between node i and node j. This represents the distance between node i and node j; Regarding mechanical correlation, at time t, the mechanical risk transmission weight between node i and node j The calculation method is as follows: ; in, Let i and j represent the acoustic emission feature vectors of node i and node j, respectively.

[0041] The final risk transmission weight is calculated as follows: .

[0042] Compared to existing techniques (such as constructing a multiphysics model by simultaneously combining electrochemical and thermal conduction equations as mentioned in the background), step S102 is not simply multi-field coupling. The difference lies in: The parameters and topology of existing physics models are usually preset and fixed. In this embodiment of the invention, the risk transmission weights between nodes are calculated in real time based on changes in different modal states, and this drives the dynamic update of the graph topology. This closed-loop "calculation-update" mechanism enables the model to reflect long-term dynamic processes such as battery aging and electrode structure evolution; The transmission of lithium plating risk among electrical, thermal, and mechanical fields is highly dynamic and nonlinear. This invention provides a method for quantifying this complex coupling relationship by combining feature engineering based on the intrinsic mechanism of the physical field with dynamic calculation rules, rather than simply using parameter weighting, through a risk transmission weighting method that utilizes electrochemical, thermal, and mechanical factors.

[0043] S103. Using a pre-defined cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, the causal-driven coupled dynamic graph is jointly modeled to extract cross-modal causal association information used to characterize the triggering and amplification relationship of lithium plating risk between different physical field states.

[0044] In step S103, the preset cross-modal feature extraction network is implemented based on a cross-modal graph attention network, wherein: From the same modal perspective, the cross-modal graph attention network extracts the time-dependent features of the electrochemical mode through the sequence feature modeling unit, the spatial distribution features of the thermal mode through the spatial feature modeling unit, and the structural features of the mechanical mode through the frequency domain feature modeling unit. From a cross-modal perspective, the cross-modal graph attention network, based on the neighborhood node structure in the causal-driven coupled dynamic graph, performs weighted fusion of different modal features to obtain cross-modal causal association information for characterizing the triggering and amplification relationship of lithium plating risk.

[0045] In this embodiment of the invention, for the time-dependent features of electrochemical modes, the input hidden layer of the coupled dynamic graph is set as a bidirectional long short-term memory network (LSTM) with 128 neurons. This network effectively extracts time-dependent features by learning the changes in parameters such as overpotential at different time steps. It accurately captures the dynamic change trend of overpotential over time.

[0046] To assess the spatial distribution characteristics of thermal modes, a CNN network with three convolutional layers and two pooling layers was constructed, with kernel sizes set to 3×3 and 5×5, respectively. Through these convolutional and pooling operations, temperature distribution feature maps were gradually extracted. It clearly shows the distribution of abnormal temperature areas on the surface of the battery cell.

[0047] For the frequency domain structural features of mechanical modes, a one-dimensional convolutional neural network (CNN) with a kernel size of 5 is used to extract the frequency domain and structural features. It effectively captures the characteristic information of mechanical stress and electrode deformation.

[0048] From a cross-modal perspective, eight attention heads are set up, and the attention weights are calculated in the following way: ; Where v is the input feature vector, The trainingable weight matrix is ​​used to concatenate the outputs of each attention head to obtain the node features. This enables in-depth mining of causal relationship information between different physical field nodes.

[0049] Existing multimodal fusion technologies often involve simple feature-level concatenation, leading to mutual interference between features from different physical fields and a lack of physical interpretability. In this embodiment of the invention, the cross-modal graph attention network proposed in step S103 has a unique two-stage architecture: First, intra-modal decoupling. Firstly, different feature extraction units are used for the three modalities—electrochemical, thermal, and mechanical—based on their respective physical nature (LSTM for time series, CNN for spatial distribution, and 1D-CNN for frequency domain structure). This targeted design ensures that the features extracted from the original data of each modality are pure and representative, avoiding information loss caused by mixed extraction.

[0050] Second, intermodal causal weighted fusion. Utilizing a graph attention mechanism, based on the neighborhood structure of a dynamic graph, features from different modalities are selectively weighted and fused. The calculation of the attention weights, Attention(Q, K, V), is directly related to the strength of the causal relationship between nodes. This enables the model to learn "under what operating conditions, which physical field features play a dominant role in triggering lithium plating risk," for example, the weight of the thermal mode increases at low temperatures, and the weight of the electrochemical mode increases during fast charging.

[0051] The aforementioned two-stage architecture of "decoupling first, then causal weighting" is specifically designed for the causal-driven coupled dynamic graph constructed in step S102. This is not a simple patchwork of graph networks and multimodal integration, but a structured collaborative design that generates "interpretable cross-modal causal association information".

[0052] S104. Using a pre-defined risk prediction network constructed based on a spatiotemporal graph convolutional network, a spatiotemporal joint analysis is performed on the causal driving coupled dynamic graph and the cross-modal causal correlation information to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

[0053] In step S104, the preset risk prediction network is implemented based on the spatiotemporal graph convolutional network ST-GCN. The spatiotemporal graph convolutional network ST-GCN is used to jointly model the cumulative changes of battery lithium plating risk in the time dimension and the propagation process in the multi-physics topology, and output the battery lithium plating risk prediction result.

[0054] The Spatiotemporal Graph Convolutional Network (ST-GCN) is crucial in risk prediction. In this embodiment of the invention, a one-dimensional convolutional kernel of size 3 is used to extract the feature change rate of the first three time steps in the time dimension. Taking the overpotential rise rate as an example, the calculation method is as follows: ; This method can capture the trend of overpotential changes over time.

[0055] In the spatial dimension, based on the Graph Laplacian matrix ( For degree matrix, Performing graph convolution on the edge weight matrix and normalizing the result, we get: ; Multi-layer graph convolution is performed using the following method: ; Activation function Use ReLU.

[0056] Finally, the spatiotemporal features are mapped to the probability of lithium plating risk through a fully connected layer. The preset risk prediction network is optimized using a focus loss function: ; In this embodiment of the invention, , To improve the accuracy and stability of the prediction network in predicting lithium plating risk under complex working conditions.

[0057] Step S104 also includes: Based on the battery risk prediction results, the battery operation process is slowed down or terminated according to a preset risk probability threshold.

[0058] Step S104 does not simply apply an existing spatiotemporal graph convolutional network to an arbitrary graph, but rather deeply couples it with the causal-driven coupled dynamic graph constructed in step S102 and the cross-modal causal association information extracted in step S103. Specifically, the spatiotemporal graph convolutional network utilizes the temporal evolution patterns in the cross-modal causal association information in the time dimension, and utilizes the topological structure reflecting the transmission of physical field risks in the dynamic graph in the spatial dimension. This three-in-one modeling approach of graph structure, causal features, and spatiotemporal convolution makes the risk prediction results not only statistically accurate, but also physically interpretable and traceable in terms of evolutionary path.

[0059] To demonstrate the effectiveness of the method proposed in the embodiments of the present invention, the embodiments of the present invention also provide the following comparative experiments and ablation experiments as reference.

[0060] In this embodiment of the invention, 18650 commercial lithium-ion batteries (NCM811 / graphite system) were used as test subjects. Cyclic charge-discharge tests at 1C / 2C rates were conducted at -10℃ until significant lithium plating occurred. During the data acquisition process, electrochemical impedance spectroscopy (once a week), infrared thermal imaging temperature distribution (real-time), acoustic emission signals (real-time), and conventional parameters such as voltage, current, and capacity were recorded simultaneously. A total of 120 battery samples were tested. A stratified sampling strategy was adopted, with 60% of the data used for training, 20% for validation, and 20% for testing, to ensure that the distribution of lithium plating risk samples in each data subset was consistent with the overall dataset.

[0061] The surface morphology of the negative electrode of the battery after testing was observed using scanning electron microscopy (SEM). The presence and distribution density of lithium dendrites were used as the gold standard label for lithium plating risk. The degree of lithium plating was divided into three levels: no lithium plating, mild lithium plating (dendritic coverage area <30%), and severe lithium plating (dendritic coverage area ≥30%).

[0062] Evaluation metrics include accuracy, precision, recall, F1 score, AUC-ROC, average lead time, false alarm rate, and missed alarm rate.

[0063] The embodiments of the present invention are compared with the following methods: M1, which uses only electrochemical impedance spectroscopy data to predict lithium plating risk, is implemented based on a two-layer bidirectional LSTM network with 128 hidden layer neurons and an input time window length of 30 seconds.

[0064] M2 uses only infrared thermal imaging data to predict lithium plating risk. It is implemented based on a CNN network containing 3 convolutional layers and 2 pooling layers, with convolutional kernel sizes set to 3×3, 5×5 and 3×3 respectively.

[0065] M3 uses only acoustic emission signal data to predict lithium plating risk. It is based on a one-dimensional convolutional neural network and contains three convolutional layers (with kernel sizes of 5, 5, and 3).

[0066] M4 directly concatenates the features of the three modalities of electrochemical, thermal, and mechanical and then inputs them into the support vector machine (SVM) classifier, using the radial basis function (RBF).

[0067] M5 uses each cell in the battery pack as a node and constructs a static graph structure based on the spatial adjacency relationship between cells. The node feature vector is simply concatenated from three-modal features and adopts a two-layer graph convolutional network (GCN).

[0068] M6 is the method proposed in the embodiments of the present invention.

[0069] The comparative experimental results are shown in Table 1 below. It can be seen that the method proposed in this embodiment (M6) significantly outperforms various single-modal methods (M1, M2, M3) and existing mainstream fusion methods (M4, M5) in all evaluation metrics. Particularly in terms of average warning lead time, the method proposed in this embodiment achieves 67.5 minutes, more than double the 31.2 minutes of existing GNN methods. Regarding the F1 score, the method proposed in this embodiment achieves 94.4%, a 12 percentage point improvement compared to the 82.5% of the static multimodal splicing method. These results demonstrate that the battery lithium plating risk prediction method proposed in this embodiment can effectively uncover the deep coupling relationships between multiple physical fields, significantly improving the accuracy and timeliness of lithium plating risk prediction.

[0070] Table 1 Comparative Experimental Data

[0071] To further verify the independent contribution and synergistic effect of each core step in the method proposed in the embodiments of the present invention, the following four ablation settings were designed: A1. Remove adaptive topology adjustment, retain the initial structure of the dynamic graph, cross-modal attention, and spatiotemporal graph convolution, but do not perform dynamic updates of risk propagation weights.

[0072] A2. Remove cross-modal attention, retain dynamic graph construction, adaptive adjustment and spatiotemporal graph convolution, but use simple concatenation instead of attention mechanism for cross-modal feature extraction.

[0073] A3. Remove the spatiotemporal graph convolution, retain dynamic graph construction, adaptive adjustment and cross-modal attention, but use a regular graph convolutional network (GCN) instead of the spatiotemporal graph convolutional network for risk prediction.

[0074] A4. Simultaneously remove adaptive adjustment and cross-modal attention, retaining only the initial structure of the dynamic graph and the ordinary GCN.

[0075] Ablation test results as follows Figure 3 As shown in the figure, the method proposed in this embodiment of the invention is significantly superior to existing methods (M1-M5) in terms of core indicators such as accuracy and F1 score. In particular, the accuracy reaches 94.7%, which is 9.4 percentage points higher than the best existing method (M5, conventional GNN), demonstrating significant technical effectiveness.

[0076] Ablation experiments show that each core step in the proposed method of this invention makes a significant independent contribution. Removing adaptive topology adjustment (A1) reduces accuracy by 6.1 percentage points and F1 score by 6.7 percentage points; removing cross-modal attention mechanism (A2) reduces accuracy by 7.8 percentage points and F1 score by 8.9 percentage points; removing spatiotemporal graph convolution (A3) reduces accuracy by 7.3 percentage points and F1 score by 8.0 percentage points.

[0077] The proposed method in this invention exhibits significant synergistic effects among its core steps. Simultaneously removing adaptive adjustment and cross-modal attention (A4) resulted in a 12.4 percentage point decrease in accuracy and a 13.3 percentage point decrease in F1 score. This performance degradation is comparable to the sum of the individual module removal degradations (approximately 14%). Considering the nonlinear characteristics of multi-module coupling, this indicates a significant synergistic effect among the modules, rather than a simple functional accumulation.

[0078] Among these, adaptive topology adjustment has a particularly significant impact on the timeliness of early warnings. Removing this step reduced the average early warning time from 67.5 minutes to 41.3 minutes, demonstrating the unique value of the dynamic risk evolution path capture mechanism in early warning.

[0079] Another important advantage of the method proposed in this invention is that its output risk prediction results have good interpretability. Through analysis of... Figure 2 The risk transmission weights in the causal-driven coupling dynamic diagram are visualized, which can intuitively show the evolution and propagation path of lithium plating risk among electrochemical, thermal, and mechanical multiphysics fields.

[0080] The results show that the distribution of cross-modal attention weights can automatically adjust to adapt to changes in dominant risk factors under different operating conditions: in low-temperature environments, the attention weight of the thermal mode increases significantly; under high-rate fast charging conditions, the attention weight of the electrochemical mode dominates; and in the later stages of battery aging, the weight of the mechanical mode (acoustic emission signal) increases significantly. This adaptive weight allocation mechanism enables the model to dynamically focus on the most critical risk signals under different operating conditions.

[0081] The risk propagation path output by the model was compared with the SEM observation results after the actual battery was disassembled. The two were in high agreement, which further verified the physical interpretability and practical engineering application value of the method proposed in this embodiment of the invention.

[0082] In summary, this invention proposes a novel concept for predicting battery lithium plating risk based on causal driving and multi-physics coupled dynamic graphs. Unlike existing static, single-dimensional, or simply pieced-together multimodal data-driven methods, the method proposed in this invention defines the evolution of lithium plating risk as a causal propagation process that dynamically changes with time and spatial topology, including: A dynamic graph model with three coupled fields of "electricity-thermal-mechanical" was constructed. The topology of this model can be adaptively adjusted according to the risk transmission weights calculated in real time, rather than being statically preset. A dedicated cross-modal feature extraction network was designed, which has a two-stage structure of "intra-modal decoupling - inter-modal weighted fusion" to extract cross-modal causal relationships in a physically interpretable way, rather than black-box feature splicing. By deeply coupling dynamic graphs with spatiotemporal graph convolutional networks, we can achieve joint modeling of the "temporal accumulation" and "spatial propagation" of risks.

[0083] The beneficial effects achieved by this invention lie in proposing a battery lithium plating risk prediction method based on causal driving and multi-physics field coupled dynamic graphs. The beneficial effects achieved by this method include: I. Overcoming the limitations of a single data source and the lag of static models: By constructing and dynamically adjusting the dynamic graph of the coupling of the electro-thermal-mechanical three fields, the complex multi-physics interactions inside the battery and the evolution of lithium plating risk are structurally correlated in time and space, enabling the model to capture the dynamic changes of the internal structure and physical field relationships of the battery in real time, and significantly improving the predictive adaptability under complex working conditions. Second, it has achieved a qualitative leap from result prediction to path characterization: by mining the causal triggering and amplification relationship between physical fields through cross-modal feature extraction networks, and combining it with spatiotemporal graph convolutional networks for joint analysis, the output is not only a single risk probability, but also the evolution trend and propagation path of lithium plating risk, providing battery management systems with better decision information depth. Third, it generates synergistic effects: dynamic graph construction, cross-modal causal reasoning, and spatiotemporal joint analysis are deeply coupled. The dynamic graph provides a dynamic and realistic physical topological foundation for cross-modal reasoning, the causal reasoning results provide physically meaningful features for spatiotemporal convolution, and the spatiotemporal analysis, in turn, guides the update direction of the dynamic graph. This structured collaborative working mechanism effectively improves the accuracy and robustness of the entire prediction system.

[0084] This invention also provides a battery lithium plating risk prediction system 200, please refer to... Figure 4 , Figure 4 This is a schematic diagram of the battery lithium plating risk prediction system provided in an embodiment of the present invention, which includes: The data acquisition module 201 is used to acquire various types of operating data generated during battery operation, and to standardize and integrate the different operating data to obtain battery multimodal data. The modal characterization module 202 is used to construct a causal-driven coupled dynamic graph that characterizes the directed causal transmission relationship of multi-physics risks during battery operation, based on the physical field risk transmission mechanism between electrochemical, thermal, and mechanical modes, using the battery multimodal data as nodes. The edges in the causal-driven coupled dynamic graph are directed edges, the direction of which is determined based on the physical causal transmission direction between the electrochemical, thermal, and mechanical modes, and the weight of the directed edge is the risk transmission weight. Specifically, the risk transmission weights between nodes under different modes are calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weights. Feature extraction module 203 is used to jointly model the causal-driven coupled dynamic graph through a preset cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, and extract cross-modal causal association information to characterize the lithium plating risk triggering and amplification relationship between different physical field states; The risk prediction module 204 is used to perform spatiotemporal joint analysis on the causal driving coupled dynamic graph and the cross-modal causal correlation information through a preset risk prediction network constructed based on a spatiotemporal graph convolutional network, so as to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

[0085] The battery lithium plating risk prediction system 200 can implement the steps in the battery lithium plating risk prediction method in the above embodiments and achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0086] This invention also provides a computer device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. The computer device 300 includes: a memory 302, a processor 301, and a battery lithium plating risk prediction program stored in the memory 302 and capable of running on the processor 301.

[0087] The processor 301 calls the battery lithium plating risk prediction program stored in the memory 302 and executes the steps in the battery lithium plating risk prediction method provided in this embodiment of the invention. Please refer to... Figure 1 Specifically, it includes the following steps: S101. Acquire various types of operating data generated during battery operation, and standardize and integrate the different operating data to obtain battery multimodal data; S102. Using the battery multimodal data as nodes, and based on the physical field risk transmission mechanism between electrochemical, thermal, and mechanical modes, a causal-driven coupled dynamic graph is constructed to characterize the directed causal transmission relationship of multiphysical field risks during battery operation. The edges in the causal-driven coupled dynamic graph are directed edges, and the direction of the directed edges is determined based on the physical causal transmission direction between electrochemical, thermal, and mechanical modes. The weight of the directed edges is the risk transmission weight. Specifically, the risk transmission weight between nodes under different modes is calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weight. S103. Through a pre-defined cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, the causal-driven coupling dynamic graph is jointly modeled to extract cross-modal causal association information used to characterize the lithium plating risk triggering and amplification relationship between different physical field states. S104. Using a pre-defined risk prediction network constructed based on a spatiotemporal graph convolutional network, a spatiotemporal joint analysis is performed on the causal driving coupled dynamic graph and the cross-modal causal correlation information to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

[0088] Furthermore, step S101, the step of standardizing and integrating the different operational data, includes: Use spatiotemporal alignment technology to unify the data time base for different operational data; At least one of sliding window, wavelet denoising, and ROI segmentation is used to extract key features and denoise the running data. The Z-score normalization method was used to scale the different running data.

[0089] Furthermore, in step S102, the step of adaptively and dynamically adjusting the topology of the causal-driven coupled dynamic graph according to the risk transmission weight specifically includes: During battery operation, at each preset dynamic update time, the topology between different nodes in the causal-driven coupled dynamic graph is updated based on the risk transmission weight to reflect the evolution and propagation process of lithium plating risk in the battery across multiple physical fields.

[0090] Furthermore, in step S103, the preset cross-modal feature extraction network is implemented based on a cross-modal graph attention network, wherein: From the same modal perspective, the cross-modal graph attention network extracts the time-dependent features of the electrochemical mode through the sequence feature modeling unit, the spatial distribution features of the thermal mode through the spatial feature modeling unit, and the structural features of the mechanical mode through the frequency domain feature modeling unit. From a cross-modal perspective, the cross-modal graph attention network, based on the neighborhood node structure in the causal-driven coupled dynamic graph, performs weighted fusion of different modal features to obtain cross-modal causal association information for characterizing the triggering and amplification relationship of lithium plating risk.

[0091] Furthermore, in step S104, the preset risk prediction network is implemented based on a spatiotemporal graph convolutional network. The preset risk prediction network is used to jointly model the cumulative changes of battery lithium plating risk in the time dimension and the propagation process in the multi-physics topology, and output the battery lithium plating risk prediction result.

[0092] Furthermore, step S104 also includes: Based on the battery risk prediction results, the battery operation process is slowed down or terminated according to a preset risk probability threshold.

[0093] The computer device 300 provided in this embodiment of the invention can implement the steps in the battery lithium plating risk prediction method as described in the above embodiments, and can achieve the same technical effect. Refer to the description in the above embodiments, which will not be repeated here.

[0094] This invention also provides a storage medium storing a battery lithium plating risk prediction program. When the battery lithium plating risk prediction program is executed by a processor, it implements the various processes and steps in the battery lithium plating risk prediction method provided in this invention and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0095] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by instructing related hardware (such as mobile phones, computers, servers, air conditioners, or network devices) through a battery lithium plating risk prediction program. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0096] It should be noted that, in this document, 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. Unless otherwise specified, 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 that element.

[0097] The embodiments of the present invention have been described above with reference to the accompanying drawings. The disclosed embodiments are merely preferred embodiments of the present invention. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many equivalent changes in form under the guidance of the present invention without departing from the spirit and scope of the claims. All such changes are within the protection scope of the present invention.

Claims

1. A method for predicting the risk of lithium plating in batteries, characterized in that, Includes the following steps: S101. Acquire various types of operating data generated during battery operation, and standardize and integrate the different operating data to obtain battery multimodal data; S102. Using the battery multimodal data as nodes, based on the physical field risk transmission mechanism between electrochemical, thermal and mechanical modes, construct a causal-driven coupled dynamic graph to characterize the directed causal transmission relationship of multiphysical field risks during battery operation. The edges in the causal-driven coupled dynamic graph are directed edges. The direction of the directed edges is determined based on the physical causal transmission direction between electrochemical, thermal, and mechanical modes. The weight of the directed edges is the risk transmission weight. Specifically, the risk transmission weight between nodes under different modes is calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weight. S103. Through a pre-defined cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, the causal-driven coupling dynamic graph is jointly modeled to extract cross-modal causal association information used to characterize the lithium plating risk triggering and amplification relationship between different physical field states. S104. Using a pre-defined risk prediction network constructed based on a spatiotemporal graph convolutional network, a spatiotemporal joint analysis is performed on the causal driving coupled dynamic graph and the cross-modal causal correlation information to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

2. The battery lithium plating risk prediction method according to claim 1, characterized in that, Step S101, the step of standardizing and integrating the different operational data, includes: Use spatiotemporal alignment technology to unify the data time base for different operational data; At least one of sliding window, wavelet denoising, and ROI segmentation is used to extract key features and denoise the running data. The Z-score normalization method was used to scale the different running data.

3. The battery lithium plating risk prediction method according to claim 1, characterized in that, Step S102, the step of adaptively and dynamically adjusting the topology of the causal-driven coupling dynamic graph according to the risk transmission weight, specifically includes: During battery operation, at each preset dynamic update time, the topology between different nodes in the causal-driven coupled dynamic graph is updated based on the risk transmission weight to reflect the evolution and propagation process of lithium plating risk in the battery across multiple physical fields.

4. The battery lithium plating risk prediction method according to claim 1, characterized in that, In step S103, the preset cross-modal feature extraction network is implemented based on a cross-modal graph attention network, wherein: From the same modal perspective, the cross-modal graph attention network extracts the time-dependent features of the electrochemical mode through the sequence feature modeling unit, the spatial distribution features of the thermal mode through the spatial feature modeling unit, and the structural features of the mechanical mode through the frequency domain feature modeling unit. From a cross-modal perspective, the cross-modal graph attention network, based on the neighborhood node structure in the causal-driven coupled dynamic graph, performs weighted fusion of different modal features to obtain cross-modal causal association information for characterizing the triggering and amplification relationship of lithium plating risk.

5. The battery lithium plating risk prediction method according to claim 1, characterized in that, In step S104, the preset risk prediction network is implemented based on a spatiotemporal graph convolutional network. The preset risk prediction network is used to jointly model the cumulative changes of battery lithium plating risk in the time dimension and the propagation process in the multi-physics topology, and output the battery lithium plating risk prediction result.

6. The battery lithium plating risk prediction method according to claim 1, characterized in that, Step S104 also includes: Based on the battery risk prediction results, the battery operation process is slowed down or terminated according to a preset risk probability threshold.

7. A battery lithium plating risk prediction system, characterized in that, include: The data acquisition module is used to acquire various types of operational data generated during battery operation, and to standardize and integrate the different operational data to obtain battery multimodal data. The modal characterization module is used to construct a causal-driven coupled dynamic graph that characterizes the directed causal transmission relationship of multi-physical field risks during battery operation, based on the physical field risk transmission mechanism between electrochemical, thermal, and mechanical modes, using the battery multimodal data as nodes. The edges in the causal-driven coupled dynamic graph are directed edges. The direction of the directed edges is determined based on the physical causal transmission direction between electrochemical, thermal, and mechanical modes. The weight of the directed edges is the risk transmission weight. Specifically, the risk transmission weight between nodes under different modes is calculated, and the topology of the causal-driven coupled dynamic graph is adaptively and dynamically adjusted according to the risk transmission weight. The feature extraction module is used to jointly model the causal-driven coupled dynamic graph through a preset cross-modal feature extraction network dedicated to decoupling and recombining multimodal causal relationships, and extract cross-modal causal association information to characterize the lithium plating risk triggering and amplification relationship between different physical field states; The risk prediction module is used to perform spatiotemporal joint analysis on the causal driving coupled dynamic graph and the cross-modal causal correlation information through a preset risk prediction network constructed based on a spatiotemporal graph convolutional network, so as to obtain risk prediction results for characterizing the evolution trend of battery lithium plating risk.

8. A computer device, characterized in that, include: The battery lithium plating risk prediction method includes a memory, a processor, and a battery lithium plating risk prediction program stored in the memory and executable on the processor, wherein the processor, when executing the battery lithium plating risk prediction program, implements the steps of the battery lithium plating risk prediction method as described in any one of claims 1-6.

9. A storage medium, characterized in that, The storage medium stores a battery lithium plating risk prediction program, which, when executed by a processor, implements the steps in the battery lithium plating risk prediction method as described in any one of claims 1-6.

Images